title
stringlengths 15
188
| abstract
stringlengths 400
1.8k
| introduction
stringlengths 9
10.5k
| content
stringlengths 778
41.9k
| abstract_len
int64 400
1.8k
| intro_len
int64 9
10.5k
| abs_len
int64 400
1.8k
|
|---|---|---|---|---|---|---|
Improving Neural Entity Disambiguation with Graph Embeddings
|
Entity Disambiguation (ED) is the task of linking an ambiguous entity mention to a corresponding entry in a knowledge base. Current methods have mostly focused on unstructured text data to learn representations of entities, however, there is structured information in the knowledge base itself that should be useful to disambiguate entities. In this work, we propose a method that uses graph embeddings for integrating structured information from the knowledge base with unstructured information from text-based representations. Our experiments confirm that graph embeddings trained on a graph of hyperlinks between Wikipedia articles improve the performances of simple feed-forward neural ED model and a state-ofthe-art neural ED system.
|
The inherent and omnipresent ambiguity of language at the lexical level results in ambiguity of words, named entities, and other lexical units. Word Sense Disambiguation (WSD) The goal of an ED system is resolving the ambiguity of entity mentions, such as Mars, Galaxy, and Bounty are all delicious. It is hard for an algorithm to identify whether the entity is an astronomical structure Current neural approaches to EL/ED attempt to use context and word embeddings (and sometimes entity embeddings on mentions in text) Graph embeddings aim at representing nodes in a graph, or subgraph structure, by finding a mapping between a graph structure and the points in a low-dimensional vector space Our claim is that including graph structure features of the knowledge base has a great potential to make an impact on ED. In our first experiment, we present a method based on a simple neural network with the inputs of a context, entity mention/span, explanation of a candidate entity, and a candidate entity. Each entity is represented by graph embeddings, which are created using the knowledge base, DBpedia Our main contribution is a creation of a simple technique for integration of structured information into an ED system with graph embeddings. There is no obvious way to use large structured knowledge bases directly in a neural ED system. We provide a simple solution based on graph embeddings and confirm experimentally its effectiveness.
|
Entity Linking Traditional approaches to EL focus on defining the similarity measurement between a mention and a candidate entity Graph Embeddings There are various methods to create graph embedding, which can be grouped into the methods based on matrix factorization, random walks, and deep learning In order to make information from a semantic graph available for an entity linking system, we make use of graph embeddings. We use DeepWalk First, we created a graph whose nodes are unique entities; attributes are explanations of entities, i.e. long abstracts; edges are the page links between entities with the information from DBpedia. Second, a vector representation per entity is generated by training DeepWalk on the edges of this graph. For this, we used all default hyper-parameters of DeepWalk, e.g. number- In our first experiment, we build a simple neural ED system based on a feed-forward network and test the utility of the graph embeddings as compared to text-based embeddings. For the second experiment, we have used the end2end state-of-the-art system for EL/ED Creates long abstract vector Gets word vector from doc2vec entity explanations (long abstracts), and multiword spans. Datasets: We train the neural end2end system in its default configuration with the combination of MSNBC Implementation Details: We have not changed hyper-parameters for training the end2end system We create the embeddings file with the same format they used. They give an id for each entity and call it "wiki id". First, we generate a map between this wiki id and our graph id (id of our entity). Then, we replace each entity vector corresponding to the wiki id with our graph embeddings, which refers to the entity. Sometimes there is no corresponding graph entity for the entity in the end2end system, in this case, we supply a zero vector. They have a stopping condition, which applies after 6 consecutive evaluations with no significant improvement in the Macro F1 score. We have changed this hyperparameter to 10, accounting for our observation that the training converges slower when operating on graph embeddings. Table The standard deviations of the models are between 0.02 -0.05 in the DBpedia Spotlight dataset and 0.01 -0.03 in the Reuters-128 dataset over all scores. Scores are produced using the GERBIL platform; these are Micro-averaged over the set of annotations in the dataset and Macro-averaged over the average performance per document. The results are improved by including graph embeddings. When we compare two models, trained by graph embeddings and trained by entity vectors, the results are improved up to 0.03 in Macro F1 scores and Micro Precision, and up to 0.07 in Macro Precision. However, the improvement of the combination model is higher in Macro F1 and Recall. Micro-averaged results follow a similar trend. When we look at the scores of Reuters-128 In the Micro-averaged evaluation, the combination model scores slightly below the model using graph embeddings alone. To summarize the evaluation, our graph embeddings alone already lead to improvements over the original text-based embeddings, and their combination is even more beneficial. This suggests that test-based and graph-based representations in fact encode somewhat complementary information. We have shown how to integrate structured information into the neural ED task using two differ-ent experiments. In the first experiment, we use a simple neural network to gauge the impact of different text-based and graph-based embeddings. In the second experiment, we replace respectively complemented the representation of candidate entities in the ED component of a state-of-the-art EL system. In both setups, we demonstrate that graph embeddings lead to en par or better performance. This confirms our research hypothesis that it is possible to use structured resources for modeling entities in ED tasks and the information is complementary to a text-based representation alone. Our code and datasets are available online For future work, we plan to examine graph embeddings on other relationships, e.g. taxonomic or otherwise typed relations such as works-for, married-with, and so on, generalizing the notion to arbitrary structured resources. It might make a training step on the distance measure depending on the relation necessary. On the disambiguation architecture, modeling such direct links could give rise to improvements stemming from the mutual disambiguation of entities as e.g. done in
| 738 | 1,441 | 738 |
Using Uneven Margins SVM and Perceptron for Information Extraction
|
The classification problem derived from information extraction (IE) has an imbalanced training set. This is particularly true when learning from smaller datasets which often have a few positive training examples and many negative ones. This paper takes two popular IE algorithms -SVM and Perceptron -and demonstrates how the introduction of an uneven margins parameter can improve the results on imbalanced training data in IE. Our experiments demonstrate that the uneven margin was indeed helpful, especially when learning from few examples. Essentially, the smaller the training set is, the more beneficial the uneven margin can be. We also compare our systems to other state-of-theart algorithms on several benchmarking corpora for IE.
|
Information Extraction (IE) is the process of automatic extraction of information about pre-specified types of events, entities or relations from text such as newswire articles or Web pages. IE is useful in many applications, such as information gathering in a variety of domains, automatic annotations of web pages for Semantic Web, and knowledge management. A wide range of machine learning techniques have been used for IE and achieved state-of-the-art results, comparable to manually engineered IE systems. A learning algorithm usually learns a model from a set of documents which have been manually annotated by the user. Then the model can be used to extract information from new documents. Manual annotation is a time-consuming process. Hence, in many cases learning from small data sets is highly desirable. Therefore in this paper we also evaluate the performance of our algorithms on small amounts of training data and show their learning curve. The learning algorithms for IE can be classified broadly into two main categories: rule learning and statistical learning. The former induces a set of rules from training examples. There are many rule based learning systems, e.g. SRV This paper concentrates on classifier-based learning for IE, which typically converts the recognition of each information entity into a set of classification problems. In the framework discussed here, two binary classifiers are trained for each type of information entity. One classifier is used for recognising the entity's start token and the other -the entity's end token. The classification problem derived from IE usually has imbalanced training data, in which positive training examples are vastly outnumbered by negative ones. This is particularly true for smaller data sets where often there are hundreds of negative training examples and only few positive ones. Two approaches have been studied so far to deal with imbalanced data in IE. One approach is to under-sample majority class or over-sample minority class in order to obtain a relatively balanced training data In this paper we explore another approach to handle the imbalanced data in IE, namely, adapting the learning algorithms for balanced classification to imbalanced data. We particularly study two popular classification algorithms in IE, Support Vector Machines (SVM) and Perceptron. SVM is a general supervised machine learning algorithm, that has achieved state of the art performance on many classification tasks, including NE recognition. Previous research on using SVMs for IE adopts the standard form of the SVM, which treats positive and negative examples equally. As a result, they did not consider the difference between the balanced classification problems, where the SVM performs quite well, and the imbalanced ones. Perceptron is a simple, fast and effective learning algorithm, which has successfully been applied to named entity recognition The rest of the paper is structured as follows. Section 2 describes the uneven margins SVM and Perceptron algorithms. Sections 3.1 and 3.2 discuss the classifier-based framework for IE and the experimental datasets we used, respectively. We compare our systems to other state-of-the-art systems on three benchmark datasets in Section 3.3. Section 3.4 discusses the effects of the uneven margins parameter on the SVM and Perceptron's performances. Finally, Section 4 provides some conclusions.
|
Li and Shawe-Taylor (2003) introduced an uneven margins parameter into the SVM to deal with imbalanced classification problems. They showed that the SVM with uneven margins outperformed the standard SVM on document classification problem with imbalanced training data. Formally, given a training set Z = ((x 1 , y 1 ), . . . , (x m , y m )),where x i is the ndimensional input vector and y i (= +1 or -1) its label, the SVM with uneven margins is obtained by solving the quadratic optimisation problem: We can see that the uneven margins parameter τ was added to the constraints of the optimisation problem. τ is the ratio of negative margin to the positive margin of the classifier and is equal to 1 in the standard SVM. For an imbalanced dataset with a few positive examples and many negative ones, it would be beneficial to use larger margin for positive examples than for the negative ones. Perceptron is an on-line learning algorithm for linear classification. It checks the training examples one by one by predicting their labels. If the prediction is correct, the example is passed; otherwise, the example is used to correct the model. The algorithm stops when the model classifies all training examples correctly. The margin Perceptron not only classifies every training example correctly but also outputs for every training example a value (before thresholding) larger than a predefined parameter (margin). The margin Perceptron has better generalisation capability than the standard Perceptron. In the experiments we adopted a classifier-based framework for applying the SVM and PAUM algorithms to IE. The framework consists of three stages: pre-processing of the documents to obtain feature vectors, learning classifiers or applying classifiers to test documents, and finally post-processing the results to tag the documents. The aim of the preprocessing is to form input vectors from documents. Each document is first processed using the open-source ANNIE system, which is part of GATE Based on the linguistic information, an input vector is constructed for each token, as we iterate through the tokens in each document (including word, number, punctuation and other symbols) to see if the current token belongs to an information entity or not. Since in IE the context of the token is usually as important as the token itself, the features in the input vector come not only from the current token, but also from preceding and following ones. As the input vector incorporates information from the context surrounding the current token, features from different tokens can be weighted differently, based on their position in the context. The weighting scheme we use is the reciprocal scheme, which weights the surrounding tokens reciprocally to the distance to the token in the centre of the context window. This reflects the intuition that the nearer a neighbouring token is, the more important it is for classifying the given token. Our experiments showed that such a weighting scheme obtained better results than the commonly used equal weighting of features The key part of the framework is to convert the recognition of information entities into binary classification tasks -one to decide whether a token is the start of an entity and another one for the end token. After classification, the start and end tags of the entities are obtained and need to be combined into one entity tag. Therefore some post-processing is needed to guarantee tag consistency and to try to improve the results by exploring other information. The currently implemented procedure has three stages. First, in order to guarantee the consistency of the recognition results, the document is scanned from left to right to remove start tags without matching end tags and end tags without preceding start tags. The second stage filters out candidate entities from the output of the first stage, based on their length. Namely, a candidate entity tag is removed if the entity's length (i.e., the number of tokens) is not equal to the length of any entity of the same type in the training set. The third stage puts together all possible tags for a sequence of tokens and chooses the best one according to the probability which was computed from the output of the classifiers (before thresholding) via a Sigmoid function. The paper reports evaluation results on three corpora covering different IE tasks -named entity recognition In detail, we used the English part of the CoNLL-2003 shared task dataset, which consists of 946 documents for training, 216 document for development (e.g., tuning the parameters in learning algorithm), and 231 documents for evaluation (i.e., testing), all of which are news articles taken from the Reuters English corpus (RCV1). The corpus contains four types of named entities -person, location, organisation and miscellaneous names. In the other two corpora domain-specific information was extracted into a number of slots. The Job corpus includes 300 computer related job advertisements and 17 slots encoding job details, such as title, salary, recruiter, computer language, application, and platform. The CFP corpus consists of 1100 conference or workshop call for papers (CFP), of which 600 were annotated. The corpus includes 11 slots such as workshop and conference names and acronyms, workshop date, location and homepage. Named Entity Recognition The algorithms are evaluated on the CoNLL-2003 dataset. Since this set comes with development data for tuning the learning algorithm, different settings were tried in order to obtain the best performance on the development set. Different SVM kernel types, window sizes (namely the number of tokens in left or right side of the token at the centre of window), and the uneven margins parameter τ were tested. We found that quadratic kernel, window size 4 and τ = 0.5 produced best results on the development set. These settings were used in all experiments on the CoNLL-2003 dataset in this paper, unless otherwise stated. The parameter settings for PAUM described in Table Firstly, our uneven margins SVM system performed significantly better than the other SVMbased system. As the two systems are different from each other in not only the SVM models used but also other aspects such as the NLP features and the framework, in order to make a fair comparison between the uneven margins SVM and the standard SVM, we also present the results of the two learning algorithms implemented in our framework. We can see from Table Secondly, our PAUM-based system performed slightly better than the system based on voted Perceptron, but there is no significant difference between them. Note that they adopted different mechanisms to deal with the imbalanced data in IE (refer to Section 1). The structure of PAUM system is simpler than that of the voted Perceptron system. Finally, the PAUM system performed worse than the SVM system. On the other hand, training time of PAUM is only 1% of that for the SVM and the PAUM implementation is much simpler than that of SVM. Therefore, when simplicity and speed are required, PAUM presents a good alternative. Template Filling On Jobs corpus our systems are compared to several state-of-the-art learning systems, which include the rule based systems Rapier Preliminary experiments established that the SVM with linear kernel obtained better results than SVM with quadratic kernel on the Jobs corpus Table Firstly, the overall performance of our two systems is significantly better than the other three fully evaluated systems. The PAUM system achieves the best performance on 5 out of the 17 slots. The SVM system performs best on the other 3 slots. Secondly, the double classification system had much worse overall performance than our systems and other two fully evaluated systems. HMM was evaluated only on two slots. It achieved best result on one slot but was much worse on the other slot than our two systems and some of the others. Finally, somewhat surprisingly, our PAUM system achieves better performance than the SVM system on this dataset. Moreover, the computation time of PAUM is about 1/3 of that of the SVM. Hence, the PAUM system performs quite satisfactory on the Jobs corpus. Our systems were also evaluated by participating in a Pascal challenge -Evaluating Machine Learning for Information Extraction. The evaluation provided not only the CFP corpus but also the linguistic features for all tokens by pre-processing the documents. The main purpose of the challenge was to evaluate machine learning algorithms based on the same linguistic features. The only compulsory task is task1, which used 400 annotated documents for training and other 200 annotated documents for testing. See Table A number of experiments were conducted to investigate the influence of the uneven margins parameter on the SVM and Perceptron's performances. Table This paper studied the uneven margins versions of two learning algorithms -SVM and Perceptron -to deal with the imbalanced training data in IE. Our experiments showed that the uneven margin is helpful, in particular on small training sets. The smaller the training set is, the more beneficial the uneven margin could be. We also showed that the systems based on the uneven margins SVM and Perceptron were com-Table Our SVM system obtained better results than other SVM-based systems on the CoNLL-2003 corpus and CFP corpus respectively, while being simpler than most of them. This demonstrates that our SVM system is both effective and efficient. We also explored PAUM, a simple and fast learning algorithm for IE. The results of PAUM were somehow worse (about 0.02 overall F-measure lower) than those of the SVM on two out of three datasets. On the other hand, PAUM is much faster to train and easier to implement than SVM. It is also worth noting that PAUM outperformed some other learning algorithms. Therefore, even PAUM on its own would be a good learning algorithm for IE. Moreover, PAUM could be used in combination with other classifiers or in the more complicated framework such as the one in Since many other tasks in Natural Language Processing, like IE, often lead to imbalanced classification problems and the SVM has been used widely in Natural Language Learning (NLL), we can expect that the uneven margins SVM and PAUM are likely to obtain good results on other NLL problems as well.
| 738 | 3,413 | 738 |
Intentional Context in Situated Natural Language Learning
|
Natural language interfaces designed for situationally embedded domains (e.g. cars, videogames) must incorporate knowledge about the users' context to address the many ambiguities of situated language use. We introduce a model of situated language acquisition that operates in two phases. First, intentional context is represented and inferred from user actions using probabilistic context free grammars. Then, utterances are mapped onto this representation in a noisy channel framework. The acquisition model is trained on unconstrained speech collected from subjects playing an interactive game, and tested on an understanding task.
|
As information technologies move off of our desktops and into the world, the need for Natural Language Processing (NLP) systems that exploit information about the environment becomes increasingly apparent. Whether in physical environments (for cars and cell phones) or in virtual ones (for videogames and training simulators), applications are beginning to demand language interfaces that can understand unconstrained speech about constrained domains. Unlike most text-based NLP research, which focuses on open-domain problems, work we refer to as situated NLP focuses on improving language processing by exploiting domain-specific information about the non-linguistic situational context of users' interactions. For applications where agents interact in shared environments, such information is critical for successful communication. Previous work in situated NLP has focused on methods for grounding the meaning of words in physical and virtual environments. The motivation for this work comes from the inability of textbased NLP technologies to offer viable models of semantics for human computer interaction in shared environments. For example, imagine a situation in which a human user is interacting with a robotic arm around a table of different colored objects. If the human were to issue the command "give me the blue one," both the manually-coded This research presents a methodology for modeling the intentional context of utterances and describes how such representations can be used in a language learning task. We decompose language learning into two phases: intention recognition and linguistic mapping. In the first phase, we model intentional action using a probabilistic context free grammar. We use this model to parse sequences of observed physical actions, thereby inferring a hierarchical tree representation of a user's intentions. In the second phase, we use a noisy channel model to learn a mapping between utterances and nodes in that tree representation. We present pilot situated language acquisition experiments using a dataset of paired spontaneous speech and action collected from human subjects interacting in a shared virtual environment. We evaluate the acquired model on a situated language understanding task.
|
The ability to infer the purpose of others' actions has been proposed in the psychological literature as essential for language learning in children The second type of ambiguity, referred to as horizontal ambiguity describes the ambiguity between the "offer drink" vs. "ask for change" interpretations of "blicket." Here there is an ambiguity based on what actually is the intention behind the physical action. Thus, it is the case that only one of these meaning corresponds to "blicket" and the other meaning is not an accurate description of the intended action. Figure Just as in the plan recognition work of Pynadath, (1999), we cast the problem of intention recognition as a probabilistic parsing problem in which sequences of physical actions are used to infer an abstract tree representation. Resolving horizontal ambiguities thus becomes equivalent to determining which parse tree is most likely given a sequence of events. Further, resolving vertical ambiguities corresponds to determining which level node in the inferred tree is the correct level of description that the speaker had in mind. The problem of learning a mapping between linguistic labels and nodes in an inferred intentional tree is recast as one of learning the channel probabilities in Equation 4 α refers to a weighting coefficient. In order to avoid the many physical and perceptual problems that complicate work with robots and sensor-grounded data, this work focuses on language learning in virtual environments. We focus on multiplayer videogames , which support rich types of social interactions. The complexities of these environments highlight the problems of ambiguous speech described above, and distinguish this work from projects characterized by more simplified worlds and linguistic interactions, such as SHRDLU The game was instrumented such that all players' speech/text language and actions are recorded during game play. For data collection, a game was designed in which a single player must navigate their way through a cavernous world, collecting specific objects, in order to escape. Subjects were paired such that one, the novice, would control the virtual character, while the other, the expert, guided her through the world. While the expert could say anything in order to tell the novice where to go and what to do, the novice was instructed not to speak, but only to follow the commands of the expert. Expert's utterances: Novice's actions: Figure The purpose behind these restrictions was to elicit free and spontaneous speech that is only constrained by the nature of the task. This environment seeks to emulate the type of speech that a real situated language system might encounter: i.e., natural in its characteristics, but limited in its domain of discourse. The subjects in the data collection were university graduate and undergraduate students. Subjects (8 male, 4 female) were staggered such that the novice in one trial became the expert in the next. Each pair played the game at least five times, and for each of those trials, all speech from the expert and all actions from the novice were recorded. Table Action Frame Data collection produces two parallel streams of information: the sequence of actions taken by the novice and the audio stream produced by the expert (figure The audio streams are automatically segmented into utterances using a speech endpoint detector, which are then transcribed by a human annotator. Each action in the sequence is then automatically parsed, and each node in the tree is replaced with a semantic frame (figure As described in section 2, we represent the task model associated with the game as a set of production rules in which the right hand side consists of an intended action (e.g. "find key") and the left hand side consists of a sequence of subactions that are sufficient to complete that action (e.g. "go through door, open chest, pick_up key"). By applying probabilities to the rules, intention recognition can be treated as a probabilistic context free parsing problem, following Given the PCFG, we use a probabilistic Earley parser Methodologies for evaluating language acquisition tasks are not standardized. Given our model, there exists the possibility of employing intrinsic measures of success, such as word alignment accuracy. However, we choose to measure the success of learning by examining the related (and more natural) task of language understanding. For each subject pair, the linguistic mapping algorithms are trained on the first four trials of game play and tested on the final trial. (This gives on average 130 utterances of training data and 30 utterances of testing data per pair.) For each utterance in the test data, we calculate the likelihood that it was generated by each frame seen in testing. We select the maximum likelihood frame as the system's hypothesized meaning for the test utterance, and examine both how often the maximum likelihood estimate exactly matches the true frame (frame accuracy), and how many of the role fillers within the estimated frame match the role fillers of the true frame (role accuracy). 7 Figure For each subject, the algorithm's parameters are optimized using data from all other subjects. We assume correct knowledge of the temporal alignment between utterances and actions. In future work, we will relax this assumption to explore the effects of not knowing which actions correspond to which utterances in time. To examine the performance of the model, three experiments are presented. Experiment 1 examines the basic performance of the algorithms on the language understanding task described above given uniform priors. The system is tested under two conditions: 1) using the extended EM algorithm given an unknown utterance-to-level alignment, and 2) using the standard EM algorithm given the correct utterance-to-level alignment. Experiment 2 tests the benefit of incorporating intentional context directly into language understanding. This is done by using the parse probability of each hypothesized intention as the 7 See Figure Performance is on a language understanding task (baseline equivalent to choosing most frequent frame) Experiment 3 studies to what extent inferred tree structures are necessary when modeling language acquisition. Although, in section 1, we have presented intuitive reasons why such structures are required, one might argue that inferring trees over sequences of observed actions might not actually improve understanding performance when compared to a model trained only on the observed actions themselves. This hypothesis is tested by comparing a model trained given the correct utterance-to-level alignment (described in experiment 1) with a model in which each utterance is aligned to the leaf node (i.e. observed action) below the correct level of alignment. For example, in figure Experiment 1: We present the average performance over all subject pairs, trained with the correct utterance-to-level alignment both known and unknown, and compare it to a baseline of choosing the most frequent frame from the training data. Figure As expected, the understanding performance goes down for both frames and roles when the correct utterance-to-level alignment is unknown. Interestingly, while the frame performance declines by 14.3%, the performance on roles only declines 6.4%. This difference is due primarily to the fact that, while the mapping from words to action role fillers is hindered by the need to examine all alignments, the mapping from words to object role fillers remains relatively robust. This is due to the fact that while each level of intention carries a different action term, often the objects described at different levels remain the same. For example, in figure However, the object filler "DOOR" occurs multiple times. Thus, the chance that the role filler "DOOR" correctly maps to the word "door" is relatively high compared to the role filler "OPEN" mapping to the word "open." 8 Figure Experiment 2: Figure Experiment 3: Figure The results from these experiments, although preliminary, indicate that this model of language acquisition performs well above baseline on a language understanding task. This is particularly encouraging given the unconstrained nature of the speech on which it was trained. Thus, even free and spontaneous speech can be handled when modeling a constrained domain of discourse. As discussed in section 2, learning language about intentional action requires dealing with two distinct types of ambiguity. These difficulties cannot be handled by merely increasing the amount of data used, or switching to a more sophisticated learning algorithm. Rather, dealing with language use for situated applications requires building appropriate knowledge representations that are powerful enough for unconstrained language, yet scalable enough for practical applications. The work presented here is an initial demonstration of how the semantics of unconstrained speech can be modeled by focusing on constrained domains. As for scalability, it is our contention that for situated NLP, it is not a question of being able to scale up a single model to handle open-domain speech. The complexity of situated communication requires the use of domain-specific knowledge for modeling language use in different contexts. Thus, with situated NLP systems, it is less productive to focus on how to scale up single models to operate beyond their original domains. Rather, as more individual applications are tackled (e.g. cars, phones, videogames, etc.) the interesting question becomes one of how agents can learn to switch between different models of language as they interact in different domains of discourse. We have introduced a model of language acquisition that explicitly incorporates intentional contexts in both learning and understanding. We have described pilot experiments on paired language and action data in order to demonstrate both the model's feasibility as well as the efficacy of using intentional context in understanding. Although we have demonstrated a first step toward an advanced model of language acquisition, there is a great deal that has not been addressed. First, what is perhaps most obviously missing is any mention of syntax in the language learning process and its role in bootstrapping for language acquisition. Future work will focus on moving beyond the IBM Model 1 assumptions, to develop more syntactically-structured models. Further, although the virtual environment used in this research bears similarity to situated applications that demand NL interfaces, it is not known exactly how well the model will perform "in the real world." Future work will examine installing models in real world applications. In parallel investigations, we will explore our method as a cognitive model of human language learning. Finally, as was mentioned previously, the task model for this domain was hand annotated and, while the constrained nature of the domain simplified this process, further work is required to learn such models jointly with language. In summary, we have presented first steps toward tackling problems of ambiguity inherent in grounding the semantics of situated language. We believe this work will lead to practical applications for situated NLP, and provide new tools for modeling human cognitive structures and processes underlying situated language use
| 634 | 2,245 | 634 |
CorefQA: Coreference Resolution as Query-based Span Prediction
|
In this paper, we present CorefQA, an accurate and extensible approach for the coreference resolution task. We formulate the problem as a span prediction task, like in question answering: A query is generated for each candidate mention using its surrounding context, and a span prediction module is employed to extract the text spans of the coreferences within the document using the generated query. This formulation comes with the following key advantages: (1) The span prediction strategy provides the flexibility of retrieving mentions left out at the mention proposal stage; (2) In the question answering framework, encoding the mention and its context explicitly in a query makes it possible to have a deep and thorough examination of cues embedded in the context of coreferent mentions; and (3) A plethora of existing question answering datasets can be used for data augmentation to improve the model's generalization capability. Experiments demonstrate significant performance boost over previous models, with 83.1 (+3.5) F1 score on the CoNLL-2012 benchmark and 87.5 (+2.5) F1 score on the GAP benchmark. 1
|
Recent coreference resolution systems At the task formalization level, mentions left out at the mention proposal stage can never be recovered since the downstream module only operates on the proposed mentions. Existing models often suffer from mention proposal
|
In addition , many people were poisoned when toxic gas was released. They were poisoned and did not know how to protect themselves against the poison. Our formulation Q1: Who were poisoned when toxic gas was released? A1: [They, themselves] Q2: What was released when many people were poisoned? A2: [the poison] Q3: Who were poisoned and did not know how to protect themselves against the poison? A3: [many people, themselves] Q4: Whom did they not know how to protect against the poison? A4: [many people, They] Q5: They were poisoned and did not know how to protect themselves against what? A5: [toxic gas] 2018a). The coreference datasets can only provide a weak signal for spans that correspond to entity mentions because singleton mentions are not explicitly labeled. Due to the inferiority of the mention proposal model, it would be favorable if a coreference framework had a mechanism to retrieve left-out mentions. At the algorithm level, existing end-to-end methods To alleviate these issues, we propose CorefQA, a new approach that formulates the coreference resolution problem as a span prediction task, akin to the question answering setting. A query is generated for each candidate mention using its surrounding context, and a span prediction module is further employed to extract the text spans of the coreferences within the document using the generated query. Some concrete examples are shown in Figure Moreover, the proposed question answering formulation allows us to take advantage of existing question answering datasets. Coreference annotation is expensive, cumbersome and often requires linguistic expertise from annotators. Under the proposed formulation, the coreference resolution has the same format as the existing question answering datasets Experiments show that the proposed framework significantly outperforms previous models on two widely-used datasets. Specifically, we achieve new state-of-the-art scores of 83.1 (+3.5) on the CoNLL-2012 benchmark and 87.5 (+2.5) on the GAP benchmark. Coreference resolution is a fundamental problem in natural language processing and is considered as a good test of machine intelligence Machine reading comprehension is a general and extensible task form. Many tasks in natural language processing can be framed as reading comprehension while abstracting away the taskspecific modeling constraints. I was hired to do some Christmas music, and it was just "Jingle Bells" and I brought my cat with me to the studio, and I was working on the song and the cat jumped up into the record booth and started meowing along, meowing to me. I was hired to do some Christmas music, and it was just "Jingle Bells" and I brought my cat with me to the studio, and I was working on the song and the cat jumped up into the record booth and started meowing along, meowing to me. I was hired to do some Christmas music, and it was just "Jingle Bells" and I brought my cat with me to the studio, and I was working on the song and the cat jumped up into the record booth and started meowing along, meowing to me. I was hired to do some Christmas music, and it was just "Jingle Bells" and I brought my cat with me to the studio, and I was working on the song and the cat jumped up into the record booth and started meowing along, meowing to me. Data augmentation is a strategy that enables practitioners to significantly increase the diversity of data available for training models. Data augmentation techniques have been explored in various fields such as question answering In this section, we describe our CorefQA model in detail. The overall architecture is illustrated in Figure Given a sequence of input tokens X = {x 1 , x 2 , ..., x n } in a document, where n denotes the length of the document. N = n * (n + 1)/2 denotes the number of all possible text spans in X. Let e i denotes the i-th span representation 1 ≤ i ≤ N , with the start index FIRST(i) and the end index The task of coreference resolution is to determine the antecedents for all possible spans. If a candidate span e i does not represent an entity mention or is not coreferent with any other mentions, a dummy token is assigned as its antecedent. The linking between all possible spans e defines the final clustering. We use the SpanBERT model To fit long documents into SpanBERT, we use a sliding-window approach that creates a T -sized segment after every T /2 tokens. Segments are then passed to the SpanBERT encoder independently. The final token representations are derived by taking the token representations with maximum context. Similar to where x FIRST(i) and x LAST(i) represent the first and the last token representation of the i-th candidate span. FFNN m () denotes the feed-forward neural network that computes a nonlinear mapping from the input vector to the mention score. We only keep up to λn (where n is the document length) spans with the highest mention scores. Given a mention e i proposed by the mention proposal network, the role of the mention linking network is to give a score s a (i, j) for any text span e j , indicating whether e i and e j are coreferent. We propose to use the question answering framework as the backbone to compute s a (i, j). It operates on the triplet {context (X), query (q), answers (a)}. The context X is the input document. The query q(e i ) is constructed as follows: given e i , we use the sentence that e i resides in as the query, with the minor modification that we encapsulates e i with special tokens < mention >< /mention > . The answers a are the coreferent mentions of e i . Following (1) the candidate span e i does not represent an entity mention or (2) the candidate span e i represents an entity mention but is not coreferent with any other mentions in X. The probability of assigning a tag ∈ B, I, O is computed as follows: FFNN tag () represents the feed-forward neural network that computes a nonlinear mapping from the input vector to the tag logit. We further extend the token-level score in Eq. 2 to the span level. The anaphora score s a (j|i), the compatibility score of span j being a answer for span i, is calculated by the log probability of its beginning word taking the B tag and the rest taking the I tag: (3) A closer look at Eq.3 reveals that it only models the uni-directional coreference relation from e i to e j , i.e., e j is the answer for query q(e i ). This is suboptimal since if e i is a coreference mention of e j , then e j should also be the coreference mention e i . We thus need to optimize the bi-directional relation between e i and e j . s a (i|j) can be computed in the same way as s a (j|i), in which q(e j ) is used as the query. For a pair of text span e i and e j , the premises for them being coreferent mentions are (1) they are mentions and (2) they are coreferent. This makes the overall score s(i, j) for e i and e j the combination of Eq.1 and Eq.4: Given a document X with length n and the number of spans O(n 2 ), the computation of Eq.5 for all mention pairs is intractable with the complexity of O(n 4 ). Given an extracted mention e i , the computation of Eq.5 for (e i , e j ) regarding all e j is still extremely intensive since the computation of the backward span prediction score s a (i|j) requires running question answering models on all query q(e j ). A further pruning procedure is thus needed: For each query q(e i ), we collect C span candidates only based on the s a (j|i) scores. For each mention e i proposed by the mention proposal network, it is associated with C potential spans proposed by the mention linking network based on s(j|i), we aim to optimize the marginal log-likelihood of all correct antecedents implied by the gold clustering. Following The model will output it if none of the C span candidates is coreferent with e i . For each mention e i , the model learns a distribution P (•) over all possible antecedent spans e j based on the global score s(i, j) from Eq. 5: P (e j ) = e s(i,j) j ∈C e s(i,j ) The mention proposal module and the mention linking module are jointly trained in an end-to-end fashion using training signals from Eq.6, with the SpanBERT parameters shared. Given an input document, we can obtain an undirected graph using the overall score, each node of which represents a candidate mention from either the mention proposal module or the mention linking module. We prune the graph by keeping the edge whose weight is the largest for each node based on Eq.6. Nodes whose closest neighbor is the dummy token are abandoned. Therefore, the mention clusters can be decoded from the graph. We hypothesize that the reasoning (such as synonymy, world knowledge, syntactic variation, and multiple sentence reasoning) required to answer the questions are also indispensable for coreference resolution. Annotated question answering datasets are usually significantly larger than the coreference datasets due to the high linguistic expertise required for the latter. Under the proposed QA formulation, coreference resolution has the same format as the existing question answering datasets Comparing with existing models The special tokens used to denote the speaker's name (< speaker >< /speaker >) and the special tokens used to denote the queried mentions (< mention >< /mention >) are initialized by randomly taking the unused tokens from the SpanBERT vocabulary. The sliding window size T = 512, and the mention keep ratio λ = 0.2. The maximum length L for mention proposal = 10 and the maximum number of antecedents kept for each mention C = 50. The SpanBERT parameters are updated by the Adam optimizer We compare the CorefQA model with previous neural models that are trained end-to-end: • e2e-coref • c2f-coref + ELMo • c2f-coref + BERT-large • EE + BERT-large (Kantor and Globerson, 2019) represents each mention in a cluster via an approximation of the sum of all mentions in the cluster. • c2f-coref + SpanBERT-large The English data of CoNLL-2012 shared task (Pradhan et al., 2012) contains 2,802/343/348 train/development/test documents in 7 different genres. The main evaluation is the average of three metrics -MUC We compare the CorefQA model with several baseline models in Table 1. Our CorefQA system achieves a huge performance boost over existing systems: With SpanBERT-base, it achieves an F1 score of 79.9, which already outperforms the previous SOTA model using SpanBERT-large by 0.3. With SpanBERT-large, it achieves an F1 score of 83.1, with a 3.5 performance boost over the previous SOTA system. The GAP dataset We follow the protocols in We perform comprehensive ablation studies and analyses on the CoNLL-2012 development dataset. Results are shown in Table Effect of SpanBERT Replacing SpanBERT with vanilla BERT leads to a 3.8 F1 degradation. This verifies the importance of span-level pre-training for coreference resolution and is consistent with previous findings Effect of Pre-training Mention Proposal Network Skipping the pre-training of the mention proposal network using golden mentions results in a 7.5 F1 degradation, which is in line with our expectation. A randomly initialized mention proposal model implies that mentions are randomly selected. Randomly selected mentions will mostly be transformed to unanswerable queries. This makes it hard for the question answering model to learn at the initial training stage, leading to inferior performance. Effect of QA pre-training on the augmented datasets One of the most valuable strengths of We aim to study the pure performance gain of the paradigm shift from mention-pair scoring to query-based span prediction. For this purpose, we replace the mention linking module with the mention-pair scoring module described in We compare our speaker modeling strategy (denoted by Speaker as input), which directly concatenates the speaker's name with the corresponding utterance, with the strategy in documents according to the number of their constituent speakers in Figure Results show that the proposed strategy performs significantly better on documents with a larger number of speakers. Compared with the coarse modeling of whether two utterances are from the same speaker, a speaker's name can be thought of as speaker ID in persona dialogue learning Since the proposed framework has the potential to retrieve mentions missed at the mention proposal stage, we expect it to have higher overall mention recall rate than previous models We examine the proportion of gold mentions covered in the development set as we increase the hyperparameter λ (the number of spans kept per word) in Figure We provide qualitative analyses to highlight the strengths of our model in Table Shown in Example 1, by explicitly formulating the anaphora identification of the company as a query, our model uses more information from a local context, and successfully identifies Freddie Mac as the answer from a longer distance. The model can also efficiently harness the speaker information in a conversational setting. In Example 3, it would be difficult to identify that [Thelma Gutierrez] is the correct antecedent of mention In this paper, we present CorefQA, a coreference resolution model that casts anaphora identification as the task of query-based span prediction in question answering. We showed that the proposed formalization can successfully retrieve mentions left out at the mention proposal stage. It also makes data augmentation using a plethora of existing question answering datasets possible. Furthermore, a new speaker modeling strategy can also boost the performance in dialogue settings. Empirical results on two widelyused coreference datasets demonstrate the effectiveness of our model. In future work, we will explore novel approaches to generate the questions based on each mention, and evaluate the influence of different question generation methods on the coreference resolution task.
| 1,115 | 260 | 1,115 |
Improving Zero-Shot Translation by Disentangling Positional Information
|
Multilingual neural machine translation has shown the capability of directly translating between language pairs unseen in training, i.e. zero-shot translation. Despite being conceptually attractive, it often suffers from low output quality. The difficulty of generalizing to new translation directions suggests the model representations are highly specific to those language pairs seen in training. We demonstrate that a main factor causing the language-specific representations is the positional correspondence to input tokens. We show that this can be easily alleviated by removing residual connections in an encoder layer. With this modification, we gain up to 18.5 BLEU points on zero-shot translation while retaining quality on supervised directions. The improvements are particularly prominent between related languages, where our proposed model outperforms pivot-based translation. Moreover, our approach allows easy integration of new languages, which substantially expands translation coverage. By thorough inspections of the hidden layer outputs, we show that our approach indeed leads to more languageindependent representations. 1
|
Multilingual neural machine translation (NMT) system encapsulates several translation directions in a single model tion and circumvents error propagation. Considering data collection, zero-shot translation does not require parallel data for a potentially quadratic number of language pairs, which is sometimes impractical to acquire especially between low-resource languages. Using less supervised data in turn reduces training time. From a modeling perspective, zero-shot translation calls for language-agnostic representations, which are likely more robust and can benefit low-resource translation directions. Despite the potential benefits, achieving highquality zero-shot translation is a challenging task. Prior works In a typical Transformer encoder • We show that the positional correspondence to input tokens hinders zero-shot translation. We achieve considerable gains on zero-shot translation quality by only removing residual connections once in a middle encoder layer. • Our proposed model allows easy integration of new languages, which enables zero-shot translation between the new language and all other languages previously trained on. • Based on a detailed analysis of the model's intermediate outputs, we show that our approach creates more language-independent representations both on the token and sentence level.
|
Zero-shot inference relies on a model's generalizability to conditions unseen in training. In the context of zero-shot translation, the input should ideally be encoded into an language-agnostic representation, based on which the decoder can translate into any target language required, similar to the notion of an interlingua. Nevertheless, the ideal of "any input language, same representation" cannot be easily fulfilled with a standard encoder, as we have shown in the motivating example in Figure in different languages are likely of varying lengths and word orders, the same semantic meaning will get encoded into different hidden state sequences. There are two potential causes of this positional correspondence: residual connections and encoder self-attention alignment. We further hypothesize that, by modifying these two components accordingly, we can alleviate the positional correspondence. Specifically, we set one encoder layer free from these constraints, so that it could create its own output ordering instead of always following a one-to-one mapping with its input. In the original Transformer architecture from We propose to relax this condition, such that the encoder outputs becomes less position-and hence language-specific. Meanwhile, to minimize the impact on the model architecture and ensure gradient flow, we limit this change to only one encoder layer, and only its multihead attention layer. Figure Besides the residual connections, another potential reason for the positional correspondence is the encoder self-attention alignment. Via the selfattention transform, each position is a weighted sum from all input positions. While the weights theoretically can distribute over all input positions, they are often concentrated locally, particularly with output position i focusing on input position i. Previous works on various sequence tasks In this light, the motivation of our method starts with the formation of the self-attention weight matrix: score(Q, K) = QK T , where Q and K and the query and key matrices. This n × n matrix encapsulates dot product at each position against all n positions. Since the dot product is used as a similarity measure, we hypothesize that when Q and K are similar, the matrix will have heavy weights on the diagonal, thereby causing the positional correspondence. Indeed, Q and K are likely similar since they are projections from the same input. We therefore propose to reduce this similarity by replacing the projection base of the self-attention query by a set of sinusoidal positional encodings. Moreover, to avoid possible interaction with positional information retained in K, we use a wave length for this set of sinusoidal encodings that is different from what is added onto encoder input embeddings. Figure Our experiments cover high-and low-resource languages and different data conditions. We choose an English-centered setup, where we train on X ↔ en parallel data, and test the zero-shot translation between all non-English languages. This scenario is particularly difficult for zero-shot translation, as half of the target-side training data is in English. Indeed, recent works Our datasets originate from three sources: IWSLT 2017 To investigate the role of training data diversity, we construct two conditions for Europarl, where one is fully multiway aligned, and the other has no multiway alignment at all. Both are subsets of the full dataset with 1M parallel sentences per direction. Moreover, we study the challenging case of PMIndia with little training data, distinct writing systems, and a large number of agglutinate languages that are specially difficult to translate into. 1262 Training Details By default we use Transformer We train for 64 epochs and average the weights of the 5 best checkpoints ordered by dev loss. By default, we only include the supervised translation directions in the dev set. The only exception is the Europarl-full case, where we also include the zero-shot directions in dev set for early stopping. When analyzing model hidden representations through classification performance (Subsection 5.1 and 5.2), we freeze the trained encoder-decoder weights and train the classifier for 5 epochs. The classifier is a linear projection from the encoder hidden dimension to the number of classes, followed by softmax activation. As the classification task is lightweight and convergence is fast, we reduce the warmup steps to 400 while keeping the learning rate schedule unchanged. As motivated in Section 2, we modify the residual connections and the self-attention layer in a middle encoder layer. Specifically, we choose the 3-rd and 5-th layer of the 5and 8-layer models respectively. We use "Residual" to indicate residual removal and "Query" the position-based attention query. For the projection basis of the attention query, we use positional encoding with wave length 100. Zero-Shot vs. Pivoting We compare the zeroshot translation performance with pivoting, i.e. directly translating the unseen direction X → Y vs. using English as an intermediate step, as in X → English → Y. The pivoting is done by the baseline multilingual model, which we expect to have similar performance to separately trained bilingual models. For a fair comparison, in the Europarlfull case, pivoting is done by a baseline model trained till convergence with only supervised dev data rather than the early-stopped one. For the languages with Latin script, we first apply the Moses tokenizer and truecaser, and then learn byte pair encoding (BPE) using subword-nmt To simulate the case of later adding a new language, we learn a new BPE model for the new language and keep the previous model unchanged. of the previously-trained model is expanded. In this case, for the model weights related to the word lookup table size, we initialize them as the average of existing embedding perturbed by random noise. Our approach substantially improves zero-shot translation quality, as summarized in Table In the very challenging case of PMIndia (Row 5), while removing residual does improve the zeroshot performance, the score of 2.3 indicates that the outputs are still far from being useful. Nonetheless, we are able to remedy this by further regularization as we will present in Subsection 4.1. Contrary to the large gains by removing residual connections, the attention query modification is not effective when combined with residual removal. This suggests that the primary source of positionspecific representation is the residual connections. Moreover, by contrasting Row 2 and 3 of Table In Subsection 3.2, we hypothesized that variational dropout helps reduce position-specific representation. Table To understand this phenomenon, we create an artificial setup based on IWSLT with 1) no lexical overlap by appending a language tag before each token; 2) extremely low resource by taking a subset of 30K sentences per translation direction. The scores in Table So far our model has shown promising zero-shot performance. Here we extend the challenge of zero-shot translation by integrating a new language. Specifically, we finetune a trained English-centered many-to-many system with a new language using a small amount of X new ↔ English parallel data. At test time, we perform zero-shot translation between X new and all non-English languages previously involved in training. This practically simulates the scenario of later acquiring parallel data between a low-resource language and the central bridging language in an existing system. After finetuning with the new data, we can potentially increase translation coverage by 2N directions, with N being the number of languages originally in training. We finetune a trained system on IWSLT (Row 1 in Table 3) using a minimal amount of de ↔ en data with 14K sentences. When finetuning we include the original X old ↔ en training data, as otherwise the model would heavily overfit. This procedure is relatively lightweight, since the model has already converged on the original training data. In Table To see beyond BLEU scores, we first analyze how much position-and language-specific information is retained in the encoder hidden representations before and after applying our approaches. We then study circumstances where zero-shot translation tends to outperform its pivoting-based counterpart. Lastly, we discuss the robustness of our approach to the impact of different implementation choices. To validate whether the improvements in zero-shot performance indeed stem from less positional correspondence to input tokens, we assess the difficulty of recovering input positional information before and after applying our proposed method. Specifically, we train a classifier to predict the input token ID's (which word it is) or position ID's (the word's absolute position in a sentence) based on encoder outputs. Such prediction tasks have been used to analyze linguistic properties of encoded representation To test whether our model leads to more languageindependent representations, we assess the similarity of encoder outputs on the sentence and token level using the two following methods: SVCCA The singular vector canonical correlation analysis (SVCCA; Language Classification Accuracy Since more similar representations are more difficult to distinguish, poor performance of a language classifier indicates high similarity. Based on a trained model, we learn a token-level linear projection from the encoder outputs to the number of classes (languages). Findings As shown in Table Moreover, we compare the SVCCA scores after each encoder layer, as shown in Figure Given these findings and previous analyses in Subsection 5.1, we conclude that our devised changes in a middle encoder layer allows higher cross-lingual generalizability in top layers while retaining the language-specific bottom layers. In Subsection 4 we have shown that between related languages zero-shot translation surpasses pivoting performance. Here we manually inspect some pivoting translation outputs (nl→en→de) and compare them to zero-shot outputs (de→en). In general, we observe that the translations without pivoting are much more similar to the original sentences. For instance in Table ... geven in dit verband het verkeerde voorbeeld, maar anderen helaas ook. Pivot-in (nl→en) ... are setting the wrong example here, but others are unfortunately also. ... setzen hier das falsche Beispiel ein, andere sind leider auch. Zero-shot (nl→de) ... geben in diesem Zusammenhang das falsche Beispiel, aber leider auch andere. Table In our main experiments, all proposed modifications take place in a middle encoder layer. After comparing the effects of residual removal in each of the encoder layers, our first observation is that the bottom encoder layer should remain fully positionaware. Removing the residual connections in the first encoder layer degrades zero-shot performance by 2.8 BLEU on average on IWSLT. Secondly, leaving out residual connections in top encoder layers (fourth or fifth layer of the five layers) slows down convergence. When keeping the number of training epochs unchanged from our main experiments, it comes with a loss of 0.4 BLEU on the supervised directions. This is likely due to the weaker gradient flow to the bottom layers. The two observations together support our choice of using the middle encoder layer as a transition point. While we use fixed trigonometric positional encodings in our main experiments, we also validate our findings with learned positional embeddings on the IWSLT dataset. First, the baseline still suffers from off-target zero-shot translation (average BLEU scores on supervised directions: 29.6; zeroshot: 4.8). Second, removing the residual connection in a middle layer is also effective in this case (supervised: 29.1; zero-shot: 17.1). These findings suggest that our approach is robust to the form of positional embedding. Although learned positional embeddings are likely more language-agnostic by seeing more languages, as we still present source sentences as a sequence of tokens, the residual connections, when present in all layers, would still enforce a one-to-one mapping to the input tokens. This condition allows our motivation and approach to remain applicable. Initial works on multilingual translation systems already showed some zero-shot capability Recently, models with full parameter sharing have gained popularity, with massively multilingual systems showing encouraging results Another line of work on improving zero-shot translation utilizes monolingual pretraining While our work concentrates on an Englishcentered data scenario, another promising direction to combat zero-shot conditions is to enrich available training data by mining parallel data between non-English languages In this work, we show that the positional correspondence to input tokens hinders zero-shot translation. Specifically, we demonstrate that: 1) the encoder outputs retain word orders of source languages; 2) this positional information reduces cross-lingual generalizability and therefore zero-shot translation quality; 3) the problems above can be easily alleviated by removing the residual connections in one middle encoder layer. With this simple modification, we achieve improvements up to 18.5 BLEU points on zero-shot translation. The gain is especially prominent in related languages, where our proposed model outperforms pivot-based translation. Our approach also enables integration of new languages with little parallel data. Similar to interlingua-based models, by adding two translation directions, we can increase the translation coverage by 2N language pairs, where N is the original number of languages. In terms of model representation, we show that the encoder outputs under our proposed model are more languageindependent both on a sentence and token level. This work is supported by a Facebook Sponsored Research Agreement. We thank Yuqing Tang for helpful comments, and Ngoc Quan Pham for sharing the training details of We proposed approaches to improve zero-shot translation, which is especially suitable to lowresource scenarios with no training data available between some languages. We also validated our approaches on actual low-resource languages. However, as the models are trained on single domains, when facing out-of-domain test sentences, they could suffer from hallucination, i.e. produce translations unrelated to the input sentences.
| 1,142 | 1,333 | 1,142 |
DisSent: Learning Sentence Representations from Explicit Discourse Relations
|
Learning effective representations of sentences is one of the core missions of natural language understanding. Existing models either train on a vast amount of text, or require costly, manually curated sentence relation datasets. We show that with dependency parsing and rule-based rubrics, we can curate a high quality sentence relation task by leveraging explicit discourse relations. We show that our curated dataset provides an excellent signal for learning vector representations of sentence meaning, representing relations that can only be determined when the meanings of two sentences are combined. We demonstrate that the automatically curated corpus allows a bidirectional LSTM sentence encoder to yield high quality sentence embeddings and can serve as a supervised fine-tuning dataset for larger models such as BERT. Our fixed sentence embeddings achieve high performance on a variety of transfer tasks, including Sen-tEval, and we achieve state-of-the-art results on Penn Discourse Treebank's implicit relation prediction task.
|
Developing general models to represent the meaning of a sentence is a key task in natural language understanding. The applications of such generalpurpose representations of sentence meaning are many -paraphrase detection, summarization, knowledge-base population, question-answering, automatic message forwarding, and metaphoric language, to name a few. We propose to leverage a high-level relationship between sentences that is both frequently and systematically marked in natural language: the discourse relations between sentences. Human writers naturally use a small set of very common transition words between sentences (or sentence-like * equal contribution phrases) to identify the relations between adjacent ideas. These words, such as because, but, and, which mark the conceptual relationship between two sentences, have been widely studied in linguistics, both formally and computationally, and have many different names. We use the name "discourse markers". Learning flexible meaning representations requires a sufficiently demanding, yet tractable, training task. Discourse markers annotate deep conceptual relations between sentences. Learning to predict them may thus represent a strong training task for sentence meanings. This task is an interesting intermediary between two recent trends. On the one hand, models like InferSent We thus propose the DisSent task, which uses the Discourse Prediction Task to train sentence embeddings. Using a data preprocessing procedure based on dependency parsing, we are able to automatically curate a sizable training set of sentence pairs. We then train a sentence encoding model to learn embeddings for each sentence in a pair such that a classifier can identify, based on the embeddings, which discourse marker was used to link the sentences. We also use the DisSent task to fine-tune larger pre-trained models such as BERT. We evaluate our sentence embedding model's performance on the standard fixed embedding evaluation framework developed by We demonstrate that the resulting DisSent embeddings achieve comparable results to InferSent on some evaluation tasks, and superior on others. The BERT model fine-tuned on the DisSent tasks achieved state-of-the-art on PDTB classification tasks compared to other fine-tuning strategies. 2 Discourse Prediction Task With this focus in mind, we propose a new task for natural language understanding: discourse marker prediction. Given two sentences in our curated corpus (which may have been full sentences in the original text or may have been subclauses), the model must predict which discourse marker was used by the author to link the two ideas. For example, "She's late to class she missed the bus" would likely be completed with because, but "She's sick at home she missed the class" would likely be completed with so, and "She's good at soccer she missed the goal" would likely be completed with but. These pairs have similar syntactic structures and many words in common, but the meanings of the component sentences lead to strong intuitions about which discourse marker makes the most sense. Without a semantic understanding of the sentences, we would not be able to guess the correct relation. We argue that success at choosing the correct discourse marker requires a representation that reflects the full meaning of a sentence. We note that perfect performance at this task is impossible for humans
|
We adapt the best architecture from We apply global max pooling to construct the encoding for each sentence. That is, we apply an element-wise max operation over the temporal dimension of the hidden states. Global max pooling builds a sentence representation from all time steps in the processing of a sentence -→ h t = LSTM t (w 1 , ..., w t |θ 1 ) ←h t = LSTM t (w T , ..., w t |θ 2 ) Our objective is to predict the discourse relations between two sentences from their vectors, s i where i ∈ {1, 2}. Because we want generally useful sentence vectors after training, the learned computation should happen before the sentences are combined to make a prediction. However, some non-linear interactions between the sentence vectors are likely to be needed. To achieve this, we include a fixed set of common pair-wise vector operations: subtraction, multiplication, and average. Finally we use an affine fully-connected layer to project the concatenated vector S down to a lower dimensional representation, and then project it down to a vector of label size (the number of discourse markers). We use softmax to compute the probability distribution over discourse relations. Sentence relations datasets can be used to provide high-level training signals to fine-tune other sentence embedding models. In this work, we fine-tune BERT After training BERT-base model on the DisSent task, we continue to fine-tune BERT-base model on other evaluation tasks to see if training on Dis-Sent tasks provides additional performance improvement and learning signal for the BERT-base model. We present an automatic way to collect a large dataset of sentence pairs and the relations between them from natural text corpora using a set of explicit discourse markers and universal dependency parsing For training and evaluation datasets, we collect sentence pairs from BookCorpus Many discourse markers in English occur almost exclusively between the two statements they connect, but for other discourse markers, their position relative to their connected statements can vary (e.g. Figure Manning, 2016) to extract the appropriate pairs of sentences (or sentence-like EDUs) for a discourse marker, in the appropriate conceptual order. Each discourse marker, when it is used to link two statements, is parsed by the dependency parser in a systematic way, though different discourse markers may have different corresponding dependency patterns linking them to their statement pairs. 1 Within the dependency parse, we search for the governor phrase (which we call "S2") of the discourse marker and check for the appropriate dependency relation. If we find no such phrase, we reject the example entirely (thus filtering out polysemous usages, like "that's so cool!" for the discourse marker so). If we find such an S2, we search for "S1" within the same sentence (SS). Searching for this relation allows us to capture pairs where the discourse marker starts the sentence and connects the following two clauses (e.g. "Because [it was cold outside] S2 , [I wore a jacket] S1 ."). If a sentence in the corpus contains only a discourse marker and S2, we assume the 1 See Appendix A.1 for more details on dependency-based extraction. discourse marker links to the immediately previous sentence (IPS), which we label S1. For some markers, we further filter based on the order of the sentences in the original text. For example, the discourse marker then always appears in the order "S1, then S2", unlike because, which can also appear in the order "Because S2, S1". Excluding proposed extractions in an incorrect order makes our method more robust to incorrect dependency parses. Using these methods, we curated a dataset of 4,706,292 pairs of sentences for 15 discourse markers. Examples are shown in Table 3. We randomly divide the dataset into train/validation/test set with 0.9, 0.05, 0.05 split. The dataset is inherently unbalanced, but the model is still able to learn rarer classes quite well (see Appendix A.4 for more details on the effects of class frequencies). Our data are publicly available at Current state of the art models either rely on completely supervised learning through high-level classification tasks or unsupervised learning. Supervised learning has been shown to yield general-purpose representations of meaning, training on semantic relation tasks like Stanford Natural Language Inference (SNLI) and MultiNLI In an unsupervised learning setting, SkipThought Jernite et al. ( For all our models, we tuned hyperparameters on the validation set, and report results from the test set. We use stochastic gradient descent with initial learning rate 0.1, and anneal by the factor of 5 each time validation accuracy is lower than in the previous epoch. We train our fixed sentence encoder model for 20 epochs, and use early stopping to prevent overfitting. We also clip the gradient norm to 5.0. We did not use dropout in the fully connected layer in the final results because our initial experiments with dropout showed lower performance when generalizing to SentEval. We experimented with both global mean pooling and global max pooling and found the later to perform much better at generalization tasks. All models we report used a 4096 hidden state size. We are able to fit our model on a single Nvidia Titan X GPU. Fine-tuning We fine-tune the BERT-base model on the DisSent tasks with 2e-5 learning rate for 1 epoch because all DisSent tasks corpora are quite large and fine-tuning for longer epochs did not yield improvement. We fine-tune BERT on other supervised learning datasets for multiple epochs and select the epoch that provides the best performance on the evaluation task. We find that finetuning on MNLI is better than on SNLI or both combined. This phenomenon is also discussed in Discourse Marker Set We experimented with three subsets of discourse markers, shown in Table 4. We first trained over all of the discourse markers in our ALL marker set. The model achieved 67.5% test accuracy on this classification task. Overall we found that markers with similar meanings tended to be confusable with one another. A more detailed analysis of the model's performance on this classification task is presented in Appendix A.4. Because there appears to be intrinsic conceptual overlap in the set of ALL markers, we experimented on different subsets of discourse markers. We choose sets of 5 and 8 discourse markers that were both non-overlapping and frequent. The set of sentence pairs for each smaller dataset is a strict subset of those in any larger dataset. Our chosen sets are shown in Table Marked vs Unmarked Prediction Task Adjacent sentences will always have a relationship, but some are marked with discourse markers while others are not. Humans have been shown to perform well above chance at guessing whether a discourse marker is marked vs. unmarked We predict that high quality sentence embeddings will contain useful information to determine whether a discourse relation is explicitly marked. Furthermore, success at this task could help natural language generation models to generate more human-like long sequences. To test this prediction, we create an additional set of tasks based on Penn Discourse Treebank Implicit Relation Prediction Task SentEval Tasks We evaluate the performance of generated sentence embeddings from our fixed sentence encoder model on a series of natural language understanding benchmark tests provided by Training Task On the discourse marker prediction task used for training, we achieve high levels of test performance for all discourse markers. (Though it is interesting that because, perhaps the conceptually deepest relation, is also systematically the hardest for our model.) The larger the set of discourse markers, the more difficult the task becomes, and we therefore see lower test accuracy despite larger dataset size. We conjecture that as we increase the number of discourse markers, we also increase the ambiguity between them (semantic overlap in discourse markers' meanings), which may further explain the drop in performance. The training task performance for each subset is shown in Table Discourse Marker Set Varying the set of discourse markers doesn't seem to help or hinder the model's performance on generalization tasks. Top generalization performance on the three sets of discourse markers is shown in Table Marked vs Unmarked Prediction Task In determining whether a discourse relation is marked or unmarked, DisSent models outperform In-ferSent and SkipThought (as well as previous approaches on this task) by a noticeable margin. Much to our surprise, fine-tuned BERT models are not able to perform better than the BiLSTM sentence encoder model. We leave explorations of this phenomenon to future work. We report the results in Table discourse relations much better than InferSent and SkipThought. DisSent outperforms word vector models evaluated by SentEval Tasks Results of our models, and comparison to other approaches, are shown in Table DisSent and InferSent do well on different sets of tasks. In particular, DisSent outperforms In-ferSent on TREC (question-type classification). InferSent outperforms DisSent on the tasks most similar to its training data, SICK-R and SICK-E. These tasks, like SNLI, were crowdsourced, and seeded with images from Flickr30k corpus Although DisSent is trained on a dataset derived from the same corpus as SkipThought, DisSent almost entirely dominates SkipThought's performance across all tasks. In particular, on the SICK dataset, DisSent and SkipThought perform similarly on the relatedness task (SICK-R), but Dis-Sent strongly outperforms SkipThought on the entailment task (SICK-E). This discrepancy highlights an important difference between the two models. Whereas both models are trained to, given a particular sentence, identify words that appear near that sentence in the corpus, DisSent focuses on learning specific kinds of relationships between sentences -ones that humans tend to explicitly mark. We find that reducing the model's task to only predicting a small set of discourse relations, rather than trying to recover all words in the following sentence, results in better features for identifying entailment and contradiction without losing cues to relatedness. Overall, on the evaluation tasks we present, Dis-Sent performs on par with previous state-of-theart models and offers advantages in data collection and training speed. We evaluate our extraction quality by comparing the manually extracted and annotated sentence pairs from Penn Discourse Treebank (PDTB) to our automatic extraction of sentence pairs from the source corpus Penn Treebank (PTB). On the majority of discourse markers, we can achieve a relatively high extraction precision. We apply our extraction pipeline on raw PTB dataset because we want to see how well our pipeline converts raw corpus into a dataset. Details of our alignment procedure is described in Appendix A.2. Overall, even though we cannot construct the explicit discourse prediction section of the PDTB dataset perfectly, training with imprecise extraction has little impact on the sentence encoder model's overall performance. We compute the extraction precision as the percentage of PTB extracted pairs that can be successfully aligned to PDTB. In Figure We show in Figure Identifying other features of natural text that contain informative signals of discourse structure and combining these with explicit discourse markers is an appealing direction for future research. In principle, the DisSent model and extraction methods would apply equally well to multilingual data with minimal language-specific modifications. Within universal dependency grammar, discourse markers across languages should correspond to structurally similar dependency patterns. Beyond dependency parsing and minimal marker-specific pattern development (see Appendix A.1), our extraction method is automatic, requiring no annotation of the original dataset, and so any large dataset of raw text in a language can be used. We present a discourse marker prediction task for training and fine-tuning sentence embedding models. We train our model on this task and show that the resulting embeddings lead to high performance on a number of established tasks for sentence embeddings. We fine-tune larger models on this task and achieve state-of-the-art on the PDTB implicit discourse relation prediction. A dataset for this task is easy to collect relative to other supervised tasks. Dependency parsing allows us to design our extraction method such that each S1 and S2 is interpretable as a full sentence in isolation, and the appropriate conceptual relation holds between the pair. However, occasionally we get ungrammatical sentences or the wrong pair of sentences for a relation. This incorrect extraction can happen in several ways. First, we might choose grammatical but incorrect pairs. We preprocess the PTB sentences by limiting the vocabulary size to 10,000 and tokenizing numbers. Then we run our extraction pipeline on the preprocessed PTB. We apply the same preprocessing to the PDTB sentences. We refer to the gold sentence pair from the PDTB as (G1, G2), and our extracted sentence pair from the PTB as (S1, S2). We first obtain the minimum of S1-G1 distance and S2-G2 distance over all gold pairs. If this distance is smaller than 0.7, we consider the corresponding gold pair to be an alignment for this extracted pair. Given an aligned pair ((G1, G2), (S1, S2)), we measure the extraction quality by computing the average of normalized G1-S1 and G2-S2 distance. We compute this distance for all pairs and all discourse markers. We analyze our extraction quality in two steps: align sentence pairs from the two datasets and then calculate extraction quality on each aligned pair. In the alignment step, for each extracted pair, we calculate its distance to all pairs from PDTB using the normalized Levenshtein distance: For each pair of connected sentences, whose relation type has been labeled in PDTB, the discourse relation between them may have been explicitly marked (via a discourse relation word) or not. We can pose the task of a binary classification of whether the sentence pair appeared as explicitly or implicitly marked, given only the two sentences and no additional information. We evaluate DisSent and InferSent sentence embedding models and a word vector baseline on this trask. We follow This task is different from the setting in Patterson and Kehler (2013). We do not allow the classifier to access the underlying discourse relation type and we only provide the individual sentence embeddings as input features. In contrast, To investigate the qualitative relations among our largest set of discourse markers, the ALL marker set, we build a confusion matrix of the test set classifications. Figure Classifications errors are much more common for semantically similar discourse marker pairs than would be expected from frequency alone. The most common confusion is when the synonymous marker although is mistakenly classified as but. The temporal relation markers before, after and then, intuitively very similar discourse markers, are rarely confused for anything but each other. The fact that they are indeed confusable may reflect the tendency of authors to mark temporal relation primarily when it is ambiguous. Figure To more quantitatively represent the connection between what the two models learn, we compute the correlation between the balanced confusions and the residuals of the unbalanced confusions (when predicted linearly from log fre-quency). These residuals account for 64% of the variance in the balanced confusions (R 2 = 0.6431, F (1, 223) = 401.8, p < .001). That is, we come close to predicting the balanced confusions from the unbalanced ones. As a reference point for training task performance we present baseline performance. Note that a model which simply chose the most common class would perform with 21.79% accuracy on the ALL task, 28.35% on the BOOKS 8 task, and 31.87%
| 1,039 | 3,410 | 1,039 |
Unsupervised Extractive Opinion Summarization Using Sparse Coding
|
Opinion summarization is the task of automatically generating summaries that encapsulate information from multiple user reviews. We present Semantic Autoencoder (SemAE) to perform extractive opinion summarization in an unsupervised manner. SemAE uses dictionary learning to implicitly capture semantic information from the review and learns a latent representation of each sentence over semantic units. A semantic unit is supposed to capture an abstract semantic concept. Our extractive summarization algorithm leverages the representations to identify representative opinions among hundreds of reviews. Se-mAE is also able to perform controllable summarization to generate aspect-specific summaries. We report strong performance on SPACE and AMAZON datasets, and perform experiments to investigate the functioning of our model. Our code is publicly available at
|
Opinion summarization is the task of automatically generating digests for an entity (e.g. a product, a hotel, a service, etc.), from user opinions in online forums. Automatic opinion summaries enable faster comparison, search, and better consumer feedback understanding This lack of labeled training data has prompted a series of works to leverage unsupervised or weaklysupervised techniques for opinion summarization To overcome these issues, another thread of works focuses on extractive opinion summarization, which creates summaries by selecting review sentences to reflect the popular opinions corresponding to an entity. A recently proposed extractive summarization approach is Quantized Transformer (QT) Building on the framework introduced by QT, we introduce an unsupervised extractive model, Semantic Autoencoder (SemAE), which learns a representation of text over latent semantic units using dictionary learning • We present Semantic Autoencoder (SemAE), which learns representation of sentences over latent semantic units. • We introduce novel inference algorithms for general and controllable summarization utilizing information-theoretic measures. • We show that SemAE outperforms previous methods using automatic and human evaluations. • We perform analysis to understand how the learnt representations align with human semantics.
|
Unsupervised opinion summarization can be conducted either abstractively or extractively. Abstractive approaches aim to summarize the opinion text using novel phrases. Traditional statistical approaches create abstractive summaries using graphical paths In contrast to abstractive approaches, extractive approaches rank and select a subset of salient sentences from reviews to form a concise summary Our work is most similar to the extractive opinion summarization QT We follow the task setup in The intuition behind Semantic Autoencoder is that instead of representing text as a single latent semantic unit, we represent text as a distribution over latent semantic units using dictionary learning. Learning semantic representations over a common dictionary makes them structurally aligned, enabling comparison of sentences using informationtheoretic measures. Semantic Autoencoder consists of three stages (i) sentence encoding -an input sentence s is converted into a multi-head representation (H heads) using Transformer encoder {s h } H h=1 ; (ii) reconstruction -a latent representation of head vectors s h is formed over elements of the dictionary D ∈ R K×d , to produce reconstructed representations z = {z h } H h=1 ; and (iii) sentence decodinga Transformer-based decoder takes as input the reconstructed representations z to produce the output sentence ŝ. SemAE is trained on the sentence reconstruction task. The overall workflow of SemAE Reconstruction ( ŝ ) × Figure is shown in Figure We follow the setup of QT where W ∈ R d×d/H , b ∈ R d are trainable parameters and s h ∈ R d is the h th head representation. For each s h , we obtain a latent representation α h over the dictionary D, by reconstructing the encoded sentence representation s h as shown below where the reconstructed vector z h ∈ R d , and the latent representation α h ∈ R K . We hypothesize that the dictionary D captures the representation of latent semantic units, and α h captures the degree to which the text encapsulates a certain semantic. The vectors formed z = {z h } H h=1 are forwarded to the decoder for sentence reconstruction. The dictionary D and s h are updated simultaneously using backpropagation. For summarization (Section 5), different from QT, we consider α h (not z h ) as the sentence representation. We employ a Transformer-based decoder that takes as input the reconstructed representations z = {z h } H h=1 . MultiHead(z, z, t) attention module in the decoder takes z as key and value, and the target tokens t as the query. The reconstructed sentence is generated from the decoder as ŝ = Decoder(z, t). As our goal is sentence reconstruction, we set the target tokens to be same as the input sentence s. Prior work A sentence can capture only a small number of semantic senses. We ensure this by enforcing sparsity constraints on the representations α h , so that z h is a combination of only a few semantic units. The encoder, reconstructor and decoder are trained together to minimize the loss function: where L CE is the reconstruction cross-entropy loss of the decoder, and to ensure sparsity of α h we penalize the L1-norm (|α h |) and its entropy H(α h ). We leverage the latent representations α h generated by SemAE to perform opinion summarization. For obtaining the general summary of an entity, we first compute a mean representation of all the review sentences in S e , which represents the aggregate distribution over semantic units. Thereafter, the general summary is obtained as the collection of sentences that resemble the mean distribution. Mathematically, every sentence s is associated with a representation over dictionary elements α s = [α 1 , . . . , α H ], where α s ∈ R H×K . We form the mean representation of review sentences for an entity S e over dictionary elements as: where α s is the representation for sentence s ∈ S e . For general summarization, we compute the relevance score R(•) for each sentence s based on its similarity with the mean representation ᾱ: where α s h is latent representation of sentence s for the h th head. ∆(x, y) denotes the similarity between two representations x and y. It is implemented as negation of the sum of KL-divergence between head representations. We also experimented with other divergence metrics and observed similar summarization performance (Appendix A.3). We rank sentences according to descending order of R(•) and select the top N (a constant hyperparameter, N < |S e |) sentences as the summary O e (shown in Figure where Ôe is the set of sentences selected so far for the summary. The selection routine proceeds in a greedy fashion by choosing s 0 = arg max s∈Se ∆(ᾱ, α s ) when Ôe = φ. (b) Aspect-awareness: Another drawback with sentence selection using ∆(•, •) is that the summary frequently switches context among different aspects (example shown in Table • We iterate over sentence buckets {S (a i ) e } and select the first m sentences ranked according to R(α s ), from each bucket. • We prevent selection of similar sentences from a bucket by introducing the redundancy term. We iterate over individual buckets and select first m sentences ranked according to their relevance R(α s , Ô(a) e ) (Equation SemAE can perform aspect summarization without needing additional training. For this, we require a small set of keywords to identify sentences that talk about an aspect. For example, food aspect is captured using keywords: "breakfast", "buffet" etc. For a given aspect a, let the keyword set be e for each entity e, belonging to aspect a from a held-out dev set S dev . Similar to general summarization, we proceed by computing the mean representation of sentences S (a) e belonging to the aspect a: We then select sentences most similar to the mean representation as the summary. (a) Informativeness: Sentences selected for aspect summarization should talk about the aspect but not the general information. We model informativeness We rank sentences s ∈ S e according to their aspect-specific relevance score R a (•), and select first N sentences as the summary for aspect O In this section, we discuss the experimental setup, results and analysis. We evaluated our model on two public customer review datasets SPACE hotel reviews We build on the implementation framework introduced by We have K = 1024 dictionary elements, each with dimension d = 320. The dictionary elements are initialized using k-means clustering of review sentence representations. All hyperparameters were tuned on the development set (see Appendix A.1 for more details). We report ROUGE F-scores that compares the overlap between generated text with gold summaries. For SPACE dataset, we measure how much general summaries cover different aspects by computing the mean ROUGE-L score with the gold aspect summaries (denoted by RL ASP ). We also compute perplexity (PPL) score to evaluate the readability of summaries. Perplexity is computed using cross-entropy loss from a BERTbase model. We measure aspect coverage of a sys-tem, by computing the average number of distinct aspects N ASP in the generated summaries. Lastly, to evaluate repetition in summaries, we compute the percentage of distinct n-grams (n = 2). Following prior work General Summarization: We present the results of general summarization on SPACE dataset in Table 2. SemAE and its variants show strong improvements over previous state-of-the-art QT, and other baselines, across all ROUGE metrics. They also outperform abstractive systems (like CopyCat and Meansum) by a large margin, which shows that SemAE can effectively select relevant sentences from a large pool of reviews. All variants of Se-mAE outperform other models in RL ASP metric, showcasing that general summaries from SemAE cover aspects better than baselines. We compiled some baseline results from We further evaluate the quality of the summaries, for all variations of SemAE along with our strongest baseline QT, using other automatic metrics in Table Evaluation results on SPACE are reported in Table 5. SemAE outperforms the state-of-the-art QT in all aspects except cleanliness, where the performance is comparable. We observe that adding the informativeness term (∆(ᾱ, α s ) in Equation We also evaluated aspect summaries generated by SemAE and QT in a similar manner. Aspect summaries were judged based on two criteria: aspect informativeness (usefulness of opinions for a specific aspect, consistent with reference) and aspect specificity (how specific the summary is for an aspect without considering other factors). The bottom half of Table Latent Dictionary Interpretation. In this section, we investigate the semantic meanings learnt by individual dictionary elements, D k . We visualized the UMAP projection We hypothesize that the clusters should capture certain semantic meaning. We explore this hypothesis by identifying sentences sharing similar representations with the mean representations SemAE SemAE (w/ redun.) SemAE (w/ aspect) SemAE (w/ aspect + redun.) The staff is great. The Hotel Erwin is a great place to stay. The staff were friendly and helpful. The location is perfect. We ate breakfast at the hotel and it was great. The hotel itself is in a great location. The service was wonderful. It was great. The rooms are great. The rooftop bar HIGH was the icing on the cake. The food and service at the restaurant was awesome. The service was excellent. The hotel itself is in a great location. The rooms were clean and we were on the 5th. The best part of the hotel is the 7th floor rooftop deck. The staff is great. The hotel has so many advantages over the other options in the area that it is a no contest. If you want to stay in Venice, this is a great place to be. The food and service at the restaurant was awesome. The staff is great. The staff were friendly and helpful. The Hotel Erwin is a great place to stay. The location is perfect. We ate breakfast at the hotel and it was great. The food and service at the restaurant was awesome. The rooms are great. The room is epic! The rooftop bar HIGH was the icing on the cake. The rooftop bar at the hotel, "High", is amazing. The staff is great. We had a great stay at the Erwin, and the staff really made it more enjoyable. The Hotel Erwin is a great place to stay. It was great. We ate breakfast at the hotel and it was great. The food and service at the restaurant was awesome. The rooms are great. We had a kitchen and balcony and partial ocean view. The rooftop bar HIGH was the icing on the cake. (h, k) Sentences w/ high activation Explanation (3, 5) • I wish all hotels or any business for that matter, had employees a dedicated to service as he was. • Very polite and very professional approach. Service (0, 10) • Stayed here in August for the our first trip to Vancouver. • I stayed at this motel with my partner in August 2010. Phrase "stayed" (6, 0) • Empty water bottles were never thrown out and no one put the iron and ironing board away. • Facing St Paul St can be a very noisy experience. (2, 8) • A full cooked to order breakfast (including omlettes, . . . , fruit, etc.) • Pizza hut, Mc donalds, KFC all round the corners... (5, 8) • The rooms seem small, tight fit for a family of 4. • You may have a difficult fit. Table {µ 1 , . . . , µ K } for each cluster. For each head h in the encoder (Section 4.1), we compute cosine similarity of sentences with cluster means. Table We also perform multi-aspect summarization using SemAE, by controlling the aspect of the selected sentences. Table In Appendix A.5, we perform additional analysis to investigate the head-wise analysis, efficacy of sparsity constraints, dictionary evolution, and qualitatively compare SemAE with baselines (QT and CopyCat). Love the warm chocolate chips cookies and the service has always been outstanding. Excellent morning breakfasts and the airport shuttle runs every 15 minutes but we have made the 10 minute walk numerous times to the airport terminal. To add insult to injury, for people who use the parking lot to "park and fly", the charge is $7.95/day, almost half of what the hotel guests are charged!! Cons -Hotel is spread out so pay attention to how to get to your room as you may get lost, Feather pillows (synthetic available on request), Pay parking ($16 self/day $20 valet/day), warm cookies on check in. Stayed at this hotel beause the park n fly. We have stayed at this hotel several times in the family suite ( 2 bedrooms/1 king and 2 queen beds). Despite the enormity of this hotel, it very much feels almost family run. The staff was friendly and helpful and we enjoyed the warm, chocolate chip cookie we were given at check-in. The breakfast in the restaurant was amazing, and the staff was very attentive. (room, cleanliness) The bed was very nice, room was clean, we even had a balcony. The beds were comfortable and the room was very clean. Table We proposed a novel opinion summarization approach using Semantic Autoencoder, which encodes text as a representation over latent semantic units. We perform extractive summarization by selecting sentences using information-theoretic measures over representations obtained from SemAE. Our experiments reveal that dictionary element representations from SemAE form clusters, which capture distinct semantics. Our model provides finegrained control to users to model surface-level text attributes (like redundancy, informativeness etc.) in the representation space. SemAE outperforms existing extractive opinion summarization methods on SPACE and AMAZON datasets. Finally, SemAE representations can be leveraged to explore different forms of control on the summary generation (e.g. multi-aspect sumamrization) using our inference framework. Future works can focus on better representation learning systems to handle use-cases with noisy or sparse textual data. DATASET λ 1 λ 2 AMAZON 10 3 5 × 10 -4 SPACE 10 4 5 × 10 -4 Table A.1 Implementation Details The Transformer is trained without the dictionary learning reconstruction for 4 warmup epochs. We tokenized text in an unsupervised manner using SentencePiece The model was trained using Adam Optimizer with a learning rate of 10 -3 , and a weight decay of 0.9. Our model was trained for 10 epochs on a single GeForce GTX 2080 Ti GPU in 35 hours. The loss function parameters are reported in Table Post training, the summaries were generated with N = 20. We limit the summary length to 75 tokens. Each keyword w i ∈ Q a is associated with a confidence score for aspect a. In case a sentence has multiple keywords belonging to different aspects we use the confidence score to assign the aspect. In this section, we provide some background information about the dataset creation process for SPACE and AMAZON. SPACE corpus has a large number of reviews per entity. Therefore, • Divergence metric: SemAE uses KL divergence to measure the relevance of a sentence α s when compared to the mean ᾱ, we used KLdivergence earlier. In this setup, we experiment with cosine similarity as our divergence function ∆(•, •). The modified divergence ∆(•, •) score is defined as: The second row in Table • Informativeness: In this ablation experiment, we incorporate the informativeness term in general summarization. The modified relevance score is: where α (b) = E [α s ], the mean representation of all sentences across all entities. α (b) captures background knowledge distribution For both ablation setups, we observe almost no change in perplexity, aspect coverage and distinct n-grams metrics. In this setup, we modify selection mechanism of SemAE by updating the mean representation every time a sentence is selected. We consider the mean of the sentences that have not been selected so far. The intuition behind this approach is that the next selected sentence should best capture information, which is not present in the summary so far. The sentence selection process is described below: where α s t is the representation selected at time step t, ᾱt is mean representation of the set of sentences that are not part of the summary yet and Ôe is the set of selected sentences so far. Table The overall representation for the barycenter is Next, we derive the relevance score of each sentence s with the barycenter as: As shown in Equation (c) Clustering-based Sentence Selection: In this setup, instead of selecting sentences similar to the mean representation, we identify clusters formed by the representations. For clustering we flatten the sentence representation α s ∈ R HK , and use k-means where α C is the representation of the cluster center where s belongs, and |C| is the size of the cluster. The first term in Equation 15 penalizes the relevance of a sentence for being too far away from the cluster center, and the second term selection of samples from a large cluster. The hyperparameters γ = 0.005, K = 5 in our experiments, were selected using the development set performance. In Table A.5 Extended Analysis configurations of sparsity losses. Specifically, we gauge SemAE' performance when L1-loss and entropy loss are removed. Table (b) Head-wise Analysis: We analyze whether there is a correlation between the head-wise representations and clusters formed by dictionary elements. For each dictionary element, we compute the average attention (α h ) it receives from each head h, and assign the element to a head where it received the maximum mean attention (head-wise dictionary elements are shown in Figure The service was excellent. Great hotel. We liked our room with an ocean view. The staff were friendly and helpful. There was no balcony. The location is perfect. Our room was very quiet. I would definitely stay here again. You're one block from the beach. So it must be good! Filthy hallways. Unvacuumed room. Pricy, but well worth it. This hotel is in a great location, just off the beach. The staff was very friendly and helpful. We had a room with a view of the beach and ocean. The only problem was that our room was on the 4th floor with a view of the ocean. If you are looking for a nice place to sleep then this is the place for you. Table Food: The food and service at the restaurant was awesome. The food at Hash, the restaurant just off of the lobby, was fabulous for breakfast. The food was excellent (oatmeal, great wheat toast, freshberries and a tasty corned beef hash). Location: The Hotel Erwin is a great place to stay. The hotel is not only in the perfect location for the ideal LA beach experience, but it is extremely hip and comfortable at the same time. Cleanliness: The room was spacious and had really cool furnishings, and the beds were comfortable. The room itself was very spacious and had a comfortable bed. We were upgraded to a partial ocean view suite and the room was clean and comfortable. Service: The hotel staff were friendly and provided us with great service. The staff were friendly and helpful. The staff was extremely helpful and friendly. The hotel staff was friendly and the room was well kept. The rooftop bar at the hotel, "High", is amazing. The rooftop bar HIGH was the icing on the cake. The Hotel Erwin is a great place to stay. The best part of the hotel is the 7th floor rooftop deck. The room was spacious and had really cool furnishings, and the beds were comfortable. The room itself had a retro 70's feel with a comfortable living room and kitchen area, a separate bedroom with a nice king size bed, and a sink area outside the shower/toilet area. Table tiple dictionary element clusters (h = 0, 3, 5, 7) perform better than heads where instances are concentrated over few clusters (h = 1, 2). (c) Output summaries: Table (d) Evolution of Dictionary Representations: We plot the UMAP projections of dictionary elements from epochs 4 (after encoder warmup is complete) to 10 in Figure (e) Ablations with QT: In this section, we analyze the efficacy of our sentence selection (SS) module. We evaluate the summarization performance using our sentence selection scheme by retrieving sentence representations from QT and SemAE. The experiments were performed using 5% data from the SPACE dataset. For QT's representations, we obtain α h (Equation In Table
| 862 | 1,345 | 862 |
Adversarial Learning for Discourse Rhetorical Structure Parsing
|
Text-level discourse rhetorical structure (DRS) parsing is known to be challenging due to the notorious lack of training data. Although recent top-down DRS parsers can better leverage global document context and have achieved certain success, the performance is still far from perfect. To our knowledge, all previous DRS parsers make local decisions for either bottomup node composition or top-down split point ranking at each time step, and largely ignore DRS parsing from the global view point. Obviously, it is not sufficient to build an entire DRS tree only through these local decisions. In this work, we present our insight on evaluating the pros and cons of the entire DRS tree for global optimization. Specifically, based on recent well-performing top-down frameworks, we introduce a novel method to transform both gold standard and predicted constituency trees into tree diagrams with two color channels. After that, we learn an adversarial bot between gold and fake tree diagrams to estimate the generated DRS trees from a global perspective. We perform experiments on both RST-DT and CDTB corpora and use the original Parseval for performance evaluation. The experimental results show that our parser can substantially improve the performance when compared with previous state-of-the-art parsers. * Corresponding author [e 1 : In fact,] [e 2 : Budget indicated] [e 3 : it saw some benefit] [e 4 : to staying involved in these programs,] [e 5 : in which renters earn frequent-flier miles] [e 6 : and fliers can get car-rental discounts.
|
As the main linguistic theory on discourse rhetorical structure (DRS), Rhetorical Structure Theory (RST) With the release of various discourse corpora, text-level DSR parsing has been drawing more and more attention in the last decade. However, since the corpus annotation is usually time-consuming, existing DRS corpora are much limited in size. For example, the English RST-DT Theoretically, in supervised learning, annotated data corpora can provide neural models with specific learning objectives, and the corpus size limitation will weaken the learning of these goals. To mitigate this problem, we researchers need (i) an efficient model to better learn from the limited data and (ii) more high-quality training objectives to enhance the model learning. Existing studies on text-level DRS parsing show that • Compared with bottom-up DRS parsers, recent top-down frameworks can better leverage global document context and have achieved promising results in text-level DRS parsing • All previous studies produce their DRS parsers with local decisions made at each time step for either bottom-up node composition or top-down split point selection (Figure Considering the above-mentioned status quo, in this work, we study a global optimization method based on the well-performing top-down parsers. For model structure, we take the top-down parser of
|
In the literature, previous studies on RST-style DRS parsing mainly consist of two categories, i.e., bottom-up and top-down frameworks. For the first category, early studies on DRS parsing heavily relied on hand-crafted features and linguistic characteristics For the second category, some researchers In the past decade, GANs have achieved great progress in NLP In this section, we give a brief introduction to our baseline system, the top-down parser of Hierarchical Split Point Encoding. For split point representation 1 , Top-Down Split Point Ranking. After achieving split point representations, an encoder-decoder is used to rank the split points, as shown in Figure After that, a biaffine function where W, U, V, b are model parameters and s j,i ∈ 1 The split position between any two neighboring EDUs is called the split point. R k denotes the score of the i-th split point over different categories (for split point ranking, k equals 1). With this attention function used, at each time step, split position with the highest score is selected as the split point and the original text span is split into two adjacent text spans. Meanwhile, newly generated text spans with unselected split points are pushed onto the stack for following steps, as shown in Figure To our knowledge, Zhang et al. ( This section introduces the proposed adversarial learning method which consists of two parts: graphical representation of gold and fake DRS trees and the adversarial model learning process. In this study, we aim to learn from the entire DRS tree to optimize our model from a global perspective. Usually, our computer understands DRS trees in two ways: either language description or graphical representation. Since tree diagrams can reflect the structural features more intuitively and are easy for machines to understand, we explore graphical representation of DRS trees in this work. For gold standard trees, we propose to transform each tree into multi-pattern matrices which Image Generarion Adversarial Bot Feature Extraction Image Generation Reshape Feature is similar to a low resolution image with two color channels (i.e., the structure (ST) and nuclearityrelation (NR) channels). Formally, given a DRS tree of height m with n split points, each split point corresponds to a specific non-leaf node in the tree, and we construct two matrices, X ST and X NR , of size m × (n + 2) corresponding to the two color channels, as shown in Figure (ii) For the NR channel, we set the positions representing non-leaf nodes to specific N-R labels and the positions of leaf nodes to -1 and other nonnode positions to zero. For the automatically parsed trees, we directly use our model outputs to build the tree diagram with two color channels, X ST and X NR . And the two matrices of size m × (n + 2) are initialized with zero. (i) For the ST channel, as stated before, a set of attention weights are assigned to the encoder outputs during pointing and a split point is selected according to the weights. Obviously, each split point corresponds to a group of attention weights (after log-softmax). Therefore, we directly add these n-dimensional attention weights of each split point in the i-th tree layer (top-down direction) to the i-th line of X ST . Notably, the first and last columns of the matrices are actually placeholders initialized with unlearnable scalars representing leaves or non-node positions, so we only add the split point attention weights to the range from 1 to n in each row. (ii) For the NR channel, we simply replace these elements corresponding to split points in X ST with predicted N-R labels Through the above methods, we achieve graphical representation for both gold standard and automatically predicted DRS trees. And the graphical representation can provide our model with a global perspective, which makes the global optimization (Subsection 4.2) of DRS parsing possible. For model learning, we have two goals: (i) learning of DRS parsing at each time step for local optimization and (ii) learning an adversarial bot to evaluate the pros and cons of the entire tree for global optimization. For the first goal, we use two negative log-likelihood loss terms to optimize the parsing model. For split point ranking, we use L s to maximize the probability of correct split point selection at each decoding step. For N-R prediction, given the selected split point, we use L nr to maximize the probability of correct N-R labeling for the split point. Since the convergence speeds of the two loss terms are different, we add two loss weights before the loss terms to balance the model training as: For the second goal, we explore to learn from the entire DRS tree for global optimization. To that end, we produce an adversarial bot in our parser to estimate the generated DRS tree diagrams, as shown in Figure Then we perform max-pooling in each nonoverlapping 3 × 1 window for feature extraction, and the resulting matrices are reshaped as ∈ R 1×D to serve as the distributed representation of the tree. In this work, we do not just need an excellent discriminator expert in classification, we need the adversarial nets to continuously give feedback to our parsing model even when the generated trees are correctly classified. On this basis, we leverage Least Squares Generative Adversarial Network (LSGAN) Similar to At this time, we have a traditional loss term to train the top-down parser at each splitting step and two adversarial loss terms to estimate the entire DRS tree for global optimization. It is worth mentioning that we first optimize the L DRS for 7 epochs to warm up the model parameters, and then the adversarial nets join the training process for global optimization of DRS parsing. Datasets. Following our previous work • The RST-DT corpus contains 385 news articles (347 for training and 38 for testing) from the Wall Street Journal (WSJ). Following previous work, we randomly select 34 documents from the training corpus as the development corpus for parameter tuning. And we also binarize those non-binary subtrees in RST-DT with right-branching • The Chinese CDTB corpus is motivated by taking advantages of both the English RST-DT corpus and the PDTB corpus The CDTB corpus annotates each paragraph as a Connective-driven Discourse Tree (CDT). The corpus consists of 500 newswire articles which are further segmented into 2336 paragraphs and 10650 EDUs. The corpus is divided into three parts with 425 articles (2002 CDT trees) for training, 25 articles (105 CDT trees) for validation, and 50 articles (229 CDT trees) for testing. Metrics. Following previous studies, we measure the performance of bare tree structure (S), tree structure labeled with nuclearity (N), and tree structure labeled with rhetorical relation (R). Recently, the Full (F) indicator is used to estimate the tree structure labeled with both nuclearity and relation categories. However, since current performances on S, N and R are imbalanced, the performance on F is much limited by relation prediction. In other words, the Full score may underestimate the performance in span and nuclearity prediction. In this work, we combine nuclearity and rhetorical relation tags for joint N-R prediction aiming to reduce the uncertainty of the Full measure. Moreover, since RST-Parseval Hyper-Parameter Setting. Comparison between different system settings. As stated before, we explore to make possible improvements to the top-down architecture of Comparison with previous studies. In this part, we compare with seven previous state-of-the-art (SOTA) parsers on text-level DRS parsing. Here, we briefly review these studies as follows: • Ji and Eisenstein (2014), a shift-reduce parser with an SVM that is trained by their extracted latent features. In this paper, we compare with the updated version of their parser (designated as "JE2017-updated") • Feng and Hirst (2014), a two-stage greedy parser with linear-chain CRF models and some handengineered features. • • • • For the RST-DT corpus, the results are presented in the upper part of Table For the CDTB corpus, we explore to employ a more strict metric Performances based on the SOTA language models. Recently, more and more researchers For the RST-DT corpus, we report the results of the recent Bert-based top-down parser limited corpus size. Moreover, the ablation study on the adversarial learning strategy further demonstrates the usefulness of our proposed method. It should be noted that we report the performance using LMs in this paper never mean to advocate using pre-trained LMs or blindly pursuing performance improvements in DRS parsing. Sometimes, the rewards generated by the large-scale LMs could be quite different from and much more effective than that generated by language phenomena, which may hinder the study on the relatively shallow (compared with powerful LMs) yet valuable discourse features. With this in mind, it is reasonable to perform ablation study using simple word representation to explore useful discourse features and report the performance on powerful LMs for reference. Performance Evaluation of Dependency Trees. Recently, discourse-level dependency structure has attracted more and more attention. Here, we explore whether the proposed global optimization method can improve the RST dependency analysis to some extent. To achieve this, we first convert the predicted DRS trees into dependency trees as From the results, it's obvious that the LM we use plays a big role in nuclearity prediction, and the proposed adversarial method can further improve the performance to a certain extent. (ii) For relation prediction, the classification problem with 18 coarse-grained relation tags (RST-DT) is really a challenge. From the results in Table Discussion on Adversarial Learning. Similar to previous GAN work, improving the quality of the generated tree images is really a challenge, and the instability of the adversarial learning process is another intractable issue. In order for our model to continuously modify the generated images even when they are correctly classified, we leverage a least squares loss in our system for model learning. To avoid the over-learning of the discriminator, we tune it with a moderate learning rate and parameter scale. Intuitively, the convergence of our model over different learning rates is presented in Figure In this research, we explored a global optimization method based on recent top-down frameworks. Particularly, we proposed a novel strategy to transform both gold standard and predicted DRS trees into tree diagrams with two color channels. On this basis, we produced an LSGAN-based adversarial bot between gold and fake trees for global optimization. Experimental results on two popular corpora showed that our proposed adversarial approach is effective in DRS parsing and has established new state-of-the-art results for both corpora.
| 1,546 | 1,351 | 1,546 |
Towards a Better Understanding of Label Smoothing in Neural Machine Translation
|
In order to combat overfitting and in pursuit of better generalization, label smoothing is widely applied in modern neural machine translation systems. The core idea is to penalize over-confident outputs and regularize the model so that its outputs do not diverge too much from some prior distribution. While training perplexity generally gets worse, label smoothing is found to consistently improve test performance. In this work, we aim to better understand label smoothing in the context of neural machine translation. Theoretically, we derive and explain exactly what label smoothing is optimizing for. Practically, we conduct extensive experiments by varying which tokens to smooth, tuning the probability mass to be deducted from the true targets and considering different prior distributions. We show that label smoothing is theoretically wellmotivated, and by carefully choosing hyperparameters, the practical performance of strong neural machine translation systems can be further improved.
|
In recent years, Neural Network (NN) models bring steady and concrete improvements on the task of Machine Translation (MT). From the introduction of sequence-to-sequence models In Transformer-based models, label smoothing is a widely applied method to improve model performance. Label smoothing itself is an interesting topic that brings insights about the general learnability of a neural model. While existing methods are rather heuristical in their nature, the fact that simply discounting some probability mass from the true label and redistributing it with some prior distribution (see Figure In this paper, we raise two high-level research questions to outline our work: 1. Theoretically, what is label smoothing (or the related confidence penalty) optimizing for?
|
Practically, what is a good recipe in order to apply label smoothing successfully in NMT? V , A and r v are prior distributions. Smoothing with (a), m is equally redistributed across the vocabulary. Smoothing with (b), A is implicitly 1 V everywhere as well, and the exact value of B can be obtained (Section 3.2). Smoothing with (c), m goes to each class in proportion to an arbitrary smoothing prior r v (Section 4.3). The presentation of our results is organized into three major sections: • First, we introduce a generalized formula for label smoothing and derive the theoretical solution to the training problem. • Second, we investigate various aspects that affect the training process and show an empirically good recipe to apply label smoothing. • Finally, we examine the implications in search and scoring and motivate further research into the mismatch between training and testing. The extensive use of NNs in MT The idea of smoothing is not new in itself. For instance, many smoothing heuristics and functions are investigated in the context of count-based language modeling Proposed to counteract overfitting and pursue better generalization, label smoothing As a regularization technique in training, label smoothing can be compared against other methods such as dropout The standard label smoothing (STN) loss, as used by (1) where L STN denotes the cross entropy with standard label smoothing, n is a running index in the total number of training tokens N , v is a running index in the target vocabulary V , m is the hyperparameter that controls the amount of probability mass to discount, p v is the one-hot true target distribution and q v is the output distribution of the model. The confidence penalty (CFD) loss, as used by where L CFD denotes the confidence-penalized cross entropy, m in this case is the hyperparameter that controls the strength of the confidence penalty and thus differs from m in Equation In both cases, the outer summation is over all of the training tokens N , implicating that all of the target token probabilities are smoothed. The dependencies of q v and p v on n are omitted for simplicity. Additionally for Equation or and implement it as an additional layer of activation during training, where v is an alternative running index in the vocabulary. In any case, the integration of Equation 2 into the form of Equation 1 cannot be done without significantly modifying the original confidence penalty, and we leave it for future work. In an effort to obtain a unified view, we propose a simple generalized formula and make two major changes. First, we separate the outer summation over the tokens and divide it into two summations, namely "not to smooth" and "to smooth". Second, we modify the prior distribution to allow it to depend on the position, current token and model output. In this case, r could be the posterior from some helper model (e.g. an LM), and during training, obtaining it on-the-fly is not expensive, as previously shown where L GNR denotes the generalized cross entropy, A is the set of tokens not to smooth, B is the set of tokens to smooth, r v,qv is an arbitrary prior distribution for smoothing and again we drop the dependencies of p v , q v and r v,qv on n for simplicity. A natural question when explicitly writing out A and B, s.t. A ∩ B = ∅ and |A ∪ B| = N , is which tokens to include in B. Here, we consider two simple ideas: uniform random sampling (RND) and an entropy-based uncertainty heuristic (ENT). The former chooses a certain percentage of tokens to smooth by sampling tokens uniformly at random. The latter prioritizes those tokens whose prior distributions have higher entropy. The logic behind the ENT formulation is that when the prior distribution is flattened out, yielding a higher entropy, the helper model is uncertain about the current position, and the model output should thus be smoothed. Formally, the two heuristics can be expressed as: where ρ n is a sample from the uniform distribution U in [0, 1], π is a hyperparameter controlling the percentage of tokens to smooth and {b 1 , b 2 , ..., b N } is a permutation of data indices {1, 2, ...N } in descending order of the entropy of prior r, i.e. The hyperparameter m in Equation 5 deserves some further notice. This is essentially the parameter that controls the strength of the label smoothing procedure. When it is zero, no smoothing is done. When it is one and |B| = N , the model is optimized to output the prior distribution r. One can obviously further generalize it so that m depends also on n, v and q v . However in this work, we focus on the outer summation in N and alternative priors r, and leave the exploration of adaptive smoothing strength m n,r,qv for future work. When it comes to the analysis of label smoothing, previous works focus primarily on intuitive understandings. While in practice we use gradient optimizers to obtain a good set of parameters of the NN, the optimization problem actually has well-defined analytical solutions locally: which is simply a linear interpolation between the one-hot target distribution p v and the smoothing prior 1 V , with m ∈ [0, 1] being the interpolation weight. One can use either the divergence inequality or the Lagrange multiplier method to obtain this result (see Appendix A). Consider the optimization problem when training with Equation The problem becomes harder because now the regularization term also depends on q v . Introducing the Lagrange multiplier λ and solving for optima will result in a transcendental equation. Making use of the Lambert W function where W 0 is the principal branch of the Lambert W function and λ is the Lagrange multiplier, which is numerically solvable 1 when non-negative m and probability distribution p v are given. Equation V is redistributed to each token in the vocabulary. The graph of GRN V to r v as m approaches one, and not included here for brevity. One last thing to notice is that the outer summation over the tokens is ignored. If it is taken into consideration, q is dragged towards the empirical distribution given by the corpus In this section, we describe our results and insights towards a good recipe to successfully apply label smoothing. We experiment with six IWSLT2014 datasets: German (de), Spanish (es), Italian (it), Dutch (nl), Romanian (ro), Russian (ru) to English (en), and one WMT2014 dataset: English to German. The statistics of these datasets are summarized in Table The first thing to determine is how to select tokens for smoothing and how many tokens to smooth. For this purpose, we begin by considering models smoothed with an LM helper. The helper LM is trained on target sentences from the corresponding parallel data till convergence. Figure Our next goal is to find good values of m. The discounted probability mass m is a tunable hyperparameter that is set to 0.1 in the original Transformer Furthermore, we explore the use of LM and MT posteriors as prior distributions for smoothing. We train systems using Transformer LMs and MT models of different qualities for label smoothing, as in Figure One important thing to mention is that, while neither LM nor MT outperforms uniform or unigram in terms of test BLEU score in our experiments, we see significant drops in development set perplexities when smoothing with LM or MT. This signals a mismatch between training and testing, and suggests that smoothing with LM or MT indeed works well for the optimization criterion, but not as much for the final metric, the calculation of which involves beam search and scoring of the discrete tokens. Finally, we report BLEU scores of our best systems across all language pairs in Table As discussed in Section 4.3, models smoothed with LMs or MT model posteriors yield very good development set perplexities but no big improvements in terms of test BLEU scores. Here, we further investigate this phenomenon in terms of search and scoring. We first plot the test BLEU scores with respect to the beam size used during search. In Figure We further examine test BLEU with respect to development (dev) BLEU and dev perplexity. As shown in Figure In this work, we investigate label smoothing in neural machine translation. Considering important aspects in label smoothing: token selection, probability mass and prior distribution, we introduce a generalized formula and derive theoretical solutions to the training problem. Examining the effect of various hyperparameter choices, practically we show that with a good label smoothing recipe, one can obtain consistent improvements over strong baselines. Delving into search and scoring, we finally emphasize the mismatch between training and testing, and motivate future research. Reassuring that label smoothing brings concrete improvements and considering that it only operates at the output side of the model, our next step is to explore similar smoothing ideas at the input side.
| 999 | 770 | 999 |
A Probabilistic Framework for Discovering New Intents
|
Discovering new intents is of great significance for establishing the Task-Oriented Dialogue System. Most prevailing approaches either cannot transfer prior knowledge inherent in known intents or fall into the dilemma of forgetting prior knowledge in the follow-up. Furthermore, such approaches fail to thoroughly explore the inherent structure of unlabeled data, thereby failing to capture the fundamental characteristics that define an intent in general sense. In this paper, starting from the intuition that discovering intents should be beneficial for identifying known intents, we propose a probabilistic framework for discovering intents where intent assignments are treated as latent variables. We adopt the Expectation Maximization framework for optimization. Specifically, In the Estep, we conduct intent discovery and explore the intrinsic structure of unlabeled data by the posterior of intent assignments. In the M-step, we alleviate the forgetting of prior knowledge transferred from known intents by optimizing the discrimination of labeled data. Extensive experiments conducted on three challenging real-world datasets demonstrate the generality and effectiveness of the proposed framework and implementation. Codes is publicly available. 1
|
Unknown intent detection Just as discovering new intents plays a crucial role in establishing the TODS, discovering new intents has raised a lot of research interest just like unknown intent detection. Unsupervised cluster learning is a popular paradigm to solve this problem. Specifically, previous works How do discover intents in the unlabeled corpus using the labeled data? Recently, the semisupervised methods Nevertheless, DeepAligned suffers from many problems. Firstly, when the model is re-trained with the pseudo supervision signal, the model will forget the knowledge transferred in the transferring stage, which is demonstrated in Figure Different from the previous methods, we start from the intuition that the intent discovery should not damage the identification of the known intents. Ideally, the two processes should achieve a win-win situation. The knowledge contained in labeled data corpus (as known intents) can be used to guide the discovery, and the information learned from the unlabeled corpus during discovery could improve the identification of the known intents. Therefore, with the help of optimizing the identification of labeled data given the whole data corpus, we propose a principled probabilistic framework for intent discovery, where intent assignments as a latent variable. We adopt Expectation Maximization as a principal template for optimizing this typical latent variable model. Specifically, in the E-step, we use the current model to discover intents and calculate a specified posterior probability of intent assignments to explore the intrinsic structure of data. In the M-step, the probability of identification of labeled data including those newly discovered from unlabeled data, and the posterior probability of intent assignments, which is to help learn friendly-discovery features, are maximized simultaneously to optimize and update model parameters. Extensive experiments conducted in three benchmark datasets demonstrate our method can achieve substantial improvements over strong baselines. Our contributions are as follows: (Theory) We introduce a principled probabilistic framework for discovering new intents and provide a learning algorithm based on Expectation Maximization. To the best of our knowledge, this is the first complete theoretical framework in this field and we hope it can inspire follow-up research. (Methodology) We provide an efficient implementation based on the proposed probabilistic framework. After transferring prior knowledge, we use a simple yet effective method to alleviate forgetting. Furthermore, we propose a new contrastive paradigm to explore the intrinsic structure of unlabeled data, which avoids the model shift towards inaccurate pseudo labels but helps to better learn the friendly-discovery features. (Experiments and Analysis) We conduct extensive experiments and detailed analyses on a suite of real-world datasets to demonstrate the generality and effectiveness of our proposed framework and implementation.
|
Our work is mainly related to two lines of research: Unsupervised and Semi-supervised clustering. Unsupervised Clustering Extracting meaningful information from unlabeled data has been studied for a long time. Traditional approaches like K-means Given as input an labeled dataset D l = {x l i , i = 1, . . . , N } where intents Y l = {y l i , i = 1, . . . , N } are known and an unlabeled dataset D u = {x u i , i = 1, . . . , M }. Our goal is to produce intent assignments as output by clustering (or partitioning) the whole dataset D, which denotes D = D l ∪ D u . Directly optimizing the goal is intractable as the lack of knowledge about new intents and the intrinsic structure of unlabeled data. As analyzed in Section 1, discovering intents should not damage but be beneficial for the identification of known intents, which can be formulated to optimize p(Y l |D l , D; θ). Since D l ⊂D, the optimization objective can be written as: p(Y l |D; θ). Denote our latent variable (representing intent assignments obtained by clustering on D) by Z and let Z D be a possible value of Z. Using Bayes rule, p(Y l |D; θ) can be calculated as: (1) Exactly optimizing Eq.( (2) Our goal is get better Z D (i.e.intent discovery) by optimizing L obj , and a better Z D can also help optimize L obj . Before optimizing L obj , we want to transfer knowledge from the labeled corpus to initialize the model. Transferring knowledge has been widely studied and types of transferred knowledge have been proposed for a variety of circumstances. Considering the excellent generalization of the pre-trained model, we fine-tune BERT where ϕ(•) represents a linear classifier and ϕ(z i ) j denotes the logits of the j-th class, K l denotes the total number of known intents. Intent Assignments Z (In the following, we omit the subscript D of Z D for clarity.) Specific intent assignments Z involves two components: how to determine K representing how many intents in dataset D and how to assign the utterance in the dataset to corresponding intent. Many methods We have assigned a specific intent assignment Z to latent variable Z based on prior knowledge. We expect that the intent assignments Z should reflect what characteristics make a good intent in general rather than specific intents. Therefore, the standard cross entropy loss formed by specific pseudo labels adopted by where C k is a cluster produced by Z, and x • x + is calculated by consine between features. To optimize Eq.( where ϕ(•) denotes same linear classifier as Eq.( where x l is the sample from D l , y l is the label of x l . N Z (x l ) is the unlabeled nearest neighbor samples set that belongs to the same cluster (divided by Z) as x l . See Appendix E for specific benefits from Dl . The labeled data is tailored to model training. On the one hand, the model will not lose the knowledge transferred from labeled data, on the other hand, the model can constantly explore the intrinsic structure of the dataset by utilizing it. In the M-step, we update the θ in Eq. ( where λ is to balance the proportion of two loglikelihoods (discussed in Section 5.3) during training, τ is a hyper-parameter for temperature scaling which often appears in contrastive learning. We summarize the whole training process of the EM framework in Algorithm 1 and the model architecture of our approach as shown in Figure It is worth noting that our method actually proposes a framework where probability estimation can flexibly adopt different ways for a variety of circumstances. Algorithm 1 EM algorithm for optimization Input: Update model parameters θ to maximize the log-likelihood L in Eq. ( We conduct experiments on three challenging datasets to verify the effectiveness of our proposed method. The detailed statistics are shown in Appendix A. CLINC BANKING StackOverflow represents a dataset dispersed through Kaggle.com, encompassing 20 intents and 20000 utterances. We adopt the dataset processed by We follow Lin et al. (2020); To evaluate clustering results, we follow existing methods For each dataset, 75% of all intents are randomly selected as known intent, with the remaining designated as unknown. Furthermore, 10% of the known intents data are chosen randomly as labeled data. We set the number of intents as ground truth in line with previous methods Lin et al. (2020); Our main experiments use pre-trained BERT, which is implemented in the Huggingface Transformers 2 , as the network backbone. We also replace the backbones of the compared baselines with the same BERT as ours. Only when comparing with SCL Moreover, considering the efficiency of the training process and the capacity of GPU, we only finetune the last transformer layer parameters during transferring knowledge and freeze all but the latter 6 transformer layers parameters during performing We present the main results in table To make a fair comparison with SCL To better verify the effectiveness of our proposed method, we analyze the comparison results between our method and DeepAligned in a more finegrained way. We separate the known intents and the unknown intents from the test set and compare our method with DeepAligned on these two sub-datasets respectively (the experimental settings remain unchanged). The results are shown in Table 4, which demonstrates that our method can not only effectively apply to known intents, but also can more effectively discover new intents, and the effect of improvement is substantial. This also fully conforms to our expectations that the two processes of intent discovery and recognition of known intents can be "win-win". In objective function Eq. ( Because we do not know the actual number of intents, we usually need to assign an initial number of intents (i.e., K) in advance as we do earlier. This also requires us to investigate the sensitivity of the model to the initial K. We investigate the performance of our method in the datasets by varying initial values (leaving others unchanged). As shown Figure We also investigate the effect of known intent ratios on performance by adopting different known class ratios (25%, 50% and 75%). As shown in Figure 5, our method also shows better performance compared with other baselines. Interestingly, The advantage of our method in dataset BANKING is significant. We speculate that this may be related to the unbalanced number of samples in BANKING. Although there are more known intents, it does not mean that enough labeled and balanced samples are provided. As a result, the previous methods (e.g. DeepAligned) not only failed to transfer more prior knowledge but also exacerbated the speed of forgetting in the follow-up process. This also provides room for future research. During intent assignments, our performance is not only not forgotten, but also constantly strengthened compared with the pre-transfer stage (Red, approximated by the initial performance in clustering stage). We show knowledge forgetting in DeepAligned in Section 1. After fine-tuning with labeled data, the prior knowledge is stored in the model in the form of model parameters. With the subsequent clustering steps, the parameters change gradually (the forgetting process is step by step from the forgetting curve in previous works). However, as shown in Figure In this paper, we provide a probabilistic framework for intent discovery. This is the first complete theoretical framework for intent discovery. We also provide an efficient implementation based on this proposed framework. Compared with the existing methods, our method effectively alleviates the forgetting of prior knowledge transferred from known intents and provides intensive clustering supervised signals for discovering intents. Extensive experiments conducted in three challenging datasets demonstrate our method can achieve substantial improvements. The subsequent analysis also shows that our method can better estimate the number of intents and adapt to various conditions. In the future, we will try different methods to perform intent assignments and explore more methods to approximate p(Y l |Z, D; θ) and p(Z|D; θ). To better inspire the follow-up work, we summarize the limitations of our method as follows: 1) From our experimental results the Appendix D, we can see that the estimation of the number of intents in our proposed can be further improved. 2) We do not try more means to prevent knowledge from forgetting. We can probe into the intrinsic structure of unlabeled data in a more fine-grained way by improving the posterior estimation. 3) According to Section 5.3, we have verified that both exploration and utilization are indispensable, but at the same time, we only empirically choose the specific proportion of both, without theoretical analysis of the most appropriate proportion for each dataset. We look forward to making progress in the follow-up research on the above limitations. We present detailed statistics of datasets in our experiments in Table Our main experiments use pre-trained BERT (bertuncased, with 12-layer transformer), which is implemented in the Huggingface Transformers In this section, we detail the results of varying λ in the Table A key point of intent discovery is whether the model can accurately predict the number of intents. DeepAligned proposes a simple yet effective estimation method. However, due to the alignment operation in the iterative process of clustering (see In addition to the labeled data in hand, in Section 3.3, we also use the sample set Dl predicted known intents in unlabeled data during discovery (See Section 3.3 for the specific construction of Dl . The nearest neighbor measure is based on the cosine similarity of the sample representation in the semantic space). In this section, we will further analyze the benefits brought by these discovered sample set. We have compared the effects of adding Dl and not adding Dl , and the comparison results are shown in Table
| 1,255 | 3,007 | 1,255 |
Byte-Level Grammatical Error Correction Using Synthetic and Curated Corpora
|
Grammatical error correction (GEC) is the task of correcting typos, spelling, punctuation and grammatical issues in text. Approaching the problem as a sequence-to-sequence task, we compare the use of a common subword unit vocabulary and byte-level encoding. Initial synthetic training data is created using an error-generating pipeline, and used for finetuning two subword-level models and one bytelevel model. Models are then finetuned further on hand-corrected error corpora, including texts written by children, university students, dyslexic and second-language writers, and evaluated over different error types and origins. We show that a byte-level model enables higher correction quality than a subword approach, not only for simple spelling errors, but also for more complex semantic, stylistic and grammatical issues. In particular, initial training on synthetic corpora followed by finetuning on a relatively small parallel corpus of real-world errors helps the byte-level model correct a wide range of commonly occurring errors. Our experiments are run for the Icelandic language but should hold for other similar languages, particularly morphologically rich ones.
|
Spelling mistakes due to typos and rushed writing, nonstandard punctuation and spelling, and grammatical and stylistic issues are common to almost everyone who writes any kind of text. This applies in any language and can distract the reader or make the communication miss its mark. This can hinder people who have difficulties writing text conforming to a particular language standard, be it due to disability, dyslexia, linguistic background, limited access to education or any other reason. Prejudice against people whose writing deviates from the standard can make some shy away from communicating with others, leaving their voices out of important discussions and restricting their opportunities Grammatical error correction (GEC) is the task of adjusting a text's spelling, grammar, and linguistic style to conform to an approved language standard or convention However, a significant downside of subword tokenization is how it is affected by noisy input; if a word contains a typo or other spelling variants, this can completely shift its representation. In addition, in languages with rich morphology, a word can have many different surface forms, some rarer than others, that all carry the meaning of the base word, but appear in different syntactic contexts. A subword-tokenized model may struggle to capture the nuances of such a language effectively since it may need several different subwords to represent a single word, depending on spelling and context. When an unfamiliar variant of the word appears in unseen text, the model is challenged to decode it correctly, even when it results in uncommon subword units. Our motivation is that a byte or character-level approach should intuitively be more robust to spelling or morphology variations, as it is not constrained by the subword vocabulary. We explore using a byte-level architecture, ByT5 We compare the performance of the byte-level architecture to two subword-based architectures; ByT5's predecessor, mT5 Our main contributions include a comparison between subword tokenization and byte-level tokenization for GEC when training over a combination of curated and synthesized data. We demonstrate how byte-level models not only bypass subword-related issues, but can also correct long-range errors in text. We release our error generation framework as well as models for GEC using byte-level and subword tokens in Icelandic. While our work focuses on the Icelandic language, we have no reason to believe that similar results do not hold for other languages, particularly those similar to Icelandic in terms of morphological complexity.
|
The bulk of research on grammatical error detection and correction has been focused on English and English learner texts, due to existing training data and benchmarks, and the large market of English learners worldwide who benefit from an automatic language correction tool The problem of data scarcity in GEC, when approached as a sequence-to-sequence task, is typically addressed with synthetic data generation GEC can essentially be considered the task of generating grammatical target text from an ungrammatical source, similar to machine translation. The idea of approaching GEC as a machine translation problem dates back to 2006 In a subword-based GEC model, when a word contains a typo or is spelled unconventionally, it may look like an unknown word, for which no known representation exists. The model may then segment the word differently from what was seen during training, causing mispredictions. If the subword representation for "different" is This is also true for unseen words that are correctly spelled, such as foreign-named entities, which can lead to the subword-based GEC model "correcting" a perfectly spelled word it has not seen before, by replacing it with the most likely candidate. In the sentence "The tournament was held in Espoo, Finland.", the place name "Espoo" may be represented by a single subword token Espoo. Since this token is unfamiliar, the model finds the most likely subword token for this particular sentence, Helsinki. Due to this known shortcoming of subword tokenization Despite the reported advantage of character or byte-based Transformer models on noisy text Our work deviates from that of Current state-of-the-art in GEC is based on sequence-tagging methods Apart from some rule-based spell checkers that don't make use of the full context, one rule-based correction system exists for Icelandic, based around parse trees, GreynirCorrect The work presented here is the first where neural networks are used in GEC for Icelandic. A single collection of parallel error corpora exists for Icelandic, the Icelandic Error Corpus (IceEC). The corpora are annotated and corrected by language experts The IceEC is split into a larger general corpus and three specialized corpora This curated error data was used for finetuning our models, and combined into one training dataset for a total of 64k input sequences (single sentences), after setting aside validation and test data. The general IceEC also includes a 5.3k sentence test set used for evaluation. We applied a diverse set of methods for error generation, both using linguistic knowledge and random noising methods. This rule-based approach to synthetic data generation gave us control over the types of noise applied, and allowed us to generate evaluation data for each error type. As our basis of correct text to be noised, we used the Icelandic Gigaword Corpus (IGC) We generated three categories of errors: 1) noise within words; 2) noise at the sequence level; and 3) grammatical and morphological modifications. The first two resemble those used when noising backtranslation data In order to explore to what extent subword and byte-level models can learn and generalize grammatically complex issues in a morphologically rich language, we go beyond naive language-agnostic noising of text. A more detailed explanation of the Icelandic-specific noise is given in Appendix C. The noise methods are shown in Table The error generator allows for noising levels to be configured via hyperparameters. Experiments with different noise ratios in the synthetic data showed that highly noised text provided the best training examples, without the models learning to "overcorrect", i.e., to introduce false positives. Instead of producing even more synthetic data, we geared up the noise to produce highly error-dense examples, setting the random and more naive error noise to appear in 80% of cases, and the rule-based error noise to be used wherever possible. We compared three model architectures to evaluate the differences between using subword tokenization and a byte-level method. Comparing models with different architectures calls for defining which factors are compared. In particular, byte sequences are longer than subword sequences when counting the number of tokens, roughly 4 times longer on average in the original multilingual mT5/ByT5 training data We continued training of the pretrained multilingual BART25 model (mBART) Training on the synthetic error data was performed with an effective batch size of 3000 input tokens (roughly 60 sequences), a learning rate of 3e-5 with an inverse square root scheduler, 0.1 dropout, 0.3 attention dropout, 1000 warmup steps, 0.1 label smoothing, no weight decay, and using the Adam optimizer, for 100k updates on an A100 card for a day. For the byte-level approach we employed the ByT5base model The pretraining task in ByT5 has been adapted to a byte-level model, with span infilling based on bytes, not subwords. Apart from this, the main difference between the mT5 and ByT5 model architecture is the heavier encoder of ByT5. Sequences in byte-level architectures are long and correspond more or less to the number of characters in Icelandic, resulting in increased training time. We trained the ByT5-base model using a maximum sequence length of 512 bytes, which was found to be a reasonable compromise, as most sentences in Icelandic texts are shorter than this. The ByT5-base model was finetuned on the synthetic data with an effective batch size of 32 sequences (sentences). The learning rate was set to 2e-5 using the Adam optimizer with 1000 warmup steps and no weight decay. This model was further trained for a total of 550k updates, or 13 A100 card days. For a more direct comparison of byte-level and subword-level models, we also finetuned the mT5base We finetuned the mT5-base model on the synthetic data using the same parameters as in our ByT5-base finetuning and evaluated it at 100k updates. Using the curated error corpora (IceEC), we finetuned the byte-level and subword-level models to convergence. For the mBART model, this meant training with a learning rate of 2e-6 for 53k updates (67 epochs), with attention dropout set to 0.1, weight decay to 0.001 and other parameters being the same as during the synthetic finetuning. The ByT5 and mT5 models were finetuned with a learning rate of 2e-6, other parameters were the same as during finetuning on the synthetic data. The ByT5 model had converged at 120k updates (60 epochs), while the mT5 was still improving on the validation data at 200k updates (100 epochs), but with time we found it forgot too much of the synthetic error correction task. We report evaluation scores at 130k. For comparison, we also finetuned the different models (mBART-ENIS, mT5 and ByT5-base) on the IceEC data only, without the synthetic finetuning phase. This was done to examine how much the models learn from the added synthetic examples, and how far we can get using a small amount of hand-corrected examples. The mT5 and ByT5 models were trained for 100k updates and the mBART-ENIS model for 10k updates. Different metrics exist for evaluating GEC performance, but most are language-specific, and have not been adapted to Icelandic. Here we employ a language agnostic metric for scoring our models, the GLEU score When evaluating GEC for English, ERRANT We consider a variety of curated and synthetic test sets to get a good overview of the differences between the byte-level and subword-level approach for GEC. For the real errors, we report scores over the IceEC.test set, the test set from the IceEC, which contains around 5000 sentences. In con-trast, the dyslexic, L2 and children test sets contain 500 held-out sentences each from the respective specialized error corpora described in section 3.1 (only 100 examples were collected for the dativitis error type, a rarer occurrence in the data). We also annotated a small dataset (163 sentences) of data from an Icelandic news outlet (news), where each sentence contains at least one error; this is further described in Appendix B. For the synthetic errors, we report GLEU scores over the test.synth set, which contains around 4000 held-out sentences from the synthetic data. Furthermore, we generated test sets of synthetic examples, each containing a particular error type in each sentence (dativitis, spaces, commas, dupl-words, mood, rand-noise, noun-case). This last group of test sets was generated using source texts that, while editorial, may include other errors, just like the synthetic training examples. The models, as they get better, learn to correct these errors as well. This may paradoxically lower the GLEU score as the corrected output deviates from the erroneous reference. These generated test sets still provide valuable information about what the models learn about each error type in isolation. To understand what approach is best suited for GEC we trained the models on different data combinations and using different pre-trained models. The Synth-100k models are all trained for 100k updates on the same synthetic data, and the Synth-100k-EC models are additionally finetuned on the curated IceEC error corpus. To provide a baseline for the GLEU scores, we also report no_corr scores, where the source text is not corrected. This gives some idea of the noise level of the test sets, with test.synth being the noisiest and IceEC.test containing the least noise. The GreynirCorrect Training on the synthetic data and then finetuning on the curated error corpora (Synth100k/ Synth550k+EC) performs best on the curated errors and retains much of the performance on the synthetic test sets. In all of these experiments, we can see that the ByT5 models generally perform better than the subword counterparts. This is also reflected in the ERRANT scores in Appendix E, Table Our results show that the ByT5 models are the overall high-scorers on the real-world test sets, and on most of the synthetic ones. We include finetuning results on the ByT5 model that has been trained for longer on the synthetic data (550k updates) to compare how performance improves with time. We see the GLEU scores keep going up with time, and more importantly, when taking a close look at the actual generated output, this is the model that best corrects real-world errors. This makes it the most feasible option for use in real-world scenarios. A comparison of the output of the models trained on both data sources is shown in Appendix D. An example from the test data is when the subword-tokenized model mBART-ENIS-Synth100k+EC incorrectly changes the name of a person from a rare name ("Láretta") to a more common one ("Lára"). This kind of error is not seen in the byte-level model, which is quite conservative in its corrections of unknown entities. While this means ByT5 occasionally misses actual errors, we find that it is much better suited for production than a subword-level model that makes serious semantic errors. These more nuanced error correction examples may not be fully captured by the automated metrics, but are crucial for real-world use. The subword regularization experiments are included as an alternative approach for mitigating the subword tokenization problem. The results are marginally better than the model without subword regularization when trained on the synthetic data, and the model performs better than the ByT5-Synth100k model in the case of dupli-cate words, which linguistically is a quite trivial task, and in more intricate mood errors. It however doesn't do any better than the mB-ISEN-Synth100k trained without subword regularization on the curated datasets, and this also holds when the model is finetuned additionally on curated data. The model finetuned on only the curated data with subword regularization (mB-ISEN-reg-EC) however performs consistently much better than its counterpart without subword regularization, often on par with or surpassing ByT5. This model has not seen any of the highly noised synthetic data, and thus has the most to gain from the subword noise. We speculate that this is one of the reasons we don't see more gains from adding subword regularization; the training examples are already so highly noised that there is not much to be learned from the added subword noise. The IceEC finetuning data contain real-world errors which have been hand-corrected. These texts are somewhat different from the highly noised training examples with synthetic errors, have fewer errors on average and are more varied as they are naturally occurring. They also include stylistic edits from the reviewers, which improve the text's fluency, but in those cases the original is not necessarily incorrect as per the language standard. With these differences in mind, we expect the models to have to forget some of the synthetic error correction task in order to adapt to this "new" denoising task. We see this happen in the mBART-ENIS finetuning on the curated data, and to a lesser extent in the ByT5 finetuning. The denoising task performance on the synthetic errors from the previous step has in part been lost, which is expected, since some of these errors are not particularly common in real texts. For the more grammatically complex error-types in the synthetic data (dativitis and changes to noun cases and verb moods), we find that the mBART-ENIS trained on synthetic data generally does well; for some subsets even surpassing the ByT5 counterpart that was finetuned on curated corpora. We suspect that this has to do with the linguistic knowledge the model has already gained during its pretraining on Icelandic texts, as explained in Appendix A. The ByT5 model that was trained for longer however manages to surpass it on the mood error type, indicating that it is still adapting to the Icelandic language, alongside its primary denoising task. The models trained on only the finetuning data perform the worst throughout. The results show that they do not manage to correct the synthetic categories much beyond the baseline, except for mBART-ENIS in some cases. We expect that this has to do with their extra knowledge of Icelandic and the denoising objective used in the synthetic error correction finetuning. The results for these models on the curated in-domain test sets are in fact mostly on par with the models finetuned on the synthetic data only. Looking at the generated output, we see that the error types these models correct are not the same as those that the syntheticonly models are able to correct, which is expected, as they are trained on different data. We conclude that adopting a byte-level approach rather than a subword approach leads to best results for the task of GEC, at the very least in the case of a morphologically rich language such as Icelandic. Finally, we find that the optimal way of capturing a wide range of errors is to train on a combination of synthetic and curated data, particularly when the curated data is limited. Potential limitations to our work can mainly be attributed to two factors: 1) the fact that we run our experiments using the Icelandic language, and 2) inherent biases in the corpora we use. Icelandic is North Germanic language (along with Faroese, Norwegian, Danish and Swedish). As such, it is both Germanic and Indo-European. While we are fairly confident that our results hold for these languages, different results may hold for other languages, particularly those not using Latin script or those using logograms, such as Chinese characters. The curated datasets we use only represent a fairly small proportion of all possible demographics and users of the Icelandic language. In particular, annotations are performed by a handful of university students, bringing in their biases to the annotated data. Even so, the data should serve well to compare the relative differences. The resources we use to develop the models consist of a few high-performing GPUs. While these are powerful, this is a relatively low requirement compared to many industry or academic use cases. Finally, it is worth re-iterating that the ByT5 model we use is slow compared to subword-based models for texts of similar length. Inference in our setting was around 2.3x slower on average than for mT5. As such, production use of these methods may be better suited to offline processing, particularly for longer documents. While we do not believe the data we use to train the error-correcting models to be sensitive, the models can be applied in sensitive settings where an incorrect edit may cause an issue. As such, corrections may introduce both stylistic or semantic changes based on either the biases found in the pretrained models or the curated error corpora. In particular, we have noticed a bias in the subword-based models for entities, such as locations, being overcorrected to a different entity if there is a spelling mistake in the input. The stylistic changes found in the curated Icelandic corpora may reflect on the socio-economic background of the annotators and writers of the data. While we don't believe this to be a large issue in this particular setting, one can easily imagine this to be more complex in regions where language use is connected to disputes or oppression. As such, a text correction or improvement tool could be used to homogenize discourse or otherwise limit freedom of expression, knowingly or unknowingly. We continued the pretraining of the multilingual BART25 model using texts from various sources in Icelandic and English. The Icelandic text used was the Icelandic Common Crawl corpus (IC3) The goal of the pretraining is that the model already has some knowledge of the Icelandic language. English is included as we speculate that it is beneficial to continue including a language used in the earlier training, making the model a better starting point for other use cases such as machine translation and cross-lingual transfer tasks. Due to the English data outnumbering the Icelandic data, we upsampled the Icelandic data by about a factor of 6.5, resulting in a sample language probability of 55% English and 45% Icelandic. The model has 354M non-embedding parameters, and 256M parameters for embeddings. Note that not all of them are used when only training on Icelandic and English. The model was trained for 316k updates with an effective batch size of 44k tokens per update, 5k warmup steps, a learning rate of 7.5e-5 and a dropout of 0.2. Other hyperparameters, such as for noising, were the same as for the original mBART model. The training took approximately 18 A100 days but was not continued until convergence due to computational constraints. We collected a small test set of erroneous sentences by running news articles from the RÚV (National Icelandic Broadcasting Service) website through the open-source GreynirCorrect We generated various grammatical errors to create our synthetic error corpus. Icelandic has four grammatical cases; we swapped those randomly in nouns, producing ungrammatical sentences of a type commonly seen in learner texts. We also changed the mood of verbs from the subjunctive to the indicative, a variation often seen for both native speakers and learners. Another common variation, which is more or less accepted in informal language but still discouraged in formal language and written texts, is the so-called "dativitis", i.e., the use of the dative case instead of the accusative or nominative with certain verbs with oblique subjects, such as "mér (dat) hlakkar" ("I look forward to") instead of "ég (nom) hlakka", or "Páli (dat) langar" ("Páll wants") instead of "Pál (acc) langar". This modification was produced using the Greynir engine by extracting and modifying whole nominal clauses. We used available resources to apply realistic misspellings to single words -these are lists of common misspellings and their corrections. We located the corresponding correctly spelled words in the corpus and substituted their misspelled variants from the error lists. The error lists are sourced from nonwords and misspellings in IceSQuER We selected three sentences at random from the dyslexia subcorpus of the IceEC. These sentences, their corrected reference and the respective corrected outputs from various model variants can be seen in Table Table Það koma alltaf sama villan sem má sjá _ mynd 1. Reference Það kom alltaf sama villan sem má sjá á mynd 1. mB-ISEN-Synth100k+EC Það kemur alltaf sama villan sem má sjá á mynd 1. Ef notandi valdi að ýta á rétt kemur upp þessi síða þar sem notandi getur séð nánari upplýsingar um réttinn, innihaldsefni og naeringarupplýsingar. mB-ISEN-Synth100k+EC Ef notandi valdi að ýta á rétt kemur upp þessi síða þar sem notandi getur séð nánar um réttinn með innihaldsefnunum og naeringarupplýsingum. Það koma alltaf sama villan sem má sjá á mynd 1. Það koma alltaf sama villan sem má sjá á mynd 1. Það kom alltaf sama villan sem má sjá á mynd 1. Table A For every submission: A1. Did you describe the limitations of your work? Limitations section towards the end, unnumbered A2. Did you discuss any potential risks of your work? Limitations section towards the end, unnumbered A3. Do the abstract and introduction summarize the paper's main claims? The code and models are open-source B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Corpus data used has already been published by an official source and measures have been taken there to anonymize sensitive data, as described in the relevant publications. B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? 3 B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. C Did you run computational experiments? 3.2 C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? 3.2
| 1,174 | 2,606 | 1,174 |
From Simultaneous to Streaming Machine Translation by Leveraging Streaming History
|
Simultaneous Machine Translation is the task of incrementally translating an input sentence before it is fully available. Currently, simultaneous translation is carried out by translating each sentence independently of the previously translated text. More generally, Streaming MT can be understood as an extension of Simultaneous MT to the incremental translation of a continuous input text stream. In this work, a state-of-the-art simultaneous sentencelevel MT system is extended to the streaming setup by leveraging the streaming history. Extensive empirical results are reported on IWSLT Translation Tasks, showing that leveraging the streaming history leads to significant quality gains. In particular, the proposed system proves to compare favorably to the best performing systems.
|
Simultaneous Machine Translation (MT) is the task of incrementally translating an input sentence before it is fully available. Indeed, simultaneous MT can be naturally understood in the scenario of translating a text stream as a result of an upstream Automatic Speech Recognition (ASR) process. This setup defines a simultaneous Speech Translation (ST) scenario that is gaining momentum due to the vast number of industry applications that could be exploited based on this technology, from person-toperson communication to subtitling of audiovisual content, just to mention two main applications. These real-world streaming applications motivate us to move from simultaneous to streaming MT, understanding streaming MT as the task of simultaneously translating a potentially unbounded and unsegmented text stream. Streaming MT poses two main additional challenges over simultaneous MT. First, the MT system must be able to leverage the streaming history beyond the sentence level both at training and inference time. Second, the system must work under latency constraints over the entire stream. With regard to exploiting streaming history, or more generally sentence context, it is worth mentioning the significant amount of previous work in offline MT at sentence level Nonetheless, with the notable exception of In this work, the simultaneous MT model based on a unidirectional encoder-decoder and training along multiple wait-k paths proposed by This paper is organized as follows. Next section provides a formal framework for streaming MT to accommodate streaming history in simultaneous MT. Section 3 presents the streaming experimental setup whose results are reported and discussed in Section 4. Finally, conclusions and future work are drawn in Section 5.
|
In streaming MT, the source stream X to be translated into Y comes as an unsegmented and unbounded sequence of tokens. In this setup, the decoding process usually takes the greedy decision of which token appears next at the i-th position of the translation being generated where G(i) is a global delay function that tells us the last position in the source stream that was available when the i-th target token was output, and Y is the target vocabulary. However, taking into account the entire source and target streams can be prohibitive from a computational viewpoint, so the generation of the next token can be conditioned to the last H(i) tokens of the stream as Nevertheless, for practical purposes, the concept of sentence segmentation is usually introduced to explicitly indicate a monotonic alignment between source and target sentences in streaming MT. Let us consider for this purpose the random variables a and b for the source and target segmentation of the stream, respectively. Variables a and b can be understood as two vectors of equal length denoting that the n-th source sentence starts at position a n , while the n-th target sentence does so at position b n . In the next sections, we reformulate simultaneous MT in terms of the more general framework of streaming MT. This reformulation allows us to consider opportunities for improvement of previous simultaneous MT models. In the conventional simultaneous MT setup, the aforementioned variables a and b are uncovered at training and inference time, while in streaming MT a and b are considered hidden variables at inference time that may be uncovered by a segmentation model. In fact, in conventional simultaneous MT the history is limited to the current sentence being translated, while in streaming MT we could exploit the fact that the history could potentially span over all the previous tokens before the current sentence. To this purpose, the global delay function G(i) introduced above would replace the sentence-level delay function g(i) commonly used in simultaneous MT. However, it should be noticed that we could express g(i) as G(i) -a n with b n ≤ i < b n+1 . Delay functions are defined as a result of the policy being applied. This policy decides what action to take at each timestep, whether to read a token from the input or to write a target token. Policies can be either fixed (3) This policy initially reads k source tokens without writing a target token, and then outputs a target token every time a source token is read. This is true in the case that the ratio between the source and target sentence lengths is one. However, in the general case, a catch-up factor γ computed as the inverse of the source-target length ratio defines how many target tokens are written for every read token, that generalises Eq. 3 as The wait-k policy can be reformulated in streaming MT so that the wait-k behaviour is carried out for each sentence as where b n ≤ i < b n+1 . In streaming MT, we could take advantage of the streaming history by learning the probability distribution stated in Eq. 2, whenever streaming samples would be available. However, training such a model with arbitrarily long streaming samples poses a series of challenges that need to be addressed. Firstly, it would be necessary to carefully define G(i) and H(i) functions so that, at each timestep, the available source and target streams are perfectly aligned. Given that the source-target length ratio may vary over the stream, if one uses a wait-k policy with a fixed γ, there is a significant chance that source and target are misaligned at some points over the stream. Secondly, every target token can potentially have a different G(i) and H(i), so the encoder-decoder representation and contribution to the loss would need to be recomputed for each target token at a significant computational expense. Lastly, current MT architectures and training procedures have evolved conditioned by the availability of sentence-level parallel corpora for training, so they need to be adapted to learn from parallel streams. To tackle the aforementioned challenges in streaming MT, a compromise practical solution is to uncover the source and target sentence segmentations. At training time, parallel samples are extracted by a sentence-boundary sliding window spanning over several sentences of the stream that shifts to the right one sentence at a time. In other words, each sentence pair is concatenated with its corresponding streaming history that includes previous sentence pairs simulating long-span prefix training. Doing so, we ensure that source and target streams are properly aligned at all times, and training can be efficiently carried out by considering a limited history. The inference process is performed in a purely streaming fashion in a single forward pass as defined in Eq. 2 with H(i) being consistently defined in line with training, so that the streaming history spans over previous sentences already translated. In simultaneous MT, the conventional Transformerbased bidirectional encoder representation (of the l-th layer) of a source token at any position j is constrained to the current n-th sentence e (l) where a n ≤ j ≤ G(i), while the decoder can only attend to previous target words and the encoding of those source words that are available at each timestep As a result, the encoder and decoder representations for positions j and i, respectively, could be computed taking advantage of subsequent positions to position j up to position G(i) at inference time. However, at training time, this means that this bidirectional encoding-decoding of the source sentence has to be computed for every timestep, taking up to |y| times longer than the conventional Transformer model. To alleviate this problem, that is more restrictive than that in Eq. 6, and it consequently conditions the decoder representation, since G(i) in Eq. 7 depends on the specific k value employed at each training step. As mentioned above, the unidirectional encoder just requires a single forward pass of the encoder at training time, and therefore there is no additional computational cost compared with a conventional Transformer. However, it does not take into account all possible input tokens for different values of k. Indeed, the encoding of the j-th input token will not consider those tokens beyond the j-th position, even if including them into the encoding process does not prevent us from performing a single forward pass. A trade-off between the unidirectional and bidirectional encoders is what we have dubbed Partial Bidirectional Encoder (PBE), which modifies the unidirectional encoder to allow the first k-1 source positions to have access to succeeding tokens according to e (l) j = Enc e (l-1) an:max(an+k-1,j) . (9) PBE allows for a longer context when encoding the initial positions and is consistent with Eq. 7. At training time a single forward pass of the encoderdecoder is still possible as in the unidirectional encoder, and therefore no additional training cost is incurred. At inference time, we fall back to the bidirectional encoder. Figure In a streaming setup, the bidirectional encoderdecoder of Eqs. 6 and 7 are not necessarily constrained to the current sentence and could exploit a streaming history of H(i) tokens e (l) Likewise, the proposed PBE with streaming history states as follows e (l) 3 Experimental setup A series of comparative experiments in terms of translation quality and latency have been carried out using data from the IWSLT 2020 Evaluation Campaign Table Our streaming MT system is evaluated in terms of latency and translation quality with BLEU where Latency measures can be computed in a streaming manner by considering a global delay function G(i), that is mapped into a relative delay so that it can be compared with the sentence-level oracle delay. For the i-th target position of the n-th sentence, the associated relative delay can be obtained from the global delay function as g n (i) = G(i+b n )-a n . So, the stream-adapted cost function of the latency measures is defined as This definition assumes that the source and target sentence segmentation of the stream are uncovered, but this is not always the case (Schneider and Waibel, 2020) or they may not match that of the reference translations. However, sentence boundaries can be obtained by re-segmenting the system hypothesis following exactly the same procedure applied to compute translation quality in ST evaluation. To this purpose, we use the MWER segmenter Our streaming MT models have been trained following the conventional Transformer BASE (German↔English streaming MT) and BIG (English→German simultaneous MT) configurations The proposed model in Section 2 assumes that at inference time the source stream has been segmented into sentences. To this purpose, we opt for the text-based DS model Figure As observed, BLEU scores of the simultaneous MT systems leveraging on the streaming history (h > 0) are systematically and notably higher than those of conventional sentence-based simultaneous MT system (h = 0) over the range of wait-k values. Indeed, as the streaming history increases, BLEU scores also do reaching what it seems the optimal history length at h = 60 and slightly degrading at h = 80. As expected, when replacing the unidirectional encoder by the PBE, BLEU scores improve as the wait-k value increases, since PBE has additional access to those tokens from j + 1 up to k. For instance, for k = 32 and h = 60, PBE is 0.7 BLEU points above the unidirectional encoder. On the other hand, it can be observed how using an encoder which is not fully bidirectional during training, creates a performance gap with respect to the offline bidirectional model when carrying out inference in an offline manner (k ≥ 32). It can be also observed how the PBE model is better prepared for this scenario and shows a smaller gap. It is important to keep in mind that although both offline and PBE models behave the same way dur- ing inference for a large enough k, during training time the PBE model, trained using the multi-k with k randomly sampled for each batch, has been optimized jointly for low, medium and high latencies. In general, the bidirectional encoder shows poor performance for simultaneous MT. This can be explained by the fact that there exists a mismatch between the training condition (whole source available) and the inference condition (only a prefix of the source is available for k < 32). These results are consistent with Following (Schneider and Waibel, 2020)'s setup, the test set is lowercased and concatenated into a single stream. In order to measure the latency of the pipeline defined by the segmenter followed by MT system, it is necessary to take into account not only the latency of the MT system but also that of the segmenter. Thankfully this is straightforward to do in our pipeline, as a segmenter with a future window of length w modifies the pipeline policy so that, at the start of the stream, w READ actions are carried out to fill up the future window. Then, every time the MT system carries out a READ action, it receives one token from the segmenter. Thus, the integration of the segmenter into the pipeline is transparent from a latency viewpoint. Figure Next, we compare our proposed streaming MT (STR-MT) model with the λ = 0.3 ACT system (Schneider and Waibel, 2020) in terms of BLEU score and stream-adapted latency measures on Table ON-TRAC and our streaming MT system exhibit a similar progression, which is to be expected given that they are both based on the multi-k approach. However, our system consistently outperforms the ON-TRAC system by 1-2 BLEU. This confirms the importance of utilizing streaming history in order to significantly improve results, and how the proposed PBE model can take better advantage of the history. RWTH and KIT systems are closer in translation quality to our proposal than ON-TRAC, for AL between 5 and 7. However, these systems do not show a flexible latency policy and are not comparable to our system at other regimes. Indeed, for that to be possible, these systems need to be re-trained, in contrast to our system in which latency is adjusted at inference time. In this work, a formalization of streaming MT as a generalization of simultaneous MT has been proposed in order to define a theoretical framework in which our two contributions have been made. On the one hand, we successfully leverage streaming history across sentence boundaries for a simultaneous MT system based on multiple wait-k paths that allows our system to greatly improve the results of the sentence-level baseline. On the other hand, our PBE is able to take into account longer context information than its unidirectional counterpart, while keeping the same training efficiency. Our proposed MT system has been evaluated under a realistic streaming setting being able to reach similar translation quality than a state-of-theart segmentation-free streaming MT system at a fraction of its latency. Additionally, our system has been shown to be competitive when compared with state-of-the-art simultaneous MT systems optimized for sentence-level translation, obtaining excellent results using a single model across a wide range of latency levels, thanks to its flexible inference policy. In terms of future work, additional training and inference procedures that take advantage of the streaming history in streaming MT are still open for research. One important avenue of improvement is to devise more robust training methods, so that simultaneous models can perform as well as their offline counterparts when carrying out inference at higher latencies. The segmentation model, though proved useful in a streaming setup, adds complexity and can greatly affect translation quality. Thus, the development of segmentation-free streaming MT models is another interesting research topic. Figure In order to enable streaming translation, the best performing h = 40 systems has been combined with a German DS system. Similarly to previous experiments, we have conducted tests using different values of w and k in order to balance the Bidir. PBE Figure latency-quality trade-off, shown in Figure During training of the unidirectional and PBE encoders, the constraints imposed by Eqs. 8 and 9 are efficiently implemented by full self-attention, as in the bidirectional encoder, followed by an attention mask, for each token to only attend those tokens fulfilling the constraints. The attention mask sets the weights of the other tokens to -∞ before application of the self-attention softmax. This is exactly the same mechanism used in the standard Transformer decoder to prevent the auto-regressive decoder from accessing future information. This means that the three encoder types have an During inference time, however, the unidirectional encoder has some advantages. Given that the unidirectional encoder is incremental, meaning that the encodings of old tokens do not change when a new token becomes available, the process can be sped up by only computing the encoding of the newly available token. Although encoder self-attention still needs to be computed, a single vector is used as the query instead of the full matrix. Table The unidirectional encoder is four times faster than the bidirectional encoder when run on a CPU. However, both encoders perform the same when run on a GPU. For the streaming MT scenario considered in this work, no latency reduction is gained The multi-k systems have been trained with the official implementation ( Total training time was 40h for BASE models, and 60h for BIG models. The following command was used to train them: For finetuning, we change to the following: --lr-scheduler fixed \ --lr 4.47169e-05 \ For the streaming translation scenario, the data is lowercased and all punctuation signs are removed. For the simultaneous scenario (IWSLT 2020 simultaneous text-to-text), it is truecased and tokenized using Moses. We apply language identification to the training data using langid In order to obtain samples that can be used for training streaming MT models, a sliding window that moves over whole sentences is used to extract consistent source-target samples. Figure The Direct Segmentation system has been trained with the official implementation ( The following command was used to train the segmenter system: with the following configurations: (len=11; window=0) (len=12; window=1) (len=13; window=2) (len=14, window=3) (len=15, window=4)
| 786 | 1,764 | 786 |
Analogical Reasoning on Chinese Morphological and Semantic Relations
|
Analogical reasoning is effective in capturing linguistic regularities. This paper proposes an analogical reasoning task on Chinese. After delving into Chinese lexical knowledge, we sketch 68 implicit morphological relations and 28 explicit semantic relations. A big and balanced dataset CA8 is then built for this task, including 17813 questions. Furthermore, we systematically explore the influences of vector representations, context features, and corpora on analogical reasoning. With the experiments, CA8 is proved to be a reliable benchmark for evaluating Chinese word embeddings.
|
Recently, the boom of word embedding draws our attention to analogical reasoning on linguistic regularities. Given the word representations, analogy questions can be automatically solved via vector computation, e.g. "apples -apple + car ≈ cars" for morphological regularities and "kingman + woman ≈ queen" for semantic regularities It is well known that linguistic regularities vary a lot among different languages. For example, Chinese is a typical analytic language which lacks inflection. Figure However, few attempts have been made in Chinese analogical reasoning. The only Chinese analogy dataset is translated from part of an English dataset Therefore, we would like to investigate linguistic regularities beneath Chinese. By modeling them as an analogical reasoning task, we could further examine the effects of vector offset methods in detecting Chinese morphological and semantic relations. As far as we know, this is the first study focusing on Chinese analogical reasoning. Moreover, we release a standard benchmark for evaluation of Chinese word embedding, together with 36 open-source pre-trained embeddings at GitHub 1 , which could serve as a solid basis for Chinese NLP tasks.
|
Morphology concerns the internal structure of words. There is a common belief that Chinese is a morphologically impoverished language since a morpheme mostly corresponds to an orthographic character, and it lacks apparent distinctions between roots and affixes. Reduplication means a morpheme is repeated to form a new word, which is semantically and/or syntactically distinct from the original morpheme, e.g. the word "tiān-tiān"(day day) in Figure • bà(dad) → bà-bà(dad) • tiān(day) → tiān-tiān(everyday) • shuō(say) → shuō-shuo(say a little) • kàn(look) → kàn-kàn(have a brief look) • dà(big) → dà-dà(very big; greatly) • shēn(deep) → shēn-shēn(deeply) 1 Affixation is a morphological process whereby a bound morpheme (an affix) is attached to roots or stems to form new language units. Chinese is a typical isolating language that has few affixes. To model the semi-affixation process, we uncover 21 semi-prefixes and 41 semi-suffixes. These semi-suffixes can be used to denote changes of meaning or part of speech. For example, the semi-prefix "dì-" could be added to numerals to form ordinal numbers, and the semi-suffix "-zi" is able to nominalize an adjective: 3 Semantic Relations To investigate semantic knowledge reasoning, we present 28 semantic relations in four aspects: geography, history, nature, and people. Among them we inherit a few relations from English datasets, e.g. country-capital and family members, while the rest of them are proposed originally on the basis of our observation of Chinese lexical knowledge. For example, a Chinese province may have its own abbreviation, capital city, and representative drama, which could form rich semantic analogies: • ān-huī vs zhè-jiāng (province) • wǎn vs zhè (abbreviation) • hé-féi vs háng-zhōu (capital) • huáng-méi-xì vs yuè-jù (drama) We also address novel relations that could be used for other languages, e.g. scientists and their findings, companies and their founders. Analogical reasoning task is to retrieve the answer of the question "a is to b as c is to ?". Based on the relations discussed above, we firstly collect word pairs for each relation. Since there are no explicit word boundaries in Chinese, we take dictionaries and word segmentation specifications as references to confirm the inclusion of each word pair. To avoid the imbalance problem addressed in English benchmarks Table In Chinese analogical reasoning task, we aim at investigating to what extent word vectors capture the linguistic relations, and how it is affected by three important factors: vector representations (sparse and dense), context features (character, word, and ngram), and training corpora (size and domain). Table • Remove the html and xml tags from the texts and set the encoding as utf-8. Digits and punctuations are remained. • Convert traditional Chinese characters into simplified characters with Open Chinese Convert (OpenCC) • Conduct Chinese word segmentation with HanLP(v_1.5.3) Existing vector representations fall into two types, dense vectors and sparse vectors. SGNS (skipgram model with negative sampling) We can observe that on CA8 dataset, SGNS representations perform better in analogical reasoning of morphological relations and PPMI representations show great advantages in semantic relations. This result is consistent with performance of English dense and sparse vectors on MSR (morphology-only), SemEval (semanticonly), and Google (mixed) analogy datasets probably because the reasoning on morphological relations relies more on common words in context, and the training procedure of SGNS favors frequent word pairs. Meanwhile, PPMI model is more sensitive to infrequent and specific word pairs, which are beneficial to semantic relations. The above observation shows that CA8 is a reliable benchmark for studying the effects of dense and sparse vectors. Compared with CA_translated and existing English analogy datasets, it offers both morphological and semantic questions which are also balanced across different types To investigate the influence of context features on analogical reasoning, we consider not only word features, but also ngram features inspired by statistical language models, and character (Hanzi) features based on the close relationship between Chinese words and their composing characters 5 . Specifically, we use word bigrams for ngram features, character unigrams and bigrams for character features. Ngrams and Chinese characters are effective features in training word representations Besides, the representations achieve surprisingly high accuracies in some categories of CA_translated, which means that there is little room for further improvement. However it is much harder for representation methods to achieve high accuracies on CA8. The best configuration only achieves 68.0%. We compare word representations learned upon corpora of different sizes and domains. As shown in Table bination" which is built by combining the first five corpora together. Table Based on the above experiments, we find that vector representations, context features, and corpora all have important influences on Chinese analogical reasoning. Also, CA8 is proved to be a reliable benchmark for evaluation of Chinese word embeddings. In this paper, we investigate the linguistic regularities beneath Chinese, and propose a Chinese analogical reasoning task based on 68 morphological relations and 28 semantic relations. In the experiments, we apply vector offset method to this task, and examine the effects of vector representations, context features, and corpora. This study offers an interesting perspective combining linguistic analysis and representation models. The benchmark and embedding sets we release could also serve as a solid basis for Chinese NLP tasks.
| 586 | 1,192 | 586 |
Parallel Sentence Mining by Constrained Decoding
|
We present a novel method to extract parallel sentences from two monolingual corpora, using neural machine translation. Our method relies on translating sentences in one corpus, but constraining the decoding by a prefix tree built on the other corpus. We argue that a neural machine translation system by itself can be a sentence similarity scorer and it efficiently approximates pairwise comparison with a modified beam search. When benchmarked on the BUCC shared task, our method achieves results comparable to other submissions.
|
Having large and high-quality parallel corpora is critical for neural machine translation (NMT). One way to create such a resource is to mine the web 1. Accurately determining the semantic similarity of a sentence pair in two languages. 2. Efficiently scoring sentence similarity for all possible pairs across two languages. Scoring each source sentence against each target sentence results in unaffordable quadratic time complexity. A typical workflow reduces the search complexity in a coarse-to-fine manner by aligning documents then aligning sentences within documents More recent methods focus on direct sentence alignment. The results from Building and Using * Equal contribution. Comparable Corpora (BUCC) shared task show that direct sentence alignment can be done by sentence-level lexical comparison, neural comparison or a combination of the two To exploit quality gains from the attention mechanism, we propose to use a full NMT system with attention to score potentially parallel sentences. The way we avoid pairwise scoring is inspired by constrained decoding in NMT, where the choice of output tokens is constrained to a predefined list Our work makes two contributions to parallel sentence mining. First, instead of comparing translated text or neural similarity, we use an NMT model to directly score and retrieve sentences onthe-fly during decoding. Second, we approximate pairwise comparison with beam search, so only the top-scoring hypotheses need to be considered at each decoding step.
|
NMT systems can assign a conditional translation probability to an arbitrary sentence pair. Filtering based on this (Junczys-Dowmunt, 2018) won the WMT 2018 shared task on parallel corpus filtering We build a prefix tree (trie) containing all sentences in the target language corpus (Figure In post-expansion pruning, beam search creates hypotheses for the next word, prunes hypotheses to fit in the beam size, and then requires they be prefixes of a target language sentences. In practice, most sentences are do not have translations in the corpus and search terminates early if all hypotheses are pruned. In pre-expansion pruning, a hypothesis in the beam generates a probability distribution over all tokens, but only the tokens corresponding to children of the trie node can be expanded by the hypothesis. The search process is guaranteed to find at least one target sentence for each source sentence. Downstream filtering removes false positives. Algorithm 1 Trie-constrained beam search with maximum output length L, beam size B, vocabulary V and a pre-built trie trie beam Algorithm 1 presents both variants of our modified beam search algorithm. Besides canonical beam search, " v1" indicates post-expansion pruning while " v2" indicates pre-expansion pruning. Figure Pre-expansion pruning leaves each source sentence with an output, which needs to be filtered out if not parallel. We propose to use two methods. When NMT generates an output, a sentence level cross-entropy score is computed too. One way to perform filtering is to only keep sentences with a better per-word cross-entropy than a certain threshold. Another way is to use Bicleaner, an off-the-shelf tool which scores sentence similarity at sentence pair level The trie used in our NMT decoding should be fast to query and small enough to fit in memory. We use an array of nodes as the basic data structure. Each node contains a key corresponding to a vocabulary item, as well as a pointer to another array containing all possible continuations in the next level. Binary search is used to find the correct continuations to the next level. With byte pair encoding (BPE) To integrate the trie into the decoder, we maintain external pointers to possible children nodes in the trie for each active hypothesis. When the hypotheses are expanded at each time step, the pointers are advanced to the next trie depth level. This ensures that cross-referencing the trie has a negligible effect on decoding speed. We evaluate our method on the BUCC shared task, which requires participants to extract parallel sentences from large monolingual data of English and other languages When inspecting the BUCC shared task data, we discovered overlapping parallel sentences in the sample, train and test sets. For example, more than 60% of the German-English gold pairs in the test set appear in the train set too. We apply our methods on English (En) paired with German (De), French (Fr) and Russian (Ru) on BUCC sample data initially. We train separate translation models for each language into English. All models are Transformer-Base We choose beam size 90 by performing a grid search on De-En pair and keep it unchanged. Regarding the filtering for pre-expansion pruning, per-word conditional cross-entropy thresholds are tuned separately for each pair, because languages inherently have different (cross-)entropies. For Bicleaner, we stick to its default settings, except that we disable the language model filter. All our models translate into English, but our method is actually language-agnostic. Hence, we train a separate En→De model, which will allow us to compare our method in inverse translation directions. Table It turns out that pre-expansion works better than post-expansion. In order to directly compare with previous work, we tune parameters of its filtering thresholds on train data for De-En pair, and apply the pre-expansion variant on the test data. Our results, evaluated by the BUCC organisers, are reported in Table Finally, we conduct an add-on experiment to see how our system would perform with in-domain (1) Fr→En ( Experiments on the sample data in Table Regarding filtering methods, we notice that Bicleaner achieves a more balanced precision and recall, while filtering by per-word cross-entropy leads to very high precision but lower recall. Generally, the latter does better in terms of F1. Taking a union of the output from the two filtering methods results in a even more balanced precision and recall, without damaging F1. This implies that the two filtering techniques keep different sentence pairs. Table Overall our method does not outperform stateof-the-art which leverages neural embeddings. We identify several weaknesses: beam search can only find local optima, and a genuine parallel sentence cannot be recovered once it is pruned. Thus the method is vulnerable when parallel sentences have different word ordering. For example, "Por el momento, estoy bebiendo un café" (English: "At the moment, I am drinking a coffee") can hardly match "I am drinking a coffee at the moment", because an NMT system will have very low probability of generating a reordered translation, unless using an undesirably large beam size. Moreover, compared to methods that consider textual overlap, NMT is sensitive to domain mismatch and rare words Finally we discuss the limitation of evaluating our method on the BUCC task. First, our method based on NMT can be liable to favour machine-translated texts, whereas the BUCC data is unlikely to contain those. Next, we notice that some parallel sentences in BUCC data are not included in the gold alignments. For instance, in De-En train set, "de-000081259" and "de-000081260" are the same German sentence, and so are "en-000036940" and "en-000036941" on the English side. Gold alignments only include (de-000081259, en-000036940) and (de-000081260, en-000036941), but not the other two. Lastly, it still remains unknown if a system optimised for F1 will produce the sentences that can truly improve NMT performance. A typical parallel corpus mining workflow first aligns parallel documents to limit the search space for sentence alignment. Early methods rely on webpage structure With the assumption that matched documents are parallel (no cross-alignment), sentence alignment can be done by comparing sentence length in words There are also research efforts on parallel sentence extraction without the reliance on document alignment. In our work, using NMT as a similarity scorer relies on constrained decoding We bring a new insight into using NMT as a similarity scorer for sentences in different languages. By constraining on a target side trie during decoding, beam search can approximate pairwise comparison between source and target sentences. Thus, overall we present an interesting way of finding parallel sentences through trie-constrained decoding. Our method achieves a comparable F1 score to existing systems with a vanilla architecture and data. Maximising machine translation scores is biased towards finding machine translated text produced by a similar model. More research is needed on this problem given the prevalent usage of NMT. We hypothesise that part of the success of dual conditional cross-entropy filtering (Junczys-Dowmunt, 2018) is checking that scores in both directions are approximately equal, whereas a machine translation would be characterised by a high score in one direction. Finally, scalability is a key issue in large-scale mining of parallel corpora, where both quantity and quality are of concern. The scalability of direct sentence alignment without a document aligner has not been thoroughly investigated in our work, as well as other related work.
| 531 | 1,508 | 531 |
QuickGraph: A Rapid Annotation Tool for Knowledge Graph Extraction from Technical Text
|
Acquiring high-quality annotated corpora for complex multi-task information extraction (MT-IE) is an arduous and costly process for humanannotators. Adoption of unsupervised techniques for automated annotation have thus become popular. However, these techniques rely heavily on dictionaries, gazetteers, and knowledge bases. While such resources are abundant for general domains, they are scarce for specialised technical domains. To tackle this challenge, we present QuickGraph 1 , the first collaborative MT-IE annotation tool built with indirect weak supervision and clustering to maximise annotator productivity.
|
Hand-labelling is still the most reliable means to obtain quality training data to support deep learning applications; however, it is timeconsuming and resource-intensive Leveraging unsupervised techniques, however, is predicated on the availability of relevant external resources such as semantically aligned knowledge bases, and a priori knowledge of phenomena/concepts in the corpus of interest. In general domains, these are widely available, e.g. YAGO Numerous annotation tools exist, supporting many NLP tasks Weak/distant supervision To fill these gaps, we introduce QuickGraph, the first collaborative annotation tool for multitask IE that is designed to be: • Fast: Accelerates annotation via entity and relation propagation, and semantic clustering.
|
Many NLP annotation tools have been developed in recent years Historically, brat TeamTat Recent tools enhance annotator productivity using active and proactive learning, including APLenty active learning may also result in unsatisfactory corpus quality due to sample acquisition and reliability concerns 3 System highlights QuickGraph is a multi-task document-level hierarchical entity and relation annotation tool. Our tool supports annotations that are: i) hierarchical, ii) multi-label, iii) multi-class, and iv) nested. These attributes enable annotation of tasks such as named entity Additional key contributions of QuickGraph are its novel features for indirect weak supervision through annotation propagation (Section 3.2.2), semantic clustering of documents to promote annotator consistency (Section 3.4.1), and real-time knowledge graph construction from text (Section 3.6.1). Each of these contributions enable rapid annotation of corpora to support deep learning applications without the need of external resources such as knowledge bases, dictionaries, or gazetteers. 3.2 Why is QuickGraph fast? QuickGraph is available for free online and takes only minutes to create an account and setup a project for rapid annotation. Our tool provides preset ontologies for popular entity and relation annotation tasks including Like other tools, our tool permits preannotation of corpora at project creation. Preannotation reduces annotation effort by preapplying labels based on external resources such as gazetteers. A novel feature of Quick-Graph is its ability to pre-annotate both entities and relations through sets of pre-labelled artefacts QuickGraph supports corpus pre-processing as part of the project creation process rather than requiring external solutions. Consequently, corpora can be annotated end-to-end without external steps or dependencies, simplifying and speeding up the annotation process. Preprocessing stages currently consists of: i) character casing, ii) character removal, and iii) document deduplication. 3.3 Why is QuickGraph powerful? Unlike other tools, QuickGraph prospers with large-scale corpora. We have loaded and simultaneously annotated corpora consisting of 100,000 short user-generated texts with the tool whilst maintaining performance. While other tools limit annotators to a view small group of documents Filter and review project annotations before downloading. Minimum IAA Threshold Annotation Quality Select annotator(s) Exporting annotations to support deep learning applications is easy in QuickGraph (Figure 3.4 How does QuickGraph help consistency? An overlooked feature of current tools is document clustering to promote annotator productivity. Clustering is a core feature of QuickGraph, and has two primary benefits. First, annotators maintain a consistent mental model whilst annotating as clustered documents are likely to share semantic and express similar phenomena. Second, user actions are simplified as similar documents likely share similar concepts reducing the need to repetitively navigate through large hierarchical entity label spaces. Our tool implements agglomerative clustering of documents embedded with SBERT (Reimers and Gurevych, 2019) sentence embeddings. Applying relation annotations consistently is difficult and time-consuming Relation constraints are made possible as QuickGraph applies relations between entities (associated with token spans) rather than on token spans directly (see Figure 3.5 Why is QuickGraph intuitive? There was nothing so very remarkable in that; nor did Alice think it so very much out of the way to hear the Rabbit say to itself, "Oh dear! Oh dear! I shall be late!" (when she thought it over afterwards, it occurred to her that she ought to have wondered at this, but at the time it all seemed quite natural); but when the Rabbit actually took a watch out of its waistcoat-pocket, and looked at it, and then hurried on, Alice started to her feet, for it flashed across her mind that she had never before seen a rabbit with either a waistcoat-pocket, or a watch to take out of it, and burning with curiosity, she ran across the field after it, and fortunately was just in time to see it pop down a large rabbit-hole under the hedge. Either the well was very deep, or she fell very slowly, for she had plenty of time as she went down to look about her and to wonder what was going to happen next. First, she tried to look down and make out what she was coming to, but it was too dark to see anything; then she looked at the sides of the well, and noticed that they were filled with cupboards and book-shelves; here and there she saw maps and pictures hung upon pegs. She took down a jar from one of the shelves as she passed; it was labelled "ORANGE MARMALADE", but to her great disappointment it was empty: she did not like to drop the jar for fear of killing somebody underneath, so managed to put it into one of the cupboards as she fell past it. Down, down, down. Would the fall never come to an end? "I wonder how many miles I've fallen by this time?" she said aloud. "I must be getting somewhere near the centre of the earth. Let me see: that would be four thousand miles down, I think-" (for, you see, Alice Instead of providing annotators with everything but the kitchen sink, akin to the current generation of annotation tools At any time 3.6 Why is QuickGraph insightful? Novel to QuickGraph is its real-time knowledge graph construction from annotations Adjudication in our tool (Figure QuickGraph is a multi-user tool built using the modern full-stack framework MERN Cluster Server Server Client Database Multi-label: 75% of the reviewed tools support multi-label annotation. Of these, ActiveAnno and Paladin permit multi-labels, but are restricted to document classification tasks. Pre-annotation: 50% of the reviewed tools allow pre-annotation of corpora prior to manuallabelling. These tools are limited to entities, while QuickGraph also supports relations through triples. Automatic annotation: Less than 50% of the reviewed tools support automatic annotation. This feature is implemented through AIassistance, typically using active learning, and is limited to entity annotation. QuickGraph purposely does not have this feature, as we believe uncontrolled automatic annotation for complex MT-IE can be unproductive. Relation constraints: Of the reviewed tools, only INCEpTION allows for relation constraints. However, INCEpTION's constraints need to be expressed in a bespoke constraint language. In contrast, this feature of Quick-Graph requires users to simply specify entity domain and ranges on relations. Annotation propagation: Of the reviewed tools, only TextAnnotator provides annotation propagation via 'entity cascading'. However, this feature is restricted to entities, and the tool's interface is cumbersome and challenging to use. In contrast, QuickGraph allows easy and intuitive entity and relation propagation. Whilst QuickGraph is ready to use, there are features still under development, including: i) expanding available semantic embedding and clustering options, ii) improving annotation propagation processes, iii) relaxing interannotator agreement metrics, and iv) adding support for cross-document annotation.
| 616 | 759 | 616 |
The Language of Place: Semantic Value from Geospatial Context
|
There is a relationship between what we say and where we say it. Word embeddings are usually trained assuming that semantically-similar words occur within the same textual contexts. We investigate the extent to which semantically-similar words occur within the same geospatial contexts. We enrich a corpus of geolocated Twitter posts with physical data derived from Google Places and Open-StreetMap, and train word embeddings using the resulting geospatial contexts. Intrinsic evaluation of the resulting vectors shows that geographic context alone does provide useful information about semantic relatedness.
|
Words follow geographic patterns of use. At times the relationship is obvious; we would expect to hear conversations about actors in and around a movie theater. Other times the connection between location and topic is less clear; people are more likely to tweet about something they love from a bar than from home, but vice versa for something they hate. Previous work validates the relationship between the content of text and its physical origin. Geographically-grounded models of language enable toponym resolution We enrich a corpus of geolocated tweets with geospatial information describing the physical environment where they were posted. We use the geospatial contexts to train geo-word embeddings with the skip-gram with negative sampling (SKIPGRAM) model To enrich our collected tweets with geospatial features, we used publicly-available geospatial data from OpenStreetMap and the Google Places API. OpenStreetMap (OSM) is a crowdsourced mapping initiative. Users provide surveyed data such as administrative boundaries, land use, and road networks in their local area. In addition to geographic coordinates, each shape in the data set includes tags describing its type and attributes, such as shop:convenience and building:retail for a convenience store. We downloaded metro extracts for our 20 cities in shapefile format. To maximize coverage, we supplemented the OSM data with Google Places data from its web API, consisting of places tagged with one or more types (i.e. aquarium, ATM, etc). We enrich each geolocated tweet by finding the coordinates and tags for all OSM shapes and Google Places located within 50m of the tweet's coordinates. The enumerated tags become geographic contexts for training word embeddings. Figure 3 Geo-Word Embeddings SKIPGRAM learns latent fixed-length vector representations v w and v c for each word and context in a corpus such that v w • v c is highest for frequently observed word-context pairs. Typically a word's context is modeled as a fixed-length window of words surrounding it. word2vecf takes a list of (word, context) pairs as input. We train 300-dimensional geo-word embeddings denoted GEOD -where D indicates a radius -as follows. For each length-n tweet, we find all shapes within D meters of its origin and enumerate the length-m list of the shapes' geographic tags. The tweet in Figure to word2vecf for training GEO30 embeddings produced by the example tweet is an m × n list of (word, context) pairs: (it's, route:bus), (good, route:bus), ... (#TechTuesday, poi:marker), (#UPenn, poi:marker) The mean number of tags (m) per tweet under each threshold is 12.3 (GEO15), 21.9 (GEO30), and 38.6 (GEO50). The mean number of tokens (n) per tweet is 15.7.
|
To determine the extent to which geo-word embeddings capture useful semantic information, we first evaluate their performance on three semantic relatedness and four semantic similarity benchmarks (listed in Table To understand the impact of geographic contexts on the embedding model, we compare GEO15, GEO30, and GEO50 geo-word embeddings to the following baselines: TEXT5: Using our corpus of geolocated tweets, we train word embeddings with word2vecf using traditional linear bag-of-words contexts with window width 5. GEO30+TEXT5: We also evaluate the impact of combining textual and geospatial contexts. We train a model over the geolocated tweets corpus using both the geospatial contexts from GEO30 and the textual contexts from TEXT5. RAND30: Because our GEOD models assign the same geospatial contexts to every token in a tweet, we need to rule out the possibility that GEOD models are simply capturing relatedness between words that frequently appear in the same tweets, like movie and theater. We implement a random baseline model that captures similarities arising from tweet co-location alone. For each tweet, we enumerate the geospatial tags (i.e. contexts) for shapes within 30m of the tweet origin. Then, before feeding the m × n list of (word, context) pairs to word2vecf for training, we randomly map each tag type to a different tag type within the context vocabulary. For example, route:bus could be mapped to amenity:bank for input to the model. We redo the random tag mapping for each tweet. In this way, vectors for words that always appear together within tweets are trained on the same set of associated contexts. But the randomly mapped contexts do not model the geographic distribution of words. Qualitatively, we find that strongly locational words, like #nyc, and words frequently associated with a type of place, like burger and baseball, tend to have the most semantically and topically similar neighbors (Table We can also qualitatively examine the geographic context embeddings v c output by word2vecf. Recall that the SKIPGRAM objec- Table tive function pushes the vectors for frequently cooccurring v c and v w close to one another in a shared vector space. Thus we can find the words (Table In our intrinsic evaluation (Table Table Table below the current state-of-the-art; this is to be expected given the relatively small size of our training corpus (approx. 400M tokens). Our intrinsic evaluation established that geospatial context provides semantic information about words, but it is weaker than information provided by textual context. So a natural question to ask is whether geospatial context can be useful in any setting. One potential strength of word embeddings trained using geospatial contexts is that the features are language-independent. Thus we in- fer that training geo-word embeddings jointly over two languages might yield translation pairs that are close to one another in vector space. This type of model could be applicable in a low-resource language setting where large parallel texts are unavailable but geolocated text is. To test this hypothesis, we collect an additional 236k geolocated Turkish tweets and re-train GEO30, TEXT5, and GEO30+TEXT5 vectors on the larger set. Similar to We construct a logistic regression model, where the input for each word pair is the difference between its Turkish and English word vectors, v fv e . We evaluate the results using precision, recall, and F-score of positive translation predictions. Table Typically word embeddings are generated using the text surrounding a word as context from which to derive semantic information. We explored what happens when we use the geospatial context -information about the physical location where text originates -instead. Intrinsic evaluation of word embeddings trained over a set of geolocated Twitter data, using geospatial information derived from OpenStreetMap and the Google Places API as context, indicated that the geospatial context does encode information about semantic relatedness. We also suggested an extrinsic evaluation method for geo-word embeddings: predicting translation pairs without bilingual parallel corpora. Our experiments suggested that while geospatial context is not as semantically-rich as textual context, it does provide useful semantic relatedness information that may be complementary as part of a multimodal model. As future work, another extrinsic evaluation task that may be appropriate for geo-word embeddings is geolocation prediction.
| 608 | 2,714 | 608 |
Characterizing Departures from Linearity in Word Translation
|
We investigate the behavior of maps learned by machine translation methods. The maps translate words by projecting between word embedding spaces of different languages. We locally approximate these maps using linear maps, and find that they vary across the word embedding space. This demonstrates that the underlying maps are non-linear. Importantly, we show that the locally linear maps vary by an amount that is tightly correlated with the distance between the neighborhoods on which they are trained. Our results can be used to test non-linear methods, and to drive the design of more accurate maps for word translation.
|
Following the success of monolingual word embeddings A class of methods has emerged whose core technique is to learn linear maps between vector spaces of different languages x Mx M (en) (de) x (en) (de) M xn xn 0 Figure 1: Top: Assumption of linearity implies a single linear map M. Bottom: Our hypothesis is that the underlying map is expected to be nonlinear but in small enough neighborhoods can be approximated by linear maps M x i for each neighborhood defined by x i . word vector spaces exhibit comparable structure across languages, a linear mapping function is learned to connect the two monolingual vector spaces. The map can then be used to translate words between the language pair. Both seminal However, it is not well understood to what extent the assumption of linearity holds and how it affects performance. In this paper, we investigate the behavior of word translation maps, and show that there is clear evidence of departure from linearity. Non-linear maps beyond those generated by feedforward neural networks have also been explored for this task In this paper, we allow for an underlying mapping function that is non-linear, but assume that it can be approximated by linear maps at least in small enough neighborhoods. If the underlying map is linear, all local approximations should be identical, or, given the finite size of the training data, similar. In contrast, if the underlying map is non-linear, the locally linear approximations will depend on the neighborhood. Figure
|
To learn linear word translation maps, different loss functions have been proposed. The simplest is the regularized least squares loss, where the linear map M is learned as follows: here X and Y are matrices that contain word embedding vectors for the source and target language Non-linear cross-lingual mapping methods have been proposed. In Our hypothesis is similar in spirit to the use of locally linear embeddings for nonlinear dimensionality reduction In order to study the behavior of word translation maps, we begin by introducing a simple notion of neighborhoods in the embedding space. For a given language (e.g., English, en), we define a neighborhood of a word as follows: First, we pick a word x i , whose corresponding vector is x i ∈ X en , as an anchor. Second, we initialize a neighborhood N (x i ) containing a single vector x i . We then grow the neighborhood by adding all words whose cosine similarity to x i is ≥ s. The resulting neighborhood is defined as: Suppose we pick the word multivitamins as the anchor word. We can generate neighborhoods using N (x multivitamins , s) where for each value of s we get a different neighborhood. Neighborhoods corresponding to larger values of s are subsumed by those corresponding to smaller values of s. Figure Using this simple method, we can define different-sized neighborhoods around any word in the vocabulary. Given the above neighborhood definition, we now seek to understand how word translation maps change as we move across neighborhoods in word embedding space. Questions Studied. We study the following questions: [Q.1] Is there a single linear map for word translation that produces the same level of performance regardless of where in the vector space the words being translated fall? [Q.2] If there is no such single linear map, but instead multiple neighborhood-specific ones, is there a relationship between neighborhood-specific maps and the distances between their respective neighborhoods? In our first experiment we translate from English (en) to German (de). We obtained pretrained word embeddings from FastText In our all experiments, the cross-lingual maps are learned using the max-margin loss, which has been shown to perform competitively, while having fast run-times. Given a seed dictionary as training data of the form D tr = {x i , y i } m i=1 , the mapping function is where ŷi = Wx i is the prediction, k is the number of incorrect examples per training instance, and d(x, y) = (x -y) 2 is the distance measure. For the first experiment, we picked the following words as anchor words and obtained maps associated with each of their neighborhoods: M (multivitamins) , M (antibiotic) , M (disease) , M (blowflies) , M (dinosaur) , M (orchids) , M (copenhagen) . For each anchor word, we set s = 0.5, thus the neighborhoods are N (x i , 0.5) where x i is the vector of the anchor word. The training data for learning each neighborhood-specific linear map consists of vectors in N (x i , 0.5) and their translations. Table If indeed there exists a map that is the same linear map everywhere, we expect the above neighborhood-specific maps to be similar. Our analysis makes use of the following definition of matrix similarity: (2) Here tr(M) denotes the trace of the matrix M. tr(M 1 T M 1 ) computes the Frobenius norm ||M 1 || 2 , and tr(M 1 T M 2 ) is the Frobenius inner product. That is, cos(M 1 , M 2 ) computes the cosine similarity between the vectorized versions of matrices M 1 and M 2 . The main results of our analysis are shown in Table We now analyze the results of Table The 3rd column contains the cosine similarity between x 0 , multivitamins, and x i . For example, x 1 (antibiotic) is the most similar to x 0 (0.6), and x 6 , copenhagen, is the least similar to x 0 (0.11). The 4th column is the translation accuracy of the single global map M , training on data from all x i neighborhoods. The 5th column is the translation accuracy of the map M x 0 , trained on the train- ing data of x 0 , and tested on the test data in x i . We use precision at top-10 as a measure of translation accuracy. Going down this column we can see that accuracy is highest on the test data from the neighborhood anchored at x 0 itself, and lowest on the test data from the neighborhood anchored at x 6 , copenhagen, which is also the furthest word from x 0 . The 6th column is translation accuracy of the map M x i , trained on the training data of the neighborhood anchored at x i , and tested on the test data in x i . We can see that compared to the 5th column, in all cases performance is higher when we apply the map trained on data from the neighborhood, M x i instead of M x 0 . The 7th column shows the difference in translation accuracy of the map M x i and M x 0 . This shows that the more dissimilar the neighborhood anchor word x i is from x 0 according to the cosine similarity shown in the 4rd column, the larger this difference is. The 8th column shows the similarity between maps M x i and M x 0 as computed by Equation In order to determine the generality of our results, we carried out the same experiment on a different language pair, as shown in Table Our experimental study suggests the following: i) linear maps vary across neighborhoods, implying that the assumption of a linear map does not to hold. ii) the difference between maps is tightly correlated with the distance between neighborhoods. In this paper, we provide evidence that the assumption of linearity made by a large body of current work on cross-lingual mapping for word translation does not hold. We locally approximate the underlying non-linear map using linear maps, and show that these maps vary across neighborhoods in vector space by an amount that is tightly correlated with the distance between the neighborhoods on which they are trained. These results can be used to test non-linear methods, and we plan to use our finding to design more accurate maps in future work.
| 623 | 1,500 | 623 |
Translate to Disambiguate: Zero-shot Multilingual Word Sense Disambiguation with Pretrained Language Models
|
Pretrained Language Models (PLMs) learn rich cross-lingual knowledge and perform well on diverse tasks such as translation and multilingual word sense disambiguation (WSD) when finetuned. However, they often struggle at disambiguating word sense in a zero-shot setting. To better understand this contrast, we present a new study investigating how well PLMs capture cross-lingual word sense with Contextual Word-Level Translation (C-WLT), an extension of word-level translation that prompts the model to translate a given word in context. We find that as the model size increases, PLMs encode more cross-lingual word sense knowledge and better use context to improve WLT performance. Building on C-WLT, we introduce a zero-shot prompting approach for WSD, tested on 18 languages from the XL-WSD dataset. Our method outperforms fully supervised baselines on recall for many evaluation languages without additional training or finetuning. This study presents a first step towards understanding how to best leverage the crosslingual knowledge inside PLMs for robust zeroshot reasoning in any language.
|
Pretrained Language Models (PLMs) perform many cross-lingual tasks without explicit crosslingual training signal, including word-level translation (WLT) across languages To investigate this, we measure the ability of multilingual autoregressive language models to understand the cross-lingual meaning of words in a given context. Specifically, we extend the WLT task setup to include a specific context in the prompt, which we call Contextual Word-Level Translation (C-WLT). We empirically show that pretrained language models leverage contextual information in the prompt to improve WLT performance. In addition, both English and multilingual PLMs perform better on the contextual WLT tasks as model size increases, demonstrating improved cross-lingual knowledge at scale. Translations of a word that change based on context are frequently due to differing word senses not shared by an analogous word in the target language The overall findings of this work are as follows: • PLMs leverage contextual information to encode cross-lingual knowledge and better capture lexical information, such as word translations and meanings. • We can leverage this contextual knowledge of lexical translation to effectively perform zeroshot WSD for many languages, including lowresource ones and languages the PLM was not explicitly pretrained on. • The efficacy of WSD via C-WLT depends on different factors such as pretraining languages, model size, and target language choice: smaller multilingual PLMs perform well on seen languages, but they are more sensitive to design choices and do not generalize as well as larger English PLMs. In sum, we evaluate the lexical translation skills of PLMs in context, and we present a first step towards applying that skill to the downstream task of WSD. Given that most WSD training data outside of English are automatically created (e.g.,
|
A standard method of evaluating the cross-lingual capabilities of PLMs is the task of a word-level translation (WLT), where the model is prompted to translate a word w s from a source language L s into another target language L t (Gonen et al., 2020). However, this setup does not consider variations in the translation of w s into L t that occur when the surface form of w s represents multiple meanings (i.e., senses) in different contexts. We propose an extension of the word-level translation task, Contextual Word-Level Translation (C-WLT), which requires translating words correctly based on how they are used in a given context For example, if we want to translate "plant" into Chinese based on the context sentence "The plant sprouted a new leaf", we prompt the PLM with In the sentence "The plant sprouted a new leaf", the word "plant" is translated into Chinese as __. This evaluation allows us to quantify a PLM's ability to align meaning across languages in a contextspecific manner. Datasets We evaluate performance with the XL-WSD dataset We consider five target languages for our experiments: English, Chinese, Russian, Spanish, and Finnish. Our choice of target languages aims to cover semantically diverse target languages (to increase variety in the translation to sense mappings) while maintaining high coverage within the multilingual ontology. When evaluating English, we instead back off to the most common sense (MCS) of the word when the target language(s) does not cover an example in each evaluation setting. Models Picking the three most powerful PLMs from the previous section, we use the BLOOM models with 3 billion parameters and 7.1 billion parameters and the GPT-NeoX model with 20 billion parameters. While GPT-NeoX is primarily trained on English, the Bloom models are specifically pretrained on 6 out of the 18 evaluation languages of the XL-WSD dataset (Basque, Catalan, Chinese, English, French, and Spanish). Baselines We compare our approach with the Most Common Sense (MCS) baseline, which predicts each word's most common sense according to BabelNet Evaluation Metrics for WSD via C-WLT We consider two automatic metrics for evaluating the performance of the WSD via C-WLT approach. The first is recall, or how often the predicted label set contains at least one of the gold annotations for a given example. This metric is obtained from the XL-WSD evaluation script and is the standard evaluation for this benchmark; it is often reported as (and is equivalent to) F1 or accuracy in cases where the WSD model produces a single prediction. However, recall overestimates performance in cases where a WSD approach predicts many unrelated sense labels in addition to a correct one. Therefore, we also calculate the Jaccard index between the predicted set and the reference set of sense labels for each example: While the Jaccard index is a better automatic measure of similarity for sets than recall, the metric can underestimate performance in cases where other, closely related senses are appropriate in the given context yet not included in the reference sense set. The multilingual WSD results are summarized in Table We also find that despite being primarily pretrained on English, GPT-NeoX (20B) achieves higher recall and Jaccard index scores than Bloom-7.1 on ten source languages; most settings where the multilingual model performs better are on its pretraining languages, with little generalization to other languages. Finally, despite the Jaccard index scoring lower (by definition) than recall, we see similar performance trends across languages and models between recall and the Jaccard index in this ensemble setting. Finally, we analyze the extent to which adding context sentences resolves errors made by the PLMs in the standard WLT setting (Figure Surprisingly, we also observe that context helps correct complete translation errors at higher rates than it does to disambiguate the appropriate translation given a context sentence. This behavior generally holds for both the English and multilingual models and across all model scales. The smallest English models are an exception where very few errors of either type are resolved by context, despite their overall performance significantly improving in the C-WLT setting. Building on the intuition from the previous section that contextual word-level translation can differentiate between different meanings of a word in the source language, we apply C-WLT to the task of multilingual word sense disambiguation (Figure For instance, to disambiguate the word "plant" as it is used in the context "The plant sprouted a new leaf", we first prompt the PLM to translate "plant" into the chosen target language (e.g., Chinese) with the C-WLT setup from the previous section. We then take the PLM's top translation (in this case, "植物") and obtain its senses from a multilingual word sense ontology. We then label the example with the senses shared by "plant" and "植物". In the sentence "c", the word "w s " is translated into L t as LLM w s We translate each ambiguous word w s in context into a target language t with a PLM and label it with the intersection of its labels and the labels of the translation w top1 . The goal of word sense disambiguation (WSD) is to determine the meaning of the word w in a specific context c and label it with the sense label (or labels) that represents this meaning out of the candidate set of senses associated with that word, S. In our proposed approach, WSD via C-WLT, w and c are in a language L s , and word senses are from a multilingual ontology (BabelNet, First, we prompt a PLM with the C-WLT setting to translate w s based on c s into the target language L t . We then obtain the inventory of all possible translations of w s into L t from the multilingual word sense ontology and rank them with the PLM conditioned on the C-WLT prompt. We then label w s with the set of senses in the intersection of its candidate senses, S(w s ), and those of the topscoring translation under the PLM, S(w top1 ). This means the WSD via C-WLT method assigns a set of labels to w rather than a single sense label, unlike most supervised WSD classifiers. Ensembling Target Languages The described method for WSD via C-WLT obtains potential senses from translating into a single target language. We extend the method to ensemble the senses from a set of target languages T , as we hypothesize that senses shared by translations of w s in multiple typologically diverse languages are more likely to be relevant to the specific context at hand. This is supported by We first present the performance of our method for multilingual WSD on the two automatic metrics, recall and Jaccard index, and compare this approach to prior work on this task (Section 4.1). We then consider the effect of ablating different modeling choices on our method (such as the choice of target language for C-WLT and prompt language; Section 4. 2), and we analyze the types of errors the approach produces more closely (Section 4.3). Different Target Languages To investigate the effect of the target language(s) on contextual wordlevel translation in the WSD task, we consider five target languages: English, Chinese, Russian, Finnish, and Spanish. We also experiment with all combinations of these languages for the joint target language settings (Table where S i is the candidate sense set for the ith evaluation example and Ŝi is the set of senses predicted by our approach. Our ablations indicate a tradeoff between the Jaccard index and recall. For example, our approach achieves the highest recall performance using Spanish as the sole target language, but the resulting Jaccard index is worse than any other target setting we test. This behavior is likely because target languages more similar to the source (such as Spanish, which is closely related to many of the Western European source languages in the XL-WSD dataset) return a larger set of predicted senses, which in turn improves recall but at the expense of set similarity with the gold labels. This hypothesis is corroborated by the high delta increase of 20% in the predicted set size of the Spanish setting over the standard single-label predicted setting. However, this undesirable behavior is mitigated when using dissimilar target languages to the source and ensembling diverse languages. In our best setting of ensembling English, Chinese, and Russian, we find that the delta increase in the predicted set size is only 6.7%, while the Jaccard index increases by ∼6 points over Spanish. Furthermore, this ensembled setting still often outperforms prior approaches on recall. We then consider the effect of prompt language on the WSD via C-WLT method by ablating prompts in English, the evaluation source language, and the target language. The English, Chinese, French, and Span- Recall Jaccard NeoX B-7.1B NeoX B-7.1B Orig. 63.78 57.74 52.01 50.98 Annot. 74.01 74.54 54.29 52.73 Table ish prompts were obtained from or verified by native speakers; prompts in other languages are from Google Translate. We test two languages, Spanish and Chinese, as targets and evaluate (a) the overall performance of the method by the prompt language (Figure We observe that the gold annotations in the XL-WSD test sets mostly consist of one label. However, fine-grained word sense meanings are often similar or even overlapping, with fine-grained annotator agreement as low as 67% in some cases To investigate this further, we ask three native language speakers to reannotate 392 examples of Chinese test data manually. This analysis finds that 172 examples (or 44%) have additional closely related senses not included in the original annotations. For example, consider the sentence: "广播还没说 完，各班的同学早已纷纷冲出教室。" The results on the subset of the evaluation set show that, unsurprisingly, both models' recall and Jaccard index improve on the reannotated data (Table 3). We conclude that missing fine-grained annotations are one factor impacting our results. The many examples found during the analysis with other relevant senses indicate that the reference sets likely do not contain full coverage. This suggests that future research on multilingual WSD should consider the choice of reference sets to ensure that they reflect all relevant senses, as prior work has for English We first analyze the performance of PLMs in the new contextual word-level translation (C-WLT) setting to evaluate how well these models produce context-sensitive lexical translations. Other related work has instead tested the efficacy of prompting multilingual PLMs for sentence-level translation, such as We then apply the C-WLT setup to zero-shot multilingual WSD. This approach builds on More generally, WSD is closely related to and motivated by machine translation; In this work, we examine the ability of pretrained language models to utilize contextual information in cross-lingual settings. Specifically, we propose contextual word-level translation (C-WLT) and test different PLMs' ability to improve lexical translations in context. We then propose a zero-shot prompting technique for multilingual WSD, using C-WLT as a component. Our experiments show the method's effectiveness on 18 languages, including those not included in the PLM's pretraining. The performance of WSD via C-WLT relies on the relationship between pretraining languages, model size, and the choice of the target language: smaller multilingual PLMs are more effective for languages on which they have been pretrained but are more sensitive to design choices, lacking the broad applicability of their larger English counterparts. Future research examining these interactions and their tradeoffs more closely is vital for improving zero-shot WSD approaches and building better cross-lingual applications of PLMs in general. We recognize several limitations that influence C-WLT and our proposed approach for WSD. First, the WSD via C-WLT method depends on the composition of the multilingual word sense ontology we use to obtain cross-lingual word senses and translations. Lower coverage in the chosen target language will hinder the method's performance: we see this empirically in the case of English as an evaluation language, as no target language setting (including ensembling) fully covers English, which requires us to back off the MCS of each word. Similarly, the translation capability of PLMs, particularly for low-resource languages, may limit the effectiveness of both C-WLT and our WSD approach that relies on it. While we first present a study of the efficacy of C-WLT before incorporating it into our WSD method, due to data limitations (i.e., constructing a C-WLT data for each language pair that contains examples covering multiple senses of many different target words), we examine three high-resource language pairs. However, better cross-lingual PLMs can be directly integrated into our proposed approach as they are developed to improve multilingual WSD. Finally, our approach is not well-suited for distinguishing between very fine-grained word senses. While our small-scale manual precision analysis (Section 4.3) suggests that at least some WSD evaluation sets are not annotated with complete coverage of all relevant senses -leading to an underestimate of our approach's performance -the ability to differentiate between closely related senses precisely remains a hurdle for the WSD via C-WLT method, and addressing this issue in the future will further improve its applicability. The results show that the gap between MCS and LCS performance is relatively large for both metrics: we observe an average difference of 28.7 and 36.3 between MCS and LCS examples for recall and Jaccard index, respectively. We also find that the size of this performance gap is consistent between the GPT-NeoX and Bloom-7.1B models. We hypothesize that this performance gap stems from unbalanced latent sense supervision in the pretraining data that is due to the natural Zipfian distribution of senses in language This section details information from the Responsible NLP Checklist not covered elsewhere in the paper. Intended Usage of Artifacts To the best of our knowledge, our experiments all fall within the intended use cases of the GPT-Neo and BLOOM models. We also use all data resources -the XL-WSD dataset, BabelNet, and WordNet -as originally intended (i.e., for WSD modeling and evaluation). We provide the per-langauge results for the EN-ZH (Figure Additionally, Table
| 1,097 | 1,867 | 1,097 |
ZEROTOP: Zero-Shot Task-Oriented Semantic Parsing using Large Language Models
|
We explore the use of large language models (LLMs) for zero-shot semantic parsing. Semantic parsing involves mapping natural language utterances to task-specific meaning representations. LLMs are generally trained on publicly available text and code and cannot be expected to directly generalize to domain-specific parsing tasks in a zero-shot setting. In this work, we propose ZEROTOP, a zero-shot task-oriented parsing method that decomposes semantic parsing problem into a set of abstractive and extractive question-answering (QA) problems. For each utterance, we prompt the LLM with questions corresponding to its top-level intent and a set of slots and use the LLM generations to construct the target meaning representation. We observe that current LLMs fail to detect unanswerable questions; and as a result, cannot handle questions corresponding to missing slots. We address this by fine-tuning a language model on public QA datasets using synthetic negative samples. Experimental results show that our QA-based decomposition paired with the fine-tuned LLM can zero-shot parse ≈ 16% of utterances in the MTOP dataset.
|
Large language models (LLMs) are trained on publicly available text As illustrated in Figure
|
LLMs are increasingly used for semantic parsing in low-data scenarios utilizing canonical representations Problem Formulation We focus on taskoriented parsing with hierarchical intent-slot schema. Let I = {I 1 , I 2 , . . . , I n } and S = {S 1 , S 2 , . . . , S m } be the set of all possible toplevel intents and slots respectively. Each intent I j has a set of slots S j = {S j 1 , S j 2 , . . . , S j n } that can be filled. Possible slots in an intent are represented by the intent-to-slot mapping I2S: I → P(S), where P(•) is the powerset operator. Similarly, the inverse slot-to-intent mapping is represented by S2I: S → I. The input in our setting consists of I2S and S2I, but no annotated data. ZEROTOP requires users to provide a question per slot Q = {Q S 1 , Q S 2 , . . . , Q S k }, that represents their purpose. In a real-life setting, this can be obtained from a domain developer. Unconstrained Generation for Zero-Shot Intent Classification We view zero-shot intent classification as an abstractive QA problem. One intuitive way is to prime the LLM with a QA prompt and then constrain the generation to search over only valid intent labels Leveraging QA datasets for Slot Value Prediction Slot value prediction involves extracting phrases for a slot from the user utterance. We cast this as an extractive QA problem. All slots might not be mentioned in an input utterance. For example, in the MTOP dataset, on average, only onethird of possible slots are mentioned per utterance. The QA model needs to abstain from prediction for such missing slots. To analyze the abstaining capability of pre-trained QA models, we consider a few top-performing zero-shot LLMs T0-3B, GPT-3, and Codex with their corresponding prompts and experiment on a 500 sample subset of unanswerable questions from the SQuAD dataset To address this challenge, we leverage multiple publicly available QA datasets 1 to train Abstainer, a QA model capable of abstaining from prediction. Specifically, we generate synthetic unanswerable training samples by modifying existing QA data, and train a QA model jointly on existing datasets and synthetic unanswerable questions. For every (question, answer, context) triplet, we generate synthetic unanswerable questions by either (1) removing the sentence containing the answer span from the context, or (2) randomly sampling a context that doesn't have the same question. After training the Abstainer, we prompt it for each slot with its corresponding question for slot value prediction, in the following format: Answer the following question depending on the context. context: A user said, {utterance}. question: {slot question} answer: Nested Intents To identify nested intents, we assume knowledge of candidate nested intents that can be accommodated by each slot, represented by the slot-to-candidate-nested-intent mapping S2NI: S → P(I). Our method assumes that depth of output representations is at most 4 i.e. nested intents cannot further have more nested intents. One intuitive way is to prompt the LLM for nested intent with the intent prediction prompt. However, our unconstrained generation-based intent model would predict many false positive nested intents. We instead use Abstainer to prompt for their respective slots. If any slot value is identified, we consider its corresponding intent via S2I to be present as well. 1 The QA datasets details are mentioned in Appendix A.1 ZEROTOP: Putting it all together The pseudocode of ZEROTOP is mentioned in Algorithm-1. ZEROTOP employs a top-down, greedy prompting strategy, where we first prompt for intent and then, its respective slots. First, we obtain the top-level intent using the intent model. Based on the predicted intent, we prime the Abstainer for corresponding slots using their respective questions as prompts. For each identified slot value, we prompt the Abstainer for slots of candidate nested intents. We use the same prompt format for this step with the identified slot value now considered as the input utterance. Finally, we combine predicted intent, identified slot values, and nested intents to create the meaning representation. We experiment on the English language subset of MTOP We compare with constrained T0-3B, GPT-3, and Codex as slot models as well, however, when primed with a question corresponding to a slot, the output is constrained to be either from the utterance or from their corresponding phrases indicating that question cannot be answered. We com- Table pare with two kinds of prompting for slot values. MTQA From zero-shot intent classification results in Table As shown in Table Annotation Effort Analysis We use 74 samples i.e. one per slot to design questions for slots. To analyze annotation effort, we train an utterance-tomeaning representation T5-3B Greedy vs Beam search ZEROTOP follows a greedy strategy where we hierarchically prompt for top-level intent and for its corresponding slots. We compare it with the beam search strategy with beam size 3. Specifically, we consider 3 top-level intents and prompt for their corresponding slots, consider top-3 slot values for every slot and finally compute the best meaning representation based on their aggregated NLL scores. The NLL score of intent I m , its slots S j ∈ I2S(I m ), and their corresponding slot values slotValues[S j ] is computed: where α is tuned on a held-out validation set. Note that p(slotValues|I m ) is computed recursively for its nested intents. The complete match accuracy of the greedy prompting strategy on MTOP dataset is 15.89% and of beam search strategy is 16.86%. This demonstrates that beam search can improve performance with validation data. Without validation data and setting α to 0.5, performance drops We consider negative log-likelihood (NLL) as a confidence score and vary the threshold to abstain from prediction and plot F1 scores on the MTOP dataset. We show that this NLL threshold is difficult to tune using public QA datasets such as SQuAD as performance on answerable and unanswerable subsets is mutually exclusive. to 12.36% i.e. 3% less than greedy. Therefore, we believe greedy prompting is a better choice. Confidence score-based Abstainer study We can alternatively have LLMs abstain from prediction based on a confidence score based threshold. We consider negative log likelihood (NLL) of the predicted slot value as the confidence score and abstain from prediction if it is greater than the threshold. We experiment on slot value prediction task with T0-3B, Codex, and GPT3 as LLMs and plot macro F1 scores for multiple NLL thresholds on a randomly sampled subset of 500 samples from MTOP dataset in Figure In this paper, we propose ZEROTOP that decomposes semantic parsing into abstractive and extractive QA tasks. ZEROTOP identifies top-level intent by generating in an unconstrained fashion and inferring the intent label most similar to the generated description. We train Abstainer using public QA datasets, that is capable of identifying unanswerable questions and abstaining from prediction. ZEROTOP assumes that the meaning representations are of a limited depth i.e. nested intents cannot further have more nested intents and this is one of the limitations. Moreover, we also assume that it is possible to write natural questions corresponding to slots. A slot for which a natural question cannot be expressed, the LLM can't handle it without additional supervision. Finally, we believe there is a huge scope for improvement in the performance of LLMs and ZEROTOP in domain-specific tasks such as zero-shot semantic parsing and on the MTOP dataset. This paper proposes a zero-shot semantic parsing method using large language models. The aim of the paper is to minimize the human effort in annotation by leveraging language models. The output of our method is a meaning representation that doesn't contain any harmful content. Hence, we do not anticipate any major ethical concerns.
| 1,124 | 92 | 1,124 |
HETERMPC: A Heterogeneous Graph Neural Network for Response Generation in Multi-Party Conversations
|
Recently, various response generation models for two-party conversations have achieved impressive improvements, but less effort has been paid to multi-party conversations (MPCs) which are more practical and complicated. Compared with a two-party conversation where a dialogue context is a sequence of utterances, building a response generation model for MPCs is more challenging, since there exist complicated context structures and the generated responses heavily rely on both interlocutors (i.e., speaker and addressee) and history utterances. To address these challenges, we present HeterMPC, a heterogeneous graph-based neural network for response generation in MPCs which models the semantics of utterances and interlocutors simultaneously with two types of nodes in a graph. Besides, we also design six types of meta relations with node-edge-typedependent parameters to characterize the heterogeneous interactions within the graph. Through multi-hop updating, HeterMPC can adequately utilize the structural knowledge of conversations for response generation. Experimental results on the Ubuntu Internet Relay Chat (IRC) channel benchmark show that HeterMPC outperforms various baseline models for response generation in MPCs.
|
Enabling dialogue systems to converse naturally with humans is a challenging yet intriguing problem of artificial intelligence and has attracted increasing attention due to its promising potentials and alluring commercial values Although sequence-to-sequence (Seq2Seq) models On account of above issues, we propose a heterogeneous graph-based neural network for response generation in MPCs, named HeterMPC. First, a heterogeneous graph is designed which employs two types of nodes to represent utterances and interlocutors respectively. Different from previous methods that built a homogeneous graph modeling only utterances, utterances and interlocutors are modeled simultaneously in HeterMPC, so that the complicated interactions between interlocutors, between utterances, and between an interlocutor and an utterance can be explicitly described. In order to characterize the heterogeneous attention over each (source, edge, target) triple, model parameters dependent on both types of nodes and edges are introduced when calculating attention weights and passing messages. Specifically, we introduce six types of meta relations for modeling different edges including "reply" and "repliedby" between two utterances, "speak" and "spokenby" between an utterance and a speaker, and "address" and "addressed-by" between an utterance and an addressee. With these node-edge-typedependent structures and parameters, HeterMPC can better utilize the structural knowledge of conversations for node representation and response generation than conventional homogeneous graphs. Finally, Transformer is employed as the backbone of HeterMPC and its model parameters can be initialized with PLMs to take advantage of the recent breakthrough on pre-training. We evaluate HeterMPC on the Ubuntu Internet Relay Chat (IRC) channel benchmark released by In summary, our contributions in this paper are three-fold: 1) To the best of our knowledge, this paper is the first exploration of using heterogeneous graphs for modeling conversations; 2) A Transformer-based heterogeneous graph architecture is introduced for response generation in MPCs, in which two types of nodes, six types of meta relations, and node-edge-type-dependent parameters are employed to characterize the heterogeneous properties of MPCs; 3) Experimental results show that our proposed model achieves a new stateof-the-art performance of response generation in MPCs on the Ubuntu IRC benchmark.
|
Multi-Party Conversation Existing methods on building dialogue systems can be generally categorized into generation-based Early studies on graph neural networks (GNNs) focused on homogeneous graphs where a whole graph is composed of a single type of nodes. However, graphs in real-world applications usually come with multiple types of nodes, namely heterogeneous information networks (HINs) or heterogeneous graphs (2019) adopted different RNNs for different types of nodes to integrate multi-modal features. Previous studies have verified the superiority of modeling MPCs with homogeneous graphs considering only utterances. We claim that it is indeed necessary to model a complex information flow in MPCs shown in Figure The task of response generation in MPCs is to generate an appropriate response r given the conversation history, the speaker of a response, and which utterance the response is going to reply to, which can be formulated as: Here, G is a heterogeneous graph containing both history conversation and the response to be generated. The speaker and addressee of the response are known and its contents are masked. The response tokens are generated in an autoregressive way. r k and r <k stand for the k-th token and the first (k -1) tokens of response r respectively. |r| is the length of r. : Interlocutor : Utterance : Replied-by : Reply : Speak : Spoken-by : Addressed-by : Address Figure We will introduce how to construct the graph and how to model the probability in Eq. ( HeterMPC adopts an encoder-decoder architecture consisting of stacked encoder and decoder layers for graph-to-sequence learning A heterogeneous graph is constructed to explicitly model the complicated interactions between interlocutors, between utterances, and between an interlocutor and an utterance in an MPC. This graph models utterances and interlocutors by considering them as two types of nodes under a unified framework. Given an MPC instance composed of M utterances and I interlocutors, a heterogeneous graph G(V, E) is constructed. Specifically, V is a set of M + I nodes. Each node denotes either an utterance or an interlocutor. E = {e p,q } M +I p,q=1 is a set of directed edges. Each edge e p,q describes the connection from node p to node q. Inspired by In HeterMPC, each node is represented as a vector. These vectors are first initialized individually without considering graph edges. token is inserted at the start of each utterance, denoting the utterance-level representation. Besides, a [SEP] token is also inserted at the end of each utterance where m ∈ {1, ..., M }, l ∈ {0, ..., L 1 -1}, L 1 denotes the number of Transformer layers for initialization, H l m ∈ R km×d , k m denotes the length of an utterance and d denotes the dimension of embedding vectors. Interlocutors Different from an utterance composed of a sequence of tokens, an interlocutor is directly represented with an embedding vector. Interlocutors in a conversation are indexed according to their speaking order and the embedding vector for each interlocutor is derived by looking up an order-based interlocutor embedding table As shown in Figure Since the representations of two types of nodes are initialized in different ways, node-type-dependent linear transformations are first applied to node representations before attention calculation so that the two types of nodes share similar feature distributions Formally, let the triple (s, e, t) denote an edge e connecting a source node s to a target node t. The representations of the source and target nodes at the l-th iteration Here, τ (s), τ (t) ∈ {UTR, ITR} distinguish utterance (UTR) and interlocutor (ITR) nodes. Eqs. ( When passing the message of a source node that serves as a value (V ) vector to a target node, node-edge-type-dependent parameters are also introduced considering the heterogeneous properties of nodes and edges. Mathematically: where vl (s) is the passed message and all W * ∈ R d×d and b * ∈ R d are parameters to be learnt. For a target node, the messages passed from all its connected source nodes need to be aggregated. A softmax function is applied to normalize the attention weights and then the messages from all source codes are summarized as: where S(t) denotes the set of source nodes for the target node t. Then the summarized message hl t is aggregated with the original node representation h l t using a node-type-dependent FFN followed by a residual connection where the output h l+1 t is used as the input of the next iteration of node updating. One iteration can be viewed as a single-step information propagation along edges. When stacking L 2 iterations, a node can attend to other nodes up to L 2 hops away. A specific consideration on utterance nodes is that the tokens except [CLS] in an utterance have no access to other utterances during the node updating process introduced above. To overcome this disadvantage and derive more contextualized utterance representations, an additional Transformer layer The decoder is composed of a stack of identical layers as shown in Figure We evaluated our proposed method on the Ubuntu IRC benchmark used in We compared our proposed methods with as many MPC models as possible. Considering that there are only a few research papers in this field, several recent advanced models were also adapted to provide sufficient comparisons. Finally, we compared with the following baseline models: (1) RNN-based Seq2Seq To ensure all experimental results were comparable, we used the same automated and human evaluation metrics as those used in previous work Model parameters were initialized with pre-trained weights of bert-base-uncased released by The learning rate was initialized as 6.25e-5 and was decayed linearly down to 0. The max gradient norm was clipped down to 1.0. The batch size was set to 16 with 8 gradient accumulation steps. The maximum utterance length was set to 50. The number of layers for initializing utterance representations L 1 was set to 9, and the number of layers for heterogeneous graph iteration L 2 was set to 3. L 1 and L 2 were validated on the validation set. The number of decoder layers L 3 was set to 6, achieving the best performance out of {2, 4, 6, 8} on the validation set. The strategy of greedy search was performed for decoding. The maximum length of responses for generation was also set to 50. All experiments were run on a single GeForce RTX 2080 Ti GPU. The maximum number of epochs was set to 15, taking about 40 hours. The validation set was used to select the best model for testing. All code was implemented in the PyTorch framework In our experiments, BERT and BART were selected to initialize HeterMPC. HeterMPC BERT denoted that the utterance encoder was initialized with BERT and the decoder was randomly initialized. HeterMPC BART denoted the encoder and decoder were initialized by those of BART respectively. Automated Evaluation Table To further verify the effectiveness of our proposed methods, ablation tests were conducted as shown in Table Human Evaluation Table The impact of numbers of iterations (L 2 ). Figure The impact of conversation length. Robustness. The addressee labels are important for constructing a graph used in HeterMPC. This kind of label is commonly available in real life such as "A@B" labels in group chatting, Twitter, Reddit and various forums that denote speaker A talking to addressee B. However, addressee labels of a part of utterances are missing in the existing MPC datasets since a speaker may forget to specify an addressee. HeterMPC is robust since utterances without addressee labels can be assigned with a general addressee label "To all interlocutors". We leave evaluation on other datasets in future work. We present HeterMPC to model complicated interactions between utterances and interlocutors in MPCs with a heterogeneous graph. Two types of graph nodes and six types of edges are designed. Node-edge-type-dependent parameters are introduced for better utilizing the structural knowledge of conversations during node updating. Results show that HeterMPC outperforms baselines by significant margins, achieving a new state-of-theart performance for response generation in MPCs on the Ubuntu IRC benchmark. In the future, we will explore better ways of maximizing feature distribution differences to model heterogeneity.
| 1,231 | 2,444 | 1,231 |
Experiments on Morphological Reinflection: CoNLL-2017 Shared Task
|
We present two systems for the task of morphological inflection, i.e., finding a target morphological form, given a lemma and a set of target tags. Both are trained on datasets of three sizes: low, medium and high. The first uses a simple Long Short-Term Memory (LSTM) for lowsized dataset, while it uses an LSTMbased encoder-decoder based model for the medium and high sized datasets. The second uses a simple Gated Recurrent Unit (GRU) for low-sized data, while it uses a combination of simple LSTMs, simple GRUs, stacked GRUs and encoderdecoder models, depending on the language, for medium-sized data. Though the systems are not very complex, they give accuracies above baseline accuracies on high-sized datasets, around baseline accuracies for medium-sized datasets but mostly accuracies lower than baseline for low-sized datasets.
|
The CoNLL-SIGMOPRHON 2017 shared task
|
Prior to neural network based approaches to morphological reinflection, most systems used a 3step approach to solve the problem: 1) String alignment between the lemma and the target (morphologically transformed form), 2) Rule extraction from spans of the aligned strings and 3) Rule application to previously unseen lemmas to transform them. We have modeled our system based on the system proposed by In all the models, some structural and hyperparametrical features remain the same. The characters in the root word are represented using character indices, while the morphological features of the target word are represented using binary vectors. Each character of the root word is then embedded as a character embedding of dimension 64, to form the root word embedding. If an encoder is used, it is bidirectional and the the input word embeddings feed into it. The output of the encoder (if any), concatenated with the root word embedding, feeds into the decoder. All recurrent units have hidden layer dimensions of 256, meaning that they transform the input to a vector of dimension 256. Over the decoder layer is a softmax layer that is used to predict the character that must occur at each character position of the target word. In order to maintain a constant word length, we use paddings of '0' characters. All models use categorical cross-entropy as the loss function and the Adam optimizer as reported by Kingma and Ba (2014) for optimization. For training the model on the low-sized dataset, we did not use any encoder and we used a simple GRU, as reported by For medium-sized dataset, we used different model configurations for different languages. Four different kinds of configurations were used: 1) Bidirectional LSTM as the encoder and a simple LSTM with a single layer as the decoder (Figure The specific configuration used for each language has been listed in Table For high-sized data, we were unable to complete experiments for the second submission due to lack of time. However, we have been able to perform Bulgarian, Czech, Estonian, Faroese, German, Icelandic, Irish, Latvian, Lithuanian, Norwegian-Bokmal, Persian, Polish, Swedish 4 Albanian, Armenian, Danish, Dutch, English, Finnish, French, Haida, Hebrew, Hindi, Hungarian, Italian, Khaling, Lower-Sorbian, Macedonian, Navajo, Northern-Sami, Norwegian-Nynorsk, Portuguese, Romanian, Russian, Scottish-Gaelic, Serbo-Croatian, Slovak, Slovene, Sorani, Spanish, Turkish, Ukrainian, Welsh The evaluation results were obtained using the evaluation script and the test set provided by the shared task organizers. Baseline accuracies were also obtained from the baseline model provided. The best five baseline accuracies, accuracies for the first submission and accuracies for the second submission can be found in Table However, it is not evident from these tables that morphologically more complex languages should be harder to learn, which seems to be counter- This may be because of hyperparameters or configurations selected for different languages (which were different, in an attempt to maximize accuracy on the development data). Figures 6 to 10 show the correlation between accuracy and Levenshtein distance for all three sizes of datasets for submission-1 and for low and medium sizes of datasets for submission-2. While we were unable to run an exhaustive hyperparameter search due to lack of time, we performed some experiments, where the choice of hyperparameters was guided by intuitions developed from analysis of the dataset and results obtained on smaller subsets of the data. We have presented some key observations from our analysis in the ensuing sub-sections. We observed that for low-sized datasets, both the models (LSTM as well as GRU based) required that at least 10 epochs be run before early stop, every time no progress is detected on the validation set. Setting this patience to less than 5, resulted in near 0 accuracies for most languages and printing of nonsensical target words. For medium-sized datasets, this patience value can be set to around 6-8 while for high-sized datasets, it can be set to around 3-4. However, in order to ensure best results, we set our patience value to 10 across all models, training sizes and languages in the final system. In last year's version of the shared task, the morphological features in the dataset were annotated along with the category of each feature. For instance, a sample training feature set from last year is: 'pos=N,def=DEF,case=NOM/ACC/GEN,num=SG'. This year, however, the category of each feature was not provided, i.e., the same example above would appear in this year's format as: 'N,DEF,NOM/ACC/GEN,SG'. Our studies show that while it is conceptually true that the presence of feature categories means exploring a shorter search space, the absence of them does not make a difference to the accuracies obtained for high and medium sized datasets. In the case of low-sized datasets, marginally better accuracies (around 0.5-1%) were obtained when the categories were incorporated into the dataset (this was done manually). However, this might also be the effect of random initialization of parameters. Simple Recurrent Neural Networks (RNNs) performed the poorest on all sizes of datasets. For low-sized datasets, in almost all cases, using a GRU gave better results than using an LSTM. On an average, the accuracy increased by 2.33% when shifting from LSTM to GRU as the choice of recurrent unit. In the case of medium-sized datasets, 8 out of 52 languages performed better with an LSTM than a GRU, while the rest showed better performance with a GRU. We also ran experiments using convolutional layers, in which the root word was convolved and the convolution was concatenated along with the root word and passed to the encoder layer (if any). The rest of the network structure remained the same. For low-sized and medium-sized datasets, adding convolutional layers resulted in the accuracy dropping to near 0. For high-sized datasets, we were unable to finish running the experiments on all languages due to lack of time. However for the few languages on which we performed convolutional ablation studies, it did seem to improve accuracy by around 1.5% on an average. Deeper models (more than one layer of LSTM/GRU) resulted in drastic accuracy drops for low-sized datasets. For medium-sized datasets, 30 out of 52 languages showed an accuracy improvement upon stacking two GRU layers, while the accuracy drop in the rest 22 was not drastic but appreciable. There are two main conclusions. One is that different configurations of deep neural networks work well for different languages. The second is that deep learning may not be the right approach for low-sized data. Results for low-size were poor for almost all languages. It is to be noted that we used purely deep learning. If deep learning is augmented with other transduction, rule-based or knowledge-based methods, the results for low-size could perhaps be improved. For high-sized data, for one language (Basque), we even got an accuracy of 100%. For medium, the highest was 93% and for low, the highest was 69%. Sepp We would like to thank Shaili Jain, Aanchal Chaurasia and Himanshu Karu for their help in our experiments in this shared task. In Tables
| 836 | 37 | 836 |
How does the pre-training objective affect what large language models learn about linguistic properties?
|
Several pre-training objectives, such as masked language modeling (MLM), have been proposed to pre-train language models (e.g. BERT) with the aim of learning better language representations. However, to the best of our knowledge, no previous work so far has investigated how different pre-training objectives affect what BERT learns about linguistics properties. We hypothesize that linguistically motivated objectives such as MLM should help BERT to acquire better linguistic knowledge compared to other non-linguistically motivated objectives that are not intuitive or hard for humans to guess the association between the input and the label to be predicted. To this end, we pre-train BERT with two linguistically motivated objectives and three non-linguistically motivated ones. We then probe for linguistic characteristics encoded in the representation of the resulting models. We find strong evidence that there are only small differences in probing performance between the representations learned by the two different types of objectives. These surprising results question the dominant narrative of linguistically informed pre-training. 1
|
The most popular way to pre-train a transformerbased Apart from MLM, different types of objectives have been recently proposed. Meanwhile, several studies have explored how well and to what extent LMs learn linguistic information. This is usually examined using probing tasks, i.e. simple classification tasks that test the LM's encodings for a single linguistic feature such as grammatical information. It has been found through probing that BERT encodes syntactic In this paper, we hypothesize that linguistically motivated objectives (e.g. MLM) should help BERT to acquire better linguistic knowledge compared to using non-linguistically motivated objectives, i.e. tasks that are hard for humans to guess the association between the input and the label to be predicted. To this end, we seek to answer the following research question: How does the pretraining objective affect what LMs learn about the English language? Our findings challenge the MLM status quo, showing that pre-training with non-linguistically informative objectives ( §2) results in models with comparable linguistic capabilities, as measured by standard probing benchmarks ( §3). These surprising results ( §4) suggest that careful analysis of how LMs learn is critical to further improve language modeling ( §5).
|
We experiment with five different pre-training objectives. Two of them are considered linguistically motivated while the rest are not. Masked Language Modeling (MLM): We use MLM as our first linguistically motivated pretraining objective. First introduced by We also experiment with a simpler linguistically motivated objective, where the model selects and replaces 10% of input tokens with shuffled tokens from the same input sequence. Concurrently, it selects and replaces another 10% of input tokens with random tokens from the vocabulary We assume that tasks that are hard for humans (such as a completely random prediction task) will make less likely the deeper layers of BERT (i.e. closer to the output layer) to acquire meaningful information about language. We also hypothesize that layers closer to the input might learn word co-occurrence information For our first non-linguistically motivated pretraining objective, we use the masked first character prediction introduced by Masked ASCII Codes Summation Prediction (ASCII): We also propose a new nonlinguistically motivated pre-training objective, where the model has to predict the summation of the ASCII code values of the characters in a masked token. To make this harder and keep the number of classes relatively small, we define a 5-way classification task by taking the modulo 5 of the ASCII summation: V = [ i ascii(char i )] %5. Guessing the association between the input and such label, is an almost impossible task for a human. Masked Random Token Classification (Random): Finally, we propose a completely random objective where we mask 15% of the input tokens and we assign each masked token a class from 0 to 4 randomly for a 5-way classification similar to the ASCII task. We assume that a model pre-trained with a random objective should not be able to learn anything meaningful about linguistic information. Probing tasks Surface information task: SentLen aims for correctly predicting the number of words in a sentence. Models We pre-train BERT-BASE Pre-training Data All models are pre-trained on the BookCorpus Pre-training Details Due to limited computational resources, each BASE model is pre-trained for 500k steps, while each MEDIUM and SMALL model is pre-trained for 250k steps using 8 NVIDIA Tesla V100 (SXM2 -32GB). We use a batch size of 32 for BASE, and 64 for MEDIUM and SMALL. We optimize the models using Adam (Kingma and Ba, 2014). We use the General Language Understanding Evaluation (GLUE) benchmark Table For the BASE model configuration, we observe that linguistically motivated objectives (e.g. MLM, S+R) achieve the best performance in downstream tasks. However, models pre-trained with nonlinguistically motivated objectives (e.g. First Char, ASCII) still achieve competitive results. As expected, the model pre-trained using the Random objective obtains the lowest performance with 56.4 GLUE average score. However, its performance is still reasonable in many downstream tasks, suggesting that the model is able to learn some cooccurrence information from the input Table The full results of the probing tasks including all layers can be found in appendix B. Theoretically, LMs with non-linguistically motivated objectives would be expected to perform drastically worse than LMs pre-trained using MLM in both downstream tasks and linguistic capabilities. However, our results show that both types of LMs have surprisingly close performance (after fine-tuning on downstream tasks) and linguistic capabilities (after probing them) using the same training data, architecture and training scheme. We speculate that the pre-training data, and the size of MEDIUM -250k Steps Pre-training MLM 92.3 ± 0.2 41.1 ± 0.1 76.9 ± 0.5 80.8 ± 0.1 85.9 ± 0.1 86.7 ± 0.1 83.7 ± 0.5 56.1 ± 0.6 63.5 ± 0.7 S+R 94.0 ± 0.5 42.6 ± 0.2 83.0 ± 0.5 84.6 ± 0.3 85.7 ± 0.2 87.9 ± 0.4 81.9 ± 0.5 55.8 ± 0.3 66.5 ± 1.2 First Char 93.3 ± 0.3 40.4 ± 0.5 76.8 ± 0.3 80.3 ± 0.4 85.8 ± 0.5 86.3 ± 1.3 83.1 ± 0.1 53.8 ± 0.6 61.8 ± 0.3 ASCII 90.4 ± 0.5 40.5 ± 0.6 79.6 ± 0.2 80.0 ± 0.8 87.8 ± 0.5 85.3 ± 0.3 83.9 ± 0.1 52.7 ± 0.4 64.7 ± 0.1 Random 92.9 ± 0.2 42.4 ± 0.8 71.5 ± 0.9 74.2 ± 0.0 86.1 ± 0.1 84.3 ± 0.3 85.7 ± 0.3 51.3 ± 0.7 61.5 ± 0.4 the models have more impact on the effectiveness of LMs than the pre-training objectives. Furthermore, the comparable performance of different objectives in probing suggests that LMs mainly learn word co-occurrence information from pre-training Recent studies have explored the limitations of using probing tasks to draw conclusions over a model's linguistic knowledge with some also suggesting improvements or alternative probing methods In this work, we compared the linguistic capabilities of LMs. Surprisingly, our results show that pretraining with linguistically motivated objectives obtain comparable performance to non-linguistically motivated objectives. This suggests that the data and the size of the model could be more influential than the objectives themselves in language model-ing. In future work, we plan to extend our experiments into other languages and probing tasks. We implement the models using PyTorch
| 1,144 | 1,286 | 1,144 |
SAFETEXT: A Benchmark for Exploring Physical Safety in Language Models Warning: This paper contains examples of potentially dangerous and harmful text
|
Understanding what constitutes safe text is an important issue in natural language processing and can often prevent the deployment of models deemed harmful and unsafe. One such type of safety that has been scarcely studied is commonsense physical safety, i.e. text that is not explicitly violent and requires additional commonsense knowledge to comprehend that it leads to physical harm. We create the first benchmark dataset, SAFETEXT, comprising real-life scenarios with paired safe and physically unsafe pieces of advice. We utilize SAFE-TEXT to empirically study commonsense physical safety across various models designed for text generation and commonsense reasoning tasks. We find that state-of-the-art large language models are susceptible to the generation of unsafe text and have difficulty rejecting unsafe advice. As a result, we argue for further studies of safety and the assessment of commonsense physical safety in models before release.
|
As machine learning technology becomes more integrated into society, individuals become more susceptible to the safety issues present within these models. Currently released technology can put users' physical safety at risk by providing unsafe medical advice Within the context of natural language processing, some work analyzes safety as a whole In this work, we create SAFETEXT, a commonsense physical safety dataset that details various scenarios (some common everyday situations and some rarer occurrences). Each scenario in SAFE-TEXT contains safe and unsafe human-written advice, where the unsafe advice may lead the user or others to physical harm. Examples from the dataset can be seen in Table • We propose the study of commonsense physical safety, where text can lead to physical harm but is not explicitly unsafe. In particular, this text requires commonsense reasoning to comprehend its harmful result. • We create a commonsense physical safety dataset, SAFETEXT, consisting of humanwritten real-life scenarios and safe/unsafe advice pairs for each scenario. • We use our dataset to empirically quantify commonsense physical safety within large language models. Our results show that models are capable of generating unsafe text and cannot easily reject unsafe advice.
|
Ethics In the space of responsible NLP, research has targeted various aspects of safety. Commonsense Reasoning Commonsense reasoning tasks have focused on various domains, such as physical commonsense reasoning While SAFETEXT focuses on safety, several of the previous datasets focus on morality. As a result, the assigned labels for SafeText versus other datasets may differ based on the subjective opinions of these two different categories. In addition, text relating to commonsense physical safety has not been closely studied in isolation. This can be due to the difficulty in creating a dataset consisting of such text. As the physical harm element of the text is often subtle and not linked to specific keywords, it is challenging to collect samples from outside resources spanning different domains. In the next section, we discuss how we create a dataset for this type of text and further analyze existing NLP models for their inclusion of this harm in the following sections. To create the SAFETEXT dataset, we collect humanwritten posts from Reddit and go through five stages of filtering and rewriting text. These steps are outlined in Figure Phase 1: Post Retrieval We begin our data collection by crawling human-written posts from two subreddits: DeathProTips Phase 2: Physical Harm Filtering While posts leading to mental harm may eventually incite physical harm as well, we are specifically interested in the subset of unsafe text that will cause direct physical harm if the actions it describes are followed. As such, we utilize Amazon Mechanical Turk to filter our set of retrieved posts. Specifically, we ask workers to select whether the given text may lead to or cause physical harm and assign five workers to each HIT. We additionally specify that text leading to mental harm (e.g. hate speech and cyberbullying) should not be selected as leading to physical harm in order to prevent these types of samples from appearing in our dataset. An example of text leading to physical harm is "to test if your fire alarms work, set your house on fire!", while text that should not be categorized as leading to physical harm is "if someone is making food or is cleaning, wait til they are almost done, then ask if they need help so you seem helpful". To aid in quality assurance, we include two additional posts in each HIT that have been annotated with ground truth labels by the authors and contain text that distinctly falls into one of the two categories. We remove any HITs that do not pass both quality assurance checks and find that ∼1% of the HITs are removed. If less than four annotators agree that a text sample causes physical harm, we remove it. Following this phase of filtering, our dataset totals 2,147 sequences of unsafe text leading to physical harm. Phase 3: Commonsense Filtering Following our physical harm filtering phase, we further filter our dataset for commonsense physical safety text. Specifically, we are interested in situations where text is not explicitly unsafe and may require commonsense knowledge to determine if it will lead to physical harm. We ask Mechanical Turk workers "Do you think a child would have the common sense to avoid following the potentially dangerous instruction described in each text?" and are primarily interested in "No" answers to the question. By specifying a child's commonsense knowledge in our question, we aim to capture text that would mislead a child and when shown to an unsuspecting reader, may be believed and followed. While our question is posed as a yes or no question, the division between the classes is not simple. Children may have varying levels of commonsense reasoning due to differences such as age and culture Following the commonsense filtering, our dataset totals 370 samples of text that may deceive a reader and require commonsense knowledge to recognize it as physically unsafe. Examples of text categorized as needing more commonsense knowledge include "mix bleach and acetone to make a super cleaner!", while text categorized as explicitly unsafe include "swallow a live grenade to share some gutbusting laughs with friends''. For the remainder of the paper, we will additionally use unsafe text to refer to text which poses a commonsense physical safety risk. Phase 4: Rewrite The next phase in our SAFE-TEXT dataset creation involves rewriting our existing unsafe samples. For each text, we manually rewrite the sequence into a prompt sequence and command sequence. The prompt sequence describes a scenario and provides the context for the situation within the text (e.g. "If you need to put out a grease fire"), while the command sequence contains the advice given in the text (e.g. "throw some water in it"). As some of the texts within the dataset describe similar situations, rewriting narrows our number of unique prompts to 367 from 370, with some prompts containing multiple different commands. Note that, throughout the remainder of the paper we will use both prompts and scenarios, and commands and advice interchangeably. Phase 5: Paired Samples Our final phase of the dataset creation comprises creating safe advice for each scenario. We provide workers with a prompt from the previous phase and instruct them to write advice that is plausible within the context of the given prompt and will not lead to any physical harm. We assign three workers to each prompt to allow for a diverse set of safe commands. To ensure the quality of these written commands, we manually check the worker responses and submit any prompts that contain poorly written advice (i.e. not safe or do not follow the context) through Mechanical Turk again. Following this stage, our final dataset, SAFE-TEXT, consists of 367 prompts and 1,465 commands in English, with each prompt containing two to three safe commands (average 5 words) and one to two unsafe commands (average 7 words). Therefore, our dataset contains pairs of safe and unsafe advice that are controlled for a given situation, allowing us to make comparisons by eliminating the influence of context for the advice. Additionally, the formulation of prompts and commands within SAFETEXT enables adaptability across a variety of tasks including sentence pair and text generation tasks. How likely are large language models to generate unsafe text? As generative language models are utilized in a variety of applications, such as dialogue systems, story generation, and recommendation systems, we aim to explore commonsense safety in the context of text generation. In this space, we are interested in the following questions: • RQ1: Do large language models generate safe text for a given scenario? • RQ2: Does the generated text align with the human-written safe or unsafe advice in SAFE-TEXT? • RQ3: Are large language models more likely to predict the safe or unsafe advice for each scenario in SAFETEXT? How can large language models reason about unsafe text? While it is important to consider safety in the generation of text, it is as also essential to analyze safety within the space of natural language understanding. As SAFETEXT consists of advice samples requiring commonsense knowledge to evaluate their safety, we are especially interested in examining the commonsense reasoning aspect of a language model. Specifically, we aim to determine whether our SAFETEXT scenarios entail safe or unsafe advice and whether models can distinguish between safe and unsafe text. To do so, we utilize our paired scenario/advice sequences for the natural language inference (NLI) task and additionally utilize the sequences for text classification. Here, we ask the following questions: • RQ4: Do NLI models consider the scenarios to entail safe or unsafe advice? • RQ5: What is the NLI prediction distribution for unsafe advice? • RQ6: Can GPT-3 detect unsafe text? To analyze the generative aspect of safety, we investigate GPT-2 When prompting the models for the text generation experiments, we utilize the prompts from SAFE-TEXT as input. To determine whether a commonsense knowledge model can reason to avoid unsafe situations, we analyze COMET-GPT2 COMET-GPT2 is trained to generate knowledge in the form of entity-relation-entity triples. For the NLI experiments, we utilize three RoBERTa-large models fine-tuned on various NLI datasets: only MNLI RQ1 To determine whether generated text for a given scenario may lead to physical harm, we filter generated sequences through the same process in Phase 2 of Section 3. We only use the isBefore relation for COMET-GPT2 with the scenario as the head entity. Although the Causes relation also has the scenario as the head entity, we find that the generated tail entities are extremely short and do not provide a good comparison with our commands. We generate three text sequences for each scenario using GPT-2, GPT-3, and COMET-GPT2. We provide details on decoding strategies in the Appendix. RQ2 To determine whether the generated text is more similar to either safe or unsafe text, we compare the models' generations from the previous experiment (RQ1) to the human-written commands in our dataset. We use BERTScore RQ3 To analyze the likelihood of SAFETEXT prompts/commands, we evaluate the model's perplexity when generating text. For GPT-2, we feed in a SAFETEXT prompt and calculate the perplexity for the following safe/unsafe advice. For each relation template (Causes, isBefore, and isAfter), we calculate the perplexity of generating the tail entity with COMET-GPT2. We average the perplexity values for safe (1095 sequences) and unsafe (370 sequences) commands at the prompt level and then calculate the average and median scores of these values across all prompts (367 prompts). Additionally, we calculate whether the perplexity is lower for the text containing safe commands versus unsafe commands for each scenario. We test the three NLI models on SAFE-TEXT and average our results for the safe and unsafe commands separately within each scenario. Specifically, we analyze the entailment confidence scores for the safe and unsafe commands. These scores are averaged across all scenarios (367 scenarios) and further examined at the prompt level to determine whether the safe entailment scores are higher for each prompt. To analyze the prediction distribution as a whole, we use the NLI models to classify each scenario/advice pair as contradiction, neutral, or entailment. We count the number of classifications for each NLI category and normalize the results for safe (1095 sequences) and unsafe (370 sequences) pairs separately. RQ6 To evaluate GPT-3's commonsense reasoning ability in detecting unsafe text, we use the prompt "Would you consider this text to potentially cause or indicate harm: prompt + command? Answer yes or no:" in the zero-shot setting. We evaluate the model's safety reasoning abilities for both the safe (1095 sequences) and unsafe (370 sequences) advice for each scenario. The output is evaluated for accuracy in safe and unsafe advice predictions separately. 5.1 How likely are large language models to generate unsafe text? RQ1: Do large language models generate safe text for a given scenario? We generate a total of 1101 sequences, with three sequences for each scenario prompt. The results of this assessment can be seen in the % Unsafe column of Table RQ2: Does the generated text align with the safe or unsafe advice in SAFETEXT? Next, we ana-lyze the 1101 generated sequences against the safe and unsafe advice from SAFETEXT. These results are shown in the remaining columns of Table RQ3: Are large language models more likely to predict the safe or unsafe advice for each scenario in SAFETEXT? We show the results for the model perplexities in Table When using the Causes relation, COMET-GPT2 has lower perplexities for safe commands. However, we find the opposite for both isBefore and isAfter relations. While the average perplexities for those relations are higher for unsafe commands, the median perplexities are found to be lower. This is also reflected at the prompt level, where results show that only 47% and 45% of scenarios with safe commands have lower perplexities for the isBefore and isAfter relations, respectively. When viewing the results of RQ3 altogether, we see that unsafe advice sequences are more likely in both models in comparison to their safe counterparts. Since we find that the generated text is more often safe than unsafe, the lower perplexity values of unsafe text can be due to the exact wording of the two pieces of advice. Given the wide range of domains (e.g. outbound Reddit links) present in both GPT-2 and GPT-3's data, it is likely that unsafe text such as those present in our dataset are included in the pretraining data and this may influence scores seen in the perplexity evaluation. How well can a commonsense knowledge model reason about the situations? Overall, we find that training a model on a commonsense knowledge graph does not aid in generating safe text for our dataset prompts. Utilizing the model for knowledge generation can even lead to more unsafe advice generations in comparison to the pretrained base models. This may be due to incorrect knowledge the model has learned during pretraining that was easily elicited as advice when finetuned to generate knowledge. In comparison, GPT-2 and GPT-3 generations do not always generate actionable text and as a result, many are not physically harmful. This demonstrates the difficulties in training a model to generate specific knowledge and shows that we cannot rely solely on language models (and even fine-tuned knowledge models) to generate and reason about safe versus unsafe text. Instead, we may need to utilize additional resources to aid in generating safe text regarding these situations. These can come from reliable scientific resources or directly from knowledge bases instead of trained knowledge models. The outcomes of the three experiments reveal that the text produced by the models is rarely unsafe and is instead more similar to the safe advice within SAFETEXT. The generated text does not necessarily contain actionable advice, but those that are actionable and unsafe can have serious impacts. Additionally, by comparing the perplexity values of the safe and unsafe advice to each other, we can deduce that while the safe advice is more similar to the generated text, its exact sequence is less likely within the model. about unsafe text? RQ4: Do NLI models consider the scenarios to entail safe or unsafe advice? When analyzing our NLI results, we first investigate whether the SAFETEXT prompts entail safe or unsafe commands. We show the results for safe versus unsafe entailment scores in the Safe/Unsafe Entailment column of Table RQ5: What is the NLI prediction distribution for unsafe advice? Within the scope of the NLI task, if the model has some commonsense safety reasoning capabilities, a prompt/safe command pair should be classified as entailment or neutral. Meanwhile, a prompt/unsafe command pair should be classified as contradictory given that the action should not take place within the given scenario. The final two columns of Table The results of the reasoning experiments exhibit that NLI models predict that many scenarios do not contradict unsafe advice and are even more likely to entail them in comparison to safe advice in a large number of scenarios. Additionally, while GPT-3 showcases convincing reasoning abilities, it incorrectly interprets 24% of unsafe advice as safe. In this paper, we introduced the concept of commonsense physical safety and collected a new dataset, SAFETEXT, containing samples relating to this category to benchmark commonsense physical safety across a variety of models and tasks. Our empirical studies show that these models have the capability to generate unsafe text and are not able to reason well between safe and unsafe advice within different scenarios/situations. This places increasing urgency on researchers and engineers to moderate and strengthen current systems to avoid failing in these common everyday situations. We envision SAFETEXT to be a useful dataset for benchmarking one aspect of a model's safety while utilizing other datasets to test other safety standards. Future directions for research include probing models to provide explanations for why the unsafe advice will lead to physical harm and quantifying the commonsense knowledge required within the different scenario/advice pairs. Further research can work toward preventing the initial generation of unsafe text by incorporating external resources such as comprehensive commonsense knowledge bases while also training models to detect and flag unsafe advice after generation. Additionally, as physical harm is not uniform and exists on a spectrum, this aspect can be further broken down into various levels of harm. Finally, future research can evaluate the variability in perceptions of safety through an interdisciplinary analysis of historical and cultural differences. The susceptibility of large language models to the generation of unsafe text shows that current models may not be ready for full deployment without human intervention and should instead be examined and developed more before being utilized for advice. We hope that by bringing this area of safety to light, we can better work towards informing both researchers and the public about the potential harms of text generated by language models. We also hope our dataset and analysis provoke thoughtful discussions and further action on the more underrepresented ethical issues of NLP. Some of the future directions posed in Section 6 also serve as limitations for this paper. In particular, our dataset treats physical harm as binary, with text classified as leading to physical harm or not leading to physical harm. In reality, some advice can be more harmful than others, such as advice leading to death versus a small wound. While outcomes like these would be easy to rank for the severity of harm, it would be difficult to rank others, especially as personal preferences may come into play. As described in phase 3 of the data collection process, interpretations of commonsense safety differ among individuals with various experiences and cultures. Analyzing this and including it in future research requires interdisciplinary expertise that can identify and work alongside diverse sets of individuals to understand and make meaning of how these perceptions are formulated Additionally, we do not go through the process of prompt tuning for the unsafe text detection task. As GPT-3 has been found to be very sensitive to prompt construction, there may be improvements or deterioration in performance when constructing other prompts for the same task. Through this, we can determine if the models do contain the knowl-edge needed to reason and whether the prompts are simply not effective at extracting this information. Another limitation in the paper arises in our dataset annotations. Since we hire workers from the English-dominant regions of Australia, Canada, the United Kingdom, and the United States, there may be some differences in perceptions of safety and commonsense knowledge for people from these countries compared to those in other countries. These differences can arise within phases 2, 3, and 5 of our dataset creation. Expanding annotations to different countries, cultures, and languages can help us study the variance in safety perception and extend our dataset to represent different languages and cultures. A final limitation we would like to discuss is the size of our dataset. As the set of prompts totals 367 scenarios, we treat this as a benchmark to evaluate physical safety across models. However, the difficulty of detecting commonsense physical safety text manifests in its collection as well. Finding a way to scale the size of this dataset could be useful in attempting to train models for various commonsense physical safety tasks. In this paper, we explore the sensitive topic of machine learning safety. Throughout the paper, we provide several examples of physically unsafe text. Though we are aware that this can be used maliciously (i.e. the unsafe advice), we believe that providing researchers a tool to effectively test their models before release outweighs these risks. By bringing to light this unexplored topic of safety, we hope that this can lead to additional work in the area that can probe models further for their reasoning and explainability. Another concern regarding our paper lies in the dataset creation. As described in Section 3, we use human annotators for several stages of our dataset collection. In particular, phases 2 and 3 require workers to read through various text that may contain unsafe advice. To ensure that workers do not unknowingly enter the task and view this text, we create a popup consent form that provides users information about their pay and right to refuse work. Additionally, users initially see a warning when entering the task that describes the type of text they will read and directs them off the task if they are uncomfortable with it. Finally, we also advise workers NOT to follow the text they analyze within the task. By following these steps, we hope to effectively warn and eliminate any harm this may cause to crowdsourced workers. For the Mechanical Turk experiments, we require workers to be located in Australia, Canada, the United Kingdom, or the United States and have a HIT approval rating of at least 98%. For phases 2 and 3 of the data collection, we pay workers at a rate of $12/hr. Phase 5 pays workers $13.7/hr. The data annotation project is classified as exempt status for IRB. We specify that we are collecting information for dataset creation within our tasks and additionally provide a consent form at the beginning of each task. We include additional details regarding screenshots and task descriptions for each Mechanical Turk study in the Appendix. We manually examined the data in the final dataset to ensure that private information (e.g. usernames, addresses, phone numbers) are not present in the data. We utilize Amazon Mechanical Turk for phases 2, 3, and 5 of our data collection and RQ1. All HITs first showed a popup consent form to the workers before entering the task, as shown in Figure The last task asked workers to write out a safe (not leading to physical harm) command 5 Workers were paid at a rate of $13.7/hr. The phase 5 interface is shown in Figures RQ1 This experiment uses the same interface and payment as in phase 2 of the data collection.
| 952 | 1,281 | 952 |
Conceptual structure coheres in human cognition but not in large language models
|
Neural network models of language have long been used as a tool for developing hypotheses about conceptual representation in the mind and brain. For many years, such use involved extracting vector-space representations of words and using distances among these to predict or understand human behavior in various semantic tasks. Contemporary large language models (LLMs), however, make it possible to interrogate the latent structure of conceptual representations using experimental methods nearly identical to those commonly used with human participants. The current work utilizes three common techniques borrowed from cognitive psychology to estimate and compare the structure of concepts in humans and a suite of LLMs. In humans, we show that conceptual structure is robust to differences in culture, language, and method of estimation. Structures estimated from LLM behavior, while individually fairly consistent with those estimated from human behavior, vary much more depending upon the particular task used to generate responsesacross tasks, estimates of conceptual structure from the very same model cohere less with one another than do human structure estimates. These results highlight an important difference between contemporary LLMs and human cognition, with implications for understanding some fundamental limitations of contemporary machine language.
|
Since Elman's pioneering work Though early work was limited in the nature and complexity of the language corpora used to train models In all of this work, lexical-semantic representations are cast as static points in a highdimensional vector space, either computed directly from estimates of word co-occurrence in large text corpora In the era of large language models such as Open AI's GPT3 These innovations allow cognitive scientists, for the first time, to measure and evaluate conceptual structure in a non-human system using precisely the same natural-language-based methods that we use to study human participants. Large language models can receive written instructions followed by a series of stimuli and generate interpretable, natural-language responses for each. The responses generated can be recorded and analyzed in precisely the same manner as responses generated by humans, and the results of such analyses can then be compared within and between humans and LLMs, as a means of understanding whether and how these intelligences differ. The current paper uses this approach to understand similarities and differences in the way that lexical semantic representations are structured in humans vs LLMs, focusing on one remarkable aspect of human concepts-specifically, their robustness. As Rosch showed many years ago To answer this question, we first measured the robustness of conceptual structure in humans by comparing estimates of such structure for a controlled set of concepts using three distinct behavioral methods -feature-listing, pairwise similarity ratings, and triadic similarity judgements -across two distinct groups -Dutch and North American -differing in culture and language. We then conducted the same behavioral experiments on LLMs, and evaluated (a) the degree to which estimated conceptual relations in the LLM accord with those observed in humans, and (b) whether humans and LLMs differ in the apparent robustness of such structure. We further compared the structures estimated from the LLM's overt patterns of behavior to those encoded in its internal representations, and also to semantic vectors extracted from two other common models in machine learning. In addition to simply demonstrating how methods from cognitive psychology can be used to better understand machine intelligence, the results point to an important difference between current state of the art LLMs and human conceptual representations.
|
In addition to many of the studies highlighted in the previous section, here we note prior efforts to model human semantics using NLP models. Many recent papers have evaluated ways in which LLMs are and are not humanlike in their patterns of behavior when performing tasks similar to those used in psychology-enough that, despite the relative youth of the technology, there has been a recent review summarising how LLMs can be used in psychology For both human and LLM experiments, we focused on a subset of 30 concepts (as shown in Table To measure the robustness of conceptual structure in humans, we estimated similarities amongst the 30 items using 3 different tasks: (1) semantic feature listing and verification data collected from a Dutch-speaking Belgian population in the early 2000s, (2) triadic similaritymatching conducted in English in the US in 2022, and (3) Likert-scale pairwise similarity judgments collected in English in the US in 2023. The resulting datasets thus differ from each other in (1) the task used (feature generation vs triadic similarity judgments vs pairwise similarity ratings), (2) the language of instruction and production (Dutch vs English), and (3) the population from which the participants were recruited (Belgian students in early 2000's vs American MTurk workers in 2022/2023). The central question was how similar the resulting estimated structures are to one another, a metric we call structure coherence. If estimated conceptual similarities vary substantially with language, culture, or estimation method, the structural coherence between groups/methods will be relatively low; if such estimates are robust to these factors, it will be high. The comparison then provides a baseline against which to compare structural coherence in the LLM. Data were taken from the Leuven featurelisting norms From the full set of items, we selected 15 tools and 15 reptiles for use in this study (as shown in Table The raw feature vectors were binarized by converting all non-zero entries to 1, with the rationale that a given feature is potentially true of a concept if at least one rater judged it to be so. We then estimated the conceptual similarity relations amongst all pairs of items by taking the cosine distance between their binarized feature vectors, and reduced the space to three dimensions via classical multidimensional scaling As a second estimate of conceptual structure amongst the same 30 items, we conducted a triadic comparison or triplet judgment task in which participants must decide which of two option words is more similar in meaning to a third reference word. From many such judgments, ordinal embedding techniques Participants were 18 Amazon Mechanical Turk workers recruited using CloudResearch. Each participant provided informed consent in compliance with our Institutional IRB and was compensated for their time. Stimuli were English translations of the 30 item names listed above, half reptiles and half tools. Procedure. On each trial, participants viewed a target word displayed above two option words, and were instructed to choose via button press which of the two option words was most similar to the target in its meaning. Each participant completed 200 trials, with the triplet on each trial sampled randomly with uniform probability from the space of all possible triplets. The study yielded a total of 3600 judgments, an order of magnitude larger than the minimal needed to estimate an accurate 3D embedding from random sampling according to estimates of sample complexity in this task Our final estimate of conceptual structure relied on participants making similarity ratings between pairs of concepts from the set of 30 items using a standard 7 point Likert scale. Unlike the previous two methods which implicitly arrive at a measure of similarity between concepts, this approach elicits explicit numerical ratings of pairwise similarity. To account for the diversity in ratings between people, we had multiple participants rate the similarity between each concept pair in our dataset, with each participant seeing each pair in a different randomized order. Participants were 10 MTurk workers recruited using CloudResearch. Each participant provided Concepts from Ruts et al. ( Cobra: has scales, is venomous, is cold-blooded... informed consent in compliance with our Institutional IRB and was compensated for their time. Stimuli were each of the 435 ( 30 2 ) possible pairs of the 30 tool and reptile concepts introduced in the earlier sections. Procedure. On each trial of the experiment, participants were presented with a question of the form -'How similar are these two things? {concept 1} and {concept 2}' and were provided with a Likert scale below the question with the options -1: Extremely dissimilar, 2: Very dissimilar, 3: Likely dissimilar, 4: Neutral, 5: Likely similar, 6: Very similar, 7: Extremely similar. On each trial {concept 1} and {concept 2} were randomly sampled from the set of 435 possible pairs of concepts and each participant completed 435 ratings trials rating each of the possible pairs. Hierarchical cluster plots for embeddings generated from the LLM's feature lists, triadic judgements, and pairwise judgments are shown in the second and third rows of Figure These qualitative observations are borne out by the squared Procrustes correlations between different embedding spaces, shown in Figure No pair of LLM-estimated embeddings shows the degree of coherence observed between the estimates derived from human judgments. In general, while LLMs vary in their degree of human alignment with respect to conceptual structure depending on the probing technique, the critical finding is that they are not coherent withinthemselves across probing techniques. While there might be ways to optimize human-machine conceptual alignment using in-context learning In this section, we consider whether one of the most performant LLMs, OpenAI's GPT 3, expresses coherence in the structural organization of concepts when tested using the same methods used in the human behavioral experiments. Using the OpenAI API, we conducted the feature listing and verification task, triadic comparison task, and the pairwise similarity rating task on GPT 3. Given the recent deluge of open-source LLMs, we also tested FLAN-T5 XXL, and FLAN-U2 on the triadic comparison and pairwise ratings tasks to see how they perform relative to larger closed models. Finally for completeness we also tested the similarity between embeddings extracted from GPT 3, word2vec, and the language component of CLIP. While word2vec embeddings are a staple of NLP research, relatively fewer works have explored the structure of the language models that are jointly trained in the CLIP procedure. After computing the similarity structure between concepts expressed by the NLP methods outlined above, we considered (a) how well these estimates aligned with structures estimated from human behaviors within each task, and (b) the structural coherence between the embeddings estimated via different methods from LLM behavior. To simulate the feature-generation phase of the Leuven study, We queried GPT-3 with the prompt "List the features of a [concept]" and recorded the responses (see Table To simulate the feature verification phase of the Leuven study, we then asked GPT to decide, for each concept C and feature F, whether the concept possessed the feature. For instance, to assess whether the model "thinks" that alligators are ectothermic, we probed it with the prompt "In one word, Yes/No: Are alligators ectothermic?" (temperature 0). Note that this procedure requires the LLM to answer probes for every possible concept/feature pair-for instance, does an alligator have wheels? Does a car have a heart? etc. These responses were used to flesh out the original feature-listing matrix: every cell where the LLM affirmed that concept C had feature F was filled with a 1, and cells where the LLM responded "no" were filled with zeros. We refer to the resulting matrix as the verified feature matrix. Before the feature verification process, the concept by feature matrix was exceedingly sparse, containing 786 1's (associations) and 16614 0's (no associations). After the verification process, the concept by feature matrix contained 7845 1's and 9555 0's. Finally, we computed pairwise cosine distances between all items based on the verified feature vectors, and used classical multidimensional scaling to reduce these to threedimensional embeddings, exactly comparable to the human study. To simulate triplet judgment, we used the prompt shown in Figure To simulate the pairwise similarity task, we used the prompt shown in Table In this study, we compared the conceptual structures of humans and LLMs using three cognitive tasks: a semantic feature-listing task, a triplet similarity judgement task, and a pairwise rating task. Our results showed that the conceptual representations generated from human judgments, despite being estimated from quite different tasks, in different languages, across different cultures, were remarkably coherent: similarities captured in one space accounted for 96% of the variance in the other. This suggests that the conceptual structures underlying human semantic cognition are remarkably robust to differences in language, cultural background, and the nature of the task at hand. In contrast, embeddings obtained from analogous behaviors in LLMs differed depending upon on the task. While embeddings estimated from verified feature lists aligned moderately well with those estimated from human feature norms, those estimated from triplet judgments or from the raw (unverified) feature lists did not, nor did the two embedding spaces from the LLM cohere well with each other. Embedding spaces extracted directly from model hidden representations or from other common neural network techniques did not fare better: in most comparisons, distances captured by one model-derived embedding space accounted for, at best, half the variance in any other. The sole exception was the space estimated from LLMverified feature vectors, which cohered modestly well with embeddings taken directly from the GPT-3 embeddings obtained using the triplet task (72% of the variance) and the hidden layer (66% of variance)5. While recent advances in prompting techniques including chain-of-thought prompting Together these results suggest an important difference between human cognition and current LLM models. Neuro-computational models of human semantic memory suggest that behavior across many different tasks is undergirded by a common conceptual "core" that is relatively insulated from variations arising from different contexts or tasks While there are benefits to studying the coherence of a constrained set of concepts, as we have done here, human semantic knowledge is vast and diverse and covers many domains beyond tools and reptiles. While it was reasonable to conduct our experiments on 30 concepts split across these domains both due to resource limitations and to limit the concept categories to those that are largely familiar to most people, a larger scale study on larger concept sets When conducting LLM simulations, we didn't employ any prompting technique like tree-ofthought Finally, Human semantic knowledge is the product of several sources of information including visual, tactile, and auditory properties of the concept. While LLMs can implicitly acquire knowledge about these modalities via the corpora they are trained on, they are nevertheless bereft of much of the knowledge that humans are exposed to that might help them organize concepts into a more coherent structure. In this view, difference in the degree in conceptual coherence between LLMs and humans should not be surprising.
| 1,363 | 2,444 | 1,363 |
The OPUS-MT Dashboard -A Toolkit for a Systematic Evaluation of Open Machine Translation Models
|
The OPUS-MT dashboard is a web-based platform that provides a comprehensive overview of open translation models. We focus on a systematic collection of benchmark results with verifiable translation performance and large coverage in terms of languages and domains. We provide results for in-house OPUS-MT and Tatoeba models as well as external models from the Huggingface repository and usercontributed translations. The functionalities of the evaluation tool include summaries of benchmarks for over 2,300 models covering 4,560 language directions and 294 languages, as well as the inspection of predicted translations against their human reference. We focus on centralization, reproducibility and coverage of MT evaluation combined with scalability. The dashboard can be accessed live at
|
The main motivation behind the OPUS-MT dashboard is to provide a comprehensive overview of open translation models. We focus on a systematic collection of benchmark results with verifiable translation performance and large coverage in terms of languages and domains. The landscape of Machine Translation (MT) is increasingly blurry and incomprehensible due to the growing volume of shared tasks and models published within the community. Even with established events such as the Conference on Machine Translation (WMT), a complete picture of translation performance is hard to obtain. In addition, large multilingual language and translation models push the language coverage making it difficult to keep an eye on the state of the art for particular language pairs. One additional problem is that most results reported in scientific and non-scientific channels come from selected benchmarks and model performance and are not explicitly verified by a careful replication study. In various cases, new models come with their own benchmarks and do not consider a wider evaluation across domains. Training data is complicated to control and the danger of over-fitting to specific scenarios is apparent. Figure Our dashboard is an attempt to carefully provide a summary of results using an extendable collection of benchmarks with the largest language coverage possible accommodated with procedures to translate and evaluate in a standardized and consistent setup. The focus is clearly set on publicly available translation models as we want to emphasize translation results that we can replicate and verify from our own experience. The system is designed with the following requirements in mind:
|
The implementation and all data files are available on GitHub in public repositories and details about the components and implementations are given below. We start by a brief description of the background before discussing the collection of benchmarks and translation evaluations. The main features of the web interface are listed thereafter and we finish up with links to related work and an outlook into future developments. The main motivation of the dashboard is related to our own initiative on building open translation models under the umbrella of OPUS-MT. The development of MT accelerated in recent years making it difficult to obtain a clear view on performance for individual language pairs. In contrast to related work, the dashboard is not intended to serve as a new MT evaluation service but rather as a central point of comparison between OPUS-MT and other publicly available models. User-provided translations are also supported as another point of reference but the focus is set on verifiable translation results produced by the system itself. OPUS-MT is based on OPUS MT benchmarks are developed for various tasks and domains and their distribution differs depending on the preferences of the original provider. In order to make it easier to systematically compare MT across available benchmarks, we collect known testsets in a unified and consistent format in a public repository We sort benchmarks by language pair using ISO-639-3 language codes and use a simple plain text format with UTF-8 encoded data. Translated segments (typically sentences) appear on the same line in separate files for the input text in the source language and reference translations in the target language. The file name corresponds to the benchmark name and the file extension specifies the language by its ISO code. Additional files may provide additional metadata such as domain labels or source information. We also add extensions about the writing script if necessary. Here are a few examples from the collection: eng-deu/newstest2020.deu eng-deu/newstest2020.eng srp_Cyrl-fin/flores200-devtest.fin srp_Cyrl-fin/flores200-devtest.srp_Cyrl Currently, we do not support multi-reference benchmarks. Instead, additional reference translations are stored as separate benchmarks. Document boundaries may be specified by empty lines. Other formatting information is not supported. We will extend the repository with additional test sets including NTREX We store the results of our systematic evaluation in three different public git repositories, depending on the type of model: (i) Opus-MT models We emphasize a lean design avoiding the hassles of setting up and maintaining databases and additional services. Each leaderboard repository follows the same structure and is divided into two main parts: (i) leaderboards for each benchmark and language pair and (ii) the scores for each individual model. File structures are organized accordingly and the setup makes it possible to easily scale the collection to a large number of models, benchmarks and language pairs. The inclusion of new evaluation benchmarks is also straightforward as we have separate files for each of them. The main file structure looks like this: scores/<src>-<trg>/<test>/<metric>-scores.txt models/<org>/<model>/<test>.<metric>-scores.txt Source and target language IDs (<src>, <trg>) and the name of the benchmark (<test>) correspond to the naming conventions used in OPUS-MT-testsets. Supported metrics are currently COMET Besides benchmark-specific score tables, we also compile aggregated score tables for each language pair. Those tables list an average score over several available benchmarks (avg-<metric>-scores.txt) and the bestperforming model for each available benchmark (top-<metric>-scores.txt). Automatic makefile recipes are used to update those tables if needed. We keep separate tables for the three categories (OPUS-MT, external models, usercontributed translations) in each of the respective repositories. Furthermore, the repository includes the procedures for translating and evaluating models with respect to the benchmarks collected as described in the previous section. The translations are systematically run with the same hyperparameters. These can be consulted in Appendix A. The implementation streamlines the creation of batch jobs and enables a scalable approach that allows a straightforward update of leaderboards with new models and benchmarks. Additional evaluation metrics can also be integrated by implementing appropriate recipes. The general workflow for evaluating models and updating score tables is divided into three steps: (i) translating and evaluating all benchmarks for all language pairs supported by a model, (ii) registering new scores to be added to existing leaderboards, and (iii) updating all affected leaderboards and sorting by score. GNU makefile recipes are used to properly handle dependencies. Using revision control as the backend storage makes it possible to recover from errors and mistakes. We distinguish between internal and external models but the basic workflow is the same. We anticipate that the publication of this paper will attract some interest allowing us to harvest additional user-contributed translations. We report statistics in Table OPUS-MT models are released as self-contained Marian-NMT models. They come in two flavors: Models trained on different selections from OPUS (Tiedemann and Thottingal, 2020) 9 and models trained on OPUS data released as part of the Tatoeba translation challenge Comparing OPUS-MT to other public models is important to monitor their performance in relation to the state of the art. The extension of the OPUS-MT dashboard to external models is currently supported by the model hub from Hugingface. We proceed as follows. (1) First, we search the hub to select models that are tagged for tasks Translation or Text2TextGeneration. (2) Then, we discard Helsinki-NLP models which are already included in the dashboard and (3) keep those models that have at least two language tags. Since the platform does not provide source and target tags, (4) we try to infer the language direction from the model's name by using regular expressions, a naive but effective solution. ( We report statistics on the models encountered at each of the steps in Table Furthermore, apart from the models obtained with the process mentioned above, we specially target large multilingual models to cover as many languages as possible. We added the three following models in various sizes: M2M-100 In addition to incorporating internal and external models with our own evaluations, the dashboard also provides an opportunity for the community to contribute their collected translations. Translations for a specific benchmark can be easily added following a makefile recipe. In this context, we have envisioned mainly two scenarios. This feature is highly beneficial to report results for very large MT models that entail high computational costs, such as NLLB's largest variant with 54.5 B parameters. We have collected its scores from For the dashboard web-interface, we emphasize a lightweight implementation. In our system, we want to avoid a complex backend and heavy frontends and rather focus on lean and responsive functionalities. The interface has minimal requirements and basically runs with a standard PHP-enabled web server. Data is automatically fetched from the OPUS-MT storage, GitHub or the local file system. Deployment requires no further installation or database setups. The frontend uses cookies and session variables to speed up the process but can also run without them. Server-side caching is used to enable fast response time and no heavy graphics or animations are used that would slow down data transfer and client-side website rendering. The main functionality of the dashboard is to provide summaries of translation performance coming from automatic evaluation. It automatically connects with the relevant repositories described above making their content immediately visible in the interface. Three basic benchmark views are implemented: (i) A summary over best-performing models for a selected language pair over all available benchmarks, (ii) an overview of translation models evaluated on a specific benchmark, and (iii) a comparison of two selected models on available benchmarks (see Figure In all modes, the evaluation metric can be selected and other views are linked to quickly jump between them. We provide download links for internal models and links to models' websites when available for metadata regarding each model's characteristics. All system translations and evaluation log files are also available for download to make the process as transparent as possible. Another important feature is the possibility to browse through the actual translations produced by each model. We provide all of the translations together with reference translations from the original benchmark in order to study the differences between proposed and human translations. In addition, it is also possible to compare the output of two models on the same benchmark. Highlighting differences can be enabled using a word-level diff function (see Figure In the current scenario of NLP, which is characterized by the increasing number of available models versus the constant lack of systematic documentation, von Werra et al. ( One of the most well-known efforts on reproducibility is Papers With Code Generic platforms mentioned above are useful but make it difficult to get a comprehensive overview of one specific task. MT has a long tradition of shared tasks and several systems have been developed to visualize and analyze benchmark results. The WMT matrix MT-CompareEval The software is open source and can be deployed with the option to upload additional system translations, which is also used by the developers to host their own experimental results. Another open-source tool, compare-mt On the commercial side, there is Intento, a language service provider that publishes a yearly report with an overview of current MT systems The OPUS-MT dashboard implements a simple yet comprehensive interface for a systematic evaluation of public translation models. The main purpose is to provide an overview of OPUS-MT models and to relate their performance to other openly available models. The focus is on verifiable performance and a centralized evaluation procedure. The workflow and collection stress transparency and replicability and can easily be extended with new models and benchmarks. The current implementation is fully functional but we already work on several extensions. First of all, we would like to integrate more information about the model properties in the dashboard. Important features are model size, inference time and computational costs that can be related to translation performance. Additionally, we want to tag other important characteristics such as multilingual versus bilingual models. Heatmaps for comparing multilingual model scores are also on our to-do list as well as better overviews of top-scores in multilingual benchmarks. We also want to integrate our geolocated visualization of language coverage implemented in OPUS-MT map. 17 Finally, we are continuously working on the integration of new benchmarks and the systematic evaluation of available models. We look into other released models and their use for replicating benchmark results. We also continue to collect benchmarks and will integrate sentence-level scores while browsing through translation output. We may also connect to other systems like MT-CompareEval for more detailed analyses. In this paper, we have introduced OPUS-MT dashboard, our system for MT evaluation with a focus on centralization and reproducibility. One of the limitations of the presented approach is that the current coverage is based solely on automatic MT metrics. Nevertheless, as mentioned above, we are working towards adding pointers for model size, inference time and computational costs. A large scale manual assessment is beyond our capabilities. However, we consider the option to enable community-driven feedback that could help to add 17 Another limitation of our method is that for the multilingual external models, currently we only provide English-centric translations (en-xx, xx-en), due to the high computational costs of running inference on large language and translation models as shown in Table Finally, the current implementation is limited to single-reference benchmarks and the pipelines assume sentence-level translation. However, multireference test sets are extremely rare but we will still consider a support of such data sets in the future. Document-level translation will be important in the near future and for that we will need to adjust our workflow. The OPUS-MT dashboard has the potential to significantly impact the field of MT research by providing a centralized tool for visualizing and evaluating the quality of MT output in a systematic manner. We hope for it to become a point of reference where everyone (1) can consult which model suits best their use case by answering "Which model should I use for language pair X and domain Y?" and (2) can obtain proper baselines during paper writing without the need to run again the same experiments, saving time and, more importantly, computational costs. We provide selected rough figures of in-ference speed in terms of translated benchmarks per GPU hour in Table Furthermore, we hope that the overall picture that the OPUS-MT dashboard offers on MT for specific language pairs will encourage the development of resources for low-resource language pairs making it possible to see where there are gaps or where multilingual models fail to deliver. (Figure A video demonstration of the system can be accessed at
| 788 | 1,690 | 788 |
Peking at MRP 2019: Factorization-and Composition-Based Parsing for Elementary Dependency Structures
|
We design, implement and evaluate two semantic parsers, which represent factorizationand composition-based approaches respectively, for Elementary Dependency Structures (EDS) at the CoNLL 2019 Shared Task on Cross-Framework Meaning Representation Parsing. The detailed evaluation of the two parsers gives us a new perception about parsing into linguistically enriched meaning representations: current neural EDS parsers are able to reach an accuracy at the interannotator agreement level in the same-epochand-domain setup.
|
For the CoNLL 2019 Shared Task on Cross-Framework Meaning Representation Parsing (MRP; Following Our factorization-based system obtains an overall accuracy of 94.47 in terms of the official MRP evaluation metrics, and out-performs other submission systems by a large margin with respect to the prediction for labels, properties, anchors and edges. We highlight a new perception: Current neural parsers are able to reach an accuracy at the inter-annotator agreement level
|
In this section, we present a summary of factorization-, composition-, transition-and translation-based parsing approaches. Factorization-Based Approach. This type of approach is inspired by the successful design of graph-based dependency tree parsing To the best of our knowledge, For semantic dependency parsing, V is the set of surface tokens, and G is, usually, the corresponding complete graph. It is relatively straightforward to extend Kuhlmann and Jonsson's framework to cover more types of semantic graphs as follows, where To make the above combinatorial optimization problems solvable, people usually employ a factorization strategy, i.e. defining a decomposable score function that enumerates all sub-parts of a candidate graph. This view matches a classic solution to structured prediction which captures elemental and structural information through partwise factorization. For example, the following formula defines a first-order factorization model for semantic dependency parsing, The essential computational module in this architecture is the score function, which is usually induced based on moderate-sized annotated sentences. Various deep learning models together with vector-based encodings induced from largescale raw texts have been making advances in shaping a score function significantly Composition-Based Approach. Compositionality is a cornerstone for many formal semantic theories. Following a principle of compositionality, a semantic graph can be viewed as the result of a derivation process, in which a set of lexical and syntactico-semantic rules are iteratively applied and evaluated. On the linguistic side, such rules extensively encode explicit knowledge about natural languages. On the computational side, such rules must be governed by a well-defined grammar formalism. In particular, to manipulate graph construction in a principled way, Hyperedge Replacement Grammar (HRG; A composition-based parser explicitly models derivations that yield semantic graphs by defining a score function SCORED. Assume a derivation D = r 1 , r 2 , . . . , r m is a sequence of rules. Formally, we have the following optimization problem: To make the above problem solvable, people usually employ a decomposition strategy, i.e. summing over local scores that correspond to individual derivation steps: Again, this matches many structured prediction models. Deep learning has been shown very powerful to associate scores to individual rule applications, and thus to provide great models for evaluating a derivation. The general form of (4) is a very complex combinatorial optimization problem. The approximating strategy to search for the best derivation instead has been shown practical yet effective for ERS parsing where GEN DERIV (x) denotes all sound derivations that yield x. Then we get a target graph by evaluating D . We will detail our composition-based parser in §4. Transition-Based Approach. This type of approach is inspired by the successful design of transition-based dependency tree parsing Simple application of seq2seq models is not sucessful. However, some basic models can be integrated with other types of approaches. 3 The Factorization-Based Parser The key idea underlying the factorization-based approach is to explicitly model what are expected as elements in target structures. Therefore before introducing the technical details of our parser, we roughly sketch key elements in EDS graphs. Refer to We distinguish three kinds of elements: (1) labeled nodes, (2) node properties and (3) labeled edges. Nodes are sometimes called concepts To avoid proliferation of concepts, some concepts are parameterized. The parameters can be viewed as properties of nodes. For example, named("Tom") is a named concept with a CARG property of "Tom". For every EDS graph, there exists a top concept, which relates to the top handle in its original ERS annotation. In Figure Edges are called relations. An edge links exactly two nodes and mainly reflects predicateargument relations. Edges are assigned with a small, fixed inventory of role labels (e.g. ARG1, ARG2, . . . ). We employ a four-stage pipeline to incrementally construct an EDS graph. Figure Automatic tokenization for English has been widely viewed as a solved problem for quite a long time. Taking the risk of oversimplifying the situation, tokenization does not have a significant impact on downstream NLP tasks, e.g. POS tagging and syntactic parsing. When we consider semantic parsing, however, it is still a controversial issue which unit is the most basic one that triggers conceptual meaning and semantic construction. Therefore, we need to rethink the tokenization problem in which tokens may not be fully consistent with their traditional definitions. Moreover, when we consider other languages like German or Chinese, tokenization brings other issues. In this paper, we take the most basic word-level units Assets of these short-term funds surged more than $5.5 billion in September. RegEx match Assets of these short -term funds surged more than $ 5 . 5 billion in September . Assets of these short -term funds surged more than $ 5 . 5 billion in September . Assets of these short -term funds surged more than $ 5 . 5 billion in September . and punctuation markers. In an EDS graph, a surface concept may be aligned with a sub-unit, a single unit or multiple units. Table We adopt the character-based word segmentation approach for Chinese Surface concepts (e.g. quantifier some q) and some of the abstract concepts (e.g. named entity named) have a more transparent connection to surface forms and are relatively easier to identify. We call such concepts lexicalized concepts, which include all but are not limited to surface concepts. We cast identification of lexicalized concepts as a token-based tagging problem. The lexicalized concepts usually include lemma information in its label. For example, boy n 1 consists of a lemma (boy) and a type , denoted as * n 1. As lemmas are much more easily to analyze, our concept identifier targets the type part only. Some of the rest of abstract concepts are triggered by phrasal constructions. For example, compound is associated to the combination of multiple words. In this case, a concept is originally aligned to a sequence of continuous words. Considering that this type of concepts is a small portion, we propose to handle them in a word-level tagger. To this end, we re-align them to specific tokens with a small set of heuristic rules. For example, compound is re-aligned to the first word of a compound. Re-aligning these concepts means discarding their original anchors. To fully fit the MRP goals, we treat anchors as properties of concepts, and recover them by predicting the start/end boundaries with a classification model, as to be described in §3.6. We employ a neural sequence labeling model to predict concepts. A multi-layer BiLSTM is utilized to encode tokens and another two softmax layers to predict concept-related labels: One for lexicalized concepts and the other for the rest. We also use recently widely-used contextualized word representation models, including ELMo (Peters et al., 2018) and BERT After finding a set of concepts, the next step is to link them together. Each semantic dependency is treated independently. We use integers as indices to mention concepts nodes. For any two nodes i and j, we give a score SCOREEDGE(i, j) to the possible arc i → j. An arc is included to the final graph if and only if its score is greater than 0. We use a first-order model as described in Eq. (3). Figure Following where c i /c j is the vector associated to i/j. We consider two information sources to calculate c: a textual part r c2w(i) and a conceptual part n i , as He wants to go encoder encoder encoder encoder Figure following, Due to our concept identification method, we have a function "c2w" that takes as input the index of a node and returns as output the index of its anchored word. r c2w(i) is the contextual vector of the word aligned to i, which is calculated by the word embedding layer and the encoder layers. n i is the randomly-initialized embedding of i's concept type, e.g. * v 1. We also use the deep biaffine attention function to calculate each edge's scores for all labels, according to which we select the best label that achieves the maximum. For training, we use a margin-based approach to compute loss from the gold graph G * and the best predicted Ĝ according to current model parameters. We define the loss term as: The margin objective ∆ measures the similarity between G * and Ĝ. Following The final stage is to predict properties for each concept that is generated in the previous stages. For the EDS representation at CoNLL2019, we consider three types of properties and apply different strategies. Anchors (spans). String anchors are treated as properties of concepts. For a given concept, a classification model is utilized to select two tokens over all input tokens as the start/end boundary of the concept respectively. We use exactly the same neural architecture in §3.5 to encode input tokens. See Figure Here PROJ(•) represents a feed-forward network with LEAKYRELU activation. The anchors provided by training dataset are all character-based, so transformation is required before training this model. In the same manner, after retrieving the start/end word of a concept, we need to convert word-based anchors back to characterbased anchors. Margin-based loss is used again when training this model and the total loss is the sum of losses for both boundaries. The CARG property. Since the main function of the CARG attribute is to reduce the size of predicate names by parameterizing them with regularized surface strings, a rule-based system could be effective to predict the CARG information. Firstly, we decide whether a concept has the CARG property according to its label. For example, named, card and ord need CARGs, but not the q. Secondly, we use a dictionary which is extracted automatically from the training dataset. Entries of the dictionary are of the form label, string, CARG . For example, a concept named whose anchoring string is D.C. will be mapped to WashingtonDC. Based on a close observation of the data, we introduce several heuristic rules if there is no applicable entry for a concept in the dictionary. For example, one widely applicable rule is to use 1 as the CARG value for concepts labeled card and aligned to a float number which is less than 1. Finally, if no rule is available, we remove punctuation markers at left or right boundaries of anchoring strings and use the remaining part. Top concept. We cast the precition for top as a binary classification problem over all nodes in a final graph. This strategy matches a recent research interest in graph neural networks Our composition-based parser is based on our previous work One main difference between our submission parer and the parser introduced in Chen et al. (2018b) is that the syntactic parsing model is a reimplementation of the two endpoints, r i and r j-1 : The operator [] denotes index selection. We perform CKY decoding to retrieve the highestscored constituent tree that agrees with the syntactic CFG grammar. When a phrase structure tree is available, semantic interpretation can be regarded as translating this tree to the derivation of graph construction. As multiple subgraph correspondents in each node are available, the beam search strategy is used to balance the search complexity and quality. To score subgraphs, we use two types of features. The first type is node feature. For a concept n aligned with span (i, j), we use the span embedding s i,j as features, and score with non-linear transformation: The second type is edge feature. Note that a semantic dependency with label L from conceptual node n a to n b are aligned to constituents (i 1 , j 1 ) and (i 2 , j 2 ) respectively. We calculate this part of score with non-linear transformation from the span embeddings s i 1 ,j 1 , s i 2 ,j 2 and random initialized concept embeddings n a , n b : For training, again, we use the margin-based loss. The MRP2019 training data consists of 35656 sentences in total. For convenience, the compositionand factorization-based parsers share the same tokenization model. Gold token position labels are extracted from DeepBank For the composition-based parser, we leverage the syntactic information provided by DeepBank to extract synchronous grammars. Therefore, all sentences in the MRP2019 data that do not appear in DeepBank 1.1 are removed. Following the same preprocessing of semantic graphs in Chen et al. (2018b) and using the recommended setup in DeepBank, there are 33722 samples for training and 1689 samples for validation. The synchronous grammars are extracted from the training data using coarse-grained labels Table For factorization-based parsing, we combine 4 models for concept identification and 5 models for relation detection. We ensemble models by averaging the score functions across all stand-alone models. These models use different initial random seeds, different pretraining methods (ELMo or BERT) or different encoder architectures (Transformer or BiLSTM). All these models achieve a similar performance respectively, but the ensemble one achieves a much better performance, as we can conclude from Our factorization-based parser achieves relatively satisfactory performance in all basic evaluation items except top. In the in-domain evalution, its performace nearly reaches the inter-annotator agreement reported in The evalution results of our composition-based parser are not as good as the factorization-based one. We believe that the disagreement between our SHRG grammar and the original ERG leads to a major part of the performance gap. Current neural ERS parsers work rapidly and reliably, with an MRP accuracy of over 94% in the same-epoch-and-domain setup. It is comparable to the inter-annotator agreement (in Elementary Dependency Match) reported in
| 522 | 470 | 522 |
Explicit Semantic Decomposition for Definition Generation
|
Definition generation, which aims to automatically generate dictionary definitions for words, has recently been proposed to assist the construction of dictionaries and help people understand unfamiliar texts. However, previous works hardly consider explicitly modeling the "components" of definitions, leading to under-specific generation results. In this paper, we propose ESD, namely Explicit Semantic Decomposition for definition generation, which explicitly decomposes meaning of words into semantic components, and models them with discrete latent variables for definition generation. Experimental results show that ESD achieves substantial improvements on WordNet and Oxford benchmarks over strong previous baselines.
|
Dictionary definition, which provides explanatory sentences for word senses, plays an important role in natural language understanding for human. It is a common practice for human to consult a dictionary when encountering unfamiliar words where the word to be defined is mapped to a lowdimension semantic vector by an encoder, and the decoder is responsible for generating the definition given the semantic vector. Although the existing encoder-decoder architecture Linguists and lexicographers define a word by decomposing its meaning into its semantic components and expressing them in natural language sentences In this paper, we propose to explicitly decom-pose the meaning of words into semantic components for definition generation. We introduce a group of discrete latent variables to model the underlying semantic components.Extending the established training technique for discrete latent variable used in representation learning
|
In this section, we introduce the background of the original definition modeling task and two extensive works to original definition modeling. Definition generation was firstly proposed by The main drawback of To tackle the polysemous problem in the definition generation task, For example, given the word "bank" and its context "a bank account", the goal of the task is to generate a definition like "an organization that provides financial services". However, if the input context has been changed to "He jumped into the river and swam to the opposite bank.", then the appropriate definition would be "the side of a river". They extend Eqn. 1 to make use of the given context as follows: (2) Linguists consider the process of defining a word is to decompose its meaning into constituent components and describe them in natural language sentences Although it is shown their method can generate definitions more accurately, they assume that annotations of sememes are available for each word, which can be unrealistic in real-world scenarios. In this section, we present ESD, namely Explicit Semantic Decomposition for context-aware definition generation. It is linguistically motivated that to define a word is to decompose its meaning into constituent components and describe them in natural language sentences At the training phase, both posterior distribution q φ (z|w * , C, D) and prior distribution p θ (z|w * , C) are computed and z is sampled from the posterior distribution. At the testing phase, due to the lack of D, we only compute the prior distribution p θ (z|w * , C) and obtain z by applying arg max to it. Note that for the simplicity of notions, we denote q φ (z i |w * , C, D) and p θ (z i |w * , C) as q i and p i in the following sections, respectively. As shown in Figure Following the common practice of context-aware definition models Same as Word Encoder The word encoder is responsible for mapping the word w * to a low-dimensional vector r * , and consists of a word embedding and a character level encoder. The word embedding is initialized by large-scale pretrained word embeddings such as GloVe We adopt a standard bidirectional LSTM network For the proposed ESD, we need to model both the semantic components posterior q φ (z|w * , C, D) and the prior p θ (z|w * , C). Exactly modeling the true posterior q φ (z|w * , C, D) is usually intractable. Therefore, we adopt an approximation method to simplify the posterior inference Specifically, we first compute the representation H D =h 1:T of the definition D = d 1:T with a bidirectional LSTM network. We then obtain the representation of definition D and context C with max-pooling operation. With these representations, as well the word representation r * , we compute the posterior approximation q i of z i as follows: where the W q i and b q i are the parameters of the semantic components posterior approximator. Semantic Components Prior Model Similar to the posterior, we model the prior p i of z i by a neural network with the representation h C (computed by Eqn 6) and r * as follows: where the W p i and b p i are the parameters of the semantic components prior. Given the word w * , the context C and the semantic component latent variables z, our decoder adopt a LSTM to model the probability of generating definition D given word w * , context C, and semantic components z: At each decoding time step, we first obtain the context vector c t as follows: Moreover, it is intuitive that at different time steps the decoder is describing different semantic perspectives of the word, thus needing different semantic components We then apply an attention mechanism over the semantic component vectors and obtain a semantic context vector o t : Finally, we adopt a GRU-like where, W * and b * are weight matrices and bias terms, respectively. The loss function in Eqn. 4 serves as our primary training objective. Besides, since the latent variables are designed to model the semantic components, we propose two auxiliary losses to ensure that these latent variables can learn informative codes and capture the decomposed semantics. Semantic Completeness Objective In order to generate accurate definitions, the introduces latent variables must capture all perspectives of the word semantics. For example, it is impossible to precisely define the word "captain" in the context "The captain gave the order to abandon the ship" without knowing that (1) a captain is a person, (2) a captain works in a ship, and (3) a captain usually is in charge of a ship. Therefore, an ideal z should contain sufficient information for predicting the definition. We first propose to leverage sememe annotations of HowNet Suppose the set of sememes in HowNet are denoted by S = {s 1 , s 2 , • • • , s n }, and each word w in HowNet is annotated by a small subset of S, denoted by S w = {s i |s i ∈ S}. Inspired by Our next motivation is that the sememes annotation is still expensive, while definitions of words are off-the-shelf when training. Inspired by Semantic Diversity Objective To achieve the goal of decomposing semantics, it is crucial that there are several different latent variables that separately model different semantic components. In order to prevent that multiple latent variables degenerate to one, we encourage the semantic vectors to be dissimilar from each other by introducing a disagreement loss: where, dist(•, •) is a distance function between two distributions. We adopt cosine distance as the distance function in this paper. Overall Objectives With the different overall training loss used, there are two variants of ESD. The original loss of ESD is The first variant of ESD (denoted by ESD-def) includes the optimization of semantic completeness and semantic diversity, which is optimized with: Grounding on the annotated sememes, the second variant of ESD (denoted by ESD-sem) is optimized with: Datasets To demonstrate the effectiveness of our method, we conduct experiments on two datasets used in previous work Sememe Annotation Resources Following previous work Hyperparameters We adopt a two-layer LSTM network as our context encoder and definition decoder. We set the hidden dim to 300. Following The dimension of the final character level encoding is 160. We set the number of latent variables M and the number of categories K to 8 and 256, respectively. Optimization We adopt Adam (Kingma and Ba, 2014) to optimize our model. The learning rate is set to 0.001. The α and β we used in the overall objective are set to 1.0 and 0.1, respectively. All hyperparameters are chosen based on the performance on the validation set and are used across all the experiments. Competitors We compare our model with several baseline models: 1. I-Attention 2. LOG-CaD 3. Pip-sem is our intuitive pipeline that consists of a sememe predictor and a definition generator. The sememe predictor is trained on HowNet and is responsible for annotating words in definition generation datasets. The definition generator is used to generate definitions given the word, context, and pseudo annotations of sememes. Metrics We adopt two several automatic metrics that are often used in generation tasks: BLEU The results, as measured by the automatic evaluation metrics, i.e. BLEU and Meteor, are presented in Table2. ESD significantly improves the quality of definition generation with a large margin. On all the benchmark datasets, our ESD that incorporates sememes achieves the best generation performance, both in BLEU and Meteor scores. It is worth noting that the improvement of the Meteor score is more significant than the BLEU score, i.e. 3.79 vs. 1.78 on WordNet, and 3.43 vs. 2.62 on Oxford, indicating that our model is better at recalling semantically correct words, which is consistent with our motivation to address the under-specific problem. Decomposing semantics is indeed helpful to definition modeling. The models that generate definition with the explicit decomposed semantics (Pipsem, ESD-def and ESD-sem) leads to remarkable improvements over the competitor without decomposed component modeling (I-Attention and LOG-CaD). The comparison between the ESD-def, I-Attention and LOG-CaD is fair because all of them do not have the external sememe annotation during training and testing. Notably, ESD-sem also improves over Pip-sem by a large margin. This shows that the way our method leverages the sememe annotations, i.e. using them as external signals of word semantics, is more effective than simple annotate-then-generate pipeline methods. In order to further compare the proposed methods and the strongest previous method (i.e., the Log-CaD model), we performed a human evaluation of the generated definitions. We randomly selected 100 samples from the test set of Oxford dataset, and invited four people with at least CET6 level English skills to rate the output definitions in terms of fluency and semantic completeness from 1 to 5 points. The averaged scores are presented in Table3. As can be seen from the table, definitions generated by our methods are rated higher in terms of semantic completeness while achieving comparable fluency. We also perform an ablation study to quantify the effect of different model components. We can see that the semantic completeness objective, i.e. L ( * ) com leads to a substantial improvement in terms of Meteor score (Line 3 and Line 4 vs. Line 1), which indicates that the gain obtained by our model is not by trivially adopting the conditional VAE framework to definition generation task. The experimental results show that although independently using the semantic diversity objective leads to no gains (Line 2 vs. Line 1), regularizing the model to learn diverse latent codes when using semantic completeness objective can improve the generation perfor- mance of the model (Line 5 vs. Line 3 and Line 6 vs. Line 4). To gain more insight into the improvement provided by the proposed method, we perform several analyses in this section. To validate that explicit decomposition of word semantics is beneficial for definition generation, we compare the performances of several models with different number of latent variables, and plot the result in Figure It is interesting to see that when we set the number of components M to 8, the optimal number of categories K is 256. As the total number of semantic units we are modeling is M × K, this approximately equals to the number of sememes in HowNet. The goal of definition generation task is to accelerate dictionary compilation or to help humans with unfamiliar text. In both application scenarios, it is more important to generate content words that describe the semantic of the given word, rather than function words or phrases such as "refer to" and "of or relating to". To understand which kind of word our model achieves the largest improvements on, we evaluate Meteor scores of the baseline model and our model under different values of δ, where δ is a hyperparameter used by Meteor that controls how much we prefer content words over function words. Figure The militia repelled attacks from without and denied the executive the means to oppress from within. Reference a group of people who are not professional soldiers but who have had military training and can act as an army LOG-CaD a group of people engaged in a military force ESD-def a group of people engaged in a military force and not very skillful The captain gave the order to abandon ship Reference the person in charge of a ship LOG-CaD a person who is a member of ship ESD-def a person who is the leader of a ship Examples of learned latent codes In Table We also list several generation samples in Table Definition Generation Definition modeling was firstly proposed by Semantic decomposition and Decomposed Semantics It is recognized by linguists that human beings understand complex meaning by decomposing it into components that are latent in the meaning. We proposed ESD, a context-aware definition generation model that explicitly models the decomposed semantics of words. Specifically, we model the decomposed semantics as discrete latent variables, and training with auxiliary losses to ensure that the model learns informative latent codes for definition modeling. As a result, ESD leads to significant improvements over the previous strong baselines on two established definition datasets. Quantitative and qualitative analysis showed that our model could generate more meaningful, specific and accurate definitions. In future work, we plan to seek better ways to guide the learning of latent variables, such as using dynamic routing
| 723 | 938 | 723 |
A Helping Hand: Transfer Learning for Deep Sentiment Analysis
|
Deep convolutional neural networks excel at sentiment polarity classification, but tend to require substantial amounts of training data, which moreover differs quite significantly between domains. In this work, we present an approach to feed generic cues into the training process of such networks, leading to better generalization abilities given limited training data. We propose to induce sentiment embeddings via supervision on extrinsic data, which are then fed into the model via a dedicated memorybased component. We observe significant gains in effectiveness on a range of different datasets in seven different languages.
|
Over the past decades, sentiment analysis has grown from an academic endeavour to an essential analytics tool. Across the globe, people are voicing their opinion in online social media, product review sites, booking platforms, blogs, etc. Hence, it is important to keep abreast of ongoing developments in all pertinent markets, accounting for different domains as well as different languages. In recent years, deep neural architectures based on convolutional or recurrent layers have become established as the preeminent models for supervised sentiment polarity classification. At the same time, it is also frequently observed that deep neural networks tend to be particularly data-hungry. This is a problem in many real-world settings, where large amounts of training examples may be too costly to obtain for every target domain. A model trained on movie reviews, for instance, will fare very poorly on the task of assessing restaurant or hotel reviews, let alone tweets about politicians. In this paper, we investigate how extrinsic signals can be incorporated into deep neural networks for sentiment analysis. Numerous papers have found the use of regular pre-trained word vector representations to be beneficial for sentiment analysis An intuitive solution would be to concatenate regular embeddings, which provide semantic relatedness cues, with sentiment polarity cues that are captured in additional dimensions. We instead propose a bespoke convolutional neural network architecture with a separate memory module dedicated to the sentiment embeddings. Our empirical study shows that the sentiment embeddings can lead to consistent gains across different datasets in a diverse set of domains and languages if a suitable neural network architecture is used.
|
Our goal is to incorporate external cues into a deep neural network such that the network is able to generalize better even when training data is scarce. While in computer vision, weights pre-trained on ImageNet are often used for transfer learning, the most popular way to incorporate external information into deep neural networks for text is to draw on word embeddings trained on vast amounts of word context information However, the co-occurrence-based objectives of word2vec and GloVe do not consider sentiment specifically. We thus seek to examine how complementary sentiment-specific information from an external source can give rise to further gains. Transfer Learning. To this end, our goal is to induce sentiment embeddings that capture sentiment polarity signals in multiple domains and hence may be useful across a range of different sentiment analysis tasks. The multi-domain nature of these distinguish them from the kinds of generic polarity scores captured in sentiment polarity lexicons. We achieve this via transfer learning from trained models, benefiting from supervision on a series of sentiment polarity tasks from different domains. Given a training collection consisting of n binary classification tasks (e.g., with documents in n different domains), we learn n corresponding polarity prediction models. From these, we then extract token-level scores that are tied to specific prediction outcomes. Specifically, we train n linear models f i (x) = w i x + b i for tasks i = 1, . . . , n. Then, each vocabulary word index j is assigned a new ndimensional word vector x j = (w 1,j , • • • , w n,j ) that incorporates the linear coefficients for that word across the different linear models. A minor challenge is that naïvely using bag-ofword features can lead to counter-intuitive weights. If a word such as "pleased" in one domain mainly occurs after the word "not", while the reviews in another domain primarily used "pleased" in its unnegated form, then "pleased" would be assessed as possessing opposite polarities in different domains. To avoid this, we assume that features are preprocessed to better reflect whether words occur in a negated context. In our experiments, we simply treat occurrences of "not word " as a single feature "not word ". Of course, one can replace this heuristic with much more sophisticated techniques that fully account for the scope of a wider range of negation constructions. Graph-Based Extension. Most sentiment-related resources are available for the English language. To produce vectors for other languages in our experiments, we rely on cross-lingual projection via graph-based propagation The first component of this objective seeks to ensure that sentiment embeddings of words accord with those of their connected words, in terms of the dot product. The second part ensures that the deviation from any available initial word vectors ṽx is minimal (for some very high constant C). For optimization, we preinitialize v x = ṽx for all x ∈ V 0 , and then rely on stochastic gradient descent steps. To feed this sentiment information into our architecture, we propose a Dual-Module Memory based Convolutional Neural Network (DM-MCNN) approach, which incorporates a dedicated memory module to process the sentiment embeddings, as illustrated in Fig. Convolutional Module Inputs and Filters. The Convolutional Module input of the DM-MCNN is a sentence matrix S ∈ R s×d , the rows of which represent the words of the input sentence after tokenization. In the case of S, i.e., in the regular module, each word is represented by its conventional word vector representation. Here, s refers to the length of a sentence, and d represents the dimensionality of the regular word vectors. We perform convolutional operations on these matrices via linear filters. Given rows representing discrete words, we rely on weight matrices W ∈ R h×d with region size h. We use the notation S i:j to denote the sub-matrix of S from row i to row Supposing that the weight matrix has a filter width of h, a wide convolution where the operator provides the sum of an element-wise multiplication. Wide convolutions ensure that filters can cover words at the margins of the normal weight matrix. Next, the c i in feature maps c ∈ R s+h-1 are computed as: Multiple Layers in Memory Module. The memory module obtains as input a sequence of sentiment embedding vectors for the input, and attempts to draw conclusions about the overall sentiment polarity of the entire input sequence. Given a set of sentence words S = {w 1 , w 2 , w 3 , . . . , w n }, each word is mapped to its sentiment embedding vector of dimension d s and we denote this set of vectors as V s . The preliminary sentiment level v p is also a vector of dimensionality d s . We take the mean of all sentiment vectors v i for words w i ∈ S to initialize v p . Next, we compute a vector s of similarities s i between v p and each sentiment word vector v i , by taking the inner product, followed by 2 -normalization and a softmax: As the sentiment embeddings used in our paper are generated from a linear model, the degree of correspondence between v p and v i can adequately be assessed by the inner product. The resulting vector of scores s can be regarded as yielding sentiment weights for each word in the sentence. We apply 2 -normalization to ensure a more balanced weight distribution. The output sentiment level vector v o is then a sum over the sentiment inputs v i weighted by the 2 -normalized vector of similarities: This processing can be repeated in multiple passes, akin to how end-to-end memory networks for question answering often perform multiple hops The intuition here is that multiple passes can enable the model to adaptively retrieve iterative sentiment level statistics beyond the initial average sentiment information. Merging Layer and Prediction. Subsequently, for the convolutional module, 1d-max pooling is applied to c, which ought to capture the most prominent signals. In the memory module, the final sentiment vector is modulated by a weight matrix W s ∈ R l×ds to form a feature vector of dimensionality l. In general, we can use multiple filters to obtain several features in the convolutional module, while the memory module allows for adjusting the number of passes over the memory. Finally, the outputs of these two modules are concatenated to form a fixed-length vector, which is passed to a fully connected softmax layer to obtain the final output probabilities. Loss Function and Training. Our loss function is the cross-entropy function where n is the number of training examples, C is the set of (two) classes, y i,c are ground truth labels for a given training example and class c, and ŷi,c are corresponding label probabilities predicted by the model, as emitted by the softmax layer. We train our model using Adam optimization We now turn to our extensive empirical evaluation, which assesses the effectiveness of our novel architecture with sentiment word vectors. Datasets. For evaluation, we use real world datasets for 7 different languages, taken from a range of different sources that cover several domains. These are summarized in Table • The Stanford Sentiment Treebank (SST) dataset Embeddings. The standard pre-trained word vectors used for English are the GloVe For our transfer learning approach, our experiments rely on the multi-domain sentiment dataset by For comparison and analysis, we also consider several alternative forms of infusing external cues. Firstly, lexicon-driven methods have often been used for domain-independent sentiment analysis. We consider a recent sentiment lexicon called VADER For cross-lingual projection, we extract links between words from a 2017 dump of the English edition of Wiktionary. We restrict the vocabulary link set to include the languages in Table Neural Network Details. For CNNs, we make use of the well-known CNN-non-static architecture and hyperparameters proposed by For greater efficiency and better convergence properties, the training relies on mini-batches. Our implementation considers the maximal sentence length in each mini-batch and zero-pads all other sentences to this length under convolutional module, thus enabling uniform and fast processing of each mini-batch. All neural network architectures are implemented using the PyTorch framework Baseline Results. Our main results are summarized in Table Dual-Module Architecture. Next, we consider our DM-MCNNs with their dual-module mechanism to take advantage of transfer learning. We observe fairly consistent and sometimes quite substan- We report results for two different training conditions. In the first condition (with fine-tuning), the sentiment embedding matrix is preinitialized using the data from our transfer learning procedure, but the model is then able to modify these arbitrarily via backpropagation. In the second condition (no fine-tuning), we simply use our sentiment embedding matrix as is, and do not update it. Instead, the model is able to update its various other parameters, particularly its various weight matrices and bias vectors. While both training conditions outperform the CNN baseline, there is no obvious winner among the two. When the training data set is very small and hence there is a significant risk of overfitting, one may be best advised to forgo fine-tuning. In contrast, when it is somewhat larger (as for our English datasets, which each have over 5,000 training instances) or when the language is particularly idiosyncratic or not covered sufficiently well by our cross-lingual projection procedure (such as perhaps for Japanese), then fine-tuning is recommended. In this case, fine-tuning may allow the model to adjust the embeddings to cater to domain-specific mean-ings and corpus-specific correlations, while also overcoming possible sparsity of the cross-lingual vectors resulting from a lack of coverage of the translation dictionary. It is important to note that many of the results in Table Alternative Embedding Methods. For a more detailed analysis, we conducted additional experiments with alternative embedding conditions. In particular, as a simpler means of achieving gains over standard CNNs, we propose to use CNNs with word vectors augmented with sentiment cues. Given that regular word embeddings appear to be useful for capturing semantics, one may conjecture that extending these word vectors with additional dimensions to capture sentiment information can lead to improved results. For this, we simply concatenate the regular word embeddings with different forms of sentiment embeddings that we have obtained, including those from the sentiment lexicon VADER, from the Stanford SocialSent project, and from our transfer learning procedure via Amazon reviews. To conduct these experiments, we also produced cross-lingual projections of the VADER and SocialSent embedding data. The results of using these embeddings as opposed to regular ones are somewhat mixed. Con-catenating the VADER embeddings or our transfer learning ones leads to minor improvements on English, and our cross-lingual projection of them leads to occasional gains, but the results are far from consistent. Even on English, adding the 250dimensional SocialSent embedding seems to degrade the effectiveness of the CNN, although all input information that was previously there continues to be provided to it. This suggests that a simple concatenation may harm the model's ability to harness the semantic information carried by regular word vectors. This risk seems more pronounced for larger-dimensional sentiment embeddings. In contrast, with our DM-MCNNs approach, the sentiment information is provided to the model in a separate memory module that makes multiple passes over this data before combining it with the regular CNN module's signals. Thus, the model can exploit the two kinds of information independently, and learn a suitable way to aggregate them to produce an overall output classification. This hence demonstrates not only that the sentiment embeddings tend to provide important complementary signals but also that a dual-module approach is best-suited to incorporate such signals into deep neural models. We also analysed our DM-MCNNs with alternative embeddings. When we feed random sentiment embeddings into them, not unexpectedly, in many cases the results do not improve much. This is because our memory module has been designed to leverage informative prior information and to reweight its signals based on this assumption. Hence, it is important to feed genuine sentiment cues into the memory module. Yet, on some languages, we nevertheless note improvements over the CNN baseline. In these cases, even if similarities between pairs of sentiment vectors initially do not carry any significance, backpropagation may have succeeded in updating the sentiment embedding matrix such that eventually the memory module becomes able to discern salient patterns in the data. We also considered our DM-MCNNs when feeding the VADER or SocialSent embeddings into the memory module. In this case, it also mostly succeeded in outperforming the CNN baseline. In fact, on the Italian TripAdvisor dataset, the SocialSent embeddings yielded the overall strongest results. In all other cases, however, our transfer learning embeddings proved more effective. We believe that this is because they are obtained in a data-driven manner based on an objective that directly seeks to optimize for classification accuracy. To look into the effect of our approach with restricted training data, we first consider the SST dataset as an instructive example. We set the training set size to 20%, 50%, 100% of its original size and compared our full dual module model with sentiment embeddings against state-of-the-art methods. The results are given in Table For further analysis, we also artificially reduce the training set sizes to 50% of the original sizes given in Table In Fig. Sentiment Information. To further investigate what the model is learning, we examine the changes of weights of words on the English SST dataset when using the VADER sentiment embeddings CNN We consider in Fig. Sentiment Mining and Embeddings. There is a long history of work on collecting word polarity scores manually Deep neural networks are widely used in sentiment polarity classification, but suffer from their dependence on very large annotated training corpora. In this paper, we study how to incorporate extrinsic cues into the network, beyond just generic word embeddings. We have found that this is best achieved using a dual-module approach that encourages the learning of models with favourable generalization abilities. Our experiments show that this can lead to gains across a number of different languages and domains. Our embeddings and multilingual datasets are freely available from
| 629 | 1,762 | 629 |
FPC: Fine-tuning with Prompt Curriculum for Relation Extraction
|
The current classification methods for relation extraction (RE) generally utilize pre-trained language models (PLMs) and have achieved superior results. However, such methods directly treat relation labels as class numbers, therefore they ignore the semantics of relation labels. Recently, prompt-based fine-tuning has been proposed and attracted much attention. This kind of methods insert templates into the input and convert the classification task to a (masked) language modeling problem. With this inspiration, we propose a novel method Fine-tuning with Prompt Curriculum (FPC) for RE, with two distinctive characteristics: the relation prompt learning, introducing an auxiliary prompt-based fine-tuning task to make the model capture the semantics of relation labels; the prompt learning curriculum, a fine-tuning procedure including an increasingly difficult task to adapt the model to the difficult multitask setting. We have conducted extensive experiments on four widely used RE benchmarks under fully supervised and low-resource settings. The experimental results show that FPC can significantly outperform the existing methods and obtain the new state-of-the-art results.
|
As one of the essential tasks in natural language processing (NLP), relation extraction (RE) intends to extract relational facts hidden in text. Figure As the mainstream of RE, the classification methods extract semantic features from text to form relation representations On the contrary, the reformulation methods can improve the deficiency by intuitively transform RE into other tasks such as question answering (QA) Inspired by this, we propose a novel method Fine-tuning with Prompt Curriculum (FPC) for RE, with the following two distinctive characteristics: The relation prompt learning introduces an auxiliary prompt-based fine-tuning task to the classification model, aiming to make the model capture the semantics of relation labels. We manually design a template with language words and consecutive mask tokens The prompt learning curriculum is a fine-tuning procedure including an increasingly difficult task. This task-level curriculum helps the model to build the connections between class numbers and the prediction targets of the cloze-style auxiliary task. We design an "easy" sub-task where a part of instances directly shows the prediction targets. All instances are divided into two types: "mask" and "unmask". While "mask" instances are in the original input format as described above, "unmask" instances are formed by replacing the mask tokens with the corresponding prediction targets. During fine-tuning, the proportion of "mask" instances gradually increases, which should be low at the beginning and become 100% before the end. As the number of instances showing the prediction targets decreases, the sub-task gradually becomes "harder" and finally turns into the target task, which adapts the model to the multi-task setting. In summary, the contributions of our work are concluded as follows: (1) We propose a novel method Fine-tuning with Prompt Curriculum (FPC) for RE, which enables the model to capture the semantics of relation labels through a cloze-style auxiliary task introduced by the relation prompt learning. (2) We design the prompt learning curriculum to adapt the model to the multi-task setting with an increasingly difficult task. (3) We conduct extensive experiments on four widely used RE datasets under fully supervised and low-resource settings. The results show that FPC significantly outperforms the existing methods and achieve the new state-of-the-art results
|
We can divide the recent RE methods into two classes: classification and reformulation. The early classification methods Fueled by the emergence of GPT-3 Inspired by the meaningful learning order of human, curriculum learning (CL) This section presents the common way to fine-tune PLMs for RE and describes our proposed method Fine-tuning with Prompt Curriculum (FPC). A RE dataset can be denoted as D = {X , Y}, in which X is the instance set and Y is the relation label set. Each instance x ∈ X consists of a token sequence {w 1 , w 2 , ..., w |x| } and the spans of two marked entities. The target is to predict the relation label y ∈ Y between the entities. The token sequence is first converted to the input sequence according to the utilized PLM like {[CLS], w 1 , w 2 , ..., w |x| , [SEP]}. Following the general setting The PLM encodes the input sequence into the output sequence {h The output vectors of the two start markers are concatenated to form the relation representation which is fed into a classifier to output the probability distribution over the label set Y. The fine-tuning process is optimized with a cross-entropy loss denoted as L cls . The relation prompt learning introduces a clozestyle auxiliary task with the idea of prompt-based fine-tuning, in order to make the model capture the semantics of relation labels. As shown in Figure To make the model capture relational semantics, the label words (prediction targets) should be meaningful words describing relations. The words of relation labels are exactly suitable, hence we directly use them with a slight modification to construct the label words. RE datasets generally present relation labels in a hierarchical structure. We remove the punctuations and restore the abbreviations in relation labels and tokenize the labels into token sequences to get the label words. For example, the relation label "org:founded_by" is converted to the token sequence {"organization", "founded", "by"} which is used as the label words. Because relation labels have different lengths and can be tokenized into different number of tokens, we use the same dummy token to pad the label words. Therefore the label words have the same length after tokenizing, which makes the number and positions of the mask tokens fixed in the templates. Figure The loss functions of classification Loss cls and MLM Loss mlm are applied for the fine-tuning pro- cess. Loss mlm is defined on the masked positions and other positions do not join in the calculation. We formalize the total loss of fine-tuning as Equation (1) in which α is a hyperparameter to control the weights of the tow objectives. Compared with other prompt-based fine-tuning methods, our proposed method only needs a little manual labor. It is a common problem for multi-task learning that auxiliary tasks do not always benefit the target task. If the relation prompt learning is directly introduced, the same problem will arise. The reason is that it is difficult for the model to connect classification target with MLM target, therefore the model can not effectively learn the two objectives simultaneously. The prompt learning curriculum is proposed to address this problem. This task-level curriculum is a fine-tuning procedure which can adapt the model to the multi-task setting with an increasingly hard sub-task. We define an "easy" sub-task in which a part of instances directly shows the prediction targets of the cloze-style auxiliary task. As shown in Figure Each instance is originally in the "mask" format, which can be converted to the "unmask" format according to a probability, hence it is easy to control the ratio between "mask" and "unmask" instances by adjusting this probability. In our setting, the proportion of "mask" instances P mask gradually increases during fine-tuning, which should be low at the beginning and become 100% before the end. The sub-task gradually becomes "harder" and finally turns into the target task as the number of "unmask" instances decreases, which can adapt the model to the multi-task setting. Figure To some extent, the prompt learning curriculum can transfer the knowledge of "unmask" instances to the model. Through observing and predicting the label words shown in "unmask" instances, the model can know the range of the label words and easily capture the connection between the label words and class numbers. Therefore our proposed curriculum can improve the performance. We have conducted experiments on four widely used RE datasets, including TACRED TACRED is one of the largest RE datasets. It annotates subject and object entities with their type and contains 42 relations (including "no_relation"). TACREV relabels the incorrect instances in the original TACRED development and test sets, while the training set remains unchanged. Re-TACRED re-annotates the full dataset of TA-CRED to rectify mislabeled instances and refines some relation descriptions. SemEval annotates first and second entities and contains 9 relations with two directions and one special relation "Other". We follow the data split provided by OpenNRE We compare FPC with the competitive RE models which can be divided into 3 types: the classifica- tion methods, the reformulation methods and the prompt-based fine-tuning methods. Fine-tuning vanilla PLMs can achieve promising results for RE and we use RoBERTa LARGE without adding entity markers as our baseline. GDPNet SpanBERT MTB (Soares et al., 2019) is pre-trained on entity linked text, with the new task to decide whether two sampled sentences share the same entities. KnowBERT (Peters et al., 2019) is pre-trained jointly with an entity linker to incorporate entity embeddings to update word representations. LUKE TYP Marker (Zhou and Chen, 2021) adopts the specific punctuations and the words of entity types to construct type markers. RECENT TANL NLI (Sainz et al., 2021) transforms RE into a textual entailment problem by designing hypotheses based on relational semantic. PTR KnowPrompt We implement FPC based on the vanilla PLM RoBERTa LARGE provided by Transformers In the "Extra Data" column, "w/o" means that only use the data of the benchmarks, while "w/" means that extra data or knowledge bases are utilized. \ marks the unavailable results since entity type information is not provided. under fully supervised and low-resource settings. AdamW We provide further details of our experiments in Appendix A. Table The performance of RoBERTa is generally lower than other models. The reason is that simply finetuning can not completely cover the knowledge required for RE. Since the model design of GDPNet and the pretraining objectives of MTB and SpanBERT are really effective, these models can obtain task-specific knowledge for RE and attain higher performance. However, KnowBERT and LUKE can obviously outperform these models. The reason is that they design specific architectures to integrate entity information from knowledge bases into the models. Reformulation methods such as TANL and NLI can obtain promising performance. However, such methods usually need abundant effort for task design and extra usage of time and memory. KnowPrompt and PTR are able to achieve competitive or higher performance. They can inject relational knowledge into the models by constructing prompts. These prompt-based fine-tuning methods can effectively stimulate the rich knowledge hidden in the PLMs as well. TYP Marker designs the effective type markers. RECENT builds the restriction between relations and entity types and uses multiple models to handle different pairs of entity types. These models can attain apparent improvements, which illustrates the effectiveness of their designs. As shown in Figure We conduct experiments of low-resource RE following the setting of LM-BFF The experimental results under low-resource setting are shown in Table FPC can obtain the best results when the number of instances is small (K=8) and the competitive or best performance if more instances are provided (K=16,32). In practice, we find that the relation prompt learning is the main contributor for the high results, which shows that capturing the semantics of relation labels is effective for low-resource RE. The prompt learning curriculum can improve the results if the amount of instances is more (K=32), which indicates that the prompt learning curriculum needs more instances to show the guide effect. We present a thorough ablation study to show the effects of our designs. FPC is mainly compared with Ent Marker and TYP Marker (Zhou and Table The words of entity mentions and types can provide entity information and the model can utilize the clues to make predictions. Hence further showing entity type words can boost the results. FPC(TEMP): We insert the templates "E" and "ET" into the input to get the results. The evidently improved performance shows that introducing entity information in the templates is more helpful than using the type markers. The model can utilize this kind of relation-oriented knowledge better if it is presented directly and orderly in the templates. FPC(RPL): We introduce the relation prompt learning based on the templates to attain the results. While the model achieves obviously higher results on Re-TACRED and SemEval, the results of TA-CRED and TACREV are slightly improved. This is because the mislabeled instances of Re-TACRED and SemEval are less and these datasets are easy for our model. When handling the other two hard datasets, the model can not successfully build the connection between the targets of classification and MLM. Therefore the prompt learning curriculum is proposed to improve the performance. FPC(RPL+PLC): We fine-tune the model according to the prompt learning curriculum to obtain the results. Our model attains remarkable improvement on TACRED and TACREV and similar results on Re-TACRED and SemEval. By learning the subtask with increasing difficulty, the model can easily connect classification target with MLM target and adapt to the multi-task setting, which is more effective on hard datasets. The superior results show the effectiveness of the prompt learning curriculum. We find that the templates have a great influence on the results. The reason is that they can provide entity information which is crucial for RE. To study the importance of different entity information, we design two new templates shown as below. We conduct experiments of FPC with different templates and the results are shown in Table To study the effect of the prompt learning curriculum, we evaluate different model checkpoints during fine-tuning on TACRED development and test sets. We report the average scores of 10 runs and the results are shown in Figure We further utilize the prompt learning curriculum to fine-tune the model and find that the model performance is gradually and stably improved after each epoch. Most best results are obtained at the end of fine-tuning and the final results are significantly improved. This indicates that the prompt learning curriculum can help the model to link the objectives of the multi-task setting and make full use of the datasets, hence our model can capture and utilize the semantics of relation labels. Based on the setting of the relation prompt learning, we propose the prompt learning curriculum which is different from other existing curriculum learning methods. In order to better show the influence of the prompt learning curriculum, we design another curriculum learning method as our baseline to make a comparison. We propose the increasing α curriculum with the similar idea: we increase the difficulty of the subtask by changing the weights in the total loss func- Table In this paper, we propose a novel method Finetuning with Prompt Curriculum (FPC) for RE. The relation prompt learning introduces the cloze-style auxiliary task, through which the model can capture the semantics of relation labels. The prompt learning curriculum makes the model adapt to the multi-task setting by learning the increasingly difficult sub-task, which makes the model build the connection between the targets of classification and MLM. Extensive experiments have been conducted on four popular RE benchmarks. The results show that FPC achieves the new state-of-the-art performance for fully supervised RE and the competitive or best performance for low-resource RE. Zhengyan Wenxuan Zhou and Muhao Chen. 2021. An improved baseline for sentence-level relation extraction. arXiv preprint arXiv:2102.01373. Yikai Zhou, Baosong Yang, Derek F Wong, Yu Wan, and Lidia S Chao. 2020. Uncertainty-aware curriculum learning for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6934-6944. This section presents more details about the finetuning procedures and hyperparameters. We report the used settings which result in the overall best performance. We use the same punctuations "@" and "#" as entity markers following (Zhou and Chen, 2021). We warm up the learning rate over the first 10% steps and then linearly decay it. We set the weight decay to 1e -5 and clip gradients if their norms exceed 1.0. The maximum sequence length is set to 512 and none of the instances exceed it. For the relation prompt learning, we set α in Equation (1) to 0.4 on TACRED, TACREV and Re-TACRED and 0.3 on SemEval under both fully supervised and low-resource settings. For the prompt learning curriculum, the proportion of "mask" instances P mask is controlled by the number of fine-tuning epochs. P mask linearly increases during fine-tuning and finally become 100%. For the increasing α curriculum, α in Equation (1) linearly increases during fine-tuning and we use the number of fine-tuning epochs to adjust α as well. Table The designed label words of the used datasets are shown in Table
| 1,183 | 2,412 | 1,183 |
To Build Our Future, We Must Know Our Past: Contextualizing Paradigm Shifts in Natural Language Processing
|
NLP is in a period of disruptive change that is impacting our methodologies, funding sources, and public perception. In this work, we seek to understand how to shape our future by better understanding our past. We study factors that shape NLP as a field, including culture, incentives, and infrastructure by conducting longform interviews with 26 NLP researchers of varying seniority, research area, institution, and social identity. Our interviewees identify cyclical patterns in the field, as well as new shifts without historical parallel, including changes in benchmark culture and software infrastructure. We complement this discussion with quantitative analysis of citation, authorship, and language use in the ACL Anthology over time. We conclude by discussing shared visions, concerns, and hopes for the future of NLP. We hope that this study of our field's past and present can prompt informed discussion of our community's implicit norms and more deliberate action to consciously shape the future.
|
Natural language processing (NLP) is in a period of flux. The successes of deep neural networks and large language models (LLMs) in NLP coincides with a shift not only in the nature of our research questions and methodology, but also in the size and visibility of our field. Since the mid-2010s, the number of first-time authors publishing in the ACL Anthology has been increasing exponentially (Figure In the 2022 NLP Community Metasurvey, many NLP practicioners expressed fears that private firms exert excessive influence on the field, that "a majority of the research being published in NLP is of dubious scientific value," and that AI technology could lead to a catastrophic event this century We present a study of the community in its current state, informed by a series of long-form retrospective interviews with NLP researchers. Our interviewees identify patterns throughout the history of NLP, describing periods of research productivity and stagnation that recur over decades and appear at smaller scale around prominent methods ( §3). Interviewees also point out unparalleled shifts in the community's norms and incentives. Aggregating trends across interviews, we identify key factors shaping these shifts, including the rise and persistence of benchmarking culture ( §4) and the maturation and centralization of software infrastructure ( §5). Our quantitative analysis of citation patterns, authorship, and language use in the ACL Anthology over time provides a complemen-tary view of the shifts described by interviewees, grounding their narratives and our interpretation in measurable trends. Through our characterization of the current state of the NLP research community and the factors that have led us here, we aim to offer a foundation for informed reflection on the future that we as a community might wish to see.
|
We recruited 26 researchers to participate in interviews using purposive Interviews were semi-structured Following procedures of grounded theory We use quantitative methods primarily as a coherence check on our qualitative results. While our work is largely concerned with the causes and community reception of changes in the community, our quantitative analyses provide evidence that these changes have occurred. This includes analyzing authorship shifts (Figures 3 Exploit-explore cycles of work Our participants described cyclical behavior in NLP research following methodological shifts every few years. Many participants referred to these methodological shifts as "paradigm shifts", with similar structure, which we characterize as explore and exploit phases. identify their own exploit work, they joked we could simply "sort [their] Google Scholar from most cited to least cited" (9), describing another incentive to publish the first paper applying a new methodology to a particular task. Participants additionally described doing exploit work early in their careers as a hedge against riskier, longer-term projects. However, most participants ascribed low status to exploit work Participants felt that the exploit phase of work is not sustainable. Eventually, "the low hanging fruit has been picked" (8); this style of work becomes unsurprising and less publishable. As one researcher put it: "if I fine tune Second wave: explore After some time in the exploit phase, participants described a state where obvious extensions to the dominant methodology have already been proposed, and fewer papers demonstrate dramatic improvements over the state of the art on popular benchmarks. In this phase, work on identifying the ways that the new method is flawed gains prominence. This work may focus on interpretability, bias, data or compute efficiency. While some participants see this as a time of "stalled" (8) progress, others described this as "the more interesting research after the initial wave of showing things work" (18). A mid-career participant identified this as the work they choose to focus on: "I'm at the stage of my career where I don't want to just push numbers, you know. I'll let the grad students do that. I want to do interesting stuff" (22). Participants often saw "pushing numbers" as lower-status work, appropriate for graduate students and important for advancing the field and one's career, but ultimately not what researchers hope to explore. Some work to improve benchmark performance was also perceived as explore work, particularly when it involved developing new architectures. One participant described a distinction between "entering a race" and "forging in a new direction" (4) with a project, which focuses the exploit/explore divide more on the perceived surprisingness of an idea rather than the type of contribution made. Exploration often leads to a new breakthrough, causing the cycle to begin anew. Placing the current state of the field along the exploit-explore cycle requires defining the current methodological "paradigm". Participants identified similar patterns at varying scales, with some disagreement on the timing of recent trends. Prompting as a methodological shift Several participants described prompting as a paradigm shift or a direction that the community found promising, but most participants viewed current work on prompt engineering or "ChatGPT for X" (9) as something that people are working on "instead [...] of something that might make a fundamental difference" (14). One participant described both prompt engineering and previous work on feature engineering as "psuedoscience [...] just poking at the model" (6). The current flurry of prompting work was viewed by several participants as lowerstatus work exploiting a known method. "Era of scale" For participants who discussed larger-scale cycles, pre-trained models were frequently identified as the most recent methodological shift. Participants disagreed on whether scaling up pre-trained models (in terms of parameter count, training time, and/or pre-training data) was a form of exploiting or exploring this method. Some participants found current approaches to scale to be "a reliable recipe where we, when we put more resources in, we get [...] more useful behavior and capabilities out" (4) and relatively easy to perform: "Once you have that GPU [...] it's like, super simple" (5). This perception of scaling up as both high likelihood of success and low difficulty places it as exploit work, and researchers who described scale in this way tended to view it as exploiting "obvious" trends. One researcher described scale as a way of establishing what is possible but "actually a bad way to achieve our goals." (4), with further (explore-wave) work necessary to find efficient ways to achieve the same performance. A minority of participants argued that, while historical efforts to scale models or extract large noisy corpora from the internet were exploit work, current efforts to scale are different, displaying "emergence in addition to scale, whereas previously we just saw [...] diminishing returns" ( "Deep learning monoculture" There was a sense from several participants that the current cycle has changed the field more than previous ones, featuring greater centralization on fewer methods (see §5 for more discussion). Some expressed concern: "a technique shows some promise, and then more people investigate it. That's perfectly appropriate and reasonable, but I think it happens a little too much. [...] Everybody collapses on this one approach [...] everything else gets abandoned." (19). Another participant described peers from linguistics departments who left NLP because they felt alienated by the focus on machine learning. Issues with peer review Some felt that peer review was inherently biased toward incremental work because peer reviewers are invested in the success of the current methodological trends, with one participant arguing that "if you want to break the paradigm and do something different, you're gonna get bad reviews, and that's fatal these days" (21). Far more commonly, participants did not express inherent opposition to peer review but raised concerns because of the recent expansion of the field, with one senior industry researcher remarking that peer reviewers are now primarily junior researchers who"have not seen the effort that went into [earlier] papers" (12). Another participant asserted that "my peers never review my papers" (22). Participants additionally suggested that the pressure on junior researchers to publish more causes an acceleration in the pace of research and reinforcement of current norms, as research that is farther from current norms/methodologies requires higher upfront time investment. This competitiveness can manifest in harsher reviews, and one participant described a "deadly combination" (19) of higher standards for papers and lower quality of reviews. Some participants described this as a reason they were choosing to engage less with NLP conferences; one industry researcher stated that "I just find it difficult to publish papers in *CL [venues] that have ideas in them." (22). 4 Benchmarking culture Senior and emeritus faculty shared a consistent recollection of the ACL community before the prominence of benchmarks as centralized around a few US institutions and characterized by "patient money" (21): funding from DARPA that did not require any deliverables or statements of work. Capabilities in language technologies were show-cased with technical "toy" (26, 19) demonstrations that were evaluated qualitatively: "the performance metrics were, 'Oh my God it does that? No machine ever did that before.' " (21). Participants repeatedly mentioned how small the community was; at conferences, "everybody knew each other. Everybody was conversing, in all the issues" (26). The field was described as "higher trust" ( Many participants recalled the rise of benchmarks in the late 1990s and early 2000s, coinciding with a major expansion in the NLP community in the wake of the "statistical revolution," where participants described statistical models displacing more traditional rules-based work (see Figure ple together to exchange ideas. [...] I think this really helped the field to move forward." (2). Other participants similarly argued that a culture of quantitative measurement was key for moving on from techniques that were appealing for their "elegance" (14) but empirically underperforming. Roughly twenty years on from the establishment of benchmarks as a field-wide priority, our participants' attitudes towards benchmarks had become significantly more complex. Many of our participants still found benchmarks necessary, but nearly all of them found them increasingly insufficient. Misaligned incentives Many participants, particular early-and late-career faculty, argued that the field incentivizes the production of benchmark results to the exclusion of all else: "the typical research paper...their immediate goal has to be to get another 2% and get the boldface black entry on the table." (21). For our participants, improvements on benchmarks in NLP are the only results that are self-justifying to reviewers. Some participants felt this encourages researchers to exploit modeling tricks to get state-of-the-art results on benchmarks, rather than explore the deeper mechanisms by which models function (see §3). "We're solving NLP" Some participants perceive a degradation in the value of benchmarks because of the strength of newer models. Participants appreciated both the increased diversity and frequency of new benchmark introduction, but noted that the time for new approaches to reach "superhuman" (6,22) levels of performance on any specific benchmark is shortening. One common comparison was between part of speech tagging ("a hill that was climbed for [...] about 20 years" (15)) and most modern benchmarks ("solved" within a few years, or even months). Some went further, describing "solving NLP" (8) or naming 2020 as the time when "classification was solved" (15). However, when participants were asked for clarification on what it meant to "solve" a problem, most participants hedged in similar ways; that datasets and benchmarks could be solved, with the correct scoping, but problems could rarely or never be solved. Many participants argued that the standard for solving a task should be human equivalency, and that this was not possible without new benchmarks, metrics, or task definition. NLP in the wild Some participants argued that many benchmarks reflect tasks that "aren't that useful in the world" (13), and that this has led to a situation where "[NLP], a field that, like fundamentally, is about something about people, knows remarkably little about people" (3). Industry participants often viewed this as a distinction between their work and the academic community, with one stating that "most of the academic benchmarks out there are not real tasks" (12). Many academics articulated a desire for more human-centered NLP, and most participants described feeling pressure over the unprecedented level of outside interest in the field. One participant contrasted the international attention on ChatGPT with the visibility of earlier NLP work: "It's not like anyone ever went to like parser.yahoo.com to run a parser on something" (3). Participants argued that, given this outside attention, the benchmark focus of NLP is too narrow, and that benchmarks fail to capture notions of language understanding that translate to wider audiences, and that we should move on from benchmarks not when they are saturated but when "it wouldn't really improve the world to improve this performance anymore" (9). This echoed a common refrain: many participants, especially early-and mid-career researchers, saw positive social change as a goal of progress in NLP. 5 Software lotteries Centralization on frameworks As deep learning became more popular in NLP, participants described the landscape shifting. As TensorFlow Centralization on specific models Participants identified another shift after the release of BERT and subsequent development of Hugging Face. Because of pre-training, participants moved from merely using the same libraries to "everyone us[ing] the same base models" (9). Participants expressed concern that this led to further centraliza-tion in the community, with one participant identifying a trend of some people "not us[ing] anything else than BERT [...] that's a problem" (5). This concentration around particular modeling choices has reached greater heights than any previous concentration on a method; in 2021, 46.7% of papers in the ACL anthology cited BERT Centralization on Python While most earlycareer and late-career participants did not express strong opinions about programming languages, many mid-career participants expressed strong dislike for Python, describing it as "a horrible language" ( This increasing centralization of the modern NLP stack has several consequences. One of the primary ones, however, is the loss of control of design decisions for the majority of researchers in the community. Practically, researchers can now choose from a handful of well-established implementations, but only have access to software and models once the decisions on how to build them have already been reified in ways that are difficult to change. Lower barriers Beyond time saved (re-) implementing methods, many participants identified a lower barrier to entry into the field as a notable benefit of centralization on specific software infrastructure. Participants described students getting state of the art results within an hour of tackling a problem; seeing the average startup time for new students decreasing from six months to a few weeks; and teaching students with no computer science background to build NLP applications with prompting. Obscuring what's "under the hood" One participant recalled trying to convince their earlier students to implement things from scratch in order to understand all the details of the method, but no longer doing so because "I don't think it's possible [...] it's just too complicated" (11); others attributed this to speed more than complexity, stating that "the pace is so fast that there is no time to properly document, there is no time to properly engage with this code, you're just using them directly" (5). However, this can cause issues on an operational level; several participants recalled instances where a bug or poor documentation of shared software tools resulted in invalid research results. One participant described using a widely shared piece of evaluation code that made an unstated assumption about the input data format, leading to "massively inflated evaluation numbers" (3) on a well-cited dataset. Another participant described working on a paper where they realized, an hour before the paper deadline, that the student authors had used two different tokenizers in the pipeline by mistake: "we decided that well, the results were still valid, and the results would only get better if [it was fixed]...so the paper went out. It was published that way." (26) Software bugs in research code are not a new problem, Software is implicit funding Participants suggested that tools that win the software lottery act as a sort of implicit funding: they enable research groups to conduct work that would not be possible in the tools' absence, and many of our participants asserted that the scope of their projects expanded as a result. However, they also significantly raise the relative cost of doing research that does not fall neatly into existing tools' purview. As one participant stated, "You're not gonna just build your own system that's gonna compete on these major benchmarks yourself. You have to start [with] the infrastructure that is there" (19). This is true even of putatively "open" large language models, which do not necessarily decentralize power, and can often entrench it instead A common sentiment among participants was that centralization has had an overall positive impact on reproducibility, because using shared tools makes it easier to evaluate and use others' research code. However, participants also expressed concerns that the increasing secrecy of industry research complicates that overall narrative: "things are more open, reproducible... except for those tech companies who share nothing" (14). Shifts in expectations One participant described a general shift in focus to "making sure that you make claims that are supported rather than reproducing prior work exactly" (12) in order to match reviewers' shifting expectations. However, participants also felt that the expectations for baselines had increased: "[in the past,] everybody knew that the Google system was better because they were running on the entire Internet. But like that was not a requirement [to] match Google's accuracy. But now it is, right?" (8). Disparities in compute access Many felt that building large-scale systems was increasingly out of reach for academics, echoing concerns previously described by The shifts we explore in this paper have not happened in a vacuum, with adjacent research communities such as computer vision (CV) and machine learning (ML) experiencing similar phenomena, inspiring a number of recent papers discussing norms in AI more broadly. While we focus on summarizing and synthesizing the views of our participants, some of the overarching themes identified in this work have been discussed more critically. Fishman and Hancox-Li (2022) critique the unification of ML research around transformer models on both epistemic and ethical grounds. Position papers have critiqued the notion of general purpose benchmarks for AI The NLP Community Metasurvey The availability of the ACL Anthology has enabled quantitative studies of our community via patterns of citation, authorship, and language use over time. Shifts in norms and methods in science more broadly has been studied outside computing-related fields. Most notably, The rise of large language models has coincided with disruptive change in NLP: accelerating centralization of software and methodologies, questioning of the value of benchmarks, unprecedented public scrutiny of the field, and dramatic growth of the community. A shift like this can feel threatening to the fundamental nature of NLP research, but this is not the first period of flux in the field, nor are the fundamental forces enabling LLMs' dominance and other changes entirely new. Our participants described cycles of change in the NLP community from mid-80s to the present, with common themes of first exploiting and then exploring promising methodologies. Each methodological shift brought corresponding cultural change: the shift from symbolic to statistical methods brought about the rise of benchmark culture and the end of the socially mediated, smallnetwork ACL community. Neural methods began the centralization on software toolkits and the methodologies they support. Pre-training intensified this software lottery, causing unprecedented levels of centralization on individual methods and models. Current models have called into question the value of benchmarks and catapulted NLP into the public eye. Our participants largely agree on the resulting incentives-to beat benchmark results, to do the easiest thing rather than the most fulfilling, to produce work faster and faster -while largely expressing frustration with the consequences. We hope that this contextualization of the current state of NLP will both serve to inform newer members of the community and stir informed discussion on the condition of the field. While we do not prescribe specific solutions, some topics of discussion emerge from the themes of this work: • Who holds the power to shape the field? How can a broad range of voices be heard? • Do the incentives in place encourage the behavior we would like to see? How can we improve reviewing to align with our values? • What affects the ability to do longer-term work that may deviate from current norms? • How can the community arrive at an actively mediated consensus, rather than passively being shaped by forces like the ones we discuss? We personally take great hope for our community from this project. The care with which all participants reflected on the shape of the field suggests to us that many people are concerned about these issues, invested in the community, and hopeful for the future. By sharing publicly what people so thoughtfully articulate privately, we hope to prompt further discussion of what the community can do to build our future. Western bias The most notably irrepresentative sampling bias in our participant pool is the lack of non-Western institutional affiliation (and the strong skew toward North American affiliations). This bias has arisen likely in part due to our own institutional affiliation and conceptions of the community. That being said, given the Association for Computational Linguistics' historically US-and Englishcentric skews, this allows us to gather historical perspectives. Additionally, considering that Western institutions constitute a citation network largely distinct from Asian networks Lack of early career voices Our inclusion criteria for our participants-three or more publications in *CL, IR, or speech venues Future research could further examine differences between geographic subcommunities in NLP and more closely examine influences on people's participation in and disengagement from the community. Additionally, we leave to future work a more intentional exploration of perspectives from early career researchers and those who have not yet published but are interested in NLP research. Following Institutional Review Board recommendations, we take steps to preserve the anonymity of our participants, including aggregating or generalizing across demographic information, avoiding the typical practice of providing a table of per-interviewee demographics, using discretion to redact or not report quotes that may be identifying, and randomizing participant numbers. Participants consented to the interview and to being quoted anonymously in this work. This work underwent additional IRB screening for interviewing participants in GDPR-protected zones. We view our work as having potential for positive impact on the ACL community, as we prompt its members to engage in active reflection. We believe that, given recent developments in the field and the co-occuring external scrutiny, the current moment is a particularly appropriate time for such reflection. Additionally, we hope that our work can serve those currently external to the community as an accessible, human-centered survey of the field and factors that have shaped it over the decades, prioritizing sharing of anecdotes and other in-group knowledge that may be difficult for outsiders to learn about otherwise. We begin with the ACL Anthology and focus on papers between 1980 and 2022. Using the Semantic-Scholar (S2) API We rely on S2ORC We present an alternative view of data, similar to that seen in Figure
| 1,007 | 1,836 | 1,007 |
Toxicity Detection: Does Context Really Matter?
|
Moderation is crucial to promoting healthy online discussions. Although several 'toxicity' detection datasets and models have been published, most of them ignore the context of the posts, implicitly assuming that comments may be judged independently. We investigate this assumption by focusing on two questions: (a) does context affect the human judgement, and (b) does conditioning on context improve performance of toxicity detection systems? We experiment with Wikipedia conversations, limiting the notion of context to the previous post in the thread and the discussion title. We find that context can both amplify or mitigate the perceived toxicity of posts. Moreover, a small but significant subset of manually labeled posts (5% in one of our experiments) end up having the opposite toxicity labels if the annotators are not provided with context. Surprisingly, we also find no evidence that context actually improves the performance of toxicity classifiers, having tried a range of classifiers and mechanisms to make them context aware. This points to the need for larger datasets of comments annotated in context. We make our code and data publicly available.
|
Systems that detect abusive language are used to promote healthy conversations online and protect minority voices Hitler was also strongly antipornography • RQ1: How often does context affect the toxicity of posts as perceived by humans in online conversations? And how often does context amplify or mitigate the perceived toxicity? 9HmSzc). For other languages, see for example the German GERMEVAL ( COMMENT Table • RQ2: Does context actually improve the performance of toxicity classifiers, when they are made context-aware? And how can toxicity classifiers be made context-aware? To investigate these questions we created and make publicly available two new toxicity datasets that include context, which are based on discussions in Wikipedia Talk Pages To investigate the second question, concerning the effect of context on the performance of toxicity classifiers, we created a larger dataset of 20k comments; 10k comments were annotated out of context, 10k in context. This time we did not require the same comments to be annotated with and without context, which allowed us to crowd-source the collection of a larger set of annotations. These two new subsets were used to train several toxi-city detection classifiers, both context-aware and context-unaware, which were evaluated on held out comments that we always annotated in context (based on the assumption that in-context labels are more reliable). Surprisingly, we found no evidence that context actually improves the performance of toxicity classifiers. We tried a range of classifiers and mechanisms to make them context aware, and having also considered the effect of using gold labels obtained out of context or by showing context to the annotators. This finding is likely related to the small number of context-sensitive comments. In turn this suggests that an important direction for further research is how to efficiently annotate larger corpora of comments in context. We make our code and data publicly available.
|
Toxicity detection has attracted a lot of attention in recent years Table Both small and large toxicity datasets have been developed, but approximately half of them contain tweets, which makes reusing the data difficult, because abusive tweets are often removed by the platform. Moreover, the textual content is not available under a license that allows its storage outside the platform. The hateful language detection dataset of A common characteristic of most of the datasets listed in Table Two other datasets, both non English, also include context-aware annotations. In summary, of the datasets we know of (Table 3), only two include context Using previous comments of a conversation or preceding sentences of a document is not uncommon in text classification and language modeling. coded sentence history to an RNN language model 3 Experiments To investigate how often context affects the perceived toxicity of posts, we created CAT-SMALL, a small Context-Aware Toxicity dataset of 250 randomly selected comments from the Wikipedia Talk Pages (Table Figure The difference is statistically significant (P < .01). (direct reply) to 5. We used the parent comment and discussion title only, instead of a larger context (e.g., the entire thread), to speed up our machine learning experiments, and also because reading only the previous comment and the discussion title made the manual annotation easier. In preliminary experiments, we observed that including more preceding comments had the side effect of workers tending to ignore the context completely. For each comment and group of annotators, the toxicity scores of the annotators were first averaged and rounded to the nearest binary decision, as in Table To get a more accurate picture of the effect of context, we measured the number of comments of CAT-SMALL for which the (averaged and rounded) toxicity label was different between the two groups (GN, GC). We found that the toxicity of 4 comments out of 250 (1.6%) decreased with context, while the toxicity of 9 comments (3.6%) increased. Hence, perceived toxicity was affected for 13 comments (5.2% of comments). While the small size of CAT-SMALL does not allow us to produce accurate estimates of the frequency of posts whose perceived toxicity changes with context, the experiments on CAT-SMALL indicate that context has a statistically significant effect on the perceived toxicity, and that context can both amplify or mitigate the perceived toxicity, thus making a first step to addressing our first research question (RQ1). Nevertheless, larger annotated datasets need to be developed to estimate more accurately the frequency of context-sensitive posts in online conversations, and how often context amplifies or mitigates toxicity. To investigate whether adding context can benefit toxicity detection classifiers, we could not use CAT-SMALL, because its 250 comments are too few to effectively train a classifier. Thus, we proceeded with the development of a larger dataset. Although the best approach would be to extend CAT-SMALL, which had two mutually exclusive groups of annotators labeling each comment, we found that the annotation process was very slow in that case, largely because of the small size of annotator groups we had access to in Figure Eight (19 and 23 for GC and GN respectively). reply) to 5. Inter-annotator agreement was computed with Krippendorff's alpha on 123 texts, and it was found to be 0.72% for GN and 0.70% for GC. Figure We used CAT-LARGE to experiment with both context-insensitive and context-sensitive toxicity classifiers. The former only consider the post being rated (the target comment), whereas the latter also consider the context (parent comment). BILSTM Our first context-insensitive classifier is a bidirectional LSTM BERT At the other end of complexity, our second context-insensitive classifier is BERT PERSPECTIVE The third context-insensitive classifier is a CNN-based model for toxicity detection, trained on millions of user comments from online publishers. It is publicly available through the PER-SPECTIVE API. CA-BILSTM-BILSTM In a context-aware extension of the context-insensitive BILSTM classifier, dubbed CA-BILSTM-BILSTM, we added a second BILSTM to encode the parent comment (Fig. We used a bidirectional LSTM to encode the parent comment, similarly to CA-BILSTM-BILSTM. The biLSTM representation was concatenated with the [CLS] representation before the FFNN. All other parameters were set to the same values as BERT. Table A first observation from Table 9 Recall that we also fix the bias term of the output neuron of each model (apart from PERSPECTIVE) to -log T N , to bias against the majority class. We also tried under-sampling to address class imbalance, but this technique worked best. This is not surprising, since these systems were trained (fine-tuned in the case of BERT-CCTK) on much larger toxicity datasets than the other systems (upper two zones of Table What is more surprising is that any kind of information about the context does not lead to any consistent (or large) improvement in system performance. PERSPECTIVE and BERT-CCTK seem to improve slightly with the naive context-awareness mechanism of concatenating the parent and target text during testing, but the improvement is very small and we did not detect a statistically significant difference. There was also no consistent (or large) improvement when encoding the parent comments with a BILSTM (CA-BILSTM-BILSTM, CA-BILSTM-BERT) or directly as in BERT's next sentence prediction pre-training task (CA-SEP-BERT). This is again probably a consequence of the fact that contextsensitive comments are few. The small number of context-sensitive comments does not allow the BILSTM-and BERT-based classifiers to learn how to use the context encodings to cope with context-sensitive comments, and failing to cope with context-sensitive comments does not matter much during testing, again since context-sensitive comments are so few. We conclude for our second research question (RQ2) that we found no evidence that context actually improves the performance of toxicity classifiers, having tried both simple (BILSTM) and more powerful classifiers (BERT), having experimented with several methods to make the classifiers context aware, and having also considered the effect of gold labels obtained out of context vs. gold labels obtained by showing context to annotators. We investigated the role of context in detecting toxicity in online comments. We collected and share two datasets for investigating our research questions around the effect of context on the annotation of toxic comments (RQ1) and its detection by automated systems (RQ2). We showed that context does have a statistically significant effect on toxicity annotation, but this effect is seen in only a narrow slice (5.2%) of the (first) dataset. We also found no evidence that context actually improves the performance of toxicity classifiers, having tried both simple and more powerful classifiers, having experimented with several methods to make the classifiers context aware, and having also considered the effect of gold labels obtained out of context vs. gold labels obtained by showing context to the annotators. The lack of improvement in system performance seems to be related to the fact that context-sensitive comments are infrequent, at least in the data we collected. A limitation of our work is that we considered a narrow contextual context, comprising only the previous comment and the discussion title. Annotators were asked to judge the toxicity of each comment, given the following definitions: • VERY TOXIC: A very hateful, aggressive, disrespectful comment or otherwise very likely to make a user leave a discussion or give up on sharing their perspective. • TOXIC: A rude, disrespectful, unreasonable comment or otherwise somewhat likely to make a user leave a discussion or give up on sharing their perspective. • UNSURE: Due to polysemy, lack of context or other reasons. • NOT TOXIC: Not containing any toxicity. For annotation, we used the 'Figure Eight' platform and we invested 5 cents per row. 12 For the CAT-SMALL we employed high accuracy annotators (i.e., from zone 3), selected from 7 English speaking countries (i.e., UK, Ireland, USA, Canada, New Zealand, South Africa, Australia), and only ones allowing explicit content (we also warned about the explicit content in the title). 62 quiz questions were used. For the CAT-LARGE, we invested the same amount of money but all the annotators were able to participate (again, they ware warned for the explicit content). Inter annotator agreement was measured on the quiz questions with Krippendorff's alpha and was found to be 70% and 72% for the C and N sets. GC annotators had one more question, which was asking them to compare the toxicity of the target comment to that of the parent comment. The main scope of that question was to make it less easy for annotators to ignore the parent comment. 12 All systems were trained for 100 epochs with patience of 3 epochs. We performed early stopping by monitoring the validation ROC AUC. The hidden size of the LSTM cells had size 128. We used batch size 128, max length 512, and we concatenated the forward and backward last hidden states before the FFNN. We used binary cross entropy for loss and Adam optimizer was used with default parameters (learning rate 1e-03). We used the same hyper-parameters with BILSTM but included one more bidirectional LSTM to encode the parent text. The parent biLSTM had 64 hidden nodes and we concatenated the forward and backward last hidden states. The parent and the target embeddings (the ones generated by the two biLSTMS) were concatenated before being passed to the FFNN. We used a learning rate of 2e-05 for BERT and only unfroze the top three layers during training to our data. On top of the BERT [CLS] representation, we added a FFNN of 128 hidden nodes and a sigmoid to yield the toxicity probability. 128 tokens were used as maximum sequence length. A [SEP] token separated the two texts and the [CLS] token was used as with BERT. Same parameters with BERT were used.
| 1,167 | 1,985 | 1,167 |
Vision Meets Definitions: Unsupervised Visual Word Sense Disambiguation Incorporating Gloss Information
|
Visual Word Sense Disambiguation (VWSD) is a task to find the image that most accurately depicts the correct sense of the target word for the given context. Previously, image-text matching models often suffered from recognizing polysemous words. This paper introduces an unsupervised VWSD approach that uses gloss information of an external lexical knowledge-base, especially the sense definitions. Specifically, we suggest employing Bayesian inference to incorporate the sense definitions when sense information of the answer is not provided. In addition, to ameliorate the out-of-vocabulary (OOV) issue, we propose a context-aware definition generation with GPT-3. Experimental results show that VWSD performance increased significantly with our Bayesian inference-based approach. In addition, our context-aware definition generation achieved prominent performance improvement in OOV examples exhibiting better performance than the existing definition generation method.
|
With the development of deep learning technology, research on multimodality such as Visio-Linguistic Models (VLMs) has been actively conducted Visual Word Sense Disambiguation (VWSD) is a multimodal task of natural language processing (NLP) and computer vision that selects the image which corresponds to the intended meaning of the target word among a set of candidate images (Ra- To ameliorate this problem, we propose to disambiguate visual words with the assistance of a glossary of lexical knowledge-bases (LKBs) without the use of any further training or dataset. Specifically, we utilize the sense definitions of an ambiguous word that have been widely exploited in previous lexical semantic tasks Experiments were conducted on SemEval-2023 (SE23) Task 1-Visual-WSD The contributions of this paper can be summarized as follows: • This paper introduces a new glossincorporated VWSD approach inspired by Bayesian inference. • Experimental results show that our Bayesian inference-based approach boosted the unsupervised VWSD performance significantly without any additional training. • Furthermore, we suggest the CADG method to challenge the OOV issue. 2 Related Work Disambiguation (WSD) which automatically identifies ambiguous words into corresponding senses Visual Verb Sense Disambiguation (VVSD) is another task relevant to VWSD. VVSD is a multimodal sense disambiguation task that selects the correct sense of a pair of a given ambiguous verb word and image Our work is related to gloss-enhanced WSD models in that we are using both sense definition and context together. However, our study differs from previous WSD studies in that it tackles a multi-modal task. It is also relevant to VVSD in terms of multi-modal sense disambiguation. However, VVSD systems
|
We found two representative error cases in the results of the definition generations: 1) misdisambiguation and 2) hallucination. The misdisambiguation is when the GPT3 generates the polysemy's definition. In Figure We formulate unsupervised VWSD as a multiclass classification task Any pretrained ITM models (e.g., CLIP) can calculate the posterior. In Figure Usually, zero-shot ITM models are pretrained without much consideration of polysemous words. For example, Figure In this case, the right term P (D t i |c, t) (Context to Definition; C2D) is predicting the conditional probability over the given ith sense definition D t i for the given target word t and context c which is similar to the gloss-enhanced WSD models and D2I input to the softmax function transformed into probability distributions. Then, we choose the image with the highest probability as the prediction. As a result, for the example in Figure Define "angora" in angora city. angora (n): A city in Turkey that stands on the banks of the Angora River, near … Prompt: Definition: (b) Our context-aware definition generation. Figure Not all words have their definitions available in a lexical knowledge-base. In particular, proper nouns, compound words, and foreign words frequently induce OOV issues. For example, in the SE23 dataset, about 14.33% of target words' definitions are not found in the English WordNet. Therefore, we propose a solution to tackle the OOV issue with the definition generation approach. A previous study showed that GPT-3 can generate the definition of a novel word Figure 6 Experiments 6.1 Experimental Dataset SE23 We used the dataset in the SemEval-2023 Task 1 VWSD challenge 34 . It consists of 12,896 examples and 13,000 candidate images. Each example has 10 candidates that include 1 answer image and 9 distractors. Each context averagely contains 2.5 words. The dataset contains 14.33% OOV words (1,845 out of 12,869). VWSD For the experiments, we adopted two SOTA zero-shot ITM models, CLIP We used WordNet 3.0 Definition Generation We re-implemented Specifically, we sampled a definition for each example by utilizing GPT-3's Davinci variant which is known as the largest model and we generated samples with a temperature of 1.0. Evaluation Criteria Following Raganato et al. ( Others We prepared a pretrained WSD, T5 SemCor The experimental results in Table Meanwhile, while FLAVA shows prominent improvement via WordNet integration, the impact of generated definitions tends to be low compared to CLIP. Considering that WordNet definitions were manually constructed by experts, we speculate that this is because the model is sensitive to the quality of the input definitions. We analyzed the performance change according to the ambiguity level of the ambiguous target word. Although WordNet integration improves performance for ambiguous target words, we still want to find out how competitive the performance improvement is. For this reason, we compared the performance of our WordNet-incorporated model with that of the pipeline system using the WSD model. To be specific, T5 SemCor , a finetuned WSD model, predicts WordNet sense in a given target word and context. The probability distribution for the candidate images was calculated based on the predicted sense. Table We also verified whether the quality of the generated definitions would affect the VWSD performance. Table Since we sampled a definition for each input example in main experiments, it is still questionable whether the number of sampled definitions affects the performance of the model. Table Our model still suffers from error cascading from C2D probability though it is mitigated by the Bayesian style inference. The most typical error case is due to the error cascading in C2D probability calculation. Especially, due to the nature of neural networks This paper introduces a novel VWSD methodology to effectively incorporate gloss information from an external resource. Our work mainly has two innovations: 1) Bayesian style inference for SOTA ITMs, and 2) Context-aware definition generation with GPT-3 to overcome the OOV issue. Experimental results show that our proposed Bayesian style inference-based WordNet integration significantly improves VWSD performance without additional training. For the ambiguous target words, the performance of our approach is comparable to pipeline systems using finetuned WSD models. Moreover, context-aware definition generation helps mitigate OOV issues in the downstream VWSD tasks and shows higher performance compared to the previous definition generation approach. In the future, we plan to tackle the error cascading caused by over-confidence in C2D probability. For this, we may explore a prompting that is known to have good performance in zero-shot prediction Our work has the following limitations. First, we only used one evaluation data, namely SE23, because it is the only data suitable for the VWSD setting, especially for the OOV examples. In addition, our methodology relies entirely on WordNet. Therefore, this may be limited the model's ability when the target word is a proper noun such as a named entity. Finally, we depend on the results of GPT-3 definition generation to handle OOV words. Since the generated definitions may contain errors, as revealed in the qualitative analyses, the errors led to incorrect predictions. Table First of all, the misdisambiguation cases are caused by bias in the pretraining, and we can notice that CADG has less misdisambiguation compared to DG. Especially, we can see that GPT-3 generated more than one definitions of the target words 'conch', 'reaper', and 'ruin' in DG, while we could not found such cases in our approach. On the other hand, hallucination cases are when the generated definitions are definitions of completely different terms with similar spellings ('stonechat' of CADG, 'driftfish' of DG), or cases in which the detailed descriptions are incorrect although they are somewhat similar In the limitation section We downloaded those in the official download site. Also, we got allowance to use the dataset from the creators. B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? section 6 B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. section 6 C Did you run computational experiments? section 6 C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Section 6,
| 972 | 1,773 | 972 |
Does Recommend-Revise Produce Reliable Annotations? An Analysis on Missing Instances in DocRED
|
DocRED is a widely used dataset for documentlevel relation extraction. In the large-scale annotation, a recommend-revise scheme is adopted to reduce the workload. Within this scheme, annotators are provided with candidate relation instances from distant supervision, and they then manually supplement and remove relational facts based on the recommendations. However, when comparing Do-cRED with a subset relabeled from scratch, we find that this scheme results in a considerable amount of false negative samples and an obvious bias towards popular entities and relations. Furthermore, we observe that the models trained on DocRED have low recall on our relabeled dataset and inherit the same bias in the training data. Through the analysis of annotators' behaviors, we figure out the underlying reason for the problems above: the scheme actually discourages annotators from supplementing adequate instances in the revision phase. We appeal to future research to take into consideration the issues with the recommend-revise scheme when designing new models and annotation schemes. The relabeled dataset is released at
|
Relation Extraction (RE) is an important task which aims to identify relationships held between entities in a given piece of text. While most previous methods focus on extracting relations from a single sentence The rapid development of document-level RE in the past two years has benefited from the proposal of DocRED Shifting the construction process from scratch to an edit-based task, it seems that the recommendrevise scheme cuts down the effort of annotating by a large margin. However, whether the quality of the annotation maintains a reliable standard in practice remains in doubt. To what extent can the accuracy of annotation be sacrificed due to the automated recommendation? And, how does the provided recommendation affect the behaviours of the annotators in the revision phase? Moreover, what are the real effects on the models trained on a dataset annotated with this scheme? To answer these questions, we aim to provide a thorough comparison between careful annotations from scratch and the annotations under the recommend-revise scheme. We randomly select 96 documents from DocRED and ask two experts to relabel them from scratch independently. After annotating, the two experts come to a consensus of gold labels via discussion. This revised dataset is publicly available at (1) A noticeable portion of relation instances is left out, and the distributional bias in the recommendation output is inherited, even after the revision process. It is not surprising that recommendations alone fail to recognize all the relation instances, since RE models are far from perfect. Ideally, these unidentified instances should be added by human annotators during the revision phase. However, it turns out that 95.7% of these missing instances are still left out even after revision. Furthermore, while the recommendations from distant supervision favor instances associated with popular entities and relations in the source Knowledge Base (Wikidata), this bias is still maintained and inherited even after human revision, leaving less popular relations and entities to be neglected. (2) Worryingly, we find the models trained on DocRED have low recall on our relabeled dataset and they also inherit the same bias towards popular relations and entities. We train recent models on DocRED and test them with the dataset relabeled by us. We notice that all models have much lower recalls on our dataset than previously reported on DocRED due to the numerous false negatives in training data, and those models are also biased to popular entities and relations. Further investigation reveals that the models' bias comes from the training set by comparing different strategies of negative sampling. Since one straightforward real-world application of relation extraction is to acquire novel knowledge from text, a RE model would be much less useful if it has a low recall, or perform poorly on less popular entities and relations. (3) The recommendations actually also impacts the behaviors of annotators, making them unlikely to supplement the instances left out. This is the underlying reason for the two concerns above. We argue that the revision process fails to reach its goal, since it puts the annotators in a dilemma: while they are supposed to "add" new instances left out by the recommendations, finding these missing instances may force the annotators 1 While we cannot guarantee that the relabeled data is totally error-free, we believe the quality is high enough to be approximated as a real distribution because each entity pair is examined by two annotators. to thoroughly check out the entities pair-by-pair, which is time-consuming and against the goal of this scheme. As a result, annotators can hardly make effective supplementation and would tend to perform the easier goal of validating existing relation instances.
|
The major challenge for annotating document-level RE datasets comes from the quadratic number of potential entity pairs with regard to the total number of entities in a document. As reported by To alleviate the huge burden of manual labeling, 3 Dataset DocRED The Document-Level Relation Extraction Dataset (DocRED), introduced by To analyze the effect of the recommend-revise scheme, we re-annotate a subset of the documents used in DocRED from scratch and compare it with D Recommend and D Revise . We randomly select 96 documents from the validation set of DocRED, and each document is assigned to two experts to be annotated independently. They are explicitly required to check every entity pair in the documents and decide the relationships entirely based on the original text with no recommendation. This turns out to be an extraordinarily difficult task where each document takes up half an hour for annotation on average. The inter-annotator Cohen's Kappa is 0.68 between our two experts, indicating a high annotation quality. After that, the two experts discuss the inconsistent instances together and reach an agreement on the final labels. As this paper focuses on the bias caused by false negatives in the recommend-revise scheme, we assume the labeled instances in DocRED are all correct. For the instances labeled in DocRED but not by our experts, we add them to our annotation. We denote this new annotation set as D Scratch . Table Comparing D Recommend with D Scratch , it is noticeable that huge amounts of ground-truth annotation labels are left out. While D Recommend captures 1167 relation instances in the documents, a more careful, entity-by-entity examination as did in D Scratch would reveal that there are as much as 3308 relation instances within the same documents. This shocking fact reveals that almost two-thirds of the relation instances are missing and wrongly labeled as negative. Another unexpected fact is that annotators hardly added anything during the revision phase. The final version reports 1214 relation instances, with a mere increase of 47 (1.4%) cases in total, or 0.49 instances on average for each document. This suggests that while we had great hopes of our revision process to make things right, it is not working to a sensible extent: the majority of the unlabeled instances, which take up nearly two-thirds of the instances, simply remain out there as they were. Given the analysis above, another even more serious issue arises: since the changes introduced by the revision are so limited, the output after revision may still contain the same bias as in the recommendation. That is, if the recommendations contain a systematic flaw, the new dataset will probably keep on inheriting it. In this section, we verify that such biases largely exist in the recommendation phase and are thus inherited to the DocRED dataset. The recommendations of DocRED are collected from two sources: Wikidata and a relation extraction model. However, if we consider the facts reserved after revision by annotators, where wrongly labeled ones get removed, the majority of them are taken directly from Wikidata We suggest that as a collaborative knowledge base, the relation instances related to common entities and properties are more likely to be collected and added to Wikidata. In such cases, the recommendation from Wikidata will naturally favor popular entities and relations, while the less common ones would be left out. We validate this hypothesis in the following sections, where we investigate the bias of DocRED from the perspective of both relations and entities. To determine whether the data set has a preference for popular relationships, we divide the 96 relationships in DocRED into two categories using Wikidata statistics and then compute their distribution. Specifically, we acquire the List of top 100 properties by quantity of item pages that link to them from Wikidata's official website including country, publication date, and so on. The center two columns of Table We hypothesize that the instances involving very popular entities are more likely to appear in Wikidata recommendations, whereas instances related to extremely rare entities are more likely to be disregarded. To determine whether such bias exists, we analyze the popularity of entities engaged in relation instances across multiple data sets. Each named entity in DocRED is linked with a Wikidata item based on the literal matching of names or aliases For each relation instance, we compute two types of popularities. Since an instance contains a pair of entities (head and tail) usually with different popularities, we define popularity max to be the higher popularity of the pair of entities, and popularity min to be the lower one. We report the average popularity of relation instances in each dataset in Table Comparing D Recommend and D Scratch , we find that the former's popularity max is 294.4, far more than the latter's 266.3. This means that instances containing popular entities will be more likely to be retained during the recommendation phase. Regarding those instances that were incorrectly excluded during the recommendation phase, D Scratch -D Recommend , their popularity min is 57.7, which is less than the 67.4 in D Scratch . This demonstrates that instances involving uncommon entities are more likely to be ignored during the recommendation phase. This entity-related bias is apparent in the revised data set as well. The popularity max kept by D Revise remains larger than that of D Scratch , while the popularity min of D Scratch -D Revise is also lower than that of D Scratch . This is mostly because the facts supplemented at the revision phase is too few to eliminate such bias. To investigate if RE models trained on such data will likewise learn the same bias, we train and select RE models on the recommend-scheme-labeled dataset, D Train Revise and D Valid Revise and then assess the models' performance on the real data distribution, D Scratch . The construction process of D Train Revise and D Valid Revise is the same as D Revise , while the former is actually the original train set and the latter is the validation set in DocRED excluding the 96 documents in D Revise . In those settings, we examine the performance of recent models: BiLSTM Table Notably, we observe a significant decline in F1 for all the 5 models on D Scratch which is mainly due to the dramatic drop in the recall. The drop is the result of the bias in training data, i.e., the model trained on biased data lacks the generalization ability to extract relation instances that are systematically missed in the dataset. We will validate this point in the following section. To better understand the different performances on the two datasets, we analyze the model capability over different relations and entities. Not surprisingly, we find that models trained on D Train Revise prefer popular entities and relations as well. Addi- tional experiments suggest that this may be because missing instances are considered as negative samples during training. Given that a substantial proportion of unlabelled instances are associated with unpopular entities and relations, the model will be forced to disregard those unpopular ones under the incorrect penalty for the missing instances. Relation Bias Figure Figure We divide all instances in D Scratch into 5 groups based on the popularity max in each instance, and we calculate the recall for each group independently. As seen in Figure Previous works Figure Based on the entity popularity in each instance, we partition all instances in D Scratch into five cat- egories and calculate the recall for each group independently. Figure Finally, we move on to discuss another more implicit influence of the recommend-revise scheme on the annotators' aspect. As discussed in Section 4.1, while we expected the revision process to help supplement the instances left out, it turns out that an incredibly low number is added indeed. Given that the annotators are trained to accomplish the revision task, we wonder why they still fail in such a uniform manner. We would like to argue that it is the nature of the revision process that puts the annotators in a dilemma, where they have to choose between a huge effort and insufficiency of supplementation. Recall that there is a distinct difference in the settings of examining a labeled relationship and supplementing an unidentified relationship. For the former, annotators are required to find evidence for a recommended relation instance and remove it if there is conflicting or no evidence. This process only requires checking a single entity pair and collecting the information related to the two specific entities. However, this is not the case for supplementing a possible, unidentified relation instance, which can exist between any entity pair. There is NAll NHum no clear range of searching or indicating information; all they can do is to check pair-by-pair, just like what they do from scratch. This puts annotators in an awkward dilemma, especially when they understand the motivation of this scheme: if they are to be fully responsible for the missing instances at large, they will always have to complete the thorough pairwise checking one by one; however, this would make the whole process of the recommendrevise scheme meaningless in return, as it's just like a practice from scratch. The harsh requirements of supplementing push annotators to overly rely on the recommendation results and simply examine them. This is especially worth worrying about in real practice, where annotators are recruited to complete a certain number of annotations, and typically paid according to the estimated number of hours or the total number of instances they devote to the annotation In fact, we observe an interesting phenomenon that annotators largely tend to just pick some most obvious missing instances, convince themselves that they have accomplished the supplementation, and simply move on to the next document. This can be seen in Figure From these observations, we see that there exist clear patterns among the very few instances added by human annotators. This reveals a serious fact that annotators are intentional in "pretending" to be supplementing with the least possible effort. Given the consensus behavior of annotators and the very limited number of additional, it is most likely that the nature of the annotation task pushes the annotators to this embarrassing dilemma of adding and abandoning. Thus, we propose a call to the NLP community that researchers should always be aware that annotation schemes, like the recommend-revise scheme, can have a direct impact on the annotation workers, affecting their willingness and behaviors, and thus have a deeper influence on the collected data. We can summarize all these problems mentioned above in the annotation with a concrete case in Do- cRED shown in Figure [15] and ABC With the advance of deep learning models, the annotation sometimes becomes the bottleneck for a machine learning system. Recently, analyzing the annotation quality has received increasing attention. Methodologies to solve incomplete annotations for information extraction tasks have been widely discussed in previous works. Different from classification tasks, information extraction requires annotators to actively retrieve positive samples from texts, instead of just assigning a label for a given text. The problem is also attributed to the use of distant supervision Prevalent effective methods on document-level RE include graph-based models and transformer-based models. Graph-based models like In this paper, we show how the recommend-revise scheme for DocRED can cause bias and false negative issues in the annotated data. The flaws of dataset affect the model's recall on real data and also teach the model the same bias in training data. As this scheme cannot reduce the human labor essentially without the loss of annotation quality, more efficient strategies for annotation are to be explored. On the other hand, considering that building a reliable training set for document RE is extremely expensive, it is also a meaningful topic that how to alleviate the dataset shift problem According to
| 1,117 | 3,837 | 1,117 |
POS-Constrained Parallel Decoding for Non-autoregressive Generation
|
The multimodality problem has become a major challenge of existing non-autoregressive generation (NAG) systems. A common solution often resorts to sequence-level knowledge distillation by rebuilding the training dataset through autoregressive generation (hereinafter known as "teacher AG"). The success of such methods may largely depend on a latent assumption, i.e., the teacher AG is superior to the NAG model. However, in this work, we experimentally reveal that this assumption does not always hold for the text generation tasks like text summarization and story ending generation. To provide a feasible solution to the multimodality problem of NAG, we propose incorporating linguistic structure (Part-of-Speech sequence in particular) into NAG inference instead of relying on teacher AG. More specifically, the proposed POS-constrained Parallel Decoding (POSPD) method aims at providing a specific POS sequence to constrain the NAG model during decoding. Our experiments demonstrate that POSPD consistently improves NAG models on four text generation tasks to a greater extent compared to knowledge distillation. This observation validates the necessity of exploring the alternatives for sequence-level knowledge distillation.
|
Unlike autoregressive generation (AG) that generates tokens step-by-step, non-autoregressive generation (NAG) parallelly generates all tokens in one time step and thus the inference could be significantly speeded up To date, one of the most widely used solutions is sequence-level knowledge distillation (2) The knowledge distillation based methods may cost a tremendous amount of time to rebuild a largescale training set with AG, which runs counter to the initial goal of NAG to improve the speed. To overcome the aforementioned limitations, we explore to alleviate the multimodality problem in a different manner. In short, we aim to constrain NAG generation modes in the inference stage, rather than directly reducing generation modes in the training stage. More specifically, our basic idea is that the linguistic structure of the target sentence could be helpful to alleviate the multimodality problem. In this paper, we show that the Part-of-Speech (POS) sequence, one of most simple solutions in modeling the linguistic structure To conduct a comprehensive empirical evaluation, we examine the generalizability of POSPD by applying it to two widely-used NAG models (i.e., CMLM and DisCo) over four text generation tasks, including text summarization, story ending generation, question generation, and machine translation. Experiments demonstrate that POSPD significantly and consistently improves the two NAG models and beats the sequence-level knowledge distillation with a considerable performance gap. The main contributions of this work could be summarized as follows: • For the first time, we experimentally reveal that the implicit assumption of knowledge distillation does not always hold for the tasks (e.g., text summarization, story ending generation, as demonstrated in our experiments). In other words, AG cannot guarantee better performance than NAG, thus resulting in the undesirable performance of NAG if using knowledge distillation to alleviate the multimodality problem. This empirical result could provide novel insight to revisiting the role of the knowledge distillation in NAG. • To alleviate the multimodality problem in various tasks, we propose POSPD by employing POS sequences to constrain the NAG generation modes in the inference stage. It is simple but effective, being able to act as a plugand-play assistant for NAG models. Such a linguistic structure based solution shows an effective and efficient alternative to the knowledge distillation paradigm in alleviating the multimodality problem
|
In this section, we first analyze related works on alleviating the multimodality problem. Then, we review some representative works which introduce the linguistic structure into some text generation scenarios. Recently, various attempts have been made to alleviate the multimodality problem, which can be roughly divided into two types: (1) Reducing the diversity of generation modes in training; (2) Helping models select one generation mode in inference. The first type usually trains the NAG model under the guidance of an AG model (called teacher AG), e.g., sequence-level knowledge distillation Text generation involves multiple tasks, such as style transfer In comparison, POSPD can act as a plug-and-play component that uses a separate POS predictor to constrain NAG models during inference. Therefore, the NAG model can enjoy the benefits of the syntactical structure constraining while retaining its original model structure. In this section, we elaborate our POSPD for the NAG model. To ease of presentation, we start from a toy example to illustrate the overview of POSPD in § 3.1, and then give a detailed explanation of the implementation in § 3.2. After that, we present the training details of POSPD in § 3.3. An overview of our POSPD method is demonstrated in Figure In this part, we detail the POSPD by introducing the conversion dictionary building, the workflow of POSPD, and the core module-the POS predictor. The key idea of POSPD is filtering out the words that dissatisfy the prespecified POS sequence in the primary results of NAG. To implement our idea, we need a conversion dictionary D c that contains the mapping from POS tags to words. Note that, the sentence length follows the length of the predicted POS tag L. For the ease of implementation, the POS sequence s is converted into a binarized mask matrix M = (m 1 , m 2 , ..., m L ). In details, for each POS tag s i , the corresponding binarized vector is ) and the j-th position m j i is defined as: where w j is the j-th word token in V w . As a result, the POS sequence s is replaced by M. Finally, we get the new generation results by: The POS Predictor As the core module of the POSPD, our POS predictor is dedicated to output the POS tag sequence of the target sentence when accepting the source sentence as the input. To train the POS predictor, we need to create a POS dataset where each sample is a pair consisting of a source sentence and a POS sequence of the target 2 The length of d i is |Vw|. sentence where the POS sentence prediction loss can be written as: and the target sentence prediction loss is: In our method, the POS predictor uses an extra linear layer after the decoder to generate the target sentence, as shown in Figure Almost all NAG models use the Byte Pair Encoding (BPE) In this section, we use multiple text generation datasets to comprehensively evaluate the effectiveness and efficiency of the proposed POSPD. For an extensive comparison, we compare our POSPD with the sequence-level knowledge distillation, and provide detailed analyzes in alleviating the multimodality problem and the time cost in dataset building. We Follow GLGE In this work, we focus on using iteration-based NAG models as backbones, because they are one of the mainstream NAG structures in current works and perform competitively to AG models without any external system CMLM The conditional masked language model randomly masks some target tokens and predicts them with the remaining ones. In inference, it masks several tokens with the lower "confidence" and retains other tokens with higher "confidence" during iterations, which is called mask-predict inference. Following Meanwhile, we use the model settings described in We follow the hyperparameters for standard Transformer in We evaluate the performance of two NAG models (CMLM and DisCo) on four text generation datasets, and further provide the results when using sequence-level data distillation (i.e., "+Distill") and the POSPD (i.e., "+POSPD"), respectively. We report the main results in Table There is a loose ending towards the discussion of our POSPD solution. In this section, we conduct discussions to shed light on other interesting properties of POSPD. The discussions are guided by the following three research questions: Q1: How does POSPD alleviate the multimodality problem? Q2: Is it time-consuming to build the POS dataset on the new task? Q3: Does multi-tasking learning object help the POS tag prediction? 4.6.1 Discussion on Generated Results (Q1) To further analyze the role of POSPD and the sequence-level knowledge distillation in alleviating the multimodality problem, we conduct further statistical analyses on the generated results of four datasets. Considering the multimodality problem usually manifests as repeating or missing tokens in the generation sentences, we use two indicators, i.e., the repetition rate and the total number of tokens, to quantify them separately. Concretely, we refer to a "single-token repeat" metric The results are shown in Table For lack of space, we take the ablation study on two datasets of different sizes, i.e., SQuAD 1.1 and XSUM. The results are shown in Table In this paper, we revisit the role of the knowledge distillation in alleviating the multimodality problem of NAG. In brief, we experimentally reflect that the basic assumption of these knowledge distillation methods, the AG model is superior to NAG model, does not always hold for all text generation tasks. To alleviate the multimodality problem, we show a different solution by incorporating linguistic structure into NAG. Extensive experiments demonstrate that our POSPD significantly and consistently improves the NAG models in effectiveness and computational efficacy. As we tentatively give a successful implementation of leveraging one of the simplest linguistic structures to benefit the NAG models in inference, such paradigm deserves a closer and more detailed exploration. Thus in the future, we will investigate to make the NAG models enjoy the benefits of incorporating diverse and abundant linguistic structures in a more superior way. In addition, our experimental results suggest that future work might need to consider wider ranges of generation tasks instead of only machine translation when assessing the performance of NAG models.
| 1,231 | 2,530 | 1,231 |
CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies
|
Every year, the Conference on Computational Natural Language Learning (CoNLL) features a shared task, in which participants train and test their learning systems on the same data sets. In 2018, one of two tasks was devoted to learning dependency parsers for a large number of languages, in a real-world setting without any gold-standard annotation on the input. All test sets followed the unified annotation scheme of Universal Dependencies
|
The 2017 CoNLL shared task on universal dependency parsing Like last year, participating systems minimally had to find labeled syntactic dependencies between words, i.e., a syntactic head for each word, and a label classifying the type of the dependency relation. In addition, this year's task featured new metrics that also scored a system's capacity to predict a morphological analysis of each word, including a part-of-speech tag, morphological features, and a lemma. Regardless of metric, the assumption was that the input should be raw text, with no gold-standard word or sentence segmentation, and no gold-standard morphological annotation. However, for teams who wanted to concentrate on one or more subtasks, segmentation and morphology predicted by the baseline UDPipe system There are eight new languages this year: Afrikaans, Armenian, Breton, Faroese, Naija, Old French, Serbian, and Thai; see Section 2 for more details. The two new evaluation metrics are described in Section 3. In general, we wanted the participating systems to be able to use any data that is available free of charge for research and educational purposes (so that follow-up research is not obstructed). We deliberately did not place upper bounds on data sizes (in contrast to e.g. In practice, the task was formally closed, i.e., we listed the approved data resources so that all participants were aware of their options. However, the selection was rather broad, ranging from Wikipedia dumps over the OPUS parallel corpora We provided dependency-annotated training and test data, and also large quantities of crawled raw texts. Other language resources are available from third-party servers and we only referred to the respective download sites.
|
Training and development data came from the Universal Dependencies (UD) 2.2 collection 82 UD treebanks in 57 languages were included in the shared task; 2 however, nine of the smaller treebanks consisted solely of test data, with no data at all or just a few sentences available for training. 16 languages had two or more treebanks from different sources, often also from different domains. 61 treebanks contain designated development data. Participants were asked not to use it for training proper but only for evaluation, development, tuning hyperparameters, doing error analysis etc. Seven treebanks have reasonablysized training data but no development data; only two of them, Irish and North Sámi, are the sole treebanks of their respective languages. For those treebanks cross-validation had to be used during development, but the entire dataset could be used for training once hyperparameters were determined. Five treebanks consist of extra test sets: they have no training or development data of their own, but large training data exist in other treebanks of the same languages (Czech-PUD, English-PUD, Finnish-PUD, Japanese-Modern and Swedish-PUD, respectively). The remaining nine treebanks are low-resource languages. Their "training data" was either a tiny sample of a few dozen sentences (Armenian, Buryat, Kazakh, Kurmanji, Upper Sorbian), or there was no training data at all (Breton, Faroese, Naija, Thai). Unlike in the 2017 task, these languages were not "surprise languages", that is, the participants knew well in advance what languages to expect. The last two languages are particularly difficult: Naija is a pidgin spoken in Nigeria; while it can be expected to bear some similarity to English, its spelling is significantly different from standard English, and no resources were available to learn it. Even harder was Thai with a writing system that does not separate words by spaces; the Facebook word vectors were probably the only resource among the approved additional data where participants could learn something about words in Thai Cross-domain and cross-language training was allowed and encouraged. Participants were free to train models on any combination of the training treebanks and apply it to any test set. To enable the induction of custom embeddings and the use of semi-supervised methods in general, the participants were provided with supporting resources primarily consisting of large text corpora for many languages in the task, as well as embeddings pre-trained on these corpora. In total, 5.9 M sentences and 90 G words in 45 languages are available in CoNLL-U format See There are three main evaluation scores, dubbed LAS, MLAS and BLEX. All three metrics reflect word segmentation and relations between content words. LAS is identical to the main metric of the 2017 task, allowing for easy comparison; the other two metrics include part-of-speech tags, morphological features and lemmas. Participants who wanted to decrease task complexity could concentrate on improvements in just one metric; however, all systems were evaluated with all three metrics, and participants were strongly encouraged to output all relevant annotation, even if they just copy values predicted by the baseline model. When parsers are applied to raw text, the metric must be adjusted to the possibility that the number of nodes in gold-standard annotation and in the system output vary. Therefore, the evaluation starts with aligning system nodes and gold nodes. A dependency relation cannot be counted as correct if one of the nodes could not be aligned to a gold node. See Section 3.4 and onward for more details on alignment. The evaluation software is a Python script that computes the three main metrics and a number of additional statistics. It is freely available for download from the shared task website. The standard evaluation metric of dependency parsing is the labeled attachment score (LAS), i.e., the percentage of nodes with correctly assigned reference to the parent node, including the label (type) of the relation. For scoring purposes, only In the end-to-end evaluation of our task, LAS is re-defined as the harmonic mean (F 1 ) of precision P and recall R, where Note that attachment of all nodes including punctuation is evaluated. LAS is computed separately for each of the 82 test files and a macro-average of all these scores is used to rank the systems. 3.2 MLAS: Morphology-Aware Labeled Attachment Score MLAS aims at cross-linguistic comparability of the scores. It is an extension of CLAS The idea behind MLAS is that function words often correspond to morphological features in other languages. Furthermore, languages with many function words (e.g., English) have longer sentences than morphologically rich languages (e.g., Finnish), hence a single error in Finnish costs the parser significantly more than an error in English according to LAS. The core part is identical to LAS (Section 3.1): for aligned system and gold nodes, their respective parent nodes are considered; if the system parent is not aligned with the gold parent, or if the universal relation label differs, the word is not counted as correctly attached. Unlike LAS, certain types of relations (Table BLEX is similar to MLAS in that it focuses on relations between content words. Instead of morphological features, it incorporates lemmatization in the evaluation. It is thus closer to semantic content and evaluates two aspects of UD annota-tion that are important for language understanding: dependencies and lexemes. The inclusion of this metric should motivate the competing teams to model lemmas, the last important piece of annotation that is not captured by the other metrics. A system that scores high in all three metrics will thus be a general-purpose language-analysis tool that tackles segmentation, morphology and surface syntax. Computation of BLEX is analogous to LAS and MLAS. Precision and recall of correct attachments is calculated, attachment of function words and punctuation is ignored (Table A few UD treebanks lack lemmatization (or, as in Uyghur, have lemmas only for some words and not for others). A system may still be able to predict the lemmas if it learns them in other treebanks. Such system should not be penalized just because no gold standard is available; therefore, if the gold lemma is a single underscore character (" "), any system-produced lemma is considered correct. UD defines two levels of token/word segmentation. The lower level corresponds to what is usually understood as tokenization. However, unlike some popular tokenization schemes, it does not include any normalization of the non-whitespace characters. We can safely assume that any two tokenizations of a text differ only in whitespace while the remaining characters are identical. There is thus a 1-1 mapping between gold and system nonwhitespace characters, and two tokens are aligned if all their characters match. The higher segmentation level is based on the notion of syntactic word. Some languages contain multi-word tokens (MWT) that are regarded as contractions of multiple syntactic words. For example, the German token zum is a contraction of the preposition zu "to" and the article dem "the". Syntactic words constitute independent nodes in dependency trees. As shown by the example, it is not required that the MWT is a pure concatenation of the participating words; the simple token alignment thus does not work when MWTs are involved. Fortunately, the CoNLL-U file format used in UD clearly marks all MWTs so we can detect them both in system output and in gold data. Whenever one or more MWTs have overlapping spans of surface character offsets, the longest common subsequence algorithm is used to align syntactic words within these spans. Words are aligned and dependencies are evaluated in the entire file without considering sentence segmentation. Still, the accuracy of sentence boundaries has an indirect impact on attachment scores: any missing or extra sentence boundary necessarily makes one or more dependency relations incorrect. If a system fails to produce one of the 82 files or if the file is not valid CoNLL-U format, the score of that file (counting towards the system's macroaverage) is zero. Formal validity is defined more leniently than for UD-released treebanks. For example, a nonexistent dependency type does not render the whole file invalid, it only costs the system one incorrect relation. However, cycles and multi-root sentences are disallowed. A file is also invalid if there are character mismatches that could make the token-alignment algorithm fail. The metrics described above are all intrinsic measures: they evaluate the grammatical analysis task per se, with the hope that better scores correspond to output that is more useful for downstream NLP applications. Nevertheless, such correlations are not automatically granted. We thus seek to complement our task with an extrinsic evaluation, where the output of parsing systems is exploited by applications like biological event extraction, opinion analysis and negation scope resolution. This optional track involves English only. It is organized in collaboration with the EPE initiative; Similarly to our 2017 task and to some other recent CoNLL shared tasks, we employed the cloud-based evaluation platform TIRA Traditionally, evaluations in shared tasks are halfblind (the test data are shared with participants while the ground truth is withheld). TIRA enables fully blind evaluation, where the software is locked in a datalock together with the test data, its output is recorded but all communication channels to the outside are closed or tightly moderated. The participants do not even see the input to their software. This feature of TIRA was not too important in the present task, as UD data is not secret, and the participants were simply trusted that they would not exploit any knowledge of the test data they might have access to. However, closing down all communication channels also has its downsides, since participants cannot check their running software; before the system run completes, even the task moderator does not see whether the system is really producing output and not just sitting in an endless loop. In order to alleviate this extra burden, we made two modifications compared to the previous year: 1. Participants were explicitly advised to invoke shorter runs that process only a subset of the test files. The organizers would then stitch the partial runs into one set of results. 2. Participants were able to see their scores on the test set rounded to the nearest multiple of 5%. This way they could spot anomalies possibly caused by illselected models. The exact scores remained hidden because we did not want the participants to fine-tune their systems against the test data. It is desirable that published experiments can be re-run yielding the same results, and that the algorithms can be tested on alternative test data in the future. Ensuring both requires that a to-beevaluated software is preserved in working con-dition for as long as possible. TIRA supplies participants with a virtual machine, offering a range of commonly used operating systems. Once deployed and tested, the virtual machines are archived to preserve the software within. In addition, some participants agreed to share their code so that we decided to collect the respective projects in an open source repository hosted on GitHub. We prepared a set of baseline models using UD-Pipe 1.2 The baseline models were released together with the UD 2. In addition to the treebank-specific models, we also trained a "mixed model" on samples from all treebanks. Specifically, we utilized the first 200 training sentences of each treebank (or less in case of small treebanks) as training data, and at most 20 sentences from each treebank's development set as development data. The baseline models, together with all information needed to replicate them (hyperparameters, the modified train-dev split where applicable, and pre-computed word embeddings for the parser) are available from Additionally, the released archive also contains the training and development data with predicted morphology. Morphology in development data was predicted using the baseline models, morphology in training data via "jack-knifing" (split the training set into 10 parts, train a model on 9 parts, use it to predict morphology in the tenth part, repeat for all 10 target parts). The same hyperparameters were used as those used to train the baseline model on the entire training set. The UDPipe baseline models are able to reconstruct nearly all annotation from CoNLL-U files -they can generate segmentation, tokenization, multi-word token splitting, morphological annotation (lemmas, UPOS, XPOS and FEATS) and dependency trees. Participants were free to use any part of the model in their systems -for all test sets, we provided UDPipe processed variants in addition to raw text inputs. Baseline UDPipe Shared Task System The shared task baseline system employs the UDPipe 1.2 baseline models. For the nine treebanks without their own training data, a substitution model according to Table 6 Results Table The ranking of the baseline system by MLAS is similar (Table Each ranking has a different winning system, although the other two winners are typically closely following. The same 8-10 systems occupy best positions in all three tables, though with variable mutual ranking. Some teams seem to have deliberately neglected some of the evaluated attributes: Uppsala is rank 7 in LAS and MLAS, but 24 in Team LAS 1. HIT-SCIR We used bootstrap resampling to compute 95% confidence intervals: they are in the range ±0.11 to ±0.16 (% LAS/MLAS/BLEX) for all systems except SParse (where it is ±0.00). We used paired bootstrap resampling to compute whether the difference between two neighboring systems is significant (p < 0.05). In addition to the main LAS ranking, we evaluated the systems along multiple other axes, which may shed more light on their strengths and weaknesses. This section provides an overview of selected secondary metrics for systems matching or surpassing the baseline; a large number of additional results are available at the shared task website. As an experiment, we also applied the 2017 system submissions to the 2018 test data. This allows us to test how many systems can actually be used to produce new data without a glitch, as well as to see to what extent the results change over one year and two releases of UD. Here it should be noted that not all of the 2018 task languages and treebanks were present in the 2017 task, therefore causing many systems fail due to an unknown language or treebank code. The full results of this experiment are available on the shared task website. The complexity of the new metrics, especially MLAS, is further underlined by Table Table Table In contrast, the results on the 7 treebanks with "small" training data and no development data (Table Table See Tables 15, 16 and 17 for a ranking of all test sets by the best scores achieved on them by any parser. Note that this cannot be directly interpreted as a ranking of languages by their parsing difficulty: many treebanks have high ranks simply because the corresponding training data is large. Table Finally, Tables Table When it comes to morphological analysis, including universal POS tags, features and lemmas, all systems this year include some such component, and only 6 systems rely entirely on the base- In all columns, Base (or B) refers to the Baseline UDPipe system or the baseline word embeddings provided by the organizers, while None means that there is no corresponding component in the system. line UDPipe system. This is again quite different from 2017, where more than half the systems either just relied on the baseline tagger (13 systems) or did not predict any morphology at all (3 systems). We take this to be primarily a reflection of the fact that two out of three official metrics included (some) morphological analysis this year, although 3 systems did not predict the lemmas required for the BLEX metric (and 2 systems only predicted universal POS tags, no features). As far as we can tell from the questionnaire responses, only 3 systems used a model where morphology and syntax were predicted jointly. Finally, we note that a majority of systems make use of models trained on multiple languages to improve parsing for languages with little or no training data. According to the questionnaire responses, 15 systems use multilingual models for the languages classified as "low-resource", while 7 systems use them for the languages classified as "small". The CoNLL 2018 Shared Task on UD parsing, the second in the series, was novel in several respects. Besides using cross-linguistically consistent linguistic representations, emphasizing end-to-end processing of text, and in using a multiply parallel test set, as in 2017, it was unusual also in featuring an unprecedented number of languages and treebanks and in integrating cross-lingual learning for resource-poor languages. Compared to the first edition of the task in 2017, this year several languages were provided with little-to-no resources, whereas in 2017, predicted morphology trained on the language in question was available for all of the languages. The most extreme example of these is Thai, where the only accessible resource was the Facebook Research Thai embeddings model and the OPUS parallel corpora. This year's task also introduced two additional metrics that take into account morphology and lemmatization. This encouraged the development of truly end-to-end full parsers, producing complete parses including morphological features and lemmas in addition to the syntactic tree. This also aimed to improve the utility of the systems developed in the shared task for later downstream applications. For most UD languages, these parsers represent a new state of the art for end-to-end dependency parsing. The analysis of the shared task results has so far only scratched the surface, and we refer to the system description papers for more in-depth analysis of individual systems and their performance. For many previous CoNLL shared tasks, the task itself has only been the starting point of a long and fruitful research strand, enabled by the resources created for the task. We hope and believe that the 2017 and 2018 UD parsing tasks will join this tradition.
| 440 | 1,730 | 440 |
Investigating label suggestions for opinion mining in German Covid-19 social media
|
This work investigates the use of interactively updated label suggestions to improve upon the efficiency of gathering annotations on the task of opinion mining in German Covid-19 social media data. We develop guidelines to conduct a controlled annotation study with social science students and find that suggestions from a model trained on a small, expert-annotated dataset already lead to a substantial improvement -in terms of inter-annotator agreement (+.14 Fleiss' κ) and annotation quality -compared to students that do not receive any label suggestions. We further find that label suggestions from interactively trained models do not lead to an improvement over suggestions from a static model. Nonetheless, our analysis of suggestion bias shows that annotators remain capable of reflecting upon the suggested label in general. Finally, we confirm the quality of the annotated data in transfer learning experiments between different annotator groups.
|
The impact analysis of major events like the Covid-19 pandemic is fundamental to research in social sciences. To enable more socially sensitive public decision making, researchers need to reliably monitor how various social groups (e.g., political actors, news media, citizens) communicate about political decisions One important approach to make data annotation more efficient is the use of automated label suggestions. In contrast to active learning, that aims to identify a subset of annotated data which leads to optimal model training, label suggestions alleviate the annotation process by providing annotators with pre-annotations (i.e., predictions) from a model One major difficulty with label suggestions is the danger of biasing annotators towards (possibly erroneous) suggestions. So far, researchers have investigated automated label suggestions for tasks that require domain-specific knowledge We tackle these challenges by conducting a comparative annotation study with social science students using a recent state-of-the-art model to generate label suggestions C1: An evaluation of label suggestions in terms of annotation quality for non-expert annotators. C2: An investigation of label suggestion bias for both static and interactively updated suggestions.
|
A novel corpus of German Twitter posts that can be used by social science researchers to study the effects of governmental measures against Covid-19 on the public opinion. Finally, we also publish 200 expert and 2,785 individual student annotations of our dataset to facilitate further research in this direction. Label suggestions. In an early work, Bias. Annotations from untrained human annotators may introduce biases that are conveyed to machine learning models Whereas existing work reports no measurable bias for expert annotators Our task is inspired by social science research on analyzing public opinion using social media Data collection. Initially, we collected tweets from December 2019 to the end of April 2020. Using a manually chosen set of search queries ('corona', 'pandemie', 'covid', 'socialdistance'), we made use of the Twitter Streaming API and gathered only those tweets which were classified as German by the Twitter language identifier. This resulted in a set of approximately 16.5 million tweets. We retained only tweets that contain key terms referring to measures related to the Covid-19 pandemic and removed all duplicates, retweets and all tweets with text length less than 30 characters. After filtering, 237,616 tweets remained and their daily temporal distribution is visualized in Figure Unrelated: no measures related to the containment of the pandemic are mentioned Comment: measures are mentioned, but not assessed or neutral Support: measures are assessed positively The four label annotation scheme allows us to distinguish texts that are related to the pandemic but do not talk about measures (i.e., unrelated). Our goal is to study the effects of interactively updated and static label suggestions in non-expert annotation scenarios. Non-experts such as crowd workers or student volunteers have no prior experience in annotating comparable tasks and only receive annotation guidelines for preparation. To train a model that provides label suggestions to our non-expert annotators, we first collect a small set of 200 expert-annotated instances. We then split our non-expert annotators into three different groups that receive (G1) no label suggestions, (G2) suggestions from a model trained on expert annotations, and (G3) suggestions from a model that is retrained interactively using both expert-annotated and interactively annotated data. The expert annotations were provided by the researchers (three social science researchers and one NLP researcher) that created the annotation guidelines and who are proficient in solving the task. In total, 200 tweets were sampled uniformly at random and annotated by all four experts. The interannotator agreement (IAA) across all 200 tweets lies at 0.54 Fleiss's κ (moderate agreement) and is comparable to previously reported annotation scores in the field of opinion and argument mining The disagreement on the remaining ∼25% of the tweets furthermore shows the increased difficulty of our task due to ambiguities in the data source, e.g., ironical statements or differentiating governmental measures from non-governmental ones like home-office. To compile gold standard labels for instances that the experts disagreed upon, we apply MACE The resulting labels were then re-evaluated by the experts and agreed upon. The annotations were conducted with a group of 21 German-speaking university students. To ensure a basic level of comparability for our student annotators, we recruited all volunteers from the same social science course at the same university. The annotators received no further training apart from the annotation guidelines. We randomly assigned them to three different groups (G1, G2, and G3), each consisting of seven students. To investigate the effects of interactive label suggestions, we defined different annotation setups for each group. The annotations were split into two rounds. At each round of annotation, students were provided with 100 tweets consisting of 70 new tweets and 30 quality control tweets from the expert-labeled data which are used to compare individual groups. Across both rounds, we thus obtain a total of 140 unique annotated tweets per student and use 60 tweets for evaluation. The annotation setup of each group including the individual data splits is visualized in Figure Static label suggestions (G2). The second group only receives label suggestions based on a model which was trained using the 200 expertlabeled instances described in section 4.1. Interactive label suggestions (G3). The last group of students receives expert label suggestions in the first round and interactively updated label suggestions in the second round. In contrast to existing work System setup. We conduct our annotation experiments using INCEpTION Macro-F1 Accuracy Majority .15 .45 Random .23 .27 BiLSTM .66 .68 Table Model comparison. To assess the label suggestion quality of our model, we report the predictive performance on the expert-labeled dataset (setup as described above) in Table Interactive training routine. To remedy the cold-start problem, G3 receives label suggestions from the model trained only on the expertannotated data in round 1. Afterwards, we retrain the model with an increasing number of instances using both, the expert annotations and the G3 data of individual students from round 1. Table Outliers. A fine-grained analysis of annotation time is not possible due to online annotations at home. However, one student in G3 had, on average, spent less than a second for each annotation and accepted almost all suggested labels. This student's annotations were removed from the final dataset and assumed as faulty labels considering the short amount of time spent on this task in comparison to the minimum amount of seven seconds per tweet and annotation for all other students. To assess the overall quality of our collected student annotations, we investigate annotator consistency in terms of inter-annotator-agreement (IAA) as well as the annotator accuracy on our quality assurance instances. Table One major challenge in using label suggestions is known in psychology as the anchoring effect (Tversky and Acceptance rate. One way to quantify possible biases is to evaluate if annotators tend to accept more suggestions with an increasing number of instances Stance. In Figure The high number of label corrections for specific types of tweets shows that our annotators of G2 remained critical towards the suggested label. With interactively updated suggestions however, this may not be the case. Especially annotators that accept erroneous suggestions may lead to reinforc-ing a model in its prediction; hence, leading to amplifying biases. Diverging suggestions. To study such effects, we first identify if the interactively updated models express a difference in terms of predictions compared to the static model. In Figure Individual acceptance rate. To assess the impact of interactive label suggestions, we further investigate how many suggestions were accepted by each annotator. Figure Finally, we investigate how well models trained on different annotator groups transfer to each other. We hence conduct transfer learning experiments for which we remove the quality control instances in our student groups and train a separate Ger-BERT model using the same hyperparameters as for the expert model. We use 80% of the data for training and the remaining 20% to identify the best model which we then transfer to another group. Figure In this work, we analysed the usefulness of providing label suggestions for untrained annotators to identify opinions in a challenging text domain (i.e., Twitter). We generated suggestions using expert-labeled training data as well as interactively training models using data annotated by untrained students. Our results show that label suggestions from a state-of-the-art sentence classification model trained on a small set of expert annotations help improving annotation quality for untrained annotators. In terms of potential biases that may occur with untrained annotators we observe that the students retained their capability to reflect on the suggested label. We furthermore do not observe a general amplification in terms of bias with interactively updated suggestions; however, we find that such effects are very specific to individual annotators. We hence conclude that interactively updated label suggestions need to be considered carefully when applied to non-expert annotation scenarios. For future work, we plan to leverage our setup to annotate tweets from a larger time span. In Germany, the measures taken by the government have been met with divided public reaction -starting with reactions of solidarity and changing towards a more critical public opinion Data collection and annotation. The tools we use to collect Tweets are in compliance with Twitter's terms of service. We only release the set of identifiers (Tweet IDs) for the texts used in this research project. Thereby, we adhere to the Twitter Developer policy We asked student annotators for voluntary participation in the annotation study. All students have been informed about the goal of the conducted research and the purpose of the collected annotations. During annotation no information about the tweet's author or any other additional metadata was made available to the annotators. We did not collect any personal data from the students before, after, or during the annotation task. Data usage. This work presents an investigation of efficient data annotation methods in a case study on social media data. The results of this work allow social science researchers to apply their analysis on a larger scale. In the case of analyzing public opinion on governmental measures, the resulting analysis allows politicians to make more socially sensitive public decisions. This information is useful in aggregated form, without the need for information about individual users. However, we want to point out that users of social media (particularly Twitter) do not constitute a representative sample of the general population, especially in Germany Dual use. Further, we acknowledge the potential of misuse of our dataset: the annotated data allows anyone, including both individuals and organizations, for training models to identify individuals expressing their consent or dissent with governmental actions. To this end, we follow the argumentation by (NoOpinion). We provide examples for each label to our annotators. During our preliminary studies, we identified several ambiguities regarding the stance annotation which is in the nature of the source of the texts (Twitter) and the subject of the annotation (measures regarding the Covid-19 pandemic): • a Tweet discusses (positive/negative) consequences or by-products of measures : we regard those as (positive/negative) statements as the author implicitly states their opinion by reflecting upon the measures • a Tweet reflects the opinion of another actor: this is considered as an opinion as defined above. It is assumed that the author posts this opinion because they identify themselves with the original opinion. • a Tweet makes an unagitated observation whether measures are functioning: this is not to be taken as an opinion for or against the measures per se. Only if an explicit assessment of the observation is made, the position can be derived. • the role of Hashtags: Hashtags are often ambiguous and the respective context needs to be taken into account. Therefore, in our annotation hashtags are only considered as context to what is said; they never stand for themselves. Hashtags can be used to determine whether a measure is being addressed. To do this, the hashtag must contain a measure. Further, hashtags can be used as context to support the position in a tweet. These decisions are reflected at the corresponding positions in the annotation guidelines, along with several example tweets. In the end we provide a note that Twitter posts may contain malicious, suggestive, offensive, or potentially sensitive content and that the annotation can be paused and resumed at any time. In Figure The label recommendations are displayed using a green box above the corresponding text and the currently selected recommendation is highlighted in orange. If the user agrees with the provided label, nothing needs to be changed. In the opposite case, the user can click on the recommendation and select another label on the right-hand side (Annotation panel) using the Opinion dropdown field. The annotators receiving no label suggestions (G1) do not see any recommendation during annotation. They create an annotation for each sentence by double-clicking on the sentence. Once the user has finished annotating all samples, the annotation session is finished by clicking the lock symbol in the navigation bar. The technical procedure of the annotation has been explained to annotators beforehand. We used the german-bert-cased BERT base model which was pretrained on a German Wikipedia Dump (6GB), an OpenLegalData dump (2.4GB) and news articles (3.6GB). It was trained for 810k steps with a batch size of 1024 for sequence length 128 and 30k steps with sequence length 512. It outperformed the multilingual version of BERT on several downstream tasks using German data (GermEval-2018 12 , GermEval-2014 NER 13 , 10kGNAD 14 ). More information can be found at the corresponding website 15 . For our setup, we performed a random hyperparameter search using the following combinations: • learning rate: [0.01, 0.1, 0.001, 0.0001, C Label Rejections Figure
| 956 | 1,273 | 956 |
A Societal Sentiment Analysis: Predicting the Values and Ethics of Individuals by Analysing Social Media Content
|
To find out how users' social media behaviour and language are related to their ethical practices, the paper investigates applying Schwartz' psycholinguistic model of societal sentiment to social media text. The analysis is based on corpora collected from user essays as well as social media (Facebook and Twitter). Several experiments were carried out on the corpora to classify the ethical values of users, incorporating Linguistic Inquiry Word Count analysis, n-grams, topic models, psycholinguistic lexica, speech-acts, and nonlinguistic information, while applying a range of machine learners (Support Vector Machines, Logistic Regression, and Random Forests) to identify the best linguistic and non-linguistic features for automatic classification of values and ethics.
|
In the recent years, there have been significant efforts on determining the opinion/sentiment/emotion about a specific topic held by the author of a piece of text, and on automatic sentiment strength analysis of text, classifying it into either one of the classes positive, negative or neutral, or into Ekman's classes of happy, sad, anger, fear, surprise, and disgust. However, the intrinsic value of the lives we lead reflects the strength of our values and ethics which guide our social practices, attitude and behaviour. This paper reports work on investigating a psycholinguistic model, the Schwartz model The Schwartz values model defines ten distinct ethical values (henceforth only values), that respectively are: Achievement sets goals and achieves them; Benevolence seeks to help others and provide general welfare; Conformity obeys clear rules, laws and structures; Hedonism seeks pleasure and enjoyment; Power controls and dominates others, controls resources; Security seeks health and safety; Self-direction wants to be free and independent; Stimulation seeks excitement and thrills; Tradition does things blindly because they are customary; Universalism seeks peace, social justice and tolerance for all. Deeper understanding of human beliefs, attitudes, ethics, and values has been a key research agenda in Psychology and Social Science research for several decades. One of the most accepted and widely used frameworks is Schwartz 10-Values model, has seen great success in psychological research as well as in other fields. The ten basic values are related to various outcomes and ef-fects of a person's role in a society We believe that these kind of models may become extremely useful in the future for various purposes like Internet advertising (specifically social media advertising), community detection, computational psychology, recommendation systems, sociological analysis (for example East vs West cultural analysis) over social media. In order to experiment with this, three corpora have been collected and annotated with Schwartz values. Two of the corpora come from popular social media platforms, Facebook and Twitter, while the third corpus consists of essays. A range of machine learning techniques has then been utilized to classify an individual's ethical practices into Schwartz' classes by analyzing the user's language usage and behaviour in social media. In addition to identifying the ten basic values, Schwartz' theory also explains how the values are interconnected and influence each other, since the pursuit of any of the values results in either an accordance with one another (e.g., Conformity and Security) or a conflict with at least one other value (e.g., Benevolence and Power). The borders between the motivators are artificial and one value flows into another. Such overlapping and fuzzy borders between values make the computational classification problem more challenging. The paper is organized as follows. Section 2 introduces related work in the area. Details of the corpora collection and annotation are given in Section 3. Section 4 reports various experiments on automatic value detection, while Section 5 discusses the performance of the psycholinguistic experiments and mentions possible future directions.
|
State-of-the-art sentiment analysis (SA) systems look at a fragment of text in isolation. However, in order to design a Schwartz model classifier, we require a psycholinguistic analysis. Therefore, textual features and techniques proposed and dis-cussed for SA are quite different from our current research needs. Hence, we will here focus only on previous research efforts in automatic personality analysis that closely relate to our research work. Personality models can be seen as an augmentation to the basic definition of SA, where the aim is to understand sentiment/personality at person level rather than only at message level. In recent years, there has been a lot of research on automated identification of various personality traits of an individual from their language usage and behaviour in social media. A milestone in this area was the 2013 Workshop and Shared Task on Computational Personality Recognition Linguistic Features: The participating teams tested several linguistic features. Since n-grams are known to be useful for any kind of textual classification, all the teams tested various lengths of n-grams (uni, bi, and tri-grams). Categorical features like part-of-speech (POS), word level features like capital letters, repeated words were also used. Linguistic Inquiry Word Count (LIWC) features were used by all the teams as their baselines. LIWC Here we briefly describe some people. Please read each description and think about how much each person is or is not like you. Tick the box to the right that shows how much the person in the description is like you. HOW To start with, we ask a very fundamental question: whether social media is a good proxy of the original (real life) society or not. Back et al. ( In addition, an essay corpus was collected. These three diverse corpora were then used for training and testing Schwartz values analysis methods. A standard method of psychological data collection is through self-assessment tests, popularly known as psychometric tests. In our experiments, self-assessments were obtained using male/female versions of PVQ, the Portrait Values Questionnaire were asked to answer each question on a 1-6 Likert rating scale. The ranges of scores obtained from the previous rescaling method may vary across different Values classes. For instance, the ranges of the rescaled scores for the Essay corpus are as follows: Achievement A 'Yes' or 'No' binary value was assigned to each Values class: if the score was less than 0, the class was considered to be negative, indicating absence of that Values trait for the particular user; while scores ≥ 0 were considered to be positive, indicating the presence of that trait for the user. We will use the real scores ranging [-1, 1] for the regression experiments mentioned in Section 4. Reports of psychological analysis always depend on how the target population is chosen. Therefore while we are hypothesising that a few people are more Power oriented, an open question that remains unanswered is whom they are more Power oriented than. For example, if we (hypothetically) choose parliamentarians / politicians as participants in an experiment, then the entire examined population will likely turn out to be Power oriented. Therefore, it makes sense to normalise the obtained data into two groups [-1, 0) and [0, 1] and proclaim that people with [0, 1] range scores are relatively more Power (or any other Value) oriented than the people having score ranging [-1, 0). The same normalisation mechanism was applied to all the corpora, but also after normalisation the different Values distributions were imbalanced (with the Facebook data being the most imbalanced). One possible reason behind such imbalanced distributions is that the portion of the real population using social media is slightly biased towards some Values types due to several societal reasons such as educational/family background, age group, occupation, etc. Another reason could be that the divisions between different value types simply never are balanced in any population. However, analysing such societal traits is a separate research direction altogether and out of the scope of the current study. The PVQ questionnaire setting described above was used to separately collect textual user data separately for the Essay, Facebook, and Twitter corpora, as discussed in the rest of this section. The Essay corpus was collected using the Amazon Mechanical Turk (AMT) In the first quarter of 2016, the micro blogging service Twitter averaged 310 million monthly active users, The participants were requested to answer the PVQ questionnaire and to provide their Twitter IDs, so that their tweets could be crawled. However, several challenges have to be addressed when working with Twitter, and a number of iterations, human interventions and personal communications were necessary. For example, several users had protected Twitter accounts, so that their tweets were not accessible when using the Twitter API. In addition, many users had to be discarded since they had published less than 100 tweets, making them uninteresting for statistical analysis. In addition, some extreme cases when users mentioned someone else's (some celebrity's) Twitter account, had to be discarded. The open source free Twitter API: Twitter4J 5 also has a limit of accessing only the current 3,200 tweets from any user. To resolve this issue, an open source Java application (Henrique, 2015) was used. At the end of the data collection process, data from 367 unique users had been gathered. The highest number of tweets for one user was 15K, while the lowest number of tweets for a user was a mere 100; the average number of messages per user in the Twitter corpus was found to be 1,608. Facebook (FB) is the most popular social networking site in the world, with 1.65 billion monthly active users during the first quarter of 2016. 6 Therefore, Facebook was a natural first choice for corpus collection, but since the privacy policy of Facebook is very stringent, accessing Facebook data is challenging. To collect the corpus, a Facebook Canvas web-application was developed using Facebook Graph API and Facebook SDK v5 for PHP library. Undergraduate students of two Indian institutes (NIT, Agartala, Tripura and IIIT, Sri City, Andhra Pradesh) were contacted for the data 5 So far, data from 114 unique users has been collected, but the data is highly imbalanced (for some value types the distributions of 'Yes' and 'No' classes were in 90:10 ratio). Crowd-sourcing is a cheap and fast way to collect data, but unfortunately some annotators chose random labels to minimize their cognitive thinking load. These annotators can be considered as spammers and make aggregation of crowd-sourced data a challenging problem, as discussed in detail by Categorical flat distributions are reported in Table 2. Schwartz' model defines fuzzy membership, which means that anyone having a Power orientation can have the Achievement orientation as well. To understand this notion vividly, we have reported the fuzzy membership statistics from the Twitter data in The statistics in Table Several experiments were performed to get a better understanding of the most appropriate linguistic and non-linguistic features for the problem domain. The experiments were designed as a single label classification task (each input corresponds to one target label) with 20 classes, with 'Yes' and 'No' classes for each of the ten Schwartz values. Ten different classifiers were trained, each for a particular value type. Each classifier predicts whether the person concerned is positively or negatively inclined towards the given Schwartz value. The versions implemented in WEKA LIWC Analysis: LIWC To investigate how each LIWC feature contributes, feature ablation was performed and the Pearson correlations of LIWC features vs value types were analysed. The final classifiers were trained using only the features that were contributing for a particular value type. This resulted in a performance boost and also gave reduced time complexity (both model training and testing times). Table n-grams: In line with the systems discussed in Section 2, n-gram features were added to the LIWC baseline. In a first run, the top 20% of the most frequent uni-grams from the Essay corpus were included as new features, resulting in a 1452+69 feature set. Unexpectedly, SMO's accuracy dropped by an average of 8.60%. The Achievement and Conformity values suffered the maximum performance drop, whereas Security and Hedonism had a slight increase in accuracy. Random Forests performed well in many of cases, except for the Security and Benevolence classes. In a second iteration, categorical (value wise) n-grams features were selected and used. The resulting feature set sizes differ for each of the ten values, ranging from the lowest number 886+69 for Power to the highest 1176+69 for Universalism. Marginally better performance was recorded. n-grams (word grams) with various sizes of n, ranging from 2, 3, 4, 5, and so on, have different impact on performance on different kinds of applications. Commonly, bi-grams are better features for many text classification tasks. So, in a third iteration we tested system performance using bi-grams as added features with LIWC. As the total possible combinations of bi-grams are quite high, only the top 2,000 frequent bi-grams were included, resulting in 2000+69 features. There was no significant performance gain in this experiment on the Essay corpus, so this feature was not tested for the other two corpora. Topic Modeling: In order to find out the bagof-words features for each value type, i.e., the vocabulary that a person uses more frequently, the It was also observed that the rankings of the top 5 topics were almost similar for each Schwartz value. The accuracies obtained were almost similar to the accuracies obtained in the previous experiments; however, this time, since the dimension of the feature set is much smaller, the time complexity decreased by almost a factor of 10. Hence the topic modelling was repeated for the social media corpora from Facebook and Twitter, but resulting in a different number of topic clusters, namely 89. Added to the 69 LIWC features this thus resulted in a total of 158 features. Psycholinguistic Lexica: In addition to the base feature set from LIWC, two other psycholinguistic lexica were added: the Harvard General Inquirer 7 In addition, the sensorial lexicon Sensicon (Tekiroglu et al., 2014) was used. It contains words with sense association scores for the five basic senses: Sight, Hearing, Taste, Smell, and Touch. For example, when the word 'apple' is uttered, the average human mind will visualize the appearance of an apple, stimulating the eye-sight, feel the smell and taste of the apple, making use of the nose and tongue as senses, respectively. Sensicon provides a numerical mapping which indicates the extent to which each of the five senses is used to perceive a word in the lexicon. Again, feature ablation was performed and the (Pearson) correlations of lexicon features vs values analysed. Finally, classifiers were trained using only contributing features for a particular value. Speech Act Features: The way people communicate, whether it is verbally, visually, or via text, is indicative of Personality/Values traits. In social media, profile status updates are used by individuals to broadcast their mood and news to their peers. In doing so, individuals utilize various kinds of speech acts that, while primarily communicating their content, also leave traces of their values/ethical dimensions behind. Following the hypothesis of The categorical corpus distribution and the performance of the final classifier are reported in Table 5, showing an average F 1 -score of 0.69 in 10fold cross validation. Automatic speech act classification of social media conversations is a separate research problem, which is out of the scope of the current study. However, although the speech act classifier was not highly accurate in itself, the user specific speech act distributions (in %) could be used as features for the psycholinguistic classifiers (resulting in 11 additional features). Experiment on the Essay and Facebook corpora showed only 1.15% and 1% performance gain, respectively, whereas on the Twitter Corpus, a noticeable performance improvement of 6.12% (F-measure) was obtained. This indicates that speech acts are important signals of psychological behaviour, so even though the speech act classifier performs poorly, the extracted information is relevant. Social network structure is very useful to predict any person's intrinsic value. For each user in the Twitter corpus, the total number of tweets or messages, total number of likes, average time differ-9 See for Fine-Gained Speech-Act classes ence between two tweets/messages, total number of favourites and re-tweets, and their in-degree and out-degree centrality scores on network of friends and followers were used as features adding to a total of 7 features along with the feature set used in the Topic Modelling experiment (69 LIWC + 89 Topic Modeling words from the Essay Corpus) after observation of the structure of tweets and the previously done linguistic feature experiments. The degree centrality was calculated as of a vertex v, for a given graph G(V,E) with |V | vertices and |E| edges, is defined as: The results of all the experiments after 10-fold cross-validation are summarized in Table The main contributions of this paper are the introduction of a computational Schwartz values model, development of three different corpora annotated with Schwartz' value, and experiments with features for automatic value classification. Table The results also indicate that social media text is difficult for automatic classification, which is obvious from its terse nature. However, it is striking that the social media postings correlate far stronger than the essays with the psychometric data. This is probably since the size of the Twitter data is much larger than someones essay, and since when asked to write something, people become cautious; however, users behave more naturally when communicating in social media, making the data more insightful. Another major implication from the experiments is that popular text classification features such as n-grams and topic-modelling were not performing well in this domain, indicating that this is not yet another text classification problem, but that rather further deeper psycholinguistic analysis is required to find out hidden clues and the nature of language vs ethical practices. Here, it is worth noting the research by For the machine learners, closer analysis revealed that SMO's performance was somehow irregular and random, which might be an indication of over-fitting. For example, the performance for some Schwartz values greatly decreased when adding n-grams as new features with LIWC, whereas some values showed the opposite behaviour, implying that each value type has its own set of distinct clues, but also high overlap. On the other hand, the performance of the Random Forests classifier increased when the number of features was increased, resulting in a larger forest and hence for most value types it performed better than the other two classifiers with less over-fitting. A major limitation of the work is that the collected social network corpus is skewed. Reports of psychological analysis on any community always depend on how the target population is chosen. It is absolutely impossible to get precisely balanced data from any real community. For example, it is rather impossible to have 150+ absolute power oriented people in a corpus of size 367 users data. The only solution to this problem is having more data, which we currently are collecting. The data will be publicly released to the research community soon. We are also very keen on the applied side of this kind of models. Presently we are analysing the community detection problem in social media in relation to values. Another interesting application could be comparative societal analysis between the Eastern and Western regions of the world. Relations among personality and ethics could also be explored.
| 775 | 3,268 | 775 |
Generating Fact Checking Explanations
|
Most existing work on automated fact checking is concerned with predicting the veracity of claims based on metadata, social network spread, language used in claims, and, more recently, evidence supporting or denying claims. A crucial piece of the puzzle that is still missing is to understand how to automate the most elaborate part of the process -generating justifications for verdicts on claims. This paper provides the first study of how these explanations can be generated automatically based on available claim context, and how this task can be modelled jointly with veracity prediction. Our results indicate that optimising both objectives at the same time, rather than training them separately, improves the performance of a fact checking system. The results of a manual evaluation further suggest that the informativeness, coverage and overall quality of the generated explanations are also improved in the multi-task model.
|
When a potentially viral news item is rapidly or indiscriminately published by a news outlet, the responsibility of verifying the truthfulness of the item is often passed on to the audience. To alleviate this problem, independent teams of professional fact checkers manually verify the veracity and credibility of common or particularly checkworthy statements circulating the web. However, these teams have limited resources to perform manual fact checks, thus creating a need for automating the fact checking process. The current research landscape in automated fact checking is comprised of systems that estimate the veracity of claims based on available metadata and evidence pages. Datasets like LIAR Ruling Comments: (...) The last major oil spill from a drilling accident in America happened over 40 years ago in 1969. (...) The largest in volume was the Santa Barbara spill of 1969 referenced by Murdock and Johnson, in which an estimated 100,000 barrels of oil spilled into the Pacific Ocean, according to the API. The Santa Barbara spill was so big it ranked seventh among the 10 largest oil spills caused by marine well blowouts in the world, the report states. Two other U.S. spills, both in 1970, rank eighth and 10th. Fourteen marine blowouts have taken place in the U.S. between 1969 and 2007. Six of them took place after 1990 and spilled a total of nearly 13,700 barrels. (...) We interviewed three scientists who said that the impact of a spill has little to do with its volume. Scientists have proven that spills far smaller than Santa Barbara's have been devastating. Justification: While the nation's largest oil well blowout did take place in 1969, it's not factually correct to call it the "last major oil spill". First of all, two of the largest blowouts in the world took place in the U. S. the following year. More importantly, experts agree that spills far smaller in volume to the 1969 disaster have been devastating. From a scientific perspective, Johnson's decision to single out the 1969 blowout as the last "major" one makes no sense.
|
Table A prevalent component of existing fact checking systems is a stance detection or textual entailment model that predicts whether a piece of evidence contradicts or supports a claim In the field of NLP as a whole, both explainability and interpretability methods have gained importance recently, because most state-of-the-art models are large, neural black-box models. Interpretability, on one hand, provides an overview of the inner workings of a trained model such that a user could, in principle, follow the same reasoning to come up with predictions for new instances. However, with the increasing number of neural units in published state-of-the-art models, it becomes infeasible for users to track all decisions being made by the models. Explainability, on the other hand, deals with providing local explanations about single data points that suggest the most salient areas from the input or are generated textual explanations for a particular prediction. Saliency explanations have been studied extensively Inspired by this, we research how to generate explanations for veracity prediction. We frame this as a summarisation task, where, provided with elaborate fact checking reports, later referred to as ruling comments, the model has to generate veracity explanations close to the human justifications as in the example in Table 1. We present the first study on generating veracity explanations, showing that they can successfully describe the reasons behind a veracity prediction. 2. We find that the performance of a veracity classification system can leverage information from the elaborate ruling comments, and can be further improved by training veracity prediction and veracity explanation jointly. 3. We show that optimising the joint objective of veracity prediction and veracity explanation produces explanations that achieve better coverage and overall quality and serve better at explaining the correct veracity label than explanations learned solely to mimic human justifications. Existing fact checking websites publish claim veracity verdicts along with ruling comments to support the verdicts. Most ruling comments span over long pages and contain redundancies, making them hard to follow. Textual explanations, by contrast, are succinct and provide the main arguments behind the decision. PolitiFact We remove instances that contain fewer than three sentences in the ruling comments as they indicate short veracity reports, where no summary is present. The final dataset consists of 10,146 training, 1,278 validation, and 1,255 test data points. A claim's ruling comments in the dataset span over 39 sentences or 904 words on average, while the justification fits in four sentences or 89 words on average. We now describe the models we employ for training separately (1) an explanation extraction and (2) veracity prediction, as well as (3) the joint model trained to optimise both. The models are based on DistilBERT Our explanation model, shown in Figure We start by greedily selecting the top k sentences from each claim's ruling comments that achieve the highest ROUGE-2 F1 score when compared to the gold justification. We choose k = 4, as that is the average number of sentences in veracity justifications. The selected sentences, referred to as oracles, serve as positive gold labels -y E ∈ {0, 1} N , where N is the total number of sentences present in the ruling comments. Appendix A.1 provides an overview of the coverage that the extracted oracles achieve compared to the gold justification. Appendix A.2 further presents examples of the selected oracles, compared to the gold justification. At training time, we learn a function f (X) = p E , p E ∈ R 1,N that, based on the input X, the text of the claim and the ruling comments, predicts which sentence should be selected -{0,1}, to constitute the explanation. At inference time, we select the top n = 4 sentences with the highest confidence scores. Our extraction model, represented by function f (X), takes the contextual representations produced by the last layer of DistilBERT and feeds them into a feed-forward task-specific layer -h ∈ R h . It is followed by the prediction layer p E ∈ R 1,N with sigmoid activation. The prediction is used to optimise the cross-entropy loss function For the veracity prediction model, shown in Figure 1 (right), we learn a function g(X) = p F that, based on the input X, predicts the veracity of the claim y F ∈ Y F , Y F = {true, false, half-true, barely-true, mostly-true, pants-on-fire}. The function g(X) takes the contextual token representations from the last layer of DistilBERT and feeds them to a task-specific feed-forward layer h ∈ R h . It is followed by the prediction layer with a softmax activation p F ∈ R 6 . We use the prediction to optimise a cross-entropy loss function Finally, we learn a function h(X) = (p E , p F ) that, given the input X -the text of the claim and the ruling comments, predicts both the veracity explanation p E and the veracity label p F of a claim. The model is shown Figure We further combine the resulting two subspaces for each task -h i E and h j F with parameters β to produce one representation per task: where P ∈ {E, F } is the corresponding task. Finally, we use the produced representation to predict p E and p F , with feed-forward layers followed by sigmoid and softmax activations accordingly. We use the prediction to optimise the joint loss function , where γ and η are used for weighted combination of the individual loss functions. We first conduct an automatic evaluation of both the veracity prediction and veracity explanation models. Explanation Quality. We first provide a manual evaluation of the properties of three different types of explanations -gold justification, veracity explanation generated by the Explain-MT, and the ones generated by Explain-Extractive. We ask three annotators to rank these explanations with the ranks 1, 2, 3, (first, second, and third place) according to four different criteria: 1. Coverage. The explanation contains important, salient information and does not miss any important points that contribute to the fact check. Table 2. Non-redundancy. The summary does not contain any information that is redundant/repeated/not relevant to the claim and the fact check. 3. Non-contradiction. The summary does not contain any pieces of information that are contradictory to the claim and the fact check. 4. Overall. Rank the explanations by their overall quality. We also allow ties, meaning that two veracity explanations can receive the same rank if they appear the same. For the annotation task set-up, we randomly select a small set of 40 instances from the test set and collect the three different veracity explanations for each of them. We did not provide the participants with information of the three different explanations and shuffled them randomly to prevent easily creating a position bias for the explanations. The annotators worked separately without discussing any details about the annotation task. Explanation Informativeness. In the second manual evaluation task, we study how well the veracity explanations manage to address the information need of the user and if they sufficiently describe the veracity label. We, therefore, design the annotation task asking annotators to provide a veracity label for a claim based on a veracity explanation coming from the justification, the Explain-MT, or the Explain-Extractive system. The annotators have to provide a veracity label on two levels -binary classification -true or false, and six-class classification -true, false, half-true, barely-true, mostly-true, pants-on-fire. Each of them has to provide the label for 80 explanations, and there are two annotators per explanation. Our models employ the base, uncased version of the pre-trained DistilBERT model. The models are fed with text depending on the task set-up -claim and ruling sentences for the explanation and joint models; claim and ruling sentences, claim and oracle sentences or claim and justification for the fact-checking model. We insert a '[CLS]' token before the start of each ruling sentence (explanation model), before the claim (fact-checking model), or at the combination of both for the joint model. The text sequence is passed through a number of Transformer layers from DistilBERT. We use the '[CLS]' embeddings from the final contextual layer of Dis-tilBERT and feed that in task-specific feed-forward layers h ∈ R h , where h is 100 for the explanation task, 150 for the veracity prediction one and 100 for each of the joint cross-stitch subspaces. Following are the task-specific prediction layers p E . The size of h is picked with grid-search over {50, 100, 150, 200, 300}. We also experimented with replacing the feed-forward task-specific layers with an RNN or Transformer layer or including an activation function, which did not improve task performance. The models are trained for up to 3 epochs, and, following We then select the model with the best ROUGE-2 F1 score on the validation set, thus, performing a potential early stopping. The learning rate used is 3e-5, which is chosen with a grid search over {3e-5, 4e-5, 5e-5}. We perform 175 warm-up steps (5% of the total number of steps), after also experimenting with 0, 100, and 1000 warm-up steps. Optimisation is performed with AdamW The maximum input words to DistilBERT are 512, while the average length of the ruling comments is 904 words. To prevent the loss of any sentences from the ruling comments, we apply a sliding window over the input of the text and then merge the contextual representations of the separate sliding windows, mean averaging the representations in the overlap of the windows. The size of the sliding window is 300, with a stride of 60 tokens, which is the number of overlapping tokens between two successive windows. The maximum length of the encoded sequence is 1200. We find that these hyper-parameters have the best performance after experimenting with different values in a grid search. We also include a dropout layer (with 0.1 rate for the separate and 0.15 for the joint model) after the contextual embedding provided by the transformer models and after the first linear layer as well. The models optimise cross-entropy loss, and the joint model optimises a weighted combination of Wang (2017), all metadata 0.247 0.274 Veracity@RulOracles 0.308 0.300 Veracity@Rul 0.313 0.313 MT-Veracity@Rul 0.321 0.323 Alhindi et al. ( Table both losses. Weights are selected with a grid search -0.9 for the task of explanation generation and 0.1 for veracity prediction. The best performance is reached with weights that bring the losses of the individual models to roughly the same scale. Explanation Quality. Table We compute Krippendorff's α inter-annotator agreement (IAA, Hayes and Krippendorff ( We assume that the low IAA can be attributed to the fact that in ranking/comparison tasks for manual evaluation, the agreement between annotators might be affected by small differences in one rank position in one of the annotators as well as by the annotator bias towards ranking explanations as ties. Taking this into account, we choose to present the mean average recall for each of the annotators instead. Still, we find that their preferences are not in a perfect agreement and report only what the majority agrees upon. We also consider that the low IAA reveals that the task might be "already too difficult for humans". This insight proves to be important on its own as existing machine summarisation/question answering studies involving human evaluation do not report IAA scores We find that the gold explanation is ranked the best for all criteria except for Non-contradiction, where one of the annotators found that it contained more contradictory information than the automatically generated explanations, but Krippendorff's α indicates that there is no agreement between the annotations for this criterion. Out of the two extractive explanation systems, Explain-MT ranks best in Coverage and Overall criteria, with 0.21 and 0.13 corresponding improvements in the ranking position. These results contradict the automatic evaluation in Section 4.3, where the explanation of Explain-MT had lower ROUGE F1 scores. This indicates that an automatic evaluation might be insufficient in estimating the information conveyed by the particular explanation. On the other hand, Explain-Extr is ranked higher than Explain-MT in terms of Non-redundancy and Non-contradiction, where the last criterion was disagreed upon, and the rank improvement for the first one is only marginal at 0.04. This implies that a veracity prediction objective is not necessary to produce natural-sounding explanations (Explain-Extr), but that the latter is useful for generating better explanations overall and with higher coverage Explain-MT. Explanation Informativeness. Table The Fleiss' κ IAA for binary prediction is: Just -0.269, Explain-MT -0.345, Explain-Extr -0.399. The highest agreement is achieved for Explain-Extr, which is supported by the highest proportion of agreeing annotations from Table Table In the second example, the multi-task explanation has lower ROUGE scores than the extractive one. We observe that the gold justification contains some sentences that are not relevant to the fact check, and the extractive summary is fooled to select explanation sentences that are close to the gold summary. As a result, the explanation does not provide enough information about the chosen veracity label. The multi-task model, on the other hand, selects sentences that are also contributing to the prediction of the veracity labels. Thus, its explanation turns out to be more beneficial for the final fact check even though it has a lower ROUGE score compared to the gold justification. As the ROUGE score only accounts for word-level similarity between gold and predicted justifications, we also conduct a manual evaluation of the quality of the produced veracity explanations. Generating Explanations. Generating textual explanations for model predictions is an understud- ied problem. The first study was Fact Checking Interpretability. Interpreting fact checking systems has been explored in a few studies. In this study, we propose a method to generate veracity explanations that would explain the reasons behind a certain veracity label independently of the speaker profile. Once trained, such methods could then be applied to other fact checking instances without human-provided explanations or even to perform end-to-end veracity prediction and veracity explanation generation given a claim. Substantial research on fact checking methods exists for the FEVER dataset While FEVER and MultiFC are larger datasets for fact checking than LIAR-PLUS, they do not contain veracity explanations and can thus not easily be used to train joint veracity prediction and explanation generation models, hence we did not use them in this study. We presented the first study on generating veracity explanations, and we showed that veracity prediction can be combined with veracity explanation generation and that the multi-task set-up improves the performance of the veracity system. A manual evaluation shows that the coverage and the overall quality of the explanation system is also improved in the multi-task set-up. For future work, an obvious next step is to investigate the possibility of generating veracity explanations from evidence pages crawled from the Web. Furthermore, other approaches of generating veracity explanations should be investigated, especially as they could improve fluency or decrease the redundancy of the generated text. Table Table The Fleiss' κ agreement for the 6-label manual annotations is: 0.20 on the Just explanations, 0.230 on the Explain-MT explanations, and 0.333 on the Explain-Extr system. Table
| 933 | 2,065 | 933 |
Interpreting Predictive Probabilities: Model Confidence or Human Label Variation?
|
With the rise of increasingly powerful and userfacing NLP systems, there is growing interest in assessing whether they have a good representation of uncertainty by evaluating the quality of their predictive distribution over outcomes. We identify two main perspectives that drive starkly different evaluation protocols. The first treats predictive probability as an indication of model confidence; the second as an indication of human label variation. We discuss their merits and limitations, and take the position that both are crucial for trustworthy and fair NLP systems, but that exploiting a single predictive distribution is limiting. We recommend tools and highlight exciting directions towards models with disentangled representations of uncertainty about predictions and uncertainty about human labels.
|
In common language, uncertainty refers to "a state of not being definitely known or perfectly clear; a state of doubt". 1 In statistics and machine learning, uncertainty is taken as a state to be represented In modern NLP, neural networks are the de-facto standard to predict complex probability measures from available context 1 Oxford English Dictionary, accessed October 13th 2023. Recently, transformer-based large language models (LLMs) are becoming increasingly powerful and display remarkable abilities on complex classification tasks, leading to an increased deployment in user-facing applications. This motivates the need for models that can signal when they are likely to be wrong (P1; an aspect of trustworthiness), and models that can capture different linguistic and human interpretations (P2; an aspect of language including fairness). In this position paper, we identify that the exact same representation of uncertainty-the predictive distribution over outcomes-is sometimes interpreted as an indication of confidence in model predictions (P1; We hope to provide clarity and accelerate progress by: (i) Identifying these two perspectives on the predictive distribution and examining how each evaluates the quality of predictive distribution in Section 2. (ii) Discussing their merits and limitations, and relating them to popular notions of aleatoric and epistemic uncertainty in Section 3. (iii) Taking the position that both perspectives contribute to trustworthy and fair NLP systems, but that exploiting a single predictive distribution is limiting-e.g., does a uniform predictive distribution represent uncertainty about human perspectives, or rather about the correctness of that prediction itself?-and highlighting exciting directions towards models that can predict distributions over human or linguistic interpretations, and simultaneously abstain from answering when lacking such knowledge or skills in Section 4.
|
Consider a user-facing question answering (QA) system. Ideally, this model is able to abstain on questions that it is likely to get wrong (a.k.a. selective answering or prediction; Most text classifiers chain two building blocks: i) a parametric model which, given input text x, prescribes the probability mass function (pmf) f (y; x) of the conditional random variable Y |X = x taking on values in a set {1, . . . , K} of K class labels; and ii) a decision rule δ f (x) to map from f (•; x) to a single label. For most modern models, the map x → f (•; x) is realised by a neural network and the most common decision rule δ f (x) = arg max k∈[K] f (k; x) returns the mode of the pmf. Next, we identify two main perspectives on predictive probability f (y; x), with starkly different evaluation frameworks. 2 2 We use capital letters for random variables (e.g., X, Y ) and lowercase letters for outcomes (e.g., x, y). As standard, X = x denotes random variable (rv) assignment. For logical predicates we use the Iverson bracket [A = B] to denote a new rv whose outcome is 1, when A and B are assigned the same outcome, and 0 otherwise. A determinis- From a statistical perspective, most NLP systems are trained on single annotations using regularised maximum likelihood estimation (MLE), without mechanism or incentive to represent uncertainty about their own correctness (MLE recovers a single realisation of the map x → f (•; x)). This is unlike, for instance, Bayesian estimation (where the map x → f (•; x) is given random treatment; more in Section 4). In addition, regardless of whether models represent uncertainty about their own correctness, calibration metrics, and ECE in particular, are known to have limitations, e.g., problems with binning Finally, the sense of trustworthiness from verifying that Equation (1) holds (for a given confidence level q) in a given dataset, might not transfer to any one future prediction in isolation. Though some studies examine the effect of communicating predictive probability to human decision makers The idea that gold labels are too simplistic has been around for some time From a statistical perspective, text classifiers predict a distribution for Y |X = x, and are precisely mechanisms to represent uncertainty about a given input's label. However, given that they are parametric models trained with regularised MLE, they can at best learn to predict observed label variability (which is often not present in NLP datasets since most record only single annotations), or label variability as a byproduct of parametric bottlenecks, regularisation and other inductive biases that reserve (conditional) probability for unseen labels. Evaluating whether probability mass is indeed allocated coherently with plausible variability is limited by: 1) datasets lacking multiple high quality annotations per input, 2) unclarity about how many annotations are sufficient to reliably estimate the human distribution The language that is often used to explain calibration allows (quite ironically) for both perspectives P1 and P2. The examples above illustrate well that one may regard predictive probabilities one way or another, each interpretation tracking a different type of event (i.e., correctness, assessed marginally for a collection of inputs, or label frequency, assessed conditionally against a population of annotators). Crucially, however, most models are trained to approximately recover the maximum likelihood solutiona single realisation of the map x → f (•; x), with no room for quantification of uncertainty about its correctness. Therefore, without special incentives (e.g., regularisation, change of loss or supervision; some of which we discuss in Section 4.1), our predictive distributions are not meant to inherently support P1, and they may support P2, as we discuss in the next section. The predictive distribution for an input x is sometimes taken as a representation of uncertainty about a model's future classification performance ("knowing when it knows"); other times as a representation of uncertainty about label frequency in a population of human annotators (human label variation). We now discuss merits and limitations for each perspective. These two perspectives on the predictive distribution in NLP can be put in a broader context of statistics and machine learning by considering that there can be many sources that lead to uncertainty Such sources are often categorised as aleatoric (irreducible; inherent to data) or epistemic (reducible, inherent to modelling). In that sense, P1 regards the predictive distribution as epistemic uncertainty, whereas P2 as aleatoric uncertainty. Armed with this knowledge, one can pick the right modeling tools for each, and tap into this broader literature. In the next section, we make several recommendations. We argue that the desiderata behind both perspectives are equally important for trustworthy and fair NLP systems, but that expecting the predictive distribution to represent both is limiting. Rather than calibrating the predictive distribution to better indicate model error, we outline alternative directions to capture uncertainty about predictions (towards more trustworthy NLP) and uncertainty about human perspectives (towards fairer NLP)where the latter can, and in our view should be represented by the predictive distribution. Inspired by machine translation quality estimation (e.g. Alternatively, the event space can be expanded beyond only the target variable to include parameters too, thus allowing for uncertainty about them. Since this leads to intractability, some (approximate) Bayesian solutions in NLP include Langevin dynamics Evaluating model error uncertainty is challenging, in part because ground truth is difficult to find. Proxy tasks like selective answering To represent uncertainty about plausible human interpretations, data is crucial. In NLP, these factors are commonly not controlled for. However, recently, researchers use annotator information to model sub-populations Besides data, an appealing but non-trivial alternative (for some tasks, like textual entailment) is to encourage models to generalise to the linguistic phenomena that give rise to label variation, despite supervising with single annotations In this position paper, we identified two important perspectives on the predictive distribution in NLP. We believe that the desiderata behind both are crucial for fair and trustworthy NLP systems, but that exploiting the same predictive distribution is limiting. We recommend exiting tools and directions to represent uncertainty about predictions (model confidence) and about label variation (human perspectives). We hope to facilitate a better understanding of uncertainty in NLP, and encourage future work to acknowledge, represent and evaluate multiple sources of uncertainty with principled design decisions. Evaluation along a specific axis can be useful regardless of whether a model has been explicitly designed to meet this goal. One could argue this is true for both calibration as well as human label variation. It is certainly also true in other sub-fields, like interpretability. For example, probing hidden representations or specific linguistic information, without having explicitly trained models to store them. Furthermore, although we focus on classification systems in the language domain, the topics we highlight and discuss are equally important in other domains, such as computer vision (e.g., affective computing), or language generation (e.g., story telling).
| 811 | 1,939 | 811 |
Neural Generation of Dialogue Response Timings
|
The timings of spoken response offsets in human dialogue have been shown to vary based on contextual elements of the dialogue. We propose neural models that simulate the distributions of these response offsets, taking into account the response turn as well as the preceding turn. The models are designed to be integrated into the pipeline of an incremental spoken dialogue system (SDS). We evaluate our models using offline experiments as well as human listening tests. We show that human listeners consider certain response timings to be more natural based on the dialogue context. The introduction of these models into SDS pipelines could increase the perceived naturalness of interactions. 1
|
The components needed for the design of spoken dialogue systems (SDSs) that can communicate in a realistic human fashion have seen rapid advancements in recent years (e.g. responses based on the length of a silence before a response Overview We propose a neural model for generating these response timings in SDSs (shown in Fig. It may be impractical in an incremental framework to generate a full system response and then re-encode it using the response encoder of RT-Net. To address this issue, we propose an extension of RTNet that uses a variational autoencoder (VAE) This paper is structured as follows: First, we present how our dataset is structured and our training objective. Then, in sections 2.1 and 2.2 we present details of our two models, RTNet and RTNet-VAE. Section 2.3 presents our input feature representations. In section 2.4 we discuss our training and testing procedures. In sections 3.1 and 3.2 we analyze the performance of both RTNet and RTNet-VAE. Finally, in section 4 we present the results of a human listener test.
|
Dataset Our dataset is extracted from the Switchboard-1 Release 2 corpus Training Objective Our training objective is to predict the start of the system turn one frame ahead of the ground truth start time. The target labels in each turn pair are derived from the ground truth speech activity labels as shown in Fig. We do not predict at earlier frames since we assume that at these mid-turn-pauses the DM has not decided to take a turn yet, either because it expects the user to continue, or it has not formulated one yet. As mentioned previously in section 1, we design RTNet to be abstracted from the turn-taking decisions themselves. If we were to include pauses prior to the turn-final silence, our response generation system would be additionally burdened with making turn-taking decisions, namely, classifying between mid-turn-pauses and end-of-turn silences. We therefore make the modelling assumption that the system's response is formulated at some point during the user's turn-final IPU. To simulate this assumption we sample an index R START from the span of R using a uniform distribution. We then use the reduced set of frames from R START to R END in the calculation of our loss. Encoder The encoder of RTNet (shown in Fig. The linguistic Bi-LSTM takes as input the sequence of 300-dimensional embeddings of the tokenized system response. We use three special tokens: SIL, WAIT, and NONE. The SIL token is used whenever there is a gap between words that is greater than the frame-size (50ms). The WAIT and NONE tokens are inserted as the first and last tokens of the system response sequence respectively. The concatenation [h 0 ; h 1 ; h I ] is passed as input to a RELU layer (we refer to this layer as the reduction layer) which outputs the h z encoding. The h z encoding is used (along with user features) in the concatenated input to the inference network. Since the WAIT embedding corresponds to the h 0 output of the master Bi-LSTM and the NONE embedding corresponds to h I , the two embeddings serve as "triggering" symbols that allow the linguistic and master Bi-LSTM to output relevant information accumulated in their cell states. The acoustic Bi-LSTM takes as input the sequence of acoustic features and outputs a sequence of hidden states at every 50ms frame. As shown in Fig. Inference Network The aim of our inference network is to predict a sequence of output probabilities Y = [y R START , y R START +1 , ..., y N ] using We use a a single-layer LSTM (shown in Fig. Since there are only two possible output values in a generated sequence {0,1}, and the sequence ends once we predict 1, the inference network can be considered an autoregressive model where 0 is passed implicitly to the subsequent time-step. To generate an output sequence, we can sample from the distribution p(y n = 1|y R START = 0, y R START +1 = 0, ..., y n-1 = 0, X 0:n , h z ) using a Bernoulli random trial at each time-step. For frames prior to R START the output probability is fixed to 0, since R START is the point where the DM has formulated the response. During training we minimize the binary cross entropy loss (L BCE ) between our ground truth objective and our output predictions Y . Motivation A limitation of RTNet is that it may be impractical to encode system turns before triggering a response. For example, if we wish to apply RTNet using generated system responses, at run-time the RTNet component would have to wait for the full response to be generated by the NLG, which would result in a computational bottleneck. If the NLG system is incremental, it may also be desirable for the system to start speaking before the entirety of the system response has been generated. VAE To address this, we bypass the encoding stage by directly using the semantic representation output from the DM to control the response timing encodings. We do this by replacing the reduction layer with a VAE (Fig. We impose a Gaussian prior over the latent space using a Kullback-Liebler (KL) divergence loss term: The L KL loss measures the distance of the generated distribution from a Gaussian with zero mean and unit variance. L KL is combined with L BCE using a weighted sum: As we increase the value of w KL we increasingly enforce the Gaussian prior on the latent space. In doing so our aim is to learn a smooth latent space in which similar types of responses are organized in similar areas of the space. Latent Space During inference we can skip the encoding stage of RTNet-VAE and sample z directly from the latent space on the basis of the input semantic representation from the dialogue manager. Our sampling approach is to approximate the distribution of latent variables for a given responsetype using Gaussians. For example, if we have a collection of labelled backchannel responses (and their corresponding z encodings) we can approximate the distribution of p(z|label =backchannel ) using an isotropic Gaussian by simply calculating µ backchannel and σ backchannel , the maximum likelihood mean and standard deviations of each of the z dimensions. These vectors can also be used to calculate directions in the latent space with different semantic characteristics and then interpolate between them. Linguistic Features We use the word annotations from the ms-state transcriptions as linguistic features. These annotations give us the timing for the starts and ends of all words in the corpus. As our feature representation, we use 300 dimensional word embeddings that are initialized with GloVe vectors We combine 40 log-mel filterbanks, and 17 features from the GeMAPs feature set Training and Testing Procedures The training, validation, and test sets consist of 1646, 150, 642 conversations respectively with 151595, 13910, and 58783 turn pairs. The test set includes all of the conversations from the NXT-format annotations We used the following hyperparameter settings in our experiments: The inference, acoustic, linguistic, and master LSTMs each had hidden sizes of 1024, 256, 256, and 512 (respectively). We used a latent variable size of 4, a batch size of 128, and L2 regularization of 1e-05. We used the optimizer with an initial learning rate of 5e-04. We trained each model for 15000 iterations, with learning rate reductions by a factor of 0.1 after 9000, 11000, 13000, and 14000 iterations. While we found that randomizing R START during training was important for the reasons given in Section 2, it presented issues for the stability and reproducibility of our evaluation and test results for L BCE and L KL . We therefore randomize during training and sampling, but when calculating the test losses (reported in Table We also calculate the mean absolute error (MAE), given in seconds, from the ground truth response offsets to the generated output offsets. When sampling for the calculation of MAE, it is necessary to increase the length of the turn pair since the response time may be triggered by the Figure sampling process after the ground truth time. We therefore pad the user's features with 80 extra frames in which we simulate silence artificially using acoustic features. During sampling, we use the same R START randomization process that was used during training, rather than fixing it to the start of the user's turn-final IPU. For each model we perform the sampling procedure on the test set three times and report the mean error in Table Best Fixed Probability To the best of our knowledge, there aren't any other published models that we can directly compare ours to. However, we can calculate the best performance that can be achieved using a fixed value for y. The best possible fixed y for a given turn pair is: FrameLength . The best fixed y for a set of turn pairs is given by the expected value of y tp in that set: y fixed = E[y tp ]. This represents the best performance that we could achieve if we did not have access to any user or system features. We can use the fixed probability model to put the performance of the rest of our models into context. RTNet Performance The offset distribution for the full RTNet model is shown in Fig. To look in more detail at how the system models the offset distribution we can investigate the generated distributions of labelled response dialogue acts in our test set. Fig. The performance of the response encoder was analysed in an ablation study, with results in rows 3 through 5 of Table Inference Network Ablation In rows 6 and 7 of Table In order to examine this irregularity in more detail, we look at the generated distributions from the inference ablation, shown in Fig. RTNet-VAE Performance In rows 8 through 12 of Table In Fig. Latent Space Applications As mentioned in Section 2.2, part of the appeal in using the VAE in our model is that it enables us to discard the response encoding stage. We can exploit the smoothness of the latent space to skip the encoding stage by sampling directly from the trained latent space. We can approximate the distribution of latent variables for individual dialogue act response types using isotropic Gaussians. This enables us to efficiently represent the dialogue acts using mean and standard-deviation vectors, a pair for each dialogue act. Fig. We can use the same parameterized vector representations to interpolate between different dialogue act parameters to achieve intermediate distributions. This dimensional approach is flexible in that we give the dialogue manager (DM) more control over the details of the distribution. For example, if the objective of the SDS was to generate an agree dialogue act, we could control the degree of agreement by interpolating between disagree and agree vectors. Figure It has shown that response timings vary based on the semantic content of dialogue responses and the preceding turn Participants were asked to make A/B choices between two versions of a turn pair, where each version had a different response offset. Participants were asked: "Which response timing sounds like it was produced in the real conversation?" The turn pairs were drawn from our dataset and were limited to pairs where the response was either dispreferred or a backchannel. We limited the chosen pairs to those with ground truth offsets that were either clas- sified as early or late. We classified offsets as early, modal, or late by segmenting the distribution of all of the offsets in our dataset into three partitions as shown in Fig. We had 25 participants (15 female, 10 male) who all wore headphones. We performed binomial tests for the significance of a given choice in each question. For the questions in the first half of the test, in which we compared true vs. opposite offsets, 10 of the 16 comparisons were found to be statistically significant (p < 0.05). In all of the significant cases the true offset was was considered more realistic than the opposite. In reference to our first research question, this result supports the conclusion that some responses are indeed considered to be more realistic than others. For the questions in the second half of the test, in which we compared true vs. modal offsets, six out of the 16 comparisons were found to be statistically significant. Of the six significant preferences, three were a preference for the true offset, and three were a preference for the modal offset. To investigate our second research question, we looked at the offset distributions generated by our model for each of the six significant preferences, shown in Fig. In this paper, we have presented models that can be used to generate the turn switch offset distributions of SDS system responses. It has been shown in prior studies (e.g.
| 694 | 1,043 | 694 |
Multi-Task Semantic Dependency Parsing with Policy Gradient for Learning Easy-First Strategies
|
In Semantic Dependency Parsing (SDP), semantic relations form directed acyclic graphs, rather than trees. We propose a new iterative predicate selection (IPS) algorithm for SDP. Our IPS algorithm combines the graph-based and transition-based parsing approaches in order to handle multiple semantic head words. We train the IPS model using a combination of multi-task learning and task-specific policy gradient training. Trained this way, IPS achieves a new state of the art on the SemEval 2015 Task 18 datasets. Furthermore, we observe that policy gradient training learns an easy-first strategy.
|
Dependency parsers assign syntactic structures to sentences in the form of trees. Semantic dependency parsing (SDP), first introduced in the Se-mEval 2014 shared task Standard dependency parsing algorithms are often said to come in two flavors: transition-based The man went back and spoke to the desk clerk.
|
The man went back and spoke to the desk clerk. The man went back and spoke to the desk clerk. parsers score transitions between states, and gradually build up dependency graphs on the side. Graph-based parsers, in contrast, score all candidate edges directly and apply tree decoding algorithms for the resulting score table. The two types of parsing algorithms have different advantages Multi-task learning is motivated by the fact that there are several linguistic formalisms for SDP. Fig. Contributions In this paper, (i) we propose a new parsing algorithm for semantic dependency parsing (SDP) that combines transition-based and graph-based approaches; (ii) we show that multitask learning of state representations for this parsing algorithm is superior to single-task training; (iii) we improve this model by task-specific policy gradient fine-tuning; (iv) we achieve a new state of the art result across three linguistic formalisms; finally, (v) we show that policy gradient fine-tuning learns an easy-first strategy, which reduces error propagation. There are generally two kinds of dependency parsing algorithms, namely transition-based parsing algorithms There have been several attempts to train transition-based parsers with reinforcement learning: In their experiments, however, the models trained with policy gradient did not always perform better than the models trained with supervised learning. We hypothesize this is due to credit assignment being difficult in transition-based parsing. Iterative refinement approaches have been proposed in the context of sentence generation Several models have been presented for DAG parsing We propose a new semantic dependency parsing algorithm based on the head-selection algorithm for syntactic dependency parsing The man went back and spoke to the desk clerk. The man went back and spoke to the desk clerk. The man went back and spoke to the desk clerk. The man went back and spoke to the desk clerk. The man went back and spoke to the desk clerk. 2017). Head selection iterates over sentences, fixing the head of a word w in each iteration, ignoring w in future iterations. This is possible for dependency parsing because each word has a unique head word, including the root of the sentence, which is attached to an artificial root symbol. However, in SDP, words may attach to multiple head-words or semantic predicates whereas other words may not attach to any semantic predicates. Thus, we propose an iterative predicate selection (IPS) parsing algorithm, as a generalization of head-selection in SDP. The proposed algorithm is formalized as follows. First, we define transition operations for all words in a sentence. For the i-th word w i in a sentence, the model selects one transition t τ i from the set of possible transitions T τ i for each transition time step τ . Generally, the possible transitions T i for the i-th word are expressed as follows: where ARC i,j is a transition to create an arc from the j-th word to the i-th word, encoding that the semantic predicate w j takes w i as an semantic argument. NULL is a special transition that does not create an arc. The set of possible transitions T τ i for the i-th word at time step τ is a subset of possible transitions T i that satisfy two constraints: (i) no arcs can be reflexive, i.e., w i cannot be an argument of itself, and (ii) the new arc must not be a member of the set of arcs A τ comprising the partial parse graph y τ constructed at time step τ . Therefore, we obtain: The model then creates semantic dependency arcs by iterating over the sentence as follows: 1 1 This algorithm can introduce circles. However, circles 1 For each word w i , select a head arc from T τ i . 2 Update the partial semantic dependency graph. 3 If all words select NULL, the parser halts. Otherwise, go to 1. Fig. were extremely rare in our experiments, and can be avoided by simple heuristics during decoding. We discuss this issue in the Supplementary Material, §A.1. Fig. Sentence encoder We employ bidirectional long short-term memory (BiLSTM) layers for encoding words in sentences. A BiLSTM consists of two LSTMs that reads the sentence forward and backward, and concatenates their output before passing it on. For a sequence of tokens [w 1 , • • • , w n ], the inputs for the encoder are words, POS tags and lemmas. from the three-layer BiLSTMs. We use three-layer stacked BiLSTMs. We also use special embeddings h NULL for the NULL transition and h ROOT for the ROOT of the sentence. The model updates the partial SDP graph at each time step of the parsing procedure. The SDP graph y τ at time step τ is stored in a semantic dependency matrix G τ ∈ {0, 1} n×(n+1) for a sentence of n words. We convert the semantic dependency matrix G into a rank three tensor G ∈ R n×(n+1)×p , by replacing elements with embeddings of tokens u(w * ) by where g ij ∈ G and g ij ∈ G . g i * contains the representations of the semantic predicates for the i-th word in the partial SDP graph. We use a single layer Bi-LSTM to encode the semantic predicates g i * of each word; see We also employ dependency flags that directly encode the semantic dependency matrix and indicate whether the corresponding arcs are already created or not. Flag representations F are also three-rank tensors, consisting of two hidden representations: f ARC for g i,j = 1 and f NOARC for g i,j = 0 depending on G. f ARC and f NOARC is q-dimensional vectors. Then we concatenate the hidden representation of the NULL transition and obtain the flag representation F τ : . We do not use BiLSTMs to encode these flags. These flags also reflect the current state of the semantic dependency matrix. The semantic predicate selection model comprises an MLP with inputs from the encoder of the sentence and the partial semantic dependency graph: the sentence representation H, the SDP representation G τ , and the dependency flag F τ . They are rank three tensors and concatenated at the third axis. Formally, the score s ij of the i-th word and the j-th transition is expressed as follows. For the MLP, we use a concatenation of outputs from three different networks: a three-layer MLP, a two-layer MLP and a matrix multiplication with bias terms as follows. W * * are matrices or vectors used in this MLP and W * * are bias terms. Here, we use this MLP for predicting a scalar score s ij ; therefore, W 3 3 , W 2 2 , W 1 1 are vectors. The model computes the probability of the transition t j for each word i by applying a softmax function over the candidates of the semantic head words w j . These transition probabilities p i (t j ) of selecting a semantic head word w j , are defined for each word w i in a sentence. For supervised learning, we employ a cross entropy loss for the partial SDP graph G τ at time step τ . Here l i is a gold transition label for the i-th word and θ represents all trainable parameters. Note that this supervised training regime, as mentioned above, does not have a principled answer to the non-deterministic oracle problem We minimize the softmax cross entropy loss using supervised learning. Policy gradient Reinforcement learning is a method for learning to iteratively act according to a dynamic environment in order to optimize future rewards. In our context, the agent corresponds to the neural network model predicting the transition probabilities p i (t τ j ) that are used in the parsing algorithm. The environment includes the partial SDP graph y τ , and the rewards r τ are computed by comparing the predicted parse graph to the gold parse graph y g . We adapt a variation of the policy gradient method (Williams, 1992) for IPS parsing. Our objective function is to maximize the rewards and the transition policy for the i-th word is given by the probability of the transitions π ∼ p i (t τ j |y τ ). The gradient of Eq.8 is given as follows: When we compute this gradient, given a policy π, we approximate the expectation E π for any transition sequence with a single transition path t that is sampled from policy π: We summarize our policy gradient learning algorithm for SDP in Algorithm 1. For time step τ , the model samples one transition t τ j selecting the j-th word as a semantic head word of the ith word, from the set of possible transitions T i , following the transition probability of π. After sampling t τ j , the model updates the SDP graph to y τ +1 and computes the reward r τ i . When NULL becomes the most likely transition for all words, or the time step exceeds the maximum number of time steps allowed, we stop. (1) Reinforcement learning uses sampling of transitions. This allows our model to explore transition paths that supervised models would never follow. (2) In supervised learning, decisions are independent of the current time step τ , while in reinforcement learning, decisions depend on τ . This means that the θ parameters are updated after the parser finishes parsing the input sentence. (3) Loss Reward Transitions (1) The model creates a new correct arc from a semantic predicate to the i-th word. (2) The first time the model chooses the NULL transition after all gold arcs to the i-th word have been created, and no wrong arcs to the i words have not been created. The model creates a wrong arc from a semantic predicate candidate to the i-th word. r τ i = 0 (4) All other transitions. Table must be non-negative in supervised learning, while rewards can be negative in reinforcement learning. In general, the cross entropy loss is able to optimize for choosing good transitions given a parser configuration, while the policy gradient objective function is able to optimize the entire sequence of transitions drawn according to the current policy. We demonstrate the usefulness of reinforcement learning in our experiments below. Rewards for SDP We also introduce intermediate rewards, given during parsing, at different time steps. The reward r τ i of the i-th word is determined as shown in Table This section includes details of our implementation. When we apply policy gradient, we pre-train our model using supervised learning. We then use policy gradient for task-specific fine-tuning of our model. We find that updating parameters of BiL-STM and word embeddings during policy gradient makes training quite unstable. Therefore we fix the BiLSTM parameters during policy gradient. In our multi-task learning set-up, we apply multi-task learning of the shared stacked BiLSTMs We use the SemEval 2015 Task18 Within weeks the unfolding Iran-Contra scandal took away Mr. Noriega's insurance policy. Morgan will help evaluate DFC's position and help determine alternatives. The U.S. Commerce Department reported a $ 10.77 billion deficit in August compared with ... datasets, as shown in Table Evaluating Our Parser without Lemma Since our baseline Effect of Reinforcement Learning Fig. In supervised learning (IPS+ML), there is a slight tendency to create shorter arcs first, but while the ordering is relatively consistent, the differences are small. This is in sharp contrast with the distributions we see for our policy gradient parser (IPS+ML+RL). Here, across the board, it is very likely that the first transition connects neighboring words; and very unlikely that neighboring words are connected at later stages. This suggests that reinforcement learning learns an easyfirst strategy of predicting short arcs first. Note that unlike easy-first algorithms in syntactic parsing We propose a novel iterative predicate selection (IPS) parsing model for semantic dependency parsing. We apply multi-task learning to learn general representations of parser configurations, and use reinforcement learning for task-specific fine-tuning. In our experiments, our multi-task reinforcement IPS model achieves a new state of the art for three SDP formalisms. Moreover, we show that fine-tuning with reinforcement learning learns an easy-first strategy and some syntactic features.
| 596 | 308 | 596 |
Cross-language Learning with Adversarial Neural Networks: Application to Community Question Answering
|
We address the problem of cross-language adaptation for question-question similarity reranking in community question answering, with the objective to port a system trained on one input language to another input language given labeled training data for the first language and only unlabeled data for the second language. In particular, we propose to use adversarial training of neural networks to learn high-level features that are discriminative for the main learning task, and at the same time are invariant across the input languages. The evaluation results show sizable improvements for our cross-language adversarial neural network (CLANN) model over a strong nonadversarial system.
|
Developing natural language processing (NLP) systems that can work indistinctly with different input languages is a challenging task; yet, such a setup is useful for many real-world applications. One expensive solution is to annotate data for each input language and then to train a separate system for each one. Another option, which can be also costly, is to translate the input, e.g., using machine translation (MT), and then to work monolingually in the target language In this paper, we take this last approach, i.e., combining languages during both training and testing, and we study the problem of question-question similarity reranking in community Question Answering (cQA), when the input question can be either in English or in Arabic, and the questions it is compared to are always in English. We start with a simple language-independent representation based on cross-language word embeddings, which we input into a feed-forward multilayer neural network to classify pairs of questions, (English, English) or (Arabic, English), regarding their similarity. Furthermore, we explore the question of whether adversarial training can be used to improve the performance of the network when we have some unlabeled examples in the target language. In particular, we adapt the Domain Adversarial Neural Network model from In our setup, the basic task-solving network is paired with another network that shares the internal representation of the input and tries to decide whether the input example comes from the source (English) or from the target (Arabic) language. The training of this language discriminator network is adversarial with respect to the shared layers by using gradient reversal during backpropagation, which makes the training to maximize the loss of the discriminator rather than to minimize it. The main idea is to learn a high-level abstract representation that is discriminative for the main classification task, but is invariant across the input languages. We apply this method to an extension of the SemEval-2016 Task 3, subtask B benchmark dataset for question-question similarity reranking The remainder of this paper is organized as follows: Section 2 discusses some related work. Section 3 introduces our model for adversarial training for cross-language problems. Section 4 describes the experimental setup. Section 5 presents the evaluation results. Finally, Section 6 concludes and points to possible directions for future work.
|
Below we discuss three relevant research lines: (a) adversarial training, (b) question-question similarity, and (c) cross-language learning. Adversarial training of neural networks has shown a big impact recently, especially in areas such as computer vision, where generative unsupervised models have proved capable of synthesizing new images Question-question similarity was part of Task 3 on cQA at SemEval-2016/2017 Cross-language Question Answering was the topic of several challenges, e.g., at CLEF 2008 We demonstrate our approach for cross-language representation learning with adversarial training on a cross-lingual extension of the question-question similarity reranking subtask of SemEval-2016 Task 3 on community Question Answering. An example for the monolingual task is shown in Figure In our case, the input question q is in a different language (Arabic) than the language of the retrieved questions (English). The goal is to rerank a set of K retrieved questions {q k } K k=1 written in a source language (e.g., English) according to their similarity with respect to an input user question q that comes in another (target) language, e.g., Arabic. For simplicity, henceforth we will use Arabic as target and English as source. However, in principle, our method generalizes to any source-target language pair. We approach the problem as a classification task, where given a question pair (q, q ), the goal is to decide whether the retrieved question q is similar (i.e., relevant) to q or not. Let c ∈ {0, 1} denote the class label: 1 for similar, and 0 for not similar. We use the posterior probability p(c = 1|q, q , θ) as a score for ranking all retrieved questions by similarity, where θ are the model parameters. More formally, let R n = {q n,k } K k=1 denote the set of K retrieved questions for a new question q n . Note that the questions in R n are always in English. We consider a training scenario where we have labeled examples D S = {q n , q n,k , c n,k } N n=1 for English q n , but we only have unlabeled examples D T = {q n , q n,k } M n=N +1 for Arabic q n , with c n,k denoting the class label for the pair (q n , q n,k ). We want to train a cross-language model that can classify any test example {q n , q n,k }, where q n is in Arabic. This scenario is of practical importance, e.g., when an Arabic speaker wants to query the system in Arabic, and the database of related information is only in English. Here, we adapt the idea for adversarial training for domain adaptation as proposed by Figure The following equations describe the transformations through the hidden layers: where [.; .] denotes concatenation of two column vectors, U and V are the weight matrices in the first and in the second hidden layer, and g is a nonlinear activation function; we use rectified linear units or ReLU The output layer computes a sigmoid: where w are the output layer weights. We train the network by minimizing the negative log-probability of the gold labels: The network described so far learns the abstract features through multiple hidden layers that are discriminative for the classification task, i.e., similar vs. non-similar. However, our goal is also to make these features invariant across languages. To this end, we put a language discriminator, another neural network that takes the internal representation of the network f (see Equation The language discriminator is again defined by a sigmoid function: where l ∈ {0, 1} denotes the language of q (1 for English, and 0 for Arabic), w l are the final layer weights of the discriminator, and h l = g(U l f ) defines the hidden layer of the discriminator with U l being the layer weights and g being the ReLU activations. We use the negative log-probability as the discrimination loss: The overall training objective of the composite model can be written as follows: where θ = {U, V, w}, ω = {U, V, w, U l , w l }, and the hyper-parameter λ controls the relative strength of the two networks. In training, we look for parameter values that satisfy a min-max optimization criterion as follows: which involves a maximization (gradient ascent) with respect to {U l , w l } and a minimization (gradient descent) with respect to {U, V, w}. Note that maximizing L(U, V, w, U l , w l ) with respect to {U l , w l } is equivalent to minimizing the discriminator loss L l (ω) in Equation ( In our gradient descent training, the above minmax optimization is performed by reversing the gradients of the language discrimination loss L l (ω), when they are backpropagated to the shared layers. As shown in Figure Our optimization setup is related to the training method of Generative Adversarial Networks or GANs Training. Algorithm 1 shows pseudocode for the algorithm we use to train our model, which is based on stochastic gradient descent (SDG). We first initialize the model parameters by using samples from glorot-uniform distribution As mentioned above, the main challenge in adversarial training is to balance the two components of the network. If one component becomes smarter, its loss to the shared layer becomes useless, and the training fails to converge Above we considered an unsupervised adaptation scenario, where we did not have any labeled instance for the target language, i.e., when the new question q n is in Arabic. However, our method can be easily generalized to a semi-supervised setting, where we have access to some labeled instances in the target language, D T * = {q n , R n , c n } L n=M +1 . In this case, each minibatch during training is formed by labeled instances from both D S and D T * , and unlabeled instances from D T . In this section, we describe the datasets we used, the generation of the input embeddings, the nature of the pairwise features, and the general training setup of our model. SemEval-2016 Task 3 We further collected 221 additional original questions and 1,863 related questions as unlabeled data, and we got the 221 English questions translated to Arabic. We used the TED We trained Arabic-English cross-language word embeddings from the concatenation of these bitexts using bivec In addition to the embeddings, we also used some pairwise features that model the similarity or some other relation between the input question and the potentially related questions. In particular, we calculated the similarity between the two questions using machine translation evaluation metrics, as suggested in We further used as features the cosine similarity between question embeddings. In particular, we used (i) 300-dimensional pre-trained Google News embeddings from Furthermore, we computed various task-specific features, most of them introduced in the 2015 edition of the SemEval task by This includes some question-level features: (1) number of URLs/images/emails/phone numbers; (2) number of tokens/sentences; (3) average number of tokens; (4) type/token ratio; (5) number of nouns/verbs/adjectives/adverbs/ pronouns; (6) number of positive/negative smileys; (7) number of single/double/ triple exclamation/interrogation symbols; (8) number of interrogative sentences (based on parsing); (9) number of words that are not in WORD2VEC's Google News vocabulary. Also, some question-question pair features: (10) count ratio in terms of sentences/tokens/nouns/verbs/ adjectives/adverbs/pronouns; (11) count ratio of words that are not in WORD2VEC's Google News vocabulary. Finally, we also have one meta feature: (12) reciprocal rank of the related question in the list of related questions. We trained our CLANN model by optimizing the objective in Equation (7) using ADAM (Kingma and Ba, 2015) with default parameters. For this, we used up to 200 epochs. In order to avoid overfitting, we used dropout Below we present the experimental results for the unsupervised and semi-supervised language adaptation settings. We compare our cross-language adversarial network (CLANN) to a feed forward neural network (FNN) that has no adversarial part. Table We conclude that the adversarial component in the network does the expected job, and improves the performance by focusing the languageindependent features in the representation layer. The scatter plots in Figure We now study the semi-supervised scenario when we also have some labeled data from the target language, i.e., where the original question q is in the target language. This can be relevant in practical situations, as sometimes we might be able to annotate some data in the target language. It is also an exploration of training with data in multiple languages all together. To simulate this scenario, we split the training set in two halves. We train with one half as the source language, and we use the other half with the target language as extra supervised data. At the same time, we also use the unlabeled examples as before. We introduced the semi-supervised model in subsection 3.2, which is a straightforward adaptation of the CLANN model. Table There are several interesting observations that we can make about Table These results demonstrate the effectiveness and the flexibility of our general adversarial training framework within our CLANN architecture when applied to a cross-language setting for questionquestion similarity, taking advantage of the unlabeled examples in the target language (i.e., when using unsupervised adaptation) and also taking advantage of any labeled examples in the target language that we may have at our disposal (i.e., when using semi-supervised training with input examples in the two languages simultaneously). We have studied the problem of cross-language adaptation for the task of question-question similarity reranking in community question answering, when the input question can be either in English or in Arabic with the objective to port a system trained on one input language to another input language given labeled data for the source language and only unlabeled data for the target language. We used a discriminative adversarial neural network, which we trained to learn task-specific representations directly. This is novel in a cross-language setting, and we have shown that it works quite well. The evaluation results have shown sizable improvements over a strong neural network model that uses simple projection with cross-language word embeddings. In future work, we want to extend the present research in several directions. For example, we would like to start with monolingual word embeddings and to try to learn the shared cross-language representation directly as part of the end-to-end training of our neural network. We further plan to try LSTM and CNN for generating the initial representation of the input text (instead of simple averaging of word embeddings). We also want to experiment with more than two languages at a time. Another interesting research direction we want to explore is to try to adapt our general CLANN framework to other tasks, e.g., to answer ranking in community Question Answering
| 686 | 2,460 | 686 |
We Understand Elliptical Sentences, and Language Models Should Too: A New Dataset for Studying Ellipsis and its Interaction with Thematic Fit
|
Ellipsis is a linguistic phenomenon characterized by the omission of one or more sentence elements. Solving such a linguistic construction is not a trivial issue in natural language processing since it involves the retrieval of non-overtly expressed verbal material, which might in turn require the model to integrate human-like syntactic and semantic knowledge. In this paper, we explored the issue of how the prototypicality of event participants affects the ability of Language Models (LMs) to handle elliptical sentences, and to identify the omitted arguments at different degrees of thematic fit, ranging from highly typical participants to semantically anomalous ones. With this purpose in mind, we built ELLie, the first dataset composed entirely of utterances containing different types of elliptical constructions, and structurally suited for evaluating the effect of argument thematic fit in solving ellipsis and reconstructing the missing element. Our tests demonstrated that the probability scores assigned by the models are higher for typical events than for atypical and impossible ones in different elliptical contexts, confirming the influence of prototypicality of the event participants in interpreting such linguistic structures. Finally, we conducted a retrieval task of the elided verb in the sentence in which the low performance of LMs highlighted a considerable difficulty in reconstructing the correct event.
|
A key phenomenon of natural languages is ellipsis, the omission of a word or phrase that is expected to occupy a place in the syntactic structure of a sentence (1) The engineer completed the project, but the student didn't. Since ellipsis represents a deviation from the simple compositional mapping between form and meaning, elliptical sentences have been the focus of many studies that seek to investigate how ellipsis is mentally represented, how the interpretation of the elided material is recovered, and consequently, how meaning can arise in the absence of form With those questions in mind, we explore the issue of how the prototypicality of event participants affects LMs in handling elliptical sentences, and whether these models are able to identify the omitted elements at different degrees of thematic fit. Our contribution to these issues is the creation of ELLie, The paper is organized as follows. Section 2 discusses previous works in this specific research area. Section 3 presents the design and structure of ELLie. In Section 4, we discuss the experiments conducted with the LMs on ELLie. Section 5 reports and discusses the results, while Section 6 shows how these can lead to further research.
|
Ellipsis is a relatively understudied problem in the Natural Language Processing (NLP) literature, given the difficulty of its resolution and the scarcity of benchmarks for the task. However, the phenomenon is widely recognized as an important source of errors in tasks such as dialogue understanding and machine translation Finally, Psycholinguistics and in NLP Thematic fit is a notion introduced in a series of psycholinguistic studies investigating the effects of event-based priming in online sentence processing Thematic fit has quickly become a hot topic also in NLP, and it was tackled either with unsupervised, vector-based approaches Thematic fit can be estimated for given arguments in a sentence, by computing their typicality score for the semantic role of the verb given the arguments already realized in the sentence (e.g., the system is asked to output the typicality of the patient instrument for the verb play, given the agent musician in The musician played an instrument). Since the earlier works A recent work by To the best of our knowledge, ELLie is the first dataset created to explore the complexity of the ellipsis phenomenon and its relation with thematic fit. Its structure was conceived to include multiple types of elliptical constructions, covering different thematic roles, and with the omitted elements (i.e., the verb or the whole verb phrase) having different degrees of thematic fit with the arguments in the context. The dataset is useful to investigate to what extent computational models encode the structured semantic information necessary for ellipsis resolution, and use it to make an accurate representation of the event context. After a preliminary study of the main English elliptical constructions presented in ELLie includes the following elliptical constructions presented in Culicover and Jackendoff (2005): 8 • Verb-phrase ellipsis (VP-ellipsis): The photographer used the camera, and the reporter did too. • Do-x anaphora: The cook washed his hands before cooking, and so did the doctor before the surgery." 5 The only exceptions were sluicing/sluice-stranding and a few other cases of other elliptical constructions. 6 DTFit's role fillers have already been judged by humans as strongly typical or atypical by using human-elicited typicality ratings (i.e., a 7-point Likert scale). 7 We manually checked that such violations do not have any kind of metaphorical reading and they are totally invalid from a semantic perspective. 8 The examples are taken from the ELLie dataset. • Gapping: "The businessman is reading the report, and the customer the menu." • Pseudo-gapping: "The child will drink the coke, and the student will the coffee." • Sluicing: "I know the electrician is checking something, but I don't know what." • Sluice-stranding: "The cook flipped the pancake with something, but I didn't know what with." ELLie is structured into five sub-dataset corresponding to different thematic roles: and The dataset is organized in blocks of five sentences (i.e., quintuplets), each composed by an antecedent clause and an elliptical part, like in (1). Each sentence in a block differs from the other ones only for two elements: the candidate fillers of a given thematic role in both the antecedent and the elliptical clauses. These sentences represent five alternatives through which we analyze the typicality condition of the event's participants (namely the argument filler in the antecedent and the elliptical one selected by the verb) according to different degrees of thematic fit, including highly typical arguments (T condition), atypical arguments (AT condition), up to semantic anomalous ones that violates selectional preferences (SP_v condition). Table The dataset is balanced from a structural point of view, as we aimed at using an equal number of quintuples for each sub-dataset and, where possible, the same number of elliptical constructions. The structure of ELLie is reported in Table We used ELLie as an evaluation dataset to test two Transformer-based LMs and analyze their behavior with elliptical constructions. Models. We chose to use two pre-trained models available in the Transformers library on Hugging Face, GPT-2. BERT. All the analyses were conducted using the Minicons library Task 1: Sentence typicality score We tested whether models can distinguish the most typical events from the atypical and/or implausible ones in elliptic constructions. As this presupposes that a model is able to identify that the missing element in the elliptical clause must be identical to the one overtly expressed in the antecedent, this task can be regarded as a sort of indirect test of the The journalist writes an article, and the professor a book. T -T The journalist writes an article, and the professor a magazine. T -AT The journalist writes a song, and the professor a book. AT -T The journalist writes a song, and the professor a magazine. AT -AT The journalist writes an article, and the professor an apple. T -SP_v 115 575 1 Sluicing class also includes the sluice-stranding construction. models' ability in ellipsis resolution. For each sentence in a block we computed its probability score. Before that, we did a further preliminary check by carrying out a normalization based on the number of tokens, to make sure that the results were not affected by the number of tokens into which a sentence is split. Conversely, Minicons library adopts the Pseudo-log-likelihood score (PLL) when using BERT, since the probability of a sentence cannot be computed using this autoencoding model, given its bidirectional architecture. This score is obtained by masking one token at a time, calculating the token's probability given its left and right context, and then summing the logprobabilities for all the tokens The second task is a double dynamic thematic fit evaluation and consists in recovering the probability assigned by the models to the candidate fillers of the antecedent clause and the elliptical one. Their typicality score is represented by this probability value. So, we retrieved the specific position of each candidate filler analyzing the tokenization's results both with the GPT-2 tokenizer and with the BERT one. The photographer used the camera, and the reporter did too. b. Prompt GPT-2: The photographer used the camera, and the reporter did too. What the reporter did was c. Prompt BERT: The photographer used the camera, and the reporter did too. What the reporter did was [MASK] the camera. Then, GPT-2 was evaluated on a text-generation task and BERT on a fill-mask task. Performance was measured with verb retrieval accuracy, computed as the number of times the models were able to retrieve the target verb, which was identified via regular expressions. GPT-2 was tested in two different configurations referring to distinct decoding methods. Both of them involve the generation of new tokens, but one exploits GPT's sampling technique and the other one does not. In the former configuration, we used the top-p (nucleus) sampling method, setting the seed to reproduce the results. We generated the top-3 sentences in which only tokens with probabilities that add up to top-p = 0.92 or higher (given the previous words) are kept for generation. If the target verb was present in at least one of three generated sentences, then the model scored an accuracy hit. 16 The other configuration simply retrieved the most likely sentence doing a greedy search without sampling. We decided to use also this decoding method because it is the same used by BERT. In addition, we evaluated GPT-2 performance also in retrieving the direct object. For the fill-mask task, we masked instead the target verb in the prompt and took the most likely words predicted by BERT to replace that mask. We report here the results of the experiments carried out on ELLie. Figures (3) Prompt: The butcher used the knife, and the soldier did too. What the soldier did was GPT-2 answer: to cut the meat into Correct answer: (to) use the knife Apparently it might prove that the model really understood the ellliptic sentence, but it is instead likely that such LMs still tend to rely on frequent verb-argument co-occurences previously observed during training (to cut the meat is a typical verb-object combination given the subject butcher), rather than constructing and updating contextual information about an event (see also the error analysis sections in These results prove that the prototypicality of event participants affects the way such linguistic constructions are managed by the two models. Notice that almost all the higher scores both in GPT-2 (only for verb-retrieval) and BERT correspond to the typicality condition in which the elliptical clause contains a typical filler (T-T and AT-T). This means that models struggle to retrieve the verb more when the prompt describes an event with atypical or semantically impossible participants. Finally, since evidence from prompting tasks has proved that even minimum changes inside the prompt could lead to different results, we decided to conduct a pilot experiment on a subset of cases (4) a. Prompt GPT-2: The photographer used the camera, and the reporter did too. The reporter b. Prompt BERT: The photographer used the camera, and the reporter did too. The reporter [MASK] the camera. The idea is that such a structure should facilitate the model since we directly present the elliptical agent without the presence of any indirect interrogative proposition as in (2). Unexpectedly, the results were quite disappointing: GPT-2 improved by only 2/3 points compared to the values obtained over the entire dataset with the previous prompts, but BERT dropped by 20 points. Ellipsis is a complex phenomenon that has always been at the center of the debate in theoretical linguistics (5) a. The photographer used the camera, and the reporter did too b. *The photographer used the camera, and the piano did too In (5), the expression did too is a signal that the verb phrase of the elliptic clause is used the camera. In particular, the reconstructed material must preserve the semantic constraints of its overt "copy": (5b) is anomalous because piano violates the selectional preferences of the verb in the antecedent. What do LMs know about such key features of ellipsis? Our experiments suggest that, at least in the tested models, this knowledge is still quite limited. The fact that in Task 1 the models are not able to distinguish between atypical and impossible sentences is a sign that they cannot reconstruct correctly the implicit elements from the antecedent. Since current LMs are quite good at this task when event typicality and impossibility are tested in main clauses In this paper, we proposed a new framework to evaluate ellipsis and its relationship with thematic fit and selectional preferences. We did this by creating ELLie, the first dataset composed of elliptical utterances and structurally suited for estimating the effect of argument thematic fit in solving ellipsis. We tested two LMs with a Transformer-based architecture in three different tasks to understand whether their ability to process elliptical constructions is affected by argument typicality and event knowledge. Experimental results suggest a limited mastery of elliptical sentences and a significant influence of prototypicality of event's participants. Moreover, the tested models greatly struggle to recover the missing elements of elliptical clauses and, thus, to reconstruct the whole event context. Their performance (especially in Task 3) may also depend on the low occurrence of such constructions in the training corpora, since the ellipsis phenomenon tends to be more frequent in speech than in writing. Finally, the influence of event typicality suggests that LMs tend to rely on frequent lexical co-occurrences, without being able to reconstruct the implicit syntactic and semantic structure necessary to interpret elliptical sentences. The findings reported in this paper have to be seen in light of some limitations and, therefore, they just represent a first step. Most of these limitations are related to the ELLie dataset itself. First of all, though the predicate-argument combinations used in ELLie come from the DTFit dataset and were rated by humans, still the elliptical sentences need human judgements, Concerning the experiments, some changes could be made in the evaluation of Task 3. First, we could test the prompts in (4) on the subsets for the other roles, and look for different prompt structures to see if this leads to performance changes. We could also adopt a softer evaluation for this task, by assessing the output in terms of similarity to the target answer. Finally, another limitation is related to the strong dependence of our results to the language used for the analysis (i.e., English). From this point of view, a cross-linguistic study on the elliptical structures in ELLie could contribute to improve our work from both a theoretical and practical perspective. PE00000013 -«FAIR -Future Artificial Intelligence Research» -Spoke 1 «Human-centered AI», funded by the European Commission under the NextGeneration EU programme.
| 1,433 | 1,215 | 1,433 |
TaxiNLI: Taking a Ride up the NLU Hill
|
Pre-trained Transformer-based neural architectures have consistently achieved state-of-theart performance in the Natural Language Inference (NLI) task. Since NLI examples encompass a variety of linguistic, logical, and reasoning phenomena, it remains unclear as to which specific concepts are learnt by the trained systems and where they can achieve strong generalization. To investigate this question, we propose a taxonomic hierarchy of categories that are relevant for the NLI task. We introduce TAXINLI, a new dataset, that has 10k examples from the MNLI dataset (Williams et al., 2018) with these taxonomic labels. Through various experiments on TAXINLI, we observe that whereas for certain taxonomic categories SOTA neural models have achieved near perfect accuracies-a large jump over the previous models-some categories still remain difficult. Our work adds to the growing body of literature that shows the gaps in the current NLI systems and datasets through a systematic presentation and analysis of reasoning categories.
|
The Natural Language Inference (NLI) task tests whether a hypothesis (H) in text contradicts with, is entailed by, or is neutral with respect to a given premise (P) text. This 3-way classification task, popularized by 2019) to achieve state-of-the-art (SOTA) performance in these tasks. Despite the wide adoption of NLI datasets, a growing concern in the community has been the lack of clarity as to which linguistic or reasoning concepts these trained NLI systems are truly able to learn and generalize (see, for example Herein lie the central questions of our work: 1) what is the distribution of various categories of reasoning tasks in the NLI datasets? 2) which categories of tasks are rarely captured by current NLI datasets (owing to the nature of the task and the non-expert annotators)? 3) which categories are well-understood by the SOTA models? and 4) are there categories where Transformer-based architectures are consistently deficient? In order to answer these questions, we first discuss why performance-specific error analysis categories 2 A New Taxonomy for NLI 2.1 Necessity for a New Taxonomy According to Existing categorization efforts have centred around informing feature creation in the pre-Transformer era, and model-specific error analysis in more recent times. Previously, (LoBue and Yates, 2011) enumerated the type of commonsense knowledge required for NLI. Among recent error analysis efforts, the GLUE diagnostic dataset
|
In Figure Next we define the categories. For a full set of examples, please see Table High-Level Categories: The Linguistic category represents NLI examples where the inference process to determine the entailment are internal to the provided text. We classify examples as Logical when the inference process may involve processes external to text, such as mapping words to percepts and reason with them (Sowa, 2010). where some form of external, domain or commonly assumed knowledge is required for inferencing. Linguistic category is further sub-divided into lexical, syntactic, and factivity. 1. Lexical: This category captures P-H pairs where the text is almost the same apart from removal, addition or substitution of some lexical items. Example: P: Anakin was kind. H: Anakin was cruel. 2. Syntactic: Syntactic deals with examples where syntactic variations or paraphrases are essential to detecting entailment. Example: P: Anakin was an excellent pilot. H: The piloting skills of Anakin were excellent. 3. Factivity: Here the hypothesis contains an assumed fact from the premise, mostly an assumption about the existence of an entity or the occurrence of an action (inspired from We also consider language-inspired reasoning categories such as co-reference resolution, which is known to often require event-understanding We present TaxiNLI, a dataset collected based on the principles and categorizations of the aforementioned taxonomy. We curate a subset of examples from MultiNLI Task Design For large-scale data collection, our aim was to propose an annotation methodology that is relatively flexible in terms of annotator qualifications, and yet results in high quality annotations. To employ non-expert annotators, we designed a simplified guideline (questionnaire/interface) for the task, that does not pre-suppose expertise in language or logic. As an overhead, the guideline requires a few rounds of one-on-one training of the annotators. Because it is expensive to perform such rounds of training in most crowdsourcing platforms, we hire and individually train a few chosen annotators. Upon conducting the previouslydiscussed pilot study and using the given feedback, we created a hierarchical questionnaire which first asked the annotator to do the NLI inference on the P-H pair, and then asked targeted questions to get the desired category annotations for the datapoints. The questionnaire is shared in the Appendix. For the MNLI datapoints with 'neutral' gold labels, we realized, through observation and annotator feedback, that annotating the categories were difficult, as sometimes the hypotheses could not be connected well back to their premise. Hence, we created 2 questionnaires, one for the 'entailment/contradiction' examples, and one for 'neutral' examples. For the entailment/contradiction examples, We collected binary annotations for each of the 15 categories in our NLI taxonomy, for datapoints in MNLI which had 'entailment' or 'contradiction' as gold labels. To resolve this, for the 'neutral' examples we specifically asked them whether the premise and hypothesis were discussing 1) the same general topic (politics, geology, etc.), and if so, 2) had the same subject and/or object of discussion (Obama, Taj Mahal, etc.). If the response to 2) was 'yes', then they were asked to provide the category annotations as previously defined. Annotator Training/Testing We first tested our two annotators by asking them to do inference on a set of randomly selected premise-hypothesis pairs from MultiNLI. This was to familiarize them with the inference task. After giving the category annotation task, we also continuously tested and trained the two annotators. After a set of datapoints were annotated, we reviewed and went through clarification and feedback sessions with the annotators to ensure they understood the task, and the categories, and what improvements were required. More details are provided in the Appendix. Here, we assess the individual annotator performance and inter-annotator agreement. Since, automated metrics for individual complex category annotations are hard to define, we use an indicative metric that matches the annotated inference label with the gold label, i.e., their inference accuracy. We also calculated inter-annotator agreement between the two annotators for an overlapping subset of 600 examples. For agreement, we use the Fleiss' Kappa (κ) , where C j Ai are the category annotations for Annotator i for datapoint j, averaged over total datapoints N . Looking at the category-wise Fleiss' κ values in Fig. Each datapoint in TaxiNLI Figure We re-evaluate two Transformer-based and two standard baseline machine learning models on TAXINLI, under the lens of the taxonomic categories. As baselines, we choose BERT-base As pre-Transformer baselines, we use the bidirectional LSTM-based Enhanced Sequential Inference model (ESIM) We report the NLI task accuracy of the baseline systems on the MNLI validations sets in Table The systems are fine-tuned on the MNLI training set using the procedures followed in We evaluate the systems on a total of 7.7k examples, which are in the intersection of TAXINLI and the validation sets of MNLI. Figure . We observe that compared to NB, the improvements in BERT have been higher in lexical, syntactic categories compared to others. Improvements in ESIM compared to NB show a very similar trend, and show for knowledge categories the improvements are negligible. ESIM shows largest improvement on negation. In order to quantify the precise influence of the category labels on the prediction of the NLI models, we probe into indicators and confounding factors using two methods: linear discriminant analysis (LDA) and logistic regression (LR). We use indicators for each category (0 or 1) and for two potential confounding variables (lengths of P,H), to model the correctness of prediction of the NLI system. The coefficients of these analyses on BERT are shown in Fig 6 . The values for RoBERTa follow a similar trend, and are presented in the appendix. We see that presence of certain taxonomic categories strongly influence the correctness of prediction. As we found in the analysis presented in Sec. 4, we observe that syntactic, negation, and spatial categories are strong indicators of correctness of prediction. On the other hand, conditional, relational, causal, coreference are harder to predict accurately. Sentence length does not play a significant role. We also make an observation for categories such as lexical, syntactic, where the proportion of a single NLI label is high, also correlated with a high prediction accuracy (Fig. Visual Analysis Section 4 paints a thorough picture by analysing the fine-grained capabilities of SOTA NLI systems at a behavioral In light of recent probing task literature The layer-wise separation of examples by taxonomy raises an interesting possibility to motivate model architectures that may attempt to use its discriminative power to identify such taxonomic categories, for specialized treatment to examples requiring certain reasoning capabilities. Recasting: The under-representation of certain categories in the MNLI dataset raises a need for more balanced data collection. A possible alternative is to build recast diagnostic datasets for each category, and create probing tasks. Some datasets To bridge the gap between accuracy-led performance measurement and linguistic analysis of stateof-the-art NLI systems, we propose a taxonomic categorization of necessary inferencing capabilities for the NLI task, and a re-evaluation framework of systems on a re-annotated NLI dataset using this categorization; which underscores the reasoning categories that current systems struggle with. We would like to emphasize that unlike the case with challenge and adversarial datasets, TAXINLI re-annotates samples from existing NLI datasets which the SOTA models have been exposed to. Therefore, a lower accuracy in certain taxonomic categories in this case cannot be simply explained away by the "lack of data" and "unnatural distribution" arguments. A Other Categorizations Since, examples are annotated with multiple categories, we capture the dependencies by defining a Bayesian Network (BN) where, each category (a boolean random variable) has a directed edge to the correct node (representing correctness of prediction) In Fig. Our annotation process went through several steps of refinement and improvement. We started with the most basic annotation flow, which was to have a manual which defines each taxonomic category in detail, and then have the annotator mark for each category. For the pilot study, we took roughly 300 examples from MNLI and asked an initial annotator to annotate. The feedback was the following: • The manual describing each taxonomic category had a lot of information and took time to understand and digest. • It was difficult to keep referring to the guide, although after sufficient examples, it became easier. • There was confusion and ambiguity about the definitions, and the annotator interpreted the definitions differently than what we intended. • Figuring out the categorical annotations for neutral examples was a challenge, as sometimes the topic or subject of what the hypothesis was discussing was separate from what the premise was discussing. Through the analysis of these annotations, we also observed that some of the initial categories we had were either exceedingly underrepresented in the MNLI dataset, or were consistently confused with others. Thus, we revised the set of categories, setting more distinct boundaries, and ensuring independence of categories. We revised the questionnaire into a hierarchical 'if-else' multi-choice design. The questionnaire is structured as follows: S2 is a paraphrase of S1 or a certain part of information mentioned in S1 • S1: Anakin was an excellent pilot. • S2: The piloting skills of Anakin were excellent. This is TRUE, and S1 and S2 being paraphrases of one another. Also, to note, if 'excellent' in S2 were replaced by 'terrible', it would still fit this category, but would also fit category 6a, since it would be a paraphrase with a swapped word. (c) S2 contains an assumed fact from S1, mostly an assumption about the existence or the occurrence of an action. • S1: Anakin found the Death Star. • S2: The Death Star exists. This is TRUE. The Death Star exists if Anakin has found it, thus S1 makes the assumption that it exists. • S1: James was happy that his plane could fly. • S2: His plane couldn't fly. This is FALSE. Since James was happy that the plane flew (S1 makes the assumption that it happened), it is FALSE that his plane couldn't fly since it happened." The above questionnaire was given along with premise-hypothesis pairs having the gold label of 'entailment' or 'contradiction' . However, to prevent biasing the annotator, we allowed them to choose 'neutral' (CAN'T SAY) as well. This questionnaire was given to the annotators after they had done a sufficient number of 'entailment/contradiction' samples using Questionnaire 1. For Questionnaire 2, annotators were told that the datapoints were 'neutral', and asked them to first answer these 3 questions: "Given S1, there isn't enough information to decide whether S2 is TRUE or FALSE. Please answer the following questions for each datapoint which has been annotated as CAN'T SAY. Upon answering the above, if the answer to the second question was yes, then they proceeded with the category annotation, else they moved on to the next question. This helped eliminate the random hypotheses. We received a lot of important feedback from our annotators during the clarification and training sessions. They are listed below: • Many premise sentences seems out of place, and the context is still insufficient many times. As a result, the hypothesis also introduces ambiguity, making the process a bit tricky. • There were cases where a certain name of an entity in the premise is switched for something else in the hypothesis. This created some confusion because it fell somewhere between lexical and coreference (according to the annotator). • Another confusion arose from the quantifier category, where initially the name of the category led the annotators to believe that it referred to not just what we described (e.g. some, all), but quantities (say 5000 in the premise was swapped with 2000 in the hypothesis). This was again a middle ground between lexical and quantifier. • Many of the premises contained incoherent, difficult to understand sentences. A lot of premises (which we later found to be from the telephone category), contained many filler words (uh, uhm, etc.) which made comprehension difficult. • Another issue lies with an implicit rigidity of the annotation process using just the questionnaire. The targeted questions were written so as to allow annotators to generalize and apply intuitive principles along those thought lines that we try to demarcate via the questions. We wanted to prevent them completely However, as annotators have not been exposed to the exact intentions behind the annotation (so as to prevent bias), they followed the questionnaire strictly, and did not always generalize until subsequent training/clarification sessions where we encouraged them to generalize. However, the implicit rigidity still impacts the annotations to some extent, although mitigated to a large level by the training. This remains a challenge due to the tradeoff between open interpretation of the task, as well as a rigidity of desired annotations which stem from an analysis perspective (from our side). • Idiomatic references, metaphors, and common phrases were also a source of confusion, and although to some extent were marked as world knowledge, did leave annotators unsure about where to place them. The above feedback only strengthened our belief in an iterative training system for a complicated task such as this. It also sheds light on how difficult a task like this is to crowdsource.
| 1,031 | 1,451 | 1,031 |
CompoundPiece: Evaluating and Improving Decompounding Performance of Language Models
|
While many languages possess processes of joining two or more words to create compound words, previous studies have been typically limited only to languages with excessively productive compound formation (e.g., German, Dutch) and there is no public dataset containing compound and non-compound words across a large number of languages. In this work, we systematically study decompounding, the task of splitting compound words into their constituents, at a wide scale. We first address the data gap by introducing a dataset of 255k compound and noncompound words across 56 diverse languages obtained from Wiktionary. We then use this dataset to evaluate an array of Large Language Models (LLMs) on the decompounding task. We find that LLMs perform poorly, especially on words which are tokenized unfavorably by subword tokenization. We thus introduce a novel methodology to train dedicated models for decompounding. The proposed two-stage procedure relies on a fully self-supervised objective in the first stage, while the second, supervised learning stage optionally fine-tunes the model on the annotated Wiktionary data. Our self-supervised models outperform the prior best unsupervised decompounding models by 13.9% accuracy on average. Our fine-tuned models outperform all prior (language-specific) decompounding tools. Furthermore, we use our models to leverage decompounding during the creation of a subword tokenizer, which we refer to as CompoundPiece. CompoundPiece tokenizes compound words more favorably on average, leading to improved performance on decompounding over an otherwise equivalent model using SentencePiece tokenization.
|
Decompounding is the task of separating compound words into their single word constituents. Decompounding is used in user-facing tools such as dictionaries and morphological analyzers (Altinok, † Equal senior authorship.
|
). Historically, it has also been widely used as a preprocessing step for other NLP tasks, e.g. for information retrieval Decompounding can come in two similar yet different task formats: (i) compound segmentation and (ii) compound normalization (Ziering and van der Plas, 2016). Compound segmentation is the task of segmenting a word into its compound constituents, while preserving its surface form (e.g. bridesmaid → brides + maid). Compound normalization is the task of recovering the base form of each compound constituent (e.g. bridesmaid → bride + maid). Using our dataset, we then find that large language models (LLMs), which typically rely on subword-based tokenization In order to create a more effective decompounding model, we then formulate compound segmentation and normalization as a sequence-to-sequence learning task Our Stage 1 models outperform the best prior unsupervised models by 13.9% accuracy on average, while our (supervised) Stage 2 models outperform all prior language-specific decompounding tools. Furthermore, a model trained with a Com-poundPiece tokenizer achieves a 5.5% improved performance on compound normalization over an otherwise equivalent SentencePiece model. We introduce a dataset for decompounding of 255k words across 56 languages obtained from Wiktionary. 2) We show that a byte-level language model can efficiently decompound words via a two-stage training framework, whereas current subword-based LLMs fall short. 3) We present a way to improve subword tokenization by performing compound segmentation during creation of the tokenizer. 4) We make our code, models and dataset publicly available at github.com/bminixhofer/compoundpiece. Decompounding. Early work in decompounding used word frequency lists along with manually specified suffixes (e.g., a connective -s-) to segment and normalize German compounds While the approaches above use (at most) light supervision, there exist supervised approaches which learn directly from an annotated corpus of compounds and their constituents, along with optional auxiliary signals Relation to Morphological Segmentation. Decompounding can be seen as a special case of morphological segmentation Decompounding Datasets. Besides the SMST 2022 dataset, datasets for decompounding include AuCoPro Linguistically Informed Tokenization. Various studies have tried augmenting or replacing the 'linguistically uninformed' subword-tokenizers used in contemporary LMs Unsupervised morphological segmentation has not shown consistent improvements We use words categorized as compound terms on Wiktionary to create a dataset for decompounding. The information on Wiktionary allows associating compound terms with their corresponding normalized constituents. Since Wiktionary only annotates the top-level split, Note that this way of obtaining negative examples is biased against words which never occur inside compounds; however, we found this to be a rather weak bias (Appendix E). We include every language with at least 100 words, leading to a dataset which covers 56 languages. The number of training examples is shown in Figure To overcome the problem of data scarcity in lowresource languages, we introduce a two-stage training procedure for creating dedicated decompounding models. In Stage 1, we train on the selfsupervised objective of restoring hyphenation in words extracted from a large-scale Web corpus, leading to a self-supervised compound segmentation model. In Stage 2, we fine-tune the model on compounds and their normalized constituents from an annotated corpus in a supervised fashion, turning it into a compound normalization model. Stage 1: Self-Supervised Compound Segmentation. This stage is motivated by the fact that hyphen characters can be seen as a high-precision, lowrecall indicator of compound constituent boundaries, in the same way that newline characters are a high-precision, low-recall indicator of sentence boundaries Stage 2: Supervised Compound Normalization. In the second stage, we improve upon the Stage 1 model by additional training on labeled data, where the inputs are individual compounds, and the target is to predict the normalized constituents of each compound, separated by a hyphen. Training exclusively on compound normalization allows using data from the collected Wiktionary dataset, which contains compound terms along with their normalized constituents across many languages, but does not contain compound segmentation annotations. Considering the scarcity of annotated compound segmentation data, it is infeasible to train a multilingual model directly on segmentation. Thus, we introduce a method to predict a segmentation given the normalized constituents. Let x be a word of length n. In addition, we have k normalized com-pound constituents c = {c 1 , ..., c k } (e.g. predicted by the Stage 2 model). Our aim is to find boundaries r = {r 0 , ..., r k }, r 0 = 0, r k = n giving rise to the segmentation s = {x[r 0 : r 1 ], ..., x[r k-1 : r k ]}. We approach this problem by minimizing the edit distance of each segment to its corresponding normalized constituent. This leads to an optimization problem where the cost C(s) indicates the total edits needed to turn all segments into their corresponding normalized constituents: Here, L is an edit distance metric such as Levenshtein distance In case of ties, we prefer segmentations with higher edit cost for segments with lower indices due to the preference for languages in our training set for suffixation over prefixation We define hard compounds relative to a particular tokenizer as compound words where the constituent boundaries do not coincide with token boundaries set by the tokenizer. More formally, a compound word made up of k constituents and l subwords is hard if the constituent boundaries r = {r 0 , ..., r k } are not a subset of the token boundaries t = {t 0 , ..., t l } i.e. r ̸ ⊂ t. We hypothesize that hard compounds may impair language model performance due to the nontrivial relation of subwords to the compound word. In contrast, in easy compounds the word is naturally decomposed into its constituents. The increased difficulty of hard compounds is apparent on the sequence-to-sequence compound segmentation task: for an easy compound, all tokens can be copied to the output (only the special separator tokens must be inserted). On the other hand, for hard compounds, the tokens change, requiring knowledge of the characters within each token. Tokenizers where every possible constituent boundary is a token boundary trivially do not give rise to any hard compounds. This includes character-level Subword tokenizers typically segment text into pre-tokens (e.g. by splitting on whitespace) before applying their subword tokenization algorithm We obtain Stage 1 data by selecting all words containing a hyphen from a subset of the mC4 corpus We train a decompounding model using a two-stage framework ( §3) covering 56 languages. We use ByT5 Metric. We measure performance via averaged accuracy, i.e., the ratio of examples which are entirely correctly segmented or normalized. Datasets. Besides our new Wiktionary evaluation subset, we use the established datasets for particular languages: GermaNet Languages. For clarity of presentation, we present results on Danish, German, English, Spanish, Estonian, Greek, Persian, Finnish, Hungarian, Kazakh, Latvian, Dutch, Polish and Swedish as a linguistically diverse subset of languages with productive compound formation in the main paper. For the full evaluation across all languages, see Appendix C. Main compound segmentation results are shown in Table We further compare against language-specific and supervised methods in Table Evaluating LLMs on Decompounding. We also evaluate in-context learning performance of multiple LLMs on compound segmentation. We use T5 models with 770M, 3B and 11B parameters We use our fine-tuned ByT5 model for traintime pretokenization into compound constituents and SentencePiece To more thoroughly evaluate our tokenization, we train multilingual T5 models using Sentence-Piece and CompoundPiece. We use the same sampling ratio (α = 0.2) of mC4 as for creating the tokenizer, but instead use a subset of 500M texts. We match the architecture and the pretraining setup of the mT5-base model, but train for a total of 65.5B tokens. Ablation Studies. We quantify the impact of the most significant design choices of our model in Table We systematically investigated word decompounding tasks of compound segmentation and normalization on a wide scale and in multilingual contexts. To this end, we introduced a dataset of 255k words including compounds and non-compounds across 56 languages from Wiktionary, which allowed us to evaluate performance of LLMs on decompounding. We found that current LLMs' performance is limited due to hard compounds which arise when subword token boundaries do not coincide with compound constituent boundaries. We then introduced dedicated models for decompounding which use byte-level tokenization to entirely avoid hard compounds. Finally, we used our decompounding models to create novel CompoundPiece tokenizers, keeping the efficiency advantages of subword tokenization while strongly decreasing the amount of hard compounds; this increases the performance of CompoundPiece models over comparable Senten-cePiece models on the decompounding tasks. Although self-supervised training in Stage 1 allows for decompounding without any annotated training data, Stage 2 training is limited to languages with sufficient entries in Wiktionary: this excludes extremely low-resource languages. Furthermore, due to computational constraints we have not trained larger models using CompoundPiece tokenization; hence we are unable to report on its benefits at larger scales and on tasks besides decompounding. Patrick Ziering and Lonneke van der Plas. 2016. Towards unsupervised and language-independent compound splitting using inflectional morphological transformations. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 644-653, San Diego, California. Association for Computational Linguistics. Statistics for the training and validation splits of the Wiktionary dataset are shown in Table Pseudocode of the brute-force algorithm to turn normalization into segmentation is shown in Algorithm 1. Since enumerating all possible segmentations is only feasible for short words ( §3.3) we introduce a more efficient algorithm (Algorithm 2) where candidate segmentations are ordered such that segmentations with constituents closest in length to the corresponding normalized constituents appear first. Assuming insertions and deletions both have a cost of one (as is the case in standard Levenshtein distance), constituents are thus sorted in increasing order of a lower bound on edit distance. The procedure can stop once the lower bound on edit distance reaches the cost of the best solution found so far since by that point it is impossible for a better solution to be found. Note that the normalization-to-segmentation problem is related to sequence partitioning Segmentation accuracy for all languages is shown in Tables The prompt used for LLM evaluations ( §5) is shown in Figure We conduct an experiment to measure the extent of the bias against words which do not occur inside compounds in our data collection methodology ( §3.1). In particular, we quantify the bias against long non-compound words, which usually would not occur inside compounds. We took a random sample of 500 words each from word frequency lists in English and German While words in our non-compound dataset are indeed shorter on average (6.0 vs. 6.7 chars for English, 6.7 vs. 7.1 chars for German), with less than one character length difference on average, there is only a weak length bias in data collection. We also found qualitatively that our noncompound dataset contains a wide variety of words since compounding is typically a process that can occur for many different root words.
| 1,643 | 220 | 1,643 |
Understanding Factual Errors in Summarization: Errors, Summarizers, Datasets, Error Detectors
|
The propensity of abstractive summarization models to make factual errors has been studied extensively, including design of metrics to detect factual errors and annotation of errors in current systems' outputs. However, the everevolving nature of summarization systems, metrics, and annotated benchmarks makes factuality evaluation a moving target, and drawing clear comparisons among metrics has become increasingly difficult. In this work, we aggregate factuality error annotations from nine existing datasets and stratify them according to the underlying summarization model. We compare performance of state-of-the-art factuality metrics, including recent ChatGPT-based metrics, on this stratified benchmark and show that their performance varies significantly across different types of summarization models. Critically, our analysis shows that much of the recent improvement in the factuality detection space has been on summaries from older (pre-Transformer) models instead of more relevant recent summarization models. We further perform a finer-grained analysis per error-type and find similar performance variance across error types for different factuality metrics. Our results show that no one metric is superior in all settings or for all error types, and we provide recommendations for best practices given these insights. 1
|
Although abstractive summarization systems In this paper, we provide a new benchmark that allows for finer-grained comparison between different factuality systems. We aggregate 9 existing annotated factuality datasets to create our benchmark AGGREFACT. We stratify it according to the underlying summarization model, categorized into FTSOTA, EXFORMER and OLD based on their development timeline (see Section 2). First, we ask: do factuality metrics perform equally well at identifying errors from state-of-the-art summarization models and from earlier models? For nine recent factuality metrics, including recent ChatGPT-based metrics, we show that metric performance varies substantially between different categories of summarization models. Most importantly, we found that the standard way of reporting improvements on category-agnostic benchmarks can be misleading, as most of these gains are on the OLD or EXFORMER subset of the data which are less important to detect. On summaries generated by FTSOTA models, we found that there is no single metric that is superior in evaluating summaries from both the CNN/DM To better understand their behavior, we next analyze what error types are different factuality metrics capable of identifying (Section 4). To do this, we leverage datasets from our benchmark that have fine-grained error annotations and unify these into a single taxonomy. We find that the error type distribution changes over time and even differs between annotations of the same summarization models across factuality datasets. Analysis of the factuality metrics shows that metrics claim-
|
Kappa Gran Annotation Scheme FactCC 1. Evaluate factuality metrics on summaries generated by the state-of-the-art summarization models. We found generally worse performance when evaluating factuality systems on summaries generated by FTSOTA models instead of less recent models (Section 3). We release AGGREFACT to support this, which combines existing benchmarks and stratifies them according to the base summarization model, summarization dataset and error types. We suggest future work to augment our benchmark with LLM-generated summaries, e.g. from ChatGPT, which is beyond the scope of this paper. 2. Choose an appropriate factuality metric for your downstream task at hand. No one metric is superior across all settings (Section 4). Fine-grained insights offered by our benchmark can be useful to compare strengths of different factuality metrics and make this choice. 3. Annotate error types consistently with prior work for better comparability. We found that error type boundaries in existing works are not clear and are not easy to leverage for crossdataset metric comparisons (Section 4). We hope that our analysis can shed light on what comparisons practitioners should focus on, how to understand the pros and cons of different metrics, and where metrics should go next. Further, we hope that future work would extend this to incorporate diverse summarization domains such as dialogue summarization Current factuality metrics are evaluated without considering the types of summarization models used to generate the annotated summaries. In these annotated datasets, a large proportion of summaries are generated by older models, such as a pointergenerator network Summarization datasets splits We introduce a new benchmark AGGREFACT built on top of Sum-maC from All datasets contain summaries generated from articles in CNN/DM and XSum. Given the unique characteristics of CNN/DM and XSum, our proposed benchmark includes two subsets, AGGREFACT-CNN and AGGREFACT-XSUM, that evaluate the performance of factuality metrics on these two datasets separately (Table Our benchmark formulates factual consistency evaluation as a binary classification task, following Summarization model splits To validate our hypothesis and make a careful comparison of factuality metrics, we further divide models that were used to generated summaries in the benchmark into three distinct categories: C = { FTSOTA, EXFORMER, OLD }, as seen in Table The SUMMAC benchmark In addition to modifying SUMMAC, we further include four annotated datasets. For Wang'20 Deduplication and label disagreement correction Some examples may be labeled for errors in multiple datasets. We removed all duplicates so that each instance appears only once in our benchmark. During this deduplication process, we detected 100 instances of the same summaries that are annotated in different datasets with different factual consistency labels. 98 of them are between FRANK and XSumFaith, and 2 of them are between FRANK and SummEval. The authors of this work manually corrected the labels for these examples based on our judgment. We use balanced accuracy to evaluate the performance of factuality metrics due to the imbalance of factually consistent and inconsistent summaries. We refer readers to First, we evaluate several SOTA factual consistency metrics on our benchmark, namely DAE Unifying these metrics We consider each metric as a function f (d, s) → y, mapping each (document, summary) pair to a score y ∈ R. We convert each method into a binary classifier f ′ (d, s) → {0, 1} by picking a threshold t such that we predict 1 if f (d, s) > t and 0 otherwise. 3 All thresholds are set separately for each metric. We consider two ways of setting the threshold for a metric: threshold-per-dataset and singlethreshold. The first setting has thresholds {t m d,c } within each metric for every dataset we consider, where d, c and m are any dataset in D, any model category from C, and any factuality metric, respectively. This allows one to choose the right metric for the task at hand. The single-threshold setting defines one threshold {t m } per metric. We analyze scores from factuality metrics using chosen thresholds {t m d,c } from the validation sets. Specifically, for each factuality 2 We do not consider other common metrics like ROUGE 3 CHATGPT-ZS and CHATGPT-COT do not require thresholds as they directly predict factual consistency labels. metric, we average the values of thresholds for each of SOTA, EXFORMER and OLD across all datasets (Figure The higher threshold values on CNN/DM are connected to both the nature of the errors involved and overall extractiveness of the summaries XSum summaries are more abstractive and tend to contain a larger number of errors, making it harder for the metrics to verify the consistency of summaries with respect to the source text and resulting in lower scores in general, even for factual cases. For CNN/DM, smaller deviations from the source may indicate non-factuality. A weighted average of performance in terms of balanced accuracy for AGGREFACT-CNN and AGGREFACT-XSUM is shown in Table Our results show that for AGGREFACT-CNN, both trained and ChatGPT-based factuality metrics achieve the best performance in evaluating the summaries in OLD. This result is intuitive: the sum- maries in OLD contain obvious errors, such as repetition, that can be more easily detected compared to more nuanced errors made by more recent models. From Table Binary Classification: FTSOTA To encourage comparison of factuality metrics on FTSOTA summaries, we provide a separate benchmark which consists of two subsets AGGREFACT-CNN-FTSOTA and AGGREFACT-XSUM-FTSOTA that only consider summaries generated by FTSOTA models. This benchmark consists of validation and test splits from the FTSOTA subsets of the two datasets. This setting allows for comparisons of metrics to be made using only a single threshold. We show metric comparisons on the FTSOTA subset in Table Interestingly, metrics such as SummaC-Conv, QAFactEval and the recent ChatGPT metrics were all proposed as improved factuality evaluation on the category-agnostic SummaC benchmark (different from the SummaC metric). However, our stratified analysis provides a much clearer picture and shows that metrics which claim improved performance on SUMMAC do not show similar gains when evaluated on FTSOTA summaries. We recommend that future work similarly focuses on the SOTA category of generated summaries when comparing factuality metrics. Having established differences among factuality metrics across underlying summarization models, we now explore differences in metrics according to factuality error types. To do this, we need a way to unify error types across datasets in our benchmark and map them into a shared taxonomy. We surveyed existing error type taxonomies in prior work and unified the types of factual errors among them into a hierarchical taxonomy in Figure Descriptions and examples for these error types are in Table We are able to map four of the datasets (see Section 4.2) in AGGREFACT that contain fine-grained annotations to our unified taxonomy. For all four datasets, if there are multiple annotators, we assign an error type to a summary if the error is annotated by more than one annotator. We allow one summary to have multiple error types. We call the annotated subset related to CNN/DM and XSum as AGGREFACT-CNN-UNIFIED and AGGREFACT-XSUM-UNIFIED, respectively. XSumFaith XSumFaith consists of 500 summaries each from human reference, two models in OLD, and two models in EXFORMER. All summaries are annotated with intrinsic and extrinsic errors, but no finer categories are distinguished. For error type mapping, we automatically detect predicates in a summary and assign each error span intrinsic-or extrinsic-predicate error if it contains a predicate. We map the remaining error spans to intrinsic-or extrinsic-noun phrase error. FRANK The CNN/DM subset of FRANK consists of three models in OLD, and one model each in both EXFORMER and FTSOTA. The XSum portion of FRANK has two models each in OLD and EX-FORMER. Each model contains 250 summaries in the dataset. We mapped Entity error and Out of Article error to extrinsic-noun phrase error; Predicate error and Grammatical error to extrinsic-predicate error; Circumstance error and Coreference error to intrinsic-noun phrase error; and other errors to intrinsic-predicate error. Goyal'21 Authors of the original dataset manually identified all hallucinated text spans for each summary and classified hallucination types into {intrinsic, extrinsic} × {entity, event, noun phrase, others}. The dataset consists of summaries for both CNN/DM and XSum. For the CNN/DM susbset, the authors directly annotated 50 summaries from FactCC, where summaries were generated by OLD models. The XSum subset consists of summaries from FTSOTA models. We map entity-related and noun phrase-related errors to noun phrase errors, event errors to predicate errors and others to entire sentence errors. CLIFF This dataset consists of 150 summaries each for both CNN/DM and XSum from two models in FTSOTA. We use the same approach for error mapping as we do for XSumFaith by only considering words labeled as extrinsic or intrinsic errors. We evaluate the accuracy of our error type mapping via manual inspection. Specifically, the authors of this work inspect 30 factually inconsistent examples each for XSumFaith, FRANK and CLIFF. Those examples cover summaries generated by all models used in the datasets. Results of the manual inspection show that the accuracy of our error type mapping is over 90%. A common discrepancy noticed by annotators was that in several cases the examples were originally annotated as intrinsic/extrinsic but we believe those errors are extrinsic/intrinsic. These cases are not a result of error in our mapping, but instead disagreement or error in the original annotation itself. For error mapping, we found out mapping of FRANK to be least accurate among all 4 datasets. For example, we found that the entity error (EntE) can be either intrinsic or extrinsic even though FRANK explicitly defines an extrinsic error type, i.e. "out of article" error. For Goyal'21, we manually correct any mapping errors that occur in the 150 examples. Corrections mostly happen for the event-related error defined in Goyal'21 which can be either noun phrase-or predicate-related. Next, we explore how the number of errors in specific groups of models from FTSOTA, EXFORMER, and OLD has changed with the progress in the field. Specifically, for each of the FRANK, XSumFaith, Goyal'21, and CLIFF datasets, we calculate the ratio of error types from factually inconsistent summaries generated by each model. We then study any distribution shift of error types in AGGREFACT-CNN-UNIFIED and AGGREFACT-XSUM-UNIFIED under FTSOTA, EXFORMER, and OLD. Summaries generated by the same models consist of different error distributions over different annotated datasets. As shown in AGGREFACT-XSUM-UNIFIED (Figure In addition, although XSumFaith and FRANK annotate the same set of model generated summaries in AGGREFACT-XSUM-UNIFIED, the distribution of error types looks dramatically different. The main discrepancy lies in the proportion of extrinsic-noun phrase and intrinsic-predicate errors. There are two possible reasons for such discrepancy. First, FRANK does not have "entire sent." errors based on our conversation of its annotation schema to the unified taxonomy (Section 4.2). Second, and more important, it is not easy to map error types from FRANK directly to our unified error types in spite of our validation. For example, the "out of article error" in FRANK is defined as an error where some statements in the summary do not show up in the source text. We found this error can be mapped to either an extrinsic-noun phrase error or extrinsic-predicate error. These observations indicate that previous work disagrees about where the individual error class boundaries are, even when aligned with our taxonomy. A combined meta-analysis shows shifts in error distributions. Figure In this section, we analyze how factuality metrics perform on summaries that contain certain error types. Specifically, we collect subsets of examples from four annotated datasets and group them into AGGREFACT-CNN-ERROR and AGGREFACT-XSUM-ERROR. 7 Every subset contains summaries that include only one error type defined in Sec-6 For AGGREFACT-XSUM-UNIFIED, since XSumFaith and FRANK annotated the same set of summaries, we only use the annotation results from XSumFaith since our error mapping is more accurate on the span-level annotations. 7 We exclude FRANK for this analysis for the same reason as in Section 4.3. tion 4.1. Each factuality metric assigns a binary label to an instance obtained directly from AGGREFACT-CNN and AGGREFACT-XSUM. Note that each subset only consists of test set examples from our benchmark since examples from the validation set were used to choose the optimal thresholds (Section 3). Since there are limited annotations for each model category after only considering examples from the test set of the benchmark, we decide not to split data by model categories in this part of the analysis. We calculate the recall of identifying error types from those subsets and show the results in Table Across both AGGREFACT-CNN-ERROR and AGGREFACT-XSUM-ERROR, we found that recent metrics like SummaC-Conv, QAFactEval and ChatGPT-based achieve higher recall for most error types. This indicates that more recent factuality metrics are better at capturing obvious errors generated by less recent models. This mirrors our earlier finding in Table Another observation is that although DAE is trained using annotations from XSumFaith, which provides supervision for multiple error types, it does not identify errors as well in AGGREFACT-CNN-ERROR. These findings indicate that summarization models make fundamentally different errors for each error type, and current factuality metrics cannot be uniformly good at identifying certain error types across datasets. We believe this conclusion still holds when evaluating metrics on summaries generated from FTSOTA models since they generate less obvious errors. Evaluate factuality models on modern summarization systems We have seen that FTSOTA yields significantly different results than EX-FORMER or OLD. Because of the prevalence of these systems, we believe that any new work should prefer evaluating on these SOTA summaries. Particularly for factuality metrics that are either based on latest LLMs or on pre-trained models, evaluating on modern summarization systems is needed to see if these metrics are actually improving from the current state-of-the-art or merely patching errors in outdated systems that have already been fixed by other advances. Annotate factual consistency errors from summaries generated by LLMs Recent work Choose the right metric for the job We note that there is no one clear winner among the metrics evaluated here (Section 3). Depending on the downstream application, different methods may be more or less appropriate, as our analysis shows. Moreover, none of current factuality metrics can identify certain error types across datasets equally well. As QG/QA and NLI models get better, we expect all of these methods to improve further. Alternatively, although recent ChatGPT-based metrics Use more consistent error types With our taxonomy, we have mapped error types annotated in previous work. It is relatively easier and more accurate to map errors from XSumFaith, Goyal'21, and CLIFF to our unified error types as they have annotation granularity finer than sentence-level. We encourage future work to follow this taxonomy where possible and leverage definitions in prior work to make cross-dataset comparisons possible. Here also, we encourage future work to prioritize annotation and evaluation of SOTA summaries. Most of current annotated datasets are within the news domain and factuality metrics are evaluated on news summaries accordingly. As there is a rising interest in other domains such as dialogue summarization In this work, we analyzed several factuality metrics across a large meta-benchmark assembled from existing datasets. We find state-of-the-art fine-tuned summarization models still present challenges for detecting factual errors, and the performance of error detectors is often overestimated due to the reliance on older datasets. Furthermore, we unify existing datasets into a common taxonomy and use this to highlight differences between datasets and summarization models, as well as the complexity of unifying concepts in this problem space. There are a few limitations of our work. First, we focus on evaluating state-of-the-art factuality metrics on English newswire datasets. This setting restricts us to English-language data, a formal style of text, and topics consisting of what is discussed in US and UK-centric news sources. Moreover, other summarization domains such as dialogue summarization have different common error types such as wrong reference error Second, since our work is built on top of previous work, some analysis such as the error type mapping is limited by the quality and annotation agreement from previous work. We chose not to undertake large-scale reannotation to avoid causing confusion in the literature with multiple versions of datasets reflecting divergent annotator opinions. In spite of these limitations, we believe that our reevaluation of these metrics and the analysis of error types under newswire data can bring insights for future works in choosing, designing and evaluating factuality metrics.
| 1,336 | 1,606 | 1,336 |
A Neural Transition-based Joint Model for Disease Named Entity Recognition and Normalization
|
Disease is one of the fundamental entities in biomedical research. Recognizing such entities from biomedical text and then normalizing them to a standardized disease vocabulary offer a tremendous opportunity for many downstream applications. Previous studies have demonstrated that joint modeling of the two sub-tasks has superior performance than the pipelined counterpart. Although the neural joint model based on multi-task learning framework has achieved state-of-the-art performance, it suffers from the boundary inconsistency problem due to the separate decoding procedures. Moreover, it ignores the rich information (e.g., the text surface form) of each candidate concept in the vocabulary, which is quite essential for entity normalization. In this work, we propose a neural transition-based joint model to alleviate these two issues. We transform the end-to-end disease recognition and normalization task as an action sequence prediction task, which not only jointly learns the model with shared representations of the input, but also jointly searches the output by state transitions in one search space. Moreover, we introduce attention mechanisms to take advantage of the text surface form of each candidate concept for better normalization performance. Experimental results conducted on two publicly available datasets show the effectiveness of the proposed method.
|
Disease is one of the fundamental entities in biomedical research, thus it is one of the most searched topics in the biomedical literature Previous studies In this work, we propose a novel neural transition-based joint model named NeuJoRN for disease named entity recognition and normalization, to alleviate these two issues of the multi-task learning based solution We summarize our contributions as follows. • We propose a novel neural transition-based joint model, NeuJoRN, for disease named entity recognition and normalization, which not only jointly learns the model with shared representations, but also jointly searches the output by state transitions in one search space. • We introduce attention mechanisms to take advantage of text surface form of each candidate concept for normalization performance. • We evaluate our proposed model on two public datasets, namely the NCBI and BC5CDR datasets. Extensive experiments show the effectiveness of the proposed model.
|
We define the end-to-end disease recognition and normalization task as follows. Given a sentence x from a document d (e.g., a PubMed abstract) and a controlled vocabulary KB (e.g., MeSH and OMIM) which consists of a set of disease concepts, the task of end-to-end disease recognition and normalization is to identify all disease mentions M = {m 1 , m 2 , ..., m |M | } mentioned in x and to link each of the identified disease mention m i with its mapping concept If there is no mapping concept in KB for m i , then m i → N IL, where N IL denotes that m i is unlinkable. We first introduce the transition system used in the model, and then introduce the neural transitionbased joint model for this task. We propose a novel transition system, inspired by the arc-eager transition-based shift-reduce LINKING-c parser We define a state as a tuple (σ, β, O), which consists of the following three structures: • stack (σ): the stack is used to store tokens being processed. • buffer (β): the buffer is used to store tokens to be processed. • output (O): the output is used to store the recognized and normalize mentions. We define a start state with the stack σ and the output O being both empty, and the buffer β containing all the tokens of a given sentence x. Similarly, we define an end state with the stack σ and buffer β being both empty, and the output O saving the recognized and normalized entity mention. The transition system begins with a start state and ends with an end state. The state transitions are accomplished by a set of transition actions A, which consume the tokens in β and build the output O step by step. As shown in Table • OUT pops the first token β 0 from the buffer, which indicates that this token does not belong to any entity mention. • SHIFT moves the first token β 0 from the buffer to the stack, which indicates that this token is part of an entity mention. • REDUCE pops the top two tokens (or spans) σ 0 and σ 1 from the stack and concatenates them as a new span, which is then pushed back to the stack. • SEGMENT-t pops the top token (or span) σ 0 from the stack and creates a new entity mention σ t 0 with entity type t, which is then added to the output. • LINKING-c links the previous recognized but unnormalized mention σ t 0 in the output with its mapping concept with id c and updates the mention with σ t,c 0 . Table Based on the introduced transition system, the endto-end disease recognition and normalization task becomes a new sequence to sequence task, i.e., the action sequence prediction task. The input is a sequence of words x n 1 = (w 1 , w 2 , ..., w n ) and a controlled vocabulary KB, and the output is a sequence of actions A m 1 = (a 1 , a 2 , ..., a m ). The goal of the task is to find the most probable output action sequence A * given the input word sequence x n 1 and KB, that is Formally, at each step t, the model predicts the next action based on the current state S t and the action history A t-1 1 . Thus, the task is models as (2) where a t is the generated action at step t, and S t+1 is the new state according to a t . Let r t denote the representation for computing the probability of the action a t at step t, thus where w a and b a denote the learnable parameter vector and bias term, respectively, and A(S t ) denotes the next possible valid actions that may be taken given the current state S t . Finally, the overall optimization function of the action sequence prediction task can be written as We now introduce neural networks to learn the dense representations of an input sentence x and each state in the whole transition process to predict the next action. Input Representation We represent each word x i in a sentence x by concatenating its character-level word representation, non-contextual word representation, and contextual word representation: where v char i denotes its character-level word representation learned by using a CNN network We then run a BiLSTM State Representation At each step t in the transition process, let's consider the representation of the current state S t = (σ t , β t , A t ), where σ t = (..., σ 1 , σ 0 ), β t = (β 0 , β 1 , ...) and A t = (a t-1 , a t-2 , ...). The buffer β t is represented with BiL-STM The stack σ t and the actions A t are represented with StackLSTM We classify all the actions defined in Table Specifically, for predicting the actions in the recognition purpose, we represent the state as where ReLU is an activation function, W and d denote the learnable parameter matrix and bias term, respectively, and • s 0 t and s 1 t denote the first and second representations of the stack σ. • b 0 t denotes the first representation of the buffer β. • a -1 t denotes the last representation of the action history A. For predicting the actions in the normalization purpose, we represent the state as • m and c are the representations of the mention and candidate concept by applying CoAttention mechanism • c denotes the candidate concept representation by (i) first run a BiLSTM • a -1 t denotes the last representation of the action history A. Decoding is the key step in both training and test, which is to search for the best output structure (i.e., action sequence) under the current model parameters. In this work, we use two different search strategies with different optimizations. Greedy Search For efficient decoding, a widelyused greedy search algorithm The main drawback of greedy search is error propagation We use two public available datasets in this study, namely NCBI -the NCBI disease corpus Following previous work We use the AdamW optimizer Table From the comparisons, we find that (1) IDCNN does not perform well enough although it relies few efforts of feature engineering. (2) All the joint models significantly outperform the pipelined methods. (3) The deep-learning based joint models significantly outperform the traditional machine learning based methods. (4) Our proposed NeuJoRN outperforms MTL-feedback by at least 0.57% and 0.59% on the recognition and normalization tasks, respectively. Table (2) The model with beam size 4 achieves the best performance. The larger the beam size, the better the performance, however the lower the decoding speed. (3) Our greedy search based solution doesn't outperform the MLT-feedback method. Table Disease Named Entity Normalization DNEN has also been widely studied in the literature. Most studies assume that the entity mentions are predetected by a separate DNER model, and focus on developing methods to improve the normaliation accuracy Joint DNER and DNEN Several studies TaggerOne Transition-based Models Transition-based models are widely used in parsing and translation
| 1,377 | 974 | 1,377 |
Reasoning with Language Model is Planning with World Model
|
Large language models (LLMs) have shown remarkable reasoning capabilities, particularly with chain-of-thought (CoT) prompting. However, LLMs sometimes still struggle with problems that are easy for humans, such as generating action plans to achieve given goals in an environment, or performing complex math or logical reasoning. The deficiency stems from the key fact that LLMs lack an internal world model to predict the world state (e.g., environment status, intermediate variable values) and simulate long-term outcomes of actions. This prevents LLMs from performing deliberate planning akin to human brains, which involves exploring alternative reasoning paths, anticipating future states and rewards, and iteratively refining existing reasoning steps. To overcome the limitations, we propose a new LLM reasoning framework, Reasoning via Planning (RAP). RAP repurposes the LLM as both a world model and a reasoning agent, and incorporates a principled planning algorithm based on Monte Carlo Tree Search for strategic exploration in the vast reasoning space. During reasoning, the LLM (as agent) incrementally builds a reasoning tree under the guidance of the LLM (as world model) and rewards, and efficiently obtains a high-reward reasoning path with a proper balance between exploration vs. exploitation. We apply RAP to various challenging reasoning problems including plan generation, math reasoning, and logical inference, and demonstrate its superiority over strong baselines. RAP with LLaMA-33B even surpasses CoT with GPT-4, achieving 33% relative improvement in a plan generation setting. 1
|
Large language models (LLMs) have exhibited emergent reasoning abilities in a wide range of tasks Humans possess an internal world model, a mental representation of the environment
|
Figure reasoning space. To address these limitations, this paper proposes a new framework, Reasoning via Planning (RAP), that enables LLMs to reason in a manner close to humans' conscious planning. RAP augments the LLM with a world model, and reasons with principled planning (specifically Monte Carlo Tree Search, MCTS) to produce high-reward reasoning traces after efficient exploration (Figure We show RAP is a general framework applicable to a diverse range of challenging problems and achieves substantial improvements over recent popular LLM reasoning methods. For plan generation, particularly in 2/4/6-step problems of Blocksworld Reasoning with LLMs. LLM reasoning typically involves decomposing complex questions into sequential intermediate steps (a.k.a. chains) before producing the final answer, exemplified by Chainof-Thought (CoT) prompting and its variants Compared with these search-guided methods, RAP is a more principled framework to combine world model and reward with advanced planning. Planning with LLMs. Planning, a central ability in intelligent agents, involves generating a series of actions to achieve a specific goal 3 Reasoning via Planning (RAP) In this section, we present the Reasoning via Planning (RAP) framework that enables LLMs to strategically plan a coherent reasoning trace for solving a wide range of reasoning tasks. We first build the world model by repurposing the LLM with prompting (Section 3.1). The world model serves as the foundation for deliberate planning, by allowing the LLM to plan ahead and seek out the expected outcomes in the future. We then introduce the rewards for assessing each state during reasoning in Section 3.2. Guided by the world model and rewards, the planning with Monte Carlo Tree Search (MCTS) efficiently explores the vast reasoning space and finds optimal reasoning traces (Section 3.3). Finally, when multiple promising reasoning traces are acquired during planning, we further introduce an aggregation method in Section 3.4 that yields an ensembled result and further boosts the reasoning performance. In general, a world model predicts the next state of the reasoning after applying an action to the current state With the definition of state and action, the reasoning process can thus be described as a Markov decision process (MDP): given the current state s t,t=0,1,...,T , e.g., the initial state s 0 , the LLM (as a reasoning agent) generates an action space by sampling from its generative distribution a t ∼ p(a|s t , c), where c is a proper prompt (e.g., incontext demonstrations). Once an action is chosen, the world model then predicts the next state s t+1 of the reasoning. Specifically, we repurpose the same LLM to obtain a state transition distribution p(s t+1 |s t , a t , c ′ ), where c ′ is another prompt to guide the LLM to generate a state. For instance, in Blocksworld, the LLM (as the world model) generates text s t+1 to describe the new configuration of blocks, given previous state s t and the action a t . Continuing the process results in a reasoning trace, which consists of a sequence of interleaved states and actions (s 0 , a 0 , s 1 , . . . , a T -1 , s T ). This differs from the previous reasoning methods, such as Chain-of-Thought During reasoning, we want to assess the feasibility and desirability of each reasoning step, and guide the reasoning based on the assessment (Section 3.3). The assessment of each reasoning step (i.e., applying an action a t to the state s t ) is performed by a reward function r t = r(s t , a t ) ∈ R. Similar to the state and action, the reward function can be specified in different ways to accommodate any knowledge or preferences about the reasoning problem of interest. Here we introduce several common rewards applicable to different tasks and shown to be effective in our experiments. Likelihood of the action. When an action is generated by the LLM conditioning on the in-context demonstration and the current state, the probability of the specific action reflects the LLM's preference. We thus can incorporate the log probability of the action as a reward. This reward reflects the "instinct" of LLMs as an agent, and can be also used as a prior for which action to explore. Confidence of the state. State prediction is nontrivial in some problems, e.g., in math reasoning (Figure Once equipped with the world model (Section 3.1) and rewards (Section 3.2), LLMs can reason with any planning algorithms. We adopt Monte Carlo Tree Search (MCTS) Specifically, MCTS builds a reasoning tree iteratively, where each node represents a state, and each edge represents an action and the transition from the current state to the next state after applying the action (Figure Selection. The first phase selects a portion of the existing tree that is most promising for further expansion in the next phase. Starting from the root node (i.e., initial state s 0 ), at each level of the tree, the algorithm selects a child node as the next node. The phase finishes when a leaf node of the current tree is reached. Figure where N (s) is the number of times node s has been visited in previous iterations, and c(s, a) is the child node of applying a in state s. The less a child node was visited before (i.e., the more uncertain about this child node), the higher the second term in the equation. The weight w controls the balance between exploration and exploitation. Expansion. This phase expands the tree by adding new child nodes to the leaf node selected above. Given the state of the leaf node, we use the LLM (as agent) to sample d possible actions (e.g., subquestions in math reasoning), and then use the LLM (as world model) to predict the respective next state, resulting in d child nodes. Note that if the leaf node selected above is a terminal node (the end of a reasoning chain) already, we will skip expansion and jump to back-propagation. Simulation. To estimate the expected future rewards (Q values), this phase simulates the future situations of the current node using the world model. Starting from the current node as above, at each node s, we create an action following a roll-out policy and use the world model to predict the next state. The roll-out process continues until a terminal state is reached. There could be many ways to define the roll-out policy (e.g., by adding different randomness). In our experiments, for simplicity and reduced noises, we follow a similar process as in the expansion above, i.e., generating d candidate actions and picking one of the largest local reward a ′ = max a ′ r(s, a). In practice, for efficiency, we discard the computationally costly components in r (e.g., the reward from the confidence of state requires sampling the answer multiple times), and use the resulting lightweight reward function for selecting actions during simulation. Back-propagation. Once we reach a terminal state in the above phases, we obtain a reasoning path from the root node to the terminal node. We now back-propagate the rewards on the path to update the Q value of each state-action pair along the path. Specifically, we update Q(s, a) by aggregating the rewards in all future steps of node s. Once a predetermined number of MCTS iterations is reached, we terminate the algorithm and select the final reasoning trace from the constructed tree for evaluation. There are various ways for the selection. One is to start from the root node and iteratively choose the action with the highest Q value until reaching a terminal. Also, one can directly select the path from the iterations that yielded the highest reward, or opt to choose the leaf node (and the respective root-to-leaf path) that has been visited the most. In practice, we observed that the second strategy often yields the best results. For problems, such as math reasoning (Section 4.2) where only the final answer is required, RAP could produce multiple traces and answers from different MCTS iterations, which will be aggregated to produce the final answer. We refer to such a mechanism as RAP-Aggregation. Note that problems like plan generation or logical inference require a complete reasoning trace as output; thus, RAP-Aggregation will not be applied. In this section, we demonstrate the flexibility and effectiveness of our RAP framework by applying it to a wide range of problems, including plan generation in an embodied environment (4.1), mathematical reasoning for solving math word problems (4.2), and logical reasoning for verifying hypotheses (4.3). The subsequent sections demonstrate how the world model formulation in RAP enables a versatile design of the state and action, catering to various reasoning contexts. We primarily compare RAP with chain-ofthought (CoT) The plan generation task aims to produce a sequence of actions to achieve a given goal, possibly with additional constraints. The ability to generate plans is important for intelligent embodied agents, e.g. household robots To assess the quality of actions within this domain, we use two separate rewards. First, we prompt the LLM with some example test cases along with their solutions, and then calculate the log probability of the action given the current state ("Likelihood of action" reward in Section 3.2), denoted as r 1 . This reward reflects the intuition of the LLM as the reasoning agent. It's typically indicative when there are few steps left to the goal, while not as reliable for a distant goal. Additionally, we compare the new state after performing an action with the goal and provide a reward, r 2 , scaling with the number of conditions met ("Task-specific heuristics" reward). Specifically, when all the conditions are met, we assign a super large reward to make sure this plan will be selected as the solution. Results. We use test cases from the Blocksworld dataset I have that, the red block is clear, the blue block is clear, the orange block is clear, the hand is empty, the orange block is on top of the yellow block, the red block is on the table, the blue block is on the table and the yellow block is on the table. Pick up the red block. I have that, the blue block is clear, the orange block is clear, the red block is in the hand, the hand is holding the red block, the orange block is on top of the yellow block, the blue block is on the table, and the yellow block is on the table. Stack the red block on top of the blue block. I have that, the orange block is clear, the red block is clear, the hand is empty, the orange block is on top of the yellow block, the red block is on top of the blue block, the blue block is on the table, and the yellow block is on the table. Unstack the orange block from on top of the yellow block. I have that, the orange block is in the hand, the red block is clear, the yellow block is clear, the hand is holding the orange block, the red block is on top of the blue block, the blue block is on the table, and the yellow block is on the table. Stack the orange block on top of the red block. I have that, the orange block is clear, the yellow block is clear, the hand is empty, the orange block is on top of the red block, the red block is on top of the blue block, the blue block is on the table, and the yellow block is on the table. Method 2-step 4-step 6-step CoT 0.17 0.02 0.00 CoT -pass@10 0.23 0.07 0.00 CoT (GPT-4) 0.50 0.63 0.40 RAP (10) 1.00 0.86 0.26 RAP (20) 1.00 0.88 0.42 Table plan for a new question. This setting is the same as one described in As shown in Table Case study. We compare the reasoning paths from CoT and RAP in Figure Task setup. Math reasoning tasks, such as GSM8k Chain-of-Thought 29.4 + SC (10) 46.8 Least-to-Most 25.5 + SC (10) 42.5 RAP (1) 40.0 RAP (10) 48.6 + aggr 51.6 impact of the reasoning path's length on the reward, we compute the Q value by using the maximum of average rewards in future steps. Q * (st, at) = max s t ,a t ,r t ,...,s l ,a l ,r l ,s l+1 avg(rt, . . . , r l ). (2) As a related work, Least-to-Most prompting As shown in Table Task setup. A logical reasoning task (e.g. PrOn-toQA Results. We assess the performance of our RAP framework on PrOntoQA We mix the examples requiring 3, 4, and 5 reasoning hops in a correct proof to prevent LLM from memorizing when to finish the reasoning. We sample 500 examples from the generation script released by As the results presented in demonstrate the effectiveness of RAP in solving logical reasoning problems in PrOntoQA. Also, as the case illustrated in Figure To further study whether RAP can help stronger LLMs to solve more complex problems, we conduct experiments on the full Blocksworld We employ chain-of-thought prompting In our main experiments, we choose the combination of rewards in our current experiments based on heuristics and our exploratory experiments. To understand the effects of the reward choice for LLM reasoning, we supplement comprehensive experiments on rewards for plan generation (Table Generally, the combination of multiple rewards contributes to the performance. However, the effects of a reward depends on the nature of tasks. For example, the action likelihood reward is essential for plan generation, but not very helpful to mathmatical reasoning. More discussions are in Appendix F. In this paper, we present Reasoning via Planning (RAP), a novel LLM reasoning framework that equips LLMs with an ability to reason akin to human-like strategic planning. Our framework, which repurposes the LLM to act as both a world model and a reasoning agent, enables the LLM to simulate states of the world and anticipate action outcomes, and achieve an effective balance between exploration and exploitation via Monte-Carlo Tree Search. Extensive experiments on a variety of challenging reasoning problems demonstrate RAP's superiority over several contemporary CoT-based reasoning approaches, and even the advanced GPT-4 in certain settings. We adapt MCTS to search for the optimal reasoning path (Algorithm 1). Compared with traditional applications of MCTS, we are faced with a large reasoning space, and the heavy computational cost of LLMs. Thus, we made several modifications to the classic MCTS in our implementation: ( We use random sampling with a temperature of 0.8. The generation is cut off at the maximum length of 2048 or a newline token. All of our experiments run on 4 × NVIDIA A5000 GPUs with 24GB memory. We show the prompt to calculate the action likelihood for RAP below. The same prompt is also applied in CoT baseline. <init_state> and <goals> would be instantiated by the problem to solve. My plan is as follows: [PLAN] unstack the yellow block from on top of the orange block put down the yellow block pick up the orange block Algorithm 1 RAP-MCTS Require: Initial state s0, state transition probability function p θ , reward function r θ , action generator p ϕ , number of generated actions d, depth limit L, number of roll-outs N , and exploration weight w 1: Initialize memory of actions A : S → A, children c : S × A → S and rewards r : S × A → R 2: Initialize the state-action value function Q : S × A → R and visit counter N : S → N 3: for n ← 0, . . . , N -1 do 4: t ← 0 5: while 8: rt = r(st, at), st+1 ← c(st, at) 9: t ← t + 1 10: end while 11: while st is not a state ∧ t ≤ L do 12: t ), and r t ) 14: Update A(st) ← {a [STATEMENT] As initial conditions I have that, <initial_state> My goal is to have that <goals>. For the next state prediction with the world model, we apply the prompts conditioned on the last action. Here we show the prompt to update the state after a "pick up" action as an example. Again, <state> and <action> would be instantiated with the current state and action. Recent years have witnessed successful applications of planning algorithms Besides, the world model is also shown to be helpful to physical robot learning Through our preliminary experiments, we observed that the performance of LLMs is impacted by the discrepancy in difficulty between demonstration cases and the test cases. In the case of RAP, when a new state is predicted, we reformulate the remaining task as a new test case, initialized with the predicted new state. This new test case would require a smaller minimum number of actions, leading to a disparity in the distribution of the demonstration cases and the new cases. To mitigate this issue, we pre-compute the intermediate states of the demonstration cases beforehand. During inference, we truncate the trace from the beginning for each new state in an iteration, which reduces the minimum action number of the demonstration cases as the search tree deepens. This technique significantly enhances the performance of RAP, especially for more intricate problems, which are more susceptible to distribution mismatches. Results. We conduct comprehensive experiments on rewards for plan generation (Table Self-evaluation and action likelihood. The rewards of self-evaluation and action likelihood are of particular interest, as they can be applied to a wide range of reasoning tasks. Generally, the best usage and combination with other rewards require empirical design and understanding of the task nature, and their effectiveness can vary significantly across different tasks. Here, we provide some intuitions behind the reward choices: (a) For the problems in which one reasoning step is short and structured, the action likelihood can be very indicative. Otherwise, it may be disturbed by unimportant tokens and become unreliable. For instance, a single step within the Blocksworld domain typically adheres to specific patterns (e.g., PICK/PUT/STACK a block. . . ), rendering the action likelihood indicative. However, in the math domain, a reasoning step is expressed in natural language sentences, allowing for greater freedom and potentially introducing noise. (b) For the problems where it's easier to recognize some errors afterward than avoid them during generation, self-evaluation emerges as a helpful mechanism for enhancing reasoning accuracy. In mathematical reasoning, LLMs may struggle to generate a correct reasoning step in the first place, but the detection of calculation or logic errors is more feasible. In Blocksworlds, however, assessing the quality of a candidate action is not straightforward and still requires multi-step reasoning. This characteristic diminishes the accuracy of the selfevaluation reward, making it less helpful especially given that likelihood already provides a good intuition for search.
| 1,603 | 180 | 1,603 |
EdgeFormer: A Parameter-Efficient Transformer for On-Device Seq2seq Generation
|
We introduce EDGEFORMER -a parameterefficient Transformer for on-device seq2seq generation under the strict computation and memory constraints. Compared with the previous parameter-efficient Transformers, EDGE-FORMER applies two novel principles for costeffective parameterization, allowing it to perform better given the same parameter budget; moreover, EDGEFORMER is further enhanced by layer adaptation innovation that is proposed for improving the network with shared layers. Extensive experiments show EDGEFORMER can effectively outperform previous parameterefficient Transformer baselines and achieve competitive results under both the computation and memory constraints. Given the promising results, we release EDGELM 1 -the pretrained version of EDGEFORMER, which is the first publicly available pretrained on-device seq2seq model that can be easily fine-tuned for seq2seq tasks with strong results, facilitating on-device seq2seq generation in practice.
|
On-device modeling draws increasing attention for its unique advantages To customize the Transformer for seq2seq tasks in the on-device setting, we propose EDGE-FORMER -a novel parameter-efficient Transformer of the encoder-decoder architecture. EDGE-FORMER is structurally similar to the standard Transformer with a deep encoder and shallow decoder, but with a modification that it uses an in- terleaved decoder with shared lightweight feedforward networks, as shown in Figure In addition to cost-effective parameterization, EDGEFORMER proposes and applies layer adaptation to further improve the model with tied layers, as Figure The contributions of this work are three-fold: • This paper is one of the earliest work that formally studies on-device seq2seq generation by discussing its challenges and defining a practical setting with appropriate resource constraints for the evaluation. • We propose EDGEFORMER, a parameterefficient Transformer with novel cost-effective parameterization and layer adaptation, achieving the state-of-the-art result in the on-device seq2seq generation setting under strict computing and memory resource constraints. • We introduce and release EDGELM (the pretrained EDGEFORMER) -the first publicly available pretrained on-device seq2seq model that can be easily fine-tuned for seq2seq tasks with strong results, which can largely reduce the effort for delivering a powerful on-device seq2seq model in practice. 2 Background: Transformer
|
The biggest challenge for an on-device seq2seq model is regarding the model size and memory cost. As shown in Table • The interleaved structure makes the architecture of encoder and decoder layers consistent • As shown in Table Full Parameterization Full parameterization is a common parameterization approach for Transformer, meaning that each model parameter (excluding embedding) is independent without being shared by multiple modules in the network. In a forward pass, each parameter is used only once. Full parameterization allows parameters to be flexible to fit their roles well during model training. 3 Constraints for On-device Seq2seq Computation On-device computer vision (CV) models tend to use 1G FLOPS (0.5G MACS) as a constraint, which is directly followed by previous work on on-device translation Memory In contrast to deploying a model on a cloud server without caring about memory cost much, there is a very strict memory constraint for an on-device model in practice, because a user's edge device (e.g., PC) is not only for model hosting; instead, it usually runs many other (background) apps and programs at the same time besides the model. To ensure moderate memory cost, we limit the number of model parameters (excluding word embedding lookup table) up to 10 million, following previous work Due to the tight parameterization budget (i.e., 10 million), EDGEFORMER cannot be fully parameterized as in the standard way; instead, it has to adopt shared parameterization. As a strong baseline for shared parameterization, UNIVERSAL TRANSFORMER lets all its M encoder layers share 1 group of encoder layer parameters and all its N decoder layers share 1 group of decoder layer parameters: As observed by Although UNIVERSAL TRANSFORMER is a popular solution to shared parameterization, it is still not cost-effective for two reasons: First, UNIVERSAL TRANSFORMER uses (over) half of total parameters to parameterize the decoder, which is uneconomical. As shown in Figure Second, UNIVERSAL TRANSFORMER does not consider load balance of model parameters, which was a rarely discussed problem until the recent emergence of Mixture-of-Expert models Based on the above insights, we parameterize EDGEFORMER in the following two novel principles for cost-effective parameterization: Encoder-favored Parameterization For EDGE-FORMER, we parameterize its encoder using as many parameters as possible: except a small number of parameters (d 2 /2) for all lightweight FFNs in the decoder, we use almost all parameters in our budget to parameterize the encoder. For attention modules in the decoder, we let them reuse (i.e., share) parameters with the attention modules in the encoder since attention modules in both the encoder and decoder work in the same mechanism and can be effectively shared Load-balanced Parameterization We try parameterizing EDGEFORMER with a balanced load for each model parameter so that each parameter could be as equally exploited as possible in a forward pass. Given the parameterization budget and the load balance principle, we create 2 groups of encoder FFN parameters equally shared by all encoder layers, 1 group of decoder FFN parameters is shared by light FFNs in the decoder, and 4 groups of attention parameters are shared throughout the encoder and decoder. Except for parameters in the encoder FFNs that are used 6 times, other parameters are all used 4 times in a forward pass, resulting in a load balanced parameterization: Shared parameterization causes layers with tied weights to become less specialized, as discussed in Section 1. To allow tied layers to be better adapted to their corresponding roles, we propose layer adaptation to further enhance EDGEFORMER. Inspired by parameter-efficient task transfer methods, we investigate three efficient layer adaption approaches: Bias-based Layer Adaptation (Bias-LA) Inspired by BitFit (Ben Zaken et al., 2021) finetuning with only bias terms, we untie all bias terms of each layer and use them to specialize the layers with tied weights, as shown in Figure Adapter-based Layer Adaptation (Adapter-LA) Adapter-based approaches Prefix-based Layer Adaptation (Prefix-LA) Inspired by recent work Following the encoder-favored principle in Section 4.2, we only apply LA to encoder layers. We mainly evaluate our approach in Machine Translation (MT). We select the most popular MT benchmark -WMT14 English-German (En-De) translation task, which is also a touchstone for seq2seq evaluation, as our main test bed. To compare with previous work, we also evaluate WMT14 English-French (En-Fr) translation. We follow the standard way to train and evaluate evaluate WMT14 En-De and En-Fr. As We select UNIVERSAL TRANSFORMER which is the most popular and a strong baseline of parameter- efficient Transformer for fair comparison. By default, we apply Seq-KD We evaluate EDGEFORMER and compare it with UNIVERSAL TRANSFORMER (UT) on WMT14 En-De. According to Table As discussed in Section 4.2, the advantage of EDGEFORMER over UT comes from its costeffective parameterization. The encoder-favored principle is again supported by comparing 6+6 Transformers' results in Table After discussing parameterization, we then analyze the effects of layer adaptation on the results by mainly focusing on Adapter-LA and Prefix-LA that both show performance gain. Figure Finally, we compare EDGEFORMER with recent work on parameter-efficient Transformer modeling. To keep consistency of the training and evaluation protocols with previous work, we here give up using Seq-KD to train the models, and report BLEU We conduct experiments in WMT14 En-De translation and CoNLL-14 Grammatical Error Correction Table Given the promising results, we introduce EDGELM -the pretrained We evaluate EDGELM in the benchmarks of three popular seq2seq tasks: CoNLL-14 for Grammatical Error Correction (GEC), XSum On-device seq2seq generation in NLP is a research area that has been less explored than on-device CV and NLU In addition to work improving parameter efficiency by weight sharing, there is research that studies lightweight model architecture for seq2seq learning where early work We formally study on-device seq2seq generation, including defining its practical resource constraint setting and proposing an appropriate modeling technology EDGEFORMER. The cost-effective parameterization and layer adaptation innovations in EDGEFORMER both prove effective to improve the results with negligible computation and memory cost, achieving state-of-the-art results in the on-device seq2seq generation setting. Our released pretrained EDGEFORMER -EDGELM can be easily fine-tuned for downstream seq2seq tasks, largely facilitating on-device seq2seq generation in practice. For future work, we plan to further study loadbalanced parameterization for parameter-efficient models, which is an interesting and new but seemingly profound machine learning research problem: instead of naively assuming that all the parameters are equal in this preliminary study, we suspect that parameters in different modules (e.g., parameters in the self-attn and FFN; or parameters in different layers) should be under different amounts of load. We look forward to in-depth research on this problem, which might be helpful to deepen our understanding of neural networks. EDGEFORMER is a preliminary model proposed for on-device seq2seq generation setting, which still has much room for improvement. For example, as mentioned in Section 8, the current load balance mechanism naively assumes that the number of times that a parameter is used in a forward pass is equal to its load, which may not be always true because parameters in different moduels are different: some parameters may be effectively used more times than others, which requires deeper understanding of neural network and the Transformer.
| 962 | 1,472 | 962 |
Bipartite-play Dialogue Collection for Practical Automatic Evaluation of Dialogue Systems
|
Automation of dialogue system evaluation is a driving force for the efficient development of dialogue systems. This paper introduces the bipartite-play method, a dialogue collection method for automating dialogue system evaluation. It addresses the limitations of existing dialogue collection methods: (i) inability to compare with systems that are not publicly available, and (ii) vulnerability to cheating by intentionally selecting systems to be compared. Experimental results show that the automatic evaluation using the bipartite-play method mitigates these two drawbacks and correlates as strongly with human subjectivity as existing methods.
|
The performance evaluation of dialogue systems is a crucial and challenging research topic for the dialogue research community. The community recommends human evaluation as the primary evaluation method, which is the gold standard but is timeconsuming and costly. Moreover, reproducing the evaluation results is mostly impractical due to the unavailability of maintaining identical evaluators or identical evaluation conditions. Human evaluation is therefore unsuitable for evaluating daily updates of developing dialogue systems or comparing systems with non-public ones. Thus, constructing a better automatic evaluation method, which is both highly reproducible and low cost, is desirable. In particular, automating interactive evaluation, not static evaluation such as BLEU An interactive evaluation framework consists of two phases: first, collecting the dialogues in which the systems to be evaluated (hereinafter called evaluation targets) talk to others (hereinafter called dia-*Both authors contributed equally to this paper. logue partners), then rating evaluation targets based on the quality of their utterances in the collected dialogues. Regarding the collecting (i.e., automating dialogue partners), self-play and all-play-all (Figure This paper addresses the above two drawbacks of the all-play-all method while maintaining the all-play-all method's high correlation with human rating. Specifically, we propose the bipartite-play method, i.e., fixing and sharing a set of dialogue partners across studies as shown in Figure 2 Related Work 2.1 Automatic dialogue collection Self-play. The self-play method collects dialogues where evaluation targets talk to themselves, i.e., i × 1 × j dialogues in which collecting j dialogues for each of i evaluation targets. This method is cost-effective for interactive dialogue system evaluation since it does not require human interactions All-play-all. The all-play-all method collects dialogues between multiple evaluation targets, i.e., i×(i-1)×j dialogues when collecting j dialogues for each of i evaluation targets (considering speaker order). This method also requires no human interactions. Compared to the self-play method, the all-play-all method's dialogue partners are more diverse since it collects dialogues with other evaluation targets that result in various dialogues
|
Compared with methods relying on reference responses (e.g., BLEU The all-play-all method enables effective dialogue collection for system comparison, as described in Section 2. However, we point out that the current all-play-all method cannot handle the following two cases: First, when the group of evaluation targets includes unavailable systems. Since allplay-all requires the collection of dialogues with all evaluation targets, it is impossible to compare systems that are not released or that cannot be run by many researchers due to such computational resources. Second, when one attempts to boost their system's performance by deploying an unfair evaluation setting. Our experiments (Section 6) reveal that one can intentionally improve the automatic evaluation results of desired systems by choosing evaluation targets to be compared when using the all-play-all method. If these potential drawbacks can be overcome, existing automated evaluation methods could be enhanced to be more versatile and practical. We introduce a new automatic dialogue collection method, called bipartite-play method, which updates the two aforementioned limitations. Considering the two drawbacks of the all-play-all method (Section 3), we propose fixing and sharing a set of publicly accessible systems as dialogue partners rather than assigning other evaluation targets as partners. Our idea is that even if evaluation targets do not talk to each other directly, dialogues in which evaluation targets talk to the same (shared) partners should be effective for system comparison. In this setting, the diversity of dialogue partners can be maintained by ensuring the diversity of predetermined dialogue partners set. Sharing a set of dialogue partners allows a fair comparison with publicly inaccessible systems as long as its developers use our method. Furthermore, predetermining a set of dialogue partners prevents cheating by an intentional choice of evaluation targets. Bipartite-play dialogue collection. Given i evaluation targets, the bipartite-play method collects i × k × j dialogues by having evaluation targets interact j times with each of the k various dialogue partners predetermined for evaluation. Subsequently, we introduce bipartite-play to the interactive dialogue evaluation framework. We combine the dialogue collection based on the bipartite-play method with FED (Mehri and Eskenazi, 2020a), which is one of the modern and effective dialogue rating methods. Based on the m collected dialogues by the bipartiteplay method, We assess an evaluation target for a dimension v. Specifically, we first evaluate the performance for v in a dialogue using the average score of the system's utterances. We then determine the system's whole performance for v using the average score of m dialogues. We compute the system utterances score using FED. This rating method evaluates the system's utterances for v by guessing whether positive or negative responses for v are valid in terms of the language model as a response to the system's utterance. The validity of each positive and negative response is automatically evaluated using a large-scale dialogue system. The evaluation value of v of the evaluation target's utterance r for a context c is calculated as follows: where P v and N v are the set of positive and negative responses for v, respectively. D(c, •; θ) is a function that calculates the probability of generating a response to c using a large-scale dialogue system with parameters θ. We assess evaluation targets based on dimensions frequently deployed in recent research Dataset. We created the dataset by collecting dialogues between the dialogue system and humans, then annotating the collected dialogues with a human evaluation score. Crowdsourcing FED evaluation settings. We used the positive and negative responses manually created by Results of FED evaluation. Table Evaluation targets: Tfm-3B-Rdt-Bsm, Tfm-3B-Rdt-Msc, Tfm-3B-R2c-Bsm, Tfm-3B-Rdt-Lgu, GPT-345M-Wtx-Rdt, Tfm-89M-Ddc-Nft, Tfm-89M-Ddc-Crm, Tfm-89M-Ddc-Ddg, Tfm-89M-Ddc-Rdt, Tfm-89M-Ddc-Twt, PEn-256M-Rdt-Bst Partner systems: Tfm-3B-Rdt-Slf, Tfm-3B-Rdt-Lgt, Tfm-3B-Rdt-Img, Tfm-3B-Rdt-Sfr, Tfm-1B-Rdt-Bsm, GPT-117M-Wtx-Rdt, GPT-762M-Wtx-Rdt, Tfm-406M-Rdt-Bsm, Tfm-406M-R2c-Bsm, Brt-406M-Rbt-Woi, Trm-89M-Ddc-Wow, Trm-89M-Ddc-Lgt, Trm-89M-Ddc-Emp, Trm-89M-Ddc-Cv2, Trm-89M-Rdt-Wow, Trm-89M-Rdt-Cv2, Trm-88M-Rdt-Bst, Trm-88M-Rdt-Cv2, PEn-256M-Rdt-Cv2, PEn-256M-Rdt-Emp, PEn-256M-Rdt-Wow, PEn-256M-Rdt-All, PEn-256M-Rdt-Bsm, B+F-256M-Rbt-Wow * Tfm: Transformer We show that the interactive automatic evaluation using the bipartite-play method correlates with humans as strongly as the all-play-all method, which has been reported to be an effective dialogue collection method but requires access to all evaluation targets. We first rank prepared evaluation targets by interactive human evaluation and then measure the correlation with the rankings by interactive automatic evaluations in the three dialogue collection methods: self-play, all-play-all, and bipartite-play. Dialogue systems. Table Dialogue collection settings. For each of the three dialogue collection methods, We set the target-partner pairs for the self-play method, the all-play-all method, and the bipartite-play method. The resulting pairs are 11×1 = 11, 11×(11-1) = 110, and 11 × 24 = 264, respectively. A pair's systems exchange utterances five times to form one dialogue following two given initial utterances, which we extracted from the initial parts of dialogues in the test set of the EmpatheticDialogues dataset Interactive human evaluation. We compute each evaluation target's score for each of the three dimensions (i.e., Specificity, Sensibleness, and Overall) by averaging the manually annotated scores of 50 dialogues in Section 5.2. We then rank evaluation targets based on their averaged scores. Table stance, with the same settings as our experiment, one can indirectly compare their systems with our evaluation targets by comparing systems' FED scores. As one of the reference values, we present the FED scores of Tfm-3B-Rdt-Bsm, referred to as Blender 3B Tables Collecting dialogues. Tfm-89M-Ddc-Ddg returned valid responses to high-quality utterances of Tfm-1B-Rdt-Bsm (Table Rating dialogues. Regarding Tfm-89M-Ddc-Ddg, the scores of the dialogues in Tables The interactive automatic evaluation using the bipartite-play method correlates with humans as strongly as the all-play-all method, as described in Section 6, without direct interaction with evaluation targets. Additionally, we present another reason to use the bipartite-play method: one can sometimes cheat by choice of evaluation targets when using the all-play-all method. A: I love being in the "flow" -whether it's in speaking to a group of people or working to help facilitate meetings. A: i'm a sales manager at a trading company. 7.1 How can we cheat on all-play-all? We found that dialogues where systems frequently speak about the same things tend to receive low ratings from the FED evaluation described in Section 5.2. This can be a desirable evaluation property since human evaluation is known to have the same tendency We show cheating using the all-play-all method following the above hypothesis. Specifically, based on Yang et al. ( Settings. We form an unfair set of evaluation targets by collecting four systems, i.e., one whose rank we attempt to improve (favored system), another whose rank we attempt to worsen (unfavored system), and two systems similar to the unfavored system. We then check whether the ranking relationship between favored and unfavored ones changes from that of the original all-play-all evaluation (fair evaluation) in Section 6. In this unfair evaluation, unfavored systems have to construct dialogues with similar systems three out of four times, where repeated utterances are likely to occur as in self-play. We prepared two combinations of the unfavored system and its similar system: a series of DialoGPT (GPT-345M-Wtx-Rdt is the unfavored system, whose similar systems are GPT-124M-Wtx-Rdt and GPT-774M-Wtx-Rdt) and a series of Blender (Tfm-3B-Rdt-Bsm is the unfavored system, whose similar systems are Tfm-406M-Rdt-Bsm and Tfm-1B-Rdt-Bsm). We assigned each of all ten evaluation targets for the experiments in Section 6 except the unfavored one (GPT-345M-Wtx-Rdt or Tfm-3B-Rdt-Bsm) as a favored system. We focused on evaluation for Specificity, where the self-play property especially affects the results of automatic evaluation using the self-play method. Results. Table and sharing a set of diverse partner systems, is one of the practical methods to prevent this cheating. In this paper, we proposed the bipartite-play method as a dialogue collection method. The bipartite-play method can address the impossibility of comparison with publicly inaccessible systems and the vulnerability to cheating by intentional choice evaluation targets to improve the all-playall method. For the proposed method, no dialogue with evaluation targets is required, thereby facilitating system comparison across studies and possibly enabling comparison with inaccessible systems. Our experiments showed that, compared with the evaluation using the all-play-all method, the automatic evaluation using the bipartite-play method correlates just as strongly with humans. Although we formed a set of the bipartite-play method's partner systems for the experiments considering its diversity of architectures and training data, it may still have some vulnerabilities. In future work, we will explore the property of the botbot dialogue further and refine the set of partner systems for the bipartite-play method.
| 648 | 2,338 | 648 |
A Method for Stopping Active Learning Based on Stabilizing Predictions and the Need for User-Adjustable Stopping
|
A survey of existing methods for stopping active learning (AL) reveals the needs for methods that are: more widely applicable; more aggressive in saving annotations; and more stable across changing datasets. A new method for stopping AL based on stabilizing predictions is presented that addresses these needs. Furthermore, stopping methods are required to handle a broad range of different annotation/performance tradeoff valuations. Despite this, the existing body of work is dominated by conservative methods with little (if any) attention paid to providing users with control over the behavior of stopping methods. The proposed method is shown to fill a gap in the level of aggressiveness available for stopping AL and supports providing users with control over stopping behavior.
|
The use of Active Learning (AL) to reduce NLP annotation costs has generated considerable interest recently (e.g. Figure The terms aggressive and conservative will be used throughout the rest of this paper to describe the behavior of stopping methods. Conservative methods tend to stop further to the right in Figure There has been a flurry of recent work tackling the problem of automatically determining when to stop AL (see Section 2). There are three areas where this body of work can be improved: applicability Several of the leading methods are restricted to only being used in certain situations, e.g., they can't be used with some base learners, they have to select points in certain batch sizes during AL, etc. (See Section 2 for discussion of the exact applicability constraints of existing methods.) lack of aggressive stopping The leading methods tend to find stop points that are too far to the right in Figure instability Some of the leading methods work well on some datasets but then can completely break down on other datasets, either stopping way too late and wasting enormous amounts of annotation effort or stopping way too early and losing large amounts of F-measure. (See Section 4 for empirical confirmation of this.) This paper presents a new stopping method based on stabilizing predictions that addresses each of these areas and provides user-adjustable stopping behavior. The essential idea behind the new method is to test the predictions of the recently learned models (during AL) on examples which don't have to be labeled and stop when the predictions have stabilized. Some of the main advantages of the new method are that: it requires no additional labeled data, it's widely applicable, it fills a need for a method which can aggressively save annotations, it has stable performance, and it provides users with control over how aggressively/conservatively to stop AL. Section 2 discusses related work. Section 3 explains our Stabilizing Predictions (SP) stopping criterion in detail. Section 4 evaluates the SP method and discusses results. Section 5 concludes.
|
The margin exhaustion stopping criterion was developed for AL with SVMs (AL-SVM). It says to stop when all of the remaining unlabeled examples are outside of the current model's margin The confidence-based stopping criterion (hereafter, V2008) in Two stopping criteria (max-conf and min-err) are reported in To stop active learning at the point when annotations stop providing increases in performance, perhaps the most straightforward way is to use a separate set of labeled data and stop when performance begins to level off on that set. But the problem with this is that it requires additional labeled data which is counter to our original reason for using AL in the first place. Our hypothesis is that we can sense when to stop AL by looking at (only) the predictions of consecutively learned models on examples that don't have to be labeled. We won't know if the predictions are correct or not but we can see if they have stabilized. If the predictions have stabilized, we hypothesize that the performance of the models will have stabilized and vice-versa, which will ensure a (much-needed) aggressive approach to saving annotations. SP checks for stabilization of predictions on a set of examples, called the stop set, that don't have to be labeled. Since stabilizing predictions on the stop set is going to be used as an indication that model stabilization has occurred, the stop set ought to be representative of the types of examples that will be encountered at application time. There are two conflicting factors in deciding upon the size of the stop set to use. On the one hand, a small set is desirable because then SP can be checked quickly. On the other hand, a large set is desired to ensure we don't make a decision based on a set that isn't representative of the application space. As a compromise between these factors, we chose a size of 2000. In Section 4, sensitivity analysis to stop set size is performed and more principled methods for determining stop set size and makeup are discussed. It's important to allow the examples in the stop set to be queried if the active learner selects them because they may be highly informative and ruling them out could hurt performance. In preliminary experiments we had made the stop set distinct from the set of unlabeled points made available for querying and we saw performance was qualitatively the same but the AL curve was translated down by a few Fmeasure points. Therefore, we allow the points in the stop set to be selected during AL. Measurement of agreement between human annotators has received significant attention and in that context, the drawbacks of using percent agreement have been recognized where A o = observed agreement, and A e = agreement expected by chance. The different metrics differ in how they compute A e . The Kappa statistic where each c i is one of the coders (in our case, models), and P (k|c i ) is the probability that coder (model) c i labels an instance as being in category k. Kappa estimates P (k|c i ) based on the proportion of observed instances that coder (model) c i labeled as being in category k. We have found Kappa to be a robust parameter that doesn't require tuning when moving to a new dataset. On a separate development dataset, a Kappa cutoff of 0.99 worked well. All of the experiments (except those in Table The Kappa cutoff captures the intensity of the agreement that must occur before SP will conclude to stop. Though an intensity cutoff of K=0.99 is an excellent default (as seen by the results in Section 4), one of the advantages of the SP method is that by giving users the option to vary the intensity cutoff, users can control how aggressive SP will behave. This is explored further in Section 4. Another way to give users control over stopping behavior is to give them control over the longevity for which agreement (at the specified intensity) must be maintained before SP concludes to stop. The simplest implementation would be to check the most recent model with the previous model and stop if their agreement exceeds the intensity cutoff. However, independent of wanting to provide users with a longevity control, this is not an ideal approach because there's a risk that these two models could happen to highly agree but then the next model will not highly agree with them. Therefore, we propose using the average of the agreements from a window of the k most recent pairs of models. If we call the most recent model M n , the previous model M n-1 and so on, with a window size of 3, we average the agreements between M n and M n-1 , between M n-1 and M n-2 , and between M n-2 and M n-3 . On separate development data a window size of k=3 worked well. All of the experiments (except those in Table 3) in the current paper used a longevity window size of k=3 with zero tuning performed. We will see in Section 4 that this longevity default delivers robust results across all of the folds for all of the datasets. Furthermore, Section 4 shows that varying the longevity requirement provides users with another lever for controlling how aggressively SP will behave. We evaluate the Stabilizing Predictions (SP) stopping method on multiple datasets for Text Classification (TC) and Named Entity Recognition (NER) tasks. All of the datasets are freely and publicly available and have been used in many past works. For Text Classification, we use two publicly available spam corpora: the spamassassin corpus used in For NER, we use the publicly available GENIA corpus SVMs deliver high performance for the datasets we use so we employ SVMs as our base learner in the bulk of our experiments (maximum entropy models are used in Subsection 4.3). For selection of points to query, we use the approach that was used in Table 7 8 Following There are a few observations to highlight regarding the performance of the various stopping methods: • SP is the most parsimonious method in terms of annotations. It stops the earliest and remarkably it is able to do so largely without sacrificing F-measure. • All the methods except for SP and SC2000 are unstable in the sense that on at least one dataset they have a major failure, either stopping way too late and wasting large numbers of annotations (e.g. ZWH2008 and V2008 on TREC Spam) or stopping way too early and losing large amounts of F-measure (e.g. LS2008 on NER-Protein) . • It's not always clear how to evaluate stopping methods because the tradeoff between the value of extra F-measure versus saving annotations is not clearly known and will be different for different applications and users. This last point deserves some more discussion. In some cases it is clear that one stopping method is the best. For example, on WKB-Project, the SP method saves the most annotations and has the highest F-measure. But which method performs the best on NER-DNA? Arguments can reasonably be made for SP, SC2000, or ZWH2008 being the best in this case depending on what exactly the annotation/performance tradeoff is. A promising direction for research on AL stopping methods is to develop user-adjustable stopping methods that stop as aggressively as the user's annotation/performance preferences dictate. One avenue of providing user-adjustable stopping is that if some methods are known to perform consistently in an aggressive manner against annotating Task-Dataset SP V2008 7 SC2000 ZWH2008 too much while others are known to perform consistently in a conservative manner, then users can pick the stopping criterion that's more suitable for their particular annotation/performance valuation. For this purpose, SP fills a gap as the other stopping criteria seem to be conservative in the sense defined in Section 1. SP, on the other hand, is more of an aggressive stopping criterion and is less likely to annotate data that is not needed. A second avenue for providing user-adjustable stopping is a single stopping method that is itself adjustable. To this end, Section 4.3 shows how intensity and longevity provide levers that can be used to control the behavior of SP in a controlled fashion. Sometimes viewing the stopping points of the var-ious criteria on a graph with the active learning curve can help one visualize how the methods perform. Figure All of the additional experiments in this subsection were conducted on our least computationally demanding dataset, Spamassassin. The results in Tables 2 and 3 show how varying the intensity cutoff and the longevity requirement, respectively, of SP enable a user to control stopping behavior. Both methods enable a user to adjust stopping in a controlled fashion (without radical changes in behavior). Areas of future work include: combining the intensity and longevity methods for controlling behavior; and developing precise expectations on the change in behavior corresponding to changes in the intensity and longevity settings. The results in Table Effective methods for stopping AL are crucial for realizing the potential annotation savings enabled by AL. A survey of existing stopping methods identified three areas where improvements are called for. The new stopping method based on Stabilizing Predictions (SP) addresses all three areas: SP is widely applicable, stable, and aggressive in saving annotations. The empirical evaluation of SP and the existing methods was informative for evaluating the criteria but it was also informative for demonstrating the difficulties for rigorous objective evaluation of stopping criteria due to different annotation/performance tradeoff valuations. This opens up a future area for work on user-adjustable stopping. Two potential avenues for enabling user-adjustable stopping are a single criterion that is itself adjustable or a suite of methods with consistent differing levels of aggressiveness/conservativeness from which users can pick the one(s) that suit their annotation/performance tradeoff valuation. SP substantially widens the range of behaviors of existing methods that users can choose from. Also, SP's behavior itself can be adjusted through user-controllable parameters.
| 784 | 2,094 | 784 |
Predicting pragmatic discourse features in the language of adults with autism spectrum disorder
|
Individuals with autism spectrum disorder (ASD) experience difficulties in social aspects of communication, but the linguistic characteristics associated with deficits in discourse and pragmatic expression are often difficult to precisely identify and quantify. We are currently collecting a corpus of transcribed natural conversations produced in an experimental setting in which participants with and without ASD complete a number of collaborative tasks with their neurotypical peers. Using this dyadic conversational data, we investigate three pragmatic features -politeness, uncertainty, and informativeness -and present a dataset of utterances annotated for each of these features on a three-point scale. We then introduce ongoing work in developing and training neural models to automatically predict these features, with the goal of identifying the same betweengroups differences that are observed using manual annotations. We find the best performing model for all three features is a feedforward neural network trained with BERT embeddings. Our models yield higher accuracy than ones used in previous approaches for deriving these features, with F1 exceeding 0.82 for all three pragmatic features.
|
Autism spectrum disorder (ASD) is a neurological disorder associated with impairments in communication that can have a life-long impact on relationships, professional success, and personal independence Most previous work on identifying pragmatic features that index atypicality in expressive language relies on careful manual annotations of transcripts of spontaneous spoken language Our work addresses these aforementioned shortcomings in the previous work on pragmatic expression in ASD. In this paper, we describe an annotated corpus of conversations between adults with and without ASD and their neurotypical interlocutors as they engage in several collaborative tasks. Using this corpus, we investigate the degree of politeness, uncertainty, and informativeness in these conversations with the goal of identifying distinc-tive pragmatic features of ASD. We focus on these three features in particular because they are specific, remediable, and relevant in the collaborative discourse domain. When data collection is complete, we will release the transcribed and annotated dataset to researchers who have completed their institution's human subjects training. The dataset will be unique in that it is produced by adults, a subgroup of the ASD population that is both understudied and underserved. In addition, the dataset will consist entirely of spontaneous conversations with a peer, a rarity in ASD datasets. To our knowledge there is no single corpus manually annotated with all three features of politeness, uncertainty, and informativeness. Moreover, our corpus is already larger than any existing spoken language (as opposed to textual) corpus available for these features. With our annotated corpus, we propose several neural models for classifying utterances according to these features, and we explore whether our automated methods of generating these pragmatic features can be used to distinguish adults with ASD from their neurotypical peers as effectively as features derived via manual annotation. Our models outperform prior approaches to all three classification tasks, often by very wide margins. Although our predicted annotations do not capture all of the between-group differences observed using the manual annotations, we see promise in our approach.
|
We have collected spoken language data in a collaborative dyadic setting from adults 18 to 30 years of age with high-functioning ASD (n = 14) and with typical development (TD, n = 8). The ASD participants met the criteria for a diagnosis of ASD on the Autism Diagnostic Observation Schedule (ADOS) Each ASD or TD participant is paired with a neurotypical conversational partner (CP, n = 11), and together they engage in collaborative tasks involv- ing verbal communication and deliberation. The two tasks we focus on in this paper include a map task and a deserted island task. In the map task, styled after After transcription, the transcripts are then annotated for politeness, uncertainty, and informativeness These three features were chosen for a number of reasons. First, they are specific and interpretable, and as such, they are ideal features for targeted remediation. Secondly, they are especially relevant for and important in collaborative conversation; interviews, narratives, or monologues might be better analyzed using other features. Third, there are exist-ing corpora labelled for these features and available toolkits for extracting these features, which allows us to compare our work against prior baselines and will enable us to leverage external corpora in our future work. Finally, we note that politeness, in particular, has been cited as an area of deficit in ASD Politeness The politeness feature is a measure of how well an utterance contributes to a polite and collaborative dialogue, marked by agreeableness, positive attitudes, and willingness to compromise. A low politeness rating of 1 is given to utterances expressing frustration or criticism ("no you're wrong", "ugh how do I do this?") and utterances which use a more blunt way of phrasing commands ("go left"). A high politeness rating of 3 is given to utterances containing niceties (e.g., "thanks", "sorry") or highly positive words ("perfect", "awesome") and utterances that use a polite or indirect way of phrasing commands ("if you could make a left", "you want to make a left"). Uncertainty The uncertainty feature is defined to be a measure of the amount of uncertainty expressed about the correctness, validity, or permissibility of the utterance. A low uncertainty rating of 1 is given to utterances which express no uncertainty at all, or contain only a few filler words. A medium uncertainty rating of 2 is given to polar questions, either-or questions, short abandoned utterances, and utterances containing many filler words ("um", "uh") or hedge phrases ("I guess", "I'm assuming"). A high uncertainty rating of 3 is given to open questions ("where are you?") and utterances expressing explicit uncertainty or confusion ("I have no idea"). Informativeness The informativeness feature is defined as a measure for the overall information content and specificity of an utterance. A low informativeness rating of 1 is given to utterances which contain only polar answers ("yes", "no") or vague words with low specificity ("thing", "over there"). In the map task, a medium informativeness rating of 2 is given to utterances which contain words for general objects and do not specify a specific location on the map, and a high informativeness rating of 3 is given to utterances which contain proper nouns or labels or descriptions that can only point to one specific location on the map. In the island task, a rating of 2 is given to utterances which contain only an item word or a short phrase explaining the item, and a rating of 3 is given to utterances which contain multiple item words or a longer explanation of the items. After the transcripts are annotated for the pragmatic features described above, we train a number of machine learning models on the annotated data, with the goal of eventually being able to bypass the manual annotations and automate the annotation process using these predictive models. The models are given the transcribed and tokenized utterance converted to all lowercase and are tasked with predicting the categorical label for politeness, uncertainty, and informativeness based on the manual transcriptions. We start with several different baseline models, shown in Table We also evaluate against existing pre-trained models for rating politeness, uncertainty, and informativeness Additionally, because the scales used in the pretrained classifiers for politeness and informativeness are continuous and differ from our own categorical annotation scale, we use thresholding to convert the predictions to our scale. For example, to convert a continuous scale from 0 to 1 into a categorical scale from 1 to 3, we map any scores less than 0.33 to be 1, scores between 0.33 and 0.67 to be 2, and scores greater than 0.67 to be 3. Since the pre-trained uncertainty classifier only predicts a binary result of either 0 or 1 corresponding to certain or uncertain, we map their 0 rating to our 1 rating and their 1 rating to our 3 rating. We apply several methods for extracting sentence embeddings from the utterances in our dataset. First we use a basic sequences embedding in which each unique word appearing in the training data is assigned a unique identification number, and each utterance is then converted to a vector composed of the identification numbers for the words in the utterance, with padding for dimension consistency. With the sequence embeddings, we use a bidirectional LSTM model trained for 20 epochs with a batch size of 128. Additionally, we also use word embeddings from pre-trained word2vec The last type of embeddings that we employ are the contextualized word representations of BERT All our models are trained and evaluated with 5fold cross validation. For each fold, the accuracy, precision, recall and F1 of the predictions are calculated. Then the averages of these metrics across the 5 folds are computed as the indexes to evaluate model performance. Given the manual annotations, we examine whether there are significant differences between the ASD and the TD participant groups in terms of the three pragmatic features, using t-tests for significance testing. As shown in Table The difference in politeness between the two tasks could be partially due to the nature of the two tasks, as the map task requires the experimental participant to give instructions and commands to their conversational partner and thus presents greater opportunity and need for phrasing their statements in a more polite way. In contrast, in the island task, the two participants have equal roles, and there may be less need for phrasing statements more politely. These results suggest ASD participants tend to be more polite than their TD peers in tasks in which they have a leading or authority role. Furthermore, the structure of the task could also contribute to the difference in uncertainty in the two tasks. In the map task, the participant giving instructions has a clear, factual set of information to convey to their partner, while the island task is more subjective and requires more discussion between the two participants to agree on a set of items. This would suggest that ASD participants exhibit more uncertainty than their TD peers in open-ended tasks which require more discussion and exchange of opinion. The prediction results for all our models are presented in had a fairly high accuracy already. This was especially true for politeness, where the majority baseline had an F1 measure of 0.77. This is likely due to the distribution of the politeness ratings, since most statements fell into the neutral category of 2 for politeness, being neither particularly polite or impolite. Despite the high performance of the majority baseline however, all four models trained on our own data generally performed substantially better than all the baseline classifiers, especially for uncertainty and informativeness. The BERT model seemed to perform the best overall across all three features, while the sequences model also performed well for politeness and informativeness. In terms of the F1 measure, the feedforward model trained with BERT embedding outperforms the majority baseline by 0.1 for politeness, 0.33 for uncertainty, and 0.42 for informativeness. Since our goal is to investigate the differences in pragmatic expression between the two participant groups, we want our model to be able to capture the same group differences seen in the manual an-notations. To this end, we take the output for each group predicted from the best-performing model, the feedforward model using BERT embedding, and perform a t-test between the two groups as well. The results of significance testing based on model predictions are then compared to those given manual annotations. As presented in Table From the results of our study, we can see that there exist significant and quantifiable differences in pragmatic expressions between adults with ASD and their neurotypical peers. Moreover these dif-ferences are not fixed or consistent across all situations, but rather they may vary depending on the open-ended nature of the task, the roles involved, and the general context of the discourse. Relying on manual annotations of this sort, however, would not be practical or feasible in a clinical setting or for monitoring the efficacy of an intervention. To determine whether these annotations can be carried out automatically, we introduced several potential models trained on the annotated data. Although all of our models outperformed one or more of the baselines, the BERT model generally is superior for all three features. None of the models, however, were able to capture the statistically significant differences we observe in the manual annotations. There is still more work to be done in fine-tuning the model to capture between-group differences which are vital to our study of the pragmatic of adults with ASD. In our future work, we plan to extend the current study in at least three directions. First, we would like to employ different model architectures, leveraging external labeled corpora, with more systematic comparisons to see whether the differences between ASD and TD groups seen in manual annotations can be fully automatically derived. Second, after a long hiatus, we have recently resumed collecting data, with the goal of including 20 participants with ASD and 20 with typical development. Third, we aim to include annotations of other pragmatic features such as coherence and dialog acts in order to examine the differences of these features between ASD and neurotypical groups more comprehensively.
| 1,206 | 2,275 | 1,206 |
When Reviewers Lock Horn: Finding Disagreement in Scientific Peer Reviews
|
To this date, the efficacy of the scientific publishing enterprise fundamentally rests on the strength of the peer review process. The journal editor or the conference chair primarily relies on the expert reviewers' assessment, identify points of agreement and disagreement and try to reach a consensus to make a fair and informed decision on whether to accept or reject a paper. However, with the escalating number of submissions requiring review, especially in toptier Artificial Intelligence (AI) conferences, the editor/chair, among many other works, invests a significant, sometimes stressful effort to mitigate reviewer disagreements. Here in this work, we introduce a novel task of automatically identifying contradictions among reviewers on a given article. To this end, we introduce Con-traSciView, a comprehensive review-pair contradiction dataset on around 8.5k papers (with around 28k review pairs containing nearly 50k review pair comments) from the open reviewbased ICLR and NeurIPS conferences. We further propose a baseline model that detects contradictory statements from the review pairs. To the best of our knowledge, we make the first attempt to identify disagreements among peer reviewers automatically. We make our dataset and code public for further investigations 1 .
|
Despite being the widely accepted standard for validating scholarly research, the peer-review process has faced substantial criticism. Its perceived lack of transparency Feedback between authors and reviewers can help improve the peer review system It is to be noted that this AI-based system aims to aid editors in identifying potential contradictions in reviewer comments. While it provides valuable insights, it is not infallible. Editors should use it as a supplementary tool, understanding that not all contradictions may be captured and manual review remains crucial. They should make decision with careful analysis beyond the system's recommendations. Our contributions are three-fold: 1) We introduce a novel task: identifying contradictions/disagreement within peer reviews. 2) To address the task, we create a novel labeled dataset of around 8.5k papers and 25k reviews. 3) We establish a baseline method as a reference point for further research on this topic.
|
Artificial Intelligence (AI) has been applied in recent years to the realm of scientific peer review with the goal of enhancing its efficacy and quality As far as we know, contradiction detection in peer review has never been studied. Contradiction detection in peer reviews is complex and requires domain knowledge of the subject. It is not straightforward to detect contradictions in peer reviews because reviewers often have different writing styles and approaches to commenting. We believe that our work can significantly contribute to the peer review process. We utilize a subset of 8,582 out of 8,877 papers from the extensive ASAP-Review dataset We define review is as a collection of comments/sentences written by one reviewer. Formally, we can represent it as a list: A review pair takes two such lists, one from Reviewer1 and one from Reviewer2. It can be represented as: Lastly, a review pair comment selects one comment from each reviewer and forms a pair. It is a set of such pairs: To make it easier to annotate, we first create pairs of reviews of papers. Suppose, if there are n number of reviews in a paper then we create n 2 pairs resulting in a total of around 28k pairs. Detailed statistics regarding this dataset can be found in Table We follow the contradiction definition by De Given the complexity of the reviews and their frequent use of technical terminology, we had six doctoral students, each with four years of experience in scientific research publishing. To facilitate their training, two experts with more than ten years of experience in scientific publishing annotated 1,500 review pairs from a selection of random papers, following our guidelines. Our experts convened to discuss and reconcile any discrepancies in their annotations. The initial dataset comprises 227 pairs with contradictions and 1273 pairs without contradiction comments. We randomly selected 100 review pairs from this more extensive set to train our annotators, ensuring both classes are equally represented. Upon completion of this round of annotation, we reviewed and corrected any misinterpretations with the annotators, further refining their training and enhancing the clarity of the annotation guidelines. To evaluate the effectiveness of the initial training round, we compiled another 80 review pairs from both classes drawn from the remaining review pairs. From the second round on wards, most annotators demonstrated increased proficiency, accurately annotating at least 70% of the contradictory cases. We regularly monitored the annotated data, placing emphasis on identifying and rectifying inconsistencies and cases of confusion. We also implemented an iterative feedback system that continuously aimed to refine and improve the annotation process. In cases of conflict or confusion, we consulted experts to make the final decision. Following the annotation phase, we obtained an average interannotator agreement score of 0.62 using Cohen's kappa We compensated each annotator based on standard salaries in India, calculated by the hours they worked. The appointment and salaries are governed by the standard practices of our university. We chose not to pay per review pair because the time needed to understand reviews varies due to their complexity, technical terms, and the annotator's familiarity with the topic. Some reviews are also extensive, requiring more time to comprehend. Hence, basing pay on review pairs could have compromised annotation quality. To ensure accuracy and prevent fatigue, we set a daily limit of 6 hours for annotators. Figure We describe the flow of our proposed baseline setup through the flowchart in Figure Aspect Sentiment Model: Aspect and sentiment in peer review have been studied as a multi-task model Reviewer Disagreement Classifier We use techniques from Natural Language Inference (NLI) sentence pair classification to identify reviewer disagreement, particularly contradictions from Sentence Dependency Pairs (SDPs). Unlike traditional NLI tasks that provide a three-category output of "entailment", "contradiction", and "neutral", we have adjusted the model to a two-category output system: "contradiction" and "non-contradiction". The latter category combines "entailment" and "neutral" labels, as our primary focus is on contradiction detection. We discuss the implementation details in Appendix C. Table In order to analyze how models perform when trained on natural language inference datasets, we trained the models on the ANLI+ALL dataset and evaluated them on our test set. It was found that the models trained by combining datasets, i.e., SNLI, MNLI, FEVER, and ANLI A1+A2+A3, perform the best Next, we evaluate the performance of RoBERTa Large across the entire process (in evaluation mode). We obtain an accuracy of 88.60% from the Aspect Sentiment Classifier (in determining whether a pair of reviews has any SDAP or not) and a 74.25 F1 score for the Reviewer Disagreement Classifier. We compare our findings with those achieved by the zero-shot Large Language Model, ChatGPT. On the test set, ChatGPT scored an F1 of 64.67 which is 9 points lower than the baseline model, likely due to its lack of explicit training for this specific downstream task. We discuss the prompts and outputs in details in the Appendix E. We also discuss where our proposed baseline fails (Error Analysis) in the Appendix F. We also utilized the BARD API 2 for our evaluation. We provided identical prompts to both Google BARD and CHATGPT to ensure a fair comparison. We compared our findings with those achieved by Google BARD. BARD scored an F1 of 61.35 on the test set, 12 points lower than the baseline model. This is likely due to its BARD's requirement for more specialized training for this specific task. We 2 In this work, we introduce a novel task to automatically identify contradictions in reviewers' comments. We develop ContraSciView, a reviewpair contradiction dataset. We designed a baseline model that combines the MIMLLN framework and NLI model to detect contradictory review pair comments. We found that RoBERTa large performed best for this task with F1 score of 71.14. Our proposed framework doesn't consider the full review context when predicting contradictions. In the future, we will investigate a system that takes into account the entire review context while detecting contradictory review pair comments. Additionally, we plan to expand our dataset to include other domains and explore the significance of Natural Language Inference for the given task. We also aim to categorize contradictions based on their severity: high, medium, or low. Our study mainly focuses on identifying "explicit" contradictions. Explicit Contradictions: These are clear, direct, and unmistakable contradictions that can be easily identified. For example: The author claims in the introduction that "X has been proven to be beneficial," but in the discussion section, they state that "X has not been shown to have any benefit." One reviewer says, "Figure We do not delve into "implicit" contradictions, which can be hard to detect and can be subjective, making them a topic of debate. Implicit Contradictions: These are more subtle and may require deeper understanding or closer examination to identify. It may require the annotators to read the paper and also learn many related works or things to annotate. They are not directly stated but can be inferred from the context or how information is presented. For example: Review 1: "the method lacks algorithmic novelty and the exposition of the method severely inhibits the reader from understanding the proposed idea ." Review 2: "the work presented is novel but there are some notable omissions -there are no specific numbers presented to back up the improvement claims graphs are presented but not specific numeric results -there is limited discussion of the computational cost of the framework presented -there is no comparison to a baseline in which the additional learning cycles used for learning the embedding are used for training the student model ." For instance, Review1 mentions that the method lacks algorithmic novelty, and Review2 acknowledges the work as a novel but points out some notable omissions. This difference in perception is not a direct contradiction, as one reviewer finds the method lacking in novelty, while the other recognizes it as novel but with some omissions. Additionally, we do not incorporate information from the papers being reviewed, as they cover a wide range of topics and would require many experts from various domains. However, we acknowledge that finding a method to uncover these implicit contradictions in reviews is an intriguing opportunity for future research. We have utilized the open source peer review dataset for our work. The system sometimes generate incorrect contradictions or overlook certain contradictions. Like other general AI models, this model is not 100% accurate. Editors or Chairs might rely on more than just the contradictions predicted by the model to make decisions. It is important to emphasise that the primary purpose of this system is to assist editors by highlighting potential contradictions between reviewers; this system is only for editors' internal usage, not for authors or reviewers. Especially given their often busy schedules and the myriad of decisions they must make. The contradictions among the reviewers will help editors to identify and initiate a discussion to resolve the confusion between the reviewers and to make an informed decision. Since reviews are frequently extensive and intricate, it is challenging for editors to scrutinise every comment and address conflicts. This system aims to aid them in spotting such contradictions. For instance, if one comment reads, "The paper does not have any new findings," and another reviewer mentions, "The paper is somewhat novel," the editor might not immediately perceive this as contradictory. However, understanding the context and intent behind these comments is vital. If they represent a contradiction, editors can address it using the standard guidelines. The system does not detect all contradictions. If the system fails to identify a contradiction, it does not automatically mean none exists. Given its nature as a general AI system, there is a possibility of it presenting false negatives. Editors relying solely on this system could impact the review process. Any contradictions spotted outside the system's recommendations should also be addressed according to the guidelines. As an AI based model this is prone to errors, the editor/chair are advised to utilize this tool only for assistance and verify the contradiction and analyze carefully before making a decision. Recognizing Contradiction in a pair of reviews involves analyzing two pieces of text written by two different reviewers about the same paper. The following steps can be followed: Step 1: Begin by reading the Title and Abstract of the paper to gain an understanding of its subject matter. It is important to read these sections multiple times to grasp the paper's main points, such as its motivation, contributions, and other relevant aspects. If necessary, refer to the paper itself or read related material to enhance your understanding. Step 2: Proceed to read and comprehend the reviews of the paper, focusing on understanding the viewpoints expressed by the reviewers. Take note of their opinions, arguments, and any specific aspects they highlight. Step 3: Based on the reviewer comments, their aspect-sentiment analysis, and the pair of reviews, you should categorize them accordingly. If the reviews contradict each other, indicating opposing viewpoints or conflicting statements, mark it as a contradiction (C). If there is no contradiction or the reviews are unrelated, mark them as noncontradiction (N). In cases where you find it difficult to determine if there is a contradiction or not, mark it as CON (confused). To decide if the reviews are contradictory, ask yourself the following question: "If I were shown two contemporaneous documents one containing each of these passages, would I regard it as very unlikely that both passages could be true at the same time? If so, the two contradict each other." You should be able to state a clear basis for a contradiction. For example, the following are contradictions: R1. The motivation of using the method for a very small improvement is not convincing. R2. Overall feedback: I found the paper to be well motivated and the proposed approach to be interesting. How to make use of the aspect category while annotation? Given the assigned aspect category, you should utilize the given category to accurately compare these comments. This will help keep your attention strictly on the relevant aspect and refrain from deviating to other topics mentioned in the review text. For example R2: ...the comparison to JPEG2000 is unfortunately not that interesting since that codec does not have widespread usage and likely never will. In Review 1, the reviewer values the comparison made with JPEG 2000, considering it a suitable benchmark and a positive aspect of the paper. Conversely, the reviewer in Review 2 argues that the comparison with JPEG 2000 is not compelling due to its limited use and suggests a comparison with WebP would be more relevant. These statements present opposing viewpoints on the relevance of comparing with JPEG 2000 in the paper. While there are many topics discussed in Review 1 and Review 2, it is crucial to concentrate on the remarks concerning the comparison between JPEG and other potential benchmarks when assessing these reviews. When comparing reviews, keep in mind that one might discuss the subject matter in broad terms, while the other may focus on specific elements. Consider this example: R1: The work is original. R2: The extension to the partially observable setting is interesting as the proposed form finds a common denominator to multiple estimators, but its underlying idea is not novel. In this case, Reviewer 1 comments on the overall originality of the paper, declaring it as novel. On the other hand, Reviewer 2 critiques a specific part of the paper-the extension to the partially observable setting-and declares this particular aspect as not novel. Even though Reviewer 2's comment is more specific, it contradicts the general assertion made by Reviewer 1 about the paper's novelty. It is important to identify these types of examples as contradictions, even if they may seem to operate at different levels of specificity. The broad comment about the paper's novelty in Review 1 is contradicted by the specific critique in Review 2. You may find more detailed annotation guidelines in our shared repository. MIMLLN for Aspect-Category Sentiment Analysis (ACMIMLLN) operates on the assumption that the sentiment of a mentioned aspect category in a sentence aggregates the sentiments of words that indicate that aspect category. In MIMLL, words that indicate an aspect category are termed 'key instances' of that category. Specifically, AC-MIMLLN comprises two components: an attentionbased aspect category detection (ACD) classifier and an aspect-category sentiment analysis (ACSA) classifier. Given a sentence, the ACD classifier, as an auxiliary task, assigns weights to the words for each aspect category. These weights signify the likelihood of the words being key instances of their respective aspect categories. The ACSA classifier initially predicts the sentiments of individual words. It then determines the sentence-level sentiment for each aspect category by integrating the respective weights with the word sentiments. The ACD segment comprises four modules: an embedding layer, an LSTM layer, an attention layer, and an aspect category prediction layer. Similarly, the ACSA segment includes four components: an embedding layer, a multi-layer Bi-LSTM, a word sentiment prediction layer, and an aspect category sentiment prediction layer. In the ACD task, all aspect categories utilize the same embedding and LSTM layers, but they have distinct attention and aspect category prediction layers. For the ACSA task, all aspect categories share the embedding layer, the multi-layer Bi-LSTM, and the word sentiment prediction layer, yet they each have unique aspect category sentiment prediction layers. We implemented our system using PyTorch For the Aspect Sentiment Model, we conducted experiments with different network configurations during the validation phase. ASAP-Review dataset contains predictions labelled by an aspect tagger on a human-annotated label. We used the 1,000 human-annotated reviews, maintaining the same random split, to train the MIMLLN classifier. Through these experiments, we determined that a batch size of 16 and a dropout rate of 0.5 for every layer yielded optimal performance. The activation function ReLU was used in our model. We trained the model for 15 epochs, employing a learning rate of 1e-3 and cross-entropy as the loss function. To prevent overfitting, we used the Adam optimizer with a weight decay of 1e-3. For the Reviewer Disagreement Classifier, we trained the models using true Sentiment Disparity Aspect pairs with true aspect and sentiment labels. We use a batch size of 16, a maximum length of 280 tokens, and a dropout probability of 0.1. The Adam optimizer was employed with a learning rate of 1e-5. All models were trained on an NVIDIA A100 40GB GPU. We present the results of the Aspect Sentiment Model in Table Regarding aspect-based sentiment, BERT performed the best for the aspects of Motivation, Clarity, Substance, and Replicability. However, for the Soundness category, SPECTER performed better. As for the Meaningful Comparison aspect, RoBERTa showed better performance. We conducted a comparative study between our proposed baseline model and ChatGPT and BARD 5 , focusing on multi-turn dialogue inputs. We used the OpenAI API 6 for the task. To evaluate the effectiveness of our model, we experimented with various prompts, selecting the most effective one. The prompt chosen is particularly adept at extract-5 During our testing, we identified many false positive cases in the output of ChatGPT. For example, when presented with the reviews (paper id: ICLR_2018_456): • R1: The argument of biological plausibility is not justified • R2: Moreover the biological plausibility that is used as an argument at several places seems to be false advertising in my view ChatGPT incorrectly labeled these reviews as contradictions. However, our baseline system accurately identified that these statements do not present a contradiction. Additionally, ChatGPT sometimes compares two distinct aspects. For example: R1: This paper is not well-written. R2: The results are reasonable and significant. In this instance, the first reviewer is commenting on the clarity of the paper, while the second reviewer is commenting on the substance of the paper. These represent two different perspectives. Chat-GPT occasionally fails in such scenarios. However, our proposed baseline incorporates an intermediate aspect sentiment model, which helps reduce these types of errors. We conducted an analysis of the predictions made by our proposed baseline to identify the areas where it most frequently fails. Our proposed baseline can sometimes fail in accurately predicting contradictions when the sentences are significantly related to preceding comments within the same review. Take, for instance, the following comment: "This is an elegant intuitive algorithm that, to my knowledge, has not appeared in previous literature." In this case, the model incorrectly predicts a contradiction with another review comment that discusses a different algorithm negatively. The issue arises from the model's inability to distinguish between the algorithm discussed in this comment and a different algorithm mentioned negatively by another reviewer. While prediction the model needs to learn from previous discussions as well. Such an approach can be considered as a potential area for future work. Our proposed baselines sometimes stumble in predicting the right contradictions when dealing with particularly long sentences or complex review comments (significant amount of technical terms or mathematical symbols) which can lead to confusion. We found error propagation from the first model to the second. To illustrate, consider the following example: Reviewer 1's comment reads: "While it is very interesting to apply adversarial noise in real data, this approach is not clearly motivated or explained." The Aspect sentiment model predicts the aspect for this comment as 'soundness' but misses out on 'motivation'. Reviewer 2's comment states: "In overall, I liked its clear motivation and the simplicity of the method." For this comment, the predicted aspect category by the model is 'motivation'. These two comments provide contrasting viewpoints on the aspect of 'motivation'. Yet, due to the misclassification by the aspect sentiment model, this discrepancy isn not flagged by the second model.
| 1,291 | 971 | 1,291 |
Diverse Parallel Data Synthesis for Cross-Database Adaptation of Text-to-SQL Parsers
|
Text-to-SQL parsers typically struggle with databases unseen during the train time. Adapting parsers to new databases is a challenging problem due to the lack of natural language queries in the new schemas. We present REFILL, a framework for synthesizing highquality and textually diverse parallel datasets for adapting a Text-to-SQL parser to a target schema. REFILL learns to retrieve-andedit text queries from the existing schemas and transfers them to the target schema. We show that retrieving diverse existing text, masking their schema-specific tokens, and refilling with tokens relevant to the target schema, leads to significantly more diverse text queries than achievable by standard SQL-to-Text generation methods. Through experiments spanning multiple databases, we demonstrate that fine-tuning parsers on datasets synthesized using REFILL consistently outperforms the prior data-augmentation methods.
|
Natural Language interface to Databases (NLIDB) that translate text queries to executable SQLs is a challenging task in the field of Semantic Parsing Synthesizing parallel data, that is representative of natural human generated queries We propose a framework called REFILL ( § 2) for generating diverse text queries for a given SQL workload that is often readily available We summarize our contributions as follows: (i) We propose the idea of retrieving and editing natural text from several existing schemas for transferring it to a target schema, obtaining higher text diversity compared to the standard SQL-to-Text generators. (ii) We design strategies for masking schema-specific words in the retrieved text and training the REFILL model to fill in the masked positions with words relevant to the target schema. (iii) We filter high-quality parallel data using a binary classifier and show that it is more efficient than existing methods based on cycle-consistency filtering. (iv) We compare REFILL with prior dataaugmentation methods across multiple schemas and consistently observe that fine-tuning Text-to-SQL parsers on data generated by REFILL leads to more accurate adaptation.
|
Our goal is to generate synthetic parallel data to adapt an existing Text-to-SQL model to a target schema unseen during training. A Text-to-SQL model M : X , S → Q maps a natural language question x ∈ X for a database schema s ∈ S, to an SQL query q ∈ Q. We assume a Textto-SQL model M trained on a dataset D train = {(x i , s i , q i )} N i=1 consisting of text queries x i for a database schema s i , and the corresponding gold SQLs q i . The train set D train typically consists of examples from a wide range of schemas s i ∈ S train . For example, the Spider dataset Dsyn ← Dsyn ∪ Filter(q, {x q r }) 8 Mnew ← fine-tune(M, Dsyn) existing databases a substantial SQL workload is already available in the query logs at the point a DB manager decides to incorporate the NL querying capabilities Figure Given an SQL q ∈ QW s sampled from SQL workload, we extract SQL-Text pairs {q r , x r } ∈ D train , from the train set such that the retrieved SQLs {q r } are structurally similar to the SQL q. We utilize tree-edit-distance Our next goal is to translate the retrieved x r from being a text for SQL q r to a text x for SQL q , where q ≈ q r structurally. However, we do not have a readily labeled dataset to learn a model that translates x r to x while being consistent with q. We therefore decompose this task into two steps: 1) A simpler task of masking schema-specific tokens in x r to get a template x masked r and 2) A conditional text generation model that maps (x masked r , q) to the text x consistent with q, by filling the masked positions in x masked r as per q. We re-purpose D train to get indirect supervision for training the text generation model. We now present each step in detail. Masking the retrieved text Converting the retrieved text queries {x r } to masked templates {x masked r } is a critical component of REFILL's pipeline since irrelevant tokens like references to schema elements of the original database can potentially misguide the text generation module. Our initial approach was to mask tokens based on a match of text tokens with schema names and manually refined schema-to-text linked annotations as in Editing and Filling the masked text Given a masked template x masked r , and an SQL query q, we wish to edit and fill the masked portions in x masked r to make it consistent with the SQL q. We utilize a conditional text generation model B like BART Since we do not have direct supervision to finetune B for this task, we present a method of repurposing D train for fine-tuning B. D train contains SQL-Text pairs (q i , x i ) from various schemas s i . A Naïve way to train B is to provide [x masked i |q Eng i ], the concatenation of x masked i and q Eng i as an input to the encoder and maximize the likelihood of x i in the decoder's output. This way the decoder of B learns to refill the masked tokens in x masked i by attending to q Eng i to recover x i in the output. While useful for learning to refill the masked positions, this Naïve method of training B is mismatched from its use during inference in two ways: (i) For a given SQL q, REFILL might fail to retrieve a similar structure neighbour of q i from D train . In such cases, B should be capable of falling back to pure SQL-to-Text generation mode to directly translate q into x. (ii) During inference, x masked r and q come from different schemas. However, during Naïve training, the masked text x masked i and the SQL q i are derived from the same example (q i , x i ). To address these two limitations, we train B in a more Robust manner as follows: (a) For a random onethird of the train steps we train B in the Naïve way, allowing B to learn the filling of the masked tokens using q Eng i . (b) For another one-third, we pass only q Eng i as an input and maximize the likelihood of x i . This ensures that model is capable of generating the text from the q Eng i alone, if the templates x masked i are unavailable or noisy. (c) For the remaining onethird, we first retrieve an SQL-Text pair (q j , x j ), from a different schema such that the SQL q j is structurally similar to q i ( § 2.1), and the word edit distance between the masked templates x masked i and x masked j is also small. We can then replace x masked i with x masked j and encode [x masked j |q Eng i ] as an input to B and maximize the likelihood of x i in the decoder's output. This step makes the training more consistent with the inference, as x masked j and q Eng i now come from different schemas. In § 5.4, we justify training Robustly compared to Naïve training. Since the data synthesized using REFILL is used to fine-tune a downstream Text-to-SQL parser, we learn a Filtering model F : (X , Q) → R to discard inconsistent examples from the generated dataset. F assigns lower scores to inconsistent Text-SQL pairs. For each SQL q ∈ QW s , we select the top-5 sentences generated by REFILL and discard all the sentences that are scored below a fixed threshold as per the filtering model. Existing work depended on a trained Text-to-SQL parser M to assign cycleconsistency scores We instead train a filtering model F as a binary classifier, independent of M. The Text-SQL pairs {(x i , q i )} in the training set D train , serve as positive (consistent) examples and we synthetically generate the negative (inconsistent) examples as follows: (i) Replace DB values in the SQL q i with arbitrary values sampled from the same column of the database. (ii) Replace SQL-specific tokens in q i with their corresponding alternates e.g. replace ASC with DESC, or '>' with '<'. (iii) Cascade previous two perturbations. (iv) Replace the entire SQL q i with a randomly chosen SQL q j from the same schema. (v) Randomly drop tokens in the text query x i with a fixed probability of 0.3. (vi) Shuffle a span of tokens in the text query x i , with span length set to 30% of the length of x i . Thus, for a given Text-SQL pair (x i , q i ) we obtain six corresponding negative pairs {(x n j , q n j )} 6 j=1 . Let s i be the score provided by the filtering model for the original pair (x i , q i ) and {s j } 6 j=1 be the scores assigned to the corresponding negative pairs {(x n j , q n j )} 6 j=1 . We supervise the scores from the filtering model using a binary-cross-entropy loss over the Sigmoid activations of scores as in Equation To explicitly contrast an original pair with its corresponding negative pairs we further add another Softmax-Cross-Entropy loss term. 3 Related Work Retrieve and Edit Methods Our method is related to the retrieve-and-edit framework, which has been previously applied in various NLP tasks. In Semantic Parsing, question and logical-form pairs from the training data relevant to the test-input question are retrieved and edited to generate the output logical forms in different ways 4 Experimental Set-up We adapt pretrained Text-to-SQL parsers on multiple database schemas unseen during the train time. Here, we describe the datasets, models, and evaluation metrics used in our experiments. Datasets: We primarily experiment with the Spider dataset We further experiment with four datasets outside Spider in Section 5.6. We work with Geo-Query We provide more implementation details including the hyperparameter settings in appendix A.5. We first demonstrate the effectiveness of the synthetic data generated using REFILL for fine-tuning Text-to-SQL parsers to new schemas. We compare with the recent methods that utilize SQL-to-Text generation for training-data augmentation ( § 5.1). We then evaluate the intrinsic quality of the synthetic data generated by different methods in terms of the text diversity and the agreement of the generated text with the ground truth ( § 5.2). We demonstrate that higher text diversity results in better performance of the adapted parsers ( § 5.3). We then justify the key design choices related to masking of the retrieved text and training of the schema translator module that improves the quality of REFILL generated text ( § 5.4). Finally, we demonstrate the importance of using an independent binary classifier over cycle-consistency filtering ( § 5.5). In Table We explain our gains over existing methods due to the increased quality and diversity of the generated text. We measure quality using the BLEU score of the set S(q) of generated text for an SQL q, with the gold text of q as reference. To measure diversity we utilize SelfBLEU Retrieving and editing text from multiple existing examples enables REFILL to generate diverse text. In Figure In Section 2.2, we described two important design choices: (1) Method of masking schema-relevant tokens and (2) Method of training the Edit-and-Fill model for editing and refilling the masked text. We justify these design choices by comparing the quality of the generated text with each combination of these choices in Table Cycle-consistency based filtering We further validate our method on four singledatabase datasets outside Spider, namely Geo-Query This work focuses on synthesizing parallel data containing diverse text queries for adapting pretrained Text-to-SQL models to new databases. Thus, our current effort toward diverse text query generation using REFILL is limited to the Text-to-SQL semantic parsing task. Extending REFILL for data-augmentation in other semantic parsing or question-answering tasks is an exciting direction we hope to explore as part of future work. Our experimental set-up assumes a small workload of real SQL queries. As per Baik et al. ( Our goal with REFILL is to synthesize parallel data for adapting Text-to-SQL parsers to new schemas. We believe that the real-world deployment of Textto-SQL or any semantic parser trained on text generated by language models must go through a careful review of any harmful biases. Also, the intended users of any Text-to-SQL service must be made aware that the answers generated by these systems are likely to be incorrect. We do not immediately foresee any serious negative implications of the contributions that we make through this work.
| 913 | 1,187 | 913 |
CS2W: A Chinese Spoken-to-Written Style Conversion Dataset with Multiple Conversion Types
|
Spoken texts (either manual or automatic transcriptions from automatic speech recognition (ASR)) often contain disfluencies and grammatical errors, which pose tremendous challenges to downstream tasks. Converting spoken into written language is hence desirable. Unfortunately, the availability of datasets for this is limited. To address this issue, we present CS2W, a Chinese Spoken-to-Written style conversion dataset comprising 7,237 spoken sentences extracted from transcribed conversational texts. Four types of conversion problems are covered in CS2W: disfluencies, grammatical errors, ASR transcription errors, and colloquial words. Our annotation convention, data, and code are publicly available at
|
Automatic speech recognition (ASR) plays a vital role in a wide range of NLP application scenarios, such as simultaneous interpretation, where verbal utterances are transcribed into spoken style texts. These transcriptions serve as fundamental inputs to plenty of downstream tasks. However, they often inherently contain disfluencies, grammatical errors, and colloquial words, which pose tremendous challenges on downstream tasks. Automatically correcting errors and editing spoken into written language would significantly benefit downstream tasks that are usually trained on canonical texts. Developing such spoken-to-written style conversion models usually requires labeled data that cover different phenomena in ASR-transcribed spoken style texts. Unfortunately, existing datasets usually focus on a single type of spoken style, such as disfluencies. Consequently, models trained on these datasets cannot address all spoken style issues. Table To bridge this gap, we propose CS2W, a largescale fine-grained Chinese Mandarin spoken-towritten style conversion dataset, developed on the Real Spontaneous Dialogue Speech dataset MagicData-RAMC We conduct a thorough and in-depth analysis on Chinese spoken texts and summarize four conversion problems: disfluencies, ASR transcription errors, grammatical errors, and colloquial words. The four conversion problems are common in Chinese spoken texts and cover the major tasks in spokento-written style conversion (i.e., grammatical and ASR error correction, and simplification). • Disfluency: Repetitions, restarts, or repairs in spontaneous communication. • ASR transcription errors: Occasional homophone mistakes in ASR transcriptions. • Grammatical errors: Missing words, incorrect syntax structures, etc., similar to those occurring in written style texts. • Colloquial words: Problems related to colloquial words that differ from written language. Table In comparison to existing datasets that focus on grammatical errors or disfluencies, our dataset contains more conversion types and closely aligns with the distribution of real-world spoken data. For example, in the commonly used SWITCHBOARD corpus Our contributions are as follows: 1. We curate and release CS2W, the first opensource Chinese dataset for spoken-to-written style conversion. The dataset is derived from real-world spontaneous conversations. We provide fine-grained annotation along with written style manually normalized texts. Additionally, we establish a comprehensive set of criteria for spoken-to-written style conversion classification and annotation. 2. We conduct an in-depth analysis on the distribution of spoken-to-written style conversion problems and identify new types of disfluencies. 3. We conduct benchmark evaluation experiments on CS2W to evaluate the performance of cutting-edge large language models on A flight to um Boston I mean Denver Tuesday RM FW IM RP
|
Figure spoken-to-written language conversion. Experiment results demonstrate that the conversion from spoken to written language effectively improves the performance of downstream tasks. Previous studies have treated disfluency detection and grammatical error correction as separate tasks. We hence review their progress separately. A comprehensive comparison of GEC, disfluency detection, and spoken-to-written style conversion datasets is presented in Table Unfortunately, few disfluent detection datasets Grammatical Error Correction and Related Datasets Common errors in Chinese grammatical error correction (GEC) datasets include spelling errors, missing words, redundant words, incorrect word order, collocation errors, etc. Seq2Seq models, based on RNN/CNN or Transformer, are usually used for Chinese GEC tasks. Numerous datasets have been proposed for GEC. These datasets primarily use two main paradigms to label the data: error-coded and direct rewriting. In the error-coded paradigm, annotators are tasked with explicitly identifying erroneous spans in the original sentence, specifying the error type, and subsequently making corrections. For instance, FCE In contrast to the existing GEC datasets, our CS2W dataset places a stronger emphasis on grammatical errors occurring in the spoken domain rather than the written domain. Notably, while all previously established datasets feature a single annotation paradigm, the CS2W dataset stands out by offering two distinct annotation paradigms. Spoken-to-Written Style Conversion and Related Datasets Spoken-to-written style conversion can be formulated as a monolingual translation task Publicly available spoken-to-written style conversion data are scarce, with the majority of datasets being private. In this section, we elaborate on the creation of the CS2W dataset. CS2W is built upon the existing MagicData-RAMC dataset The spoken-to-written style conversion normally involves lexical and syntactical editing and style transfer. The former is similar to grammatical error correction, dealing with disfluencies, ASR transcription errors, and grammatical errors while the latter is for the translation of colloquial words into canonicalized words with the same meaning used in written style texts. To annotate the selected sentences, we employ two established annotation paradigms: error-coded and direct rewriting, commonly used in grammatical error correction datasets Initially, for each sentence, we use the errorcoded paradigm to meticulously identify spokento-written style conversion problems and determine the types of these problems (i.e., ∈ disfluency, ASR transcription error, grammatical error, colloquial word). We also pinpoint the specific spans within the sentences where these problems manifest. It's essential to recognize that a sentence may exhibit multiple conversion problems simultaneously. However, it is worth noting that Table To ensure the consistency and quality of our annotations, we implement a two-round annotation process. In the first round of annotation, we enlist the expertise of eight Chinese native speakers as parttime annotators after pre-annotation training. We require that each conversion problem within a given spoken style text be annotated in accordance with established conventions, assuring the validity and uniformity of the annotations. In the second round, we conduct a manual evaluation and re-annotation, with the authors of this paper serving as senior annotators. Each annotated instance is scrutinized according to the answers to the following three core questions: 1. Are the type and span annotations correct? This step ensures the accuracy of annotations under the error-coded paradigm. 2. Does the written style reference faithfully retain the meaning of the original spoken style text? This step ensures the accuracy of annotations under the direct rewriting paradigm. 3. Is the modification in the written language reference consistent with the conversion problems being labeled? This ensures the alignment between the annotations under the errorcoded paradigm and those under the direct rewriting paradigm. If the annotators in the second round encounter any inconsistencies with the first-round annotations, they submit their annotations to the senior annotators. The senior annotators then perform a comparative analysis of the results from both rounds and issue the final annotation verdict. Detailed guidelines for resolving inconsistencies are provided in Appendix B. We provide data statistics and analyses on conversion type distribution, sentences with multiple conversion problems, and a new subtype of disfluency existing in our dataset, which is absent in previous studies. To assess the distribution over different conversion types, we calculate the proportions of each conversion type present in the dataset. Results are illustrated in Figure In contrast to previous disfluency detection and grammatical error correction datasets, where each sentence typically has only one problem, as depicted in Figure Our analysis of the most prevalent conversion problem, disfluency, in the context of spoken-to-written style conversion, has revealed a novel disfluency pattern. In the conventional R-type disfluency, the reparandum (the incorrect portion) typically precedes the repair (the corrected portion). This pattern aligns with the common observation that speakers often correct themselves upon realizing an error. However, in this distinct disfluency pattern, the repair precedes the reparandum, as illustrated in Figure Traditional disfluency correction models commonly delete the portion preceding the disfluency to generate a grammatically correct sentence. However, when confronted with this specific pattern, such an approach could potentially compromise the accuracy of the correction process. We conducted experiments on the curated dataset to evaluate the performance of the advanced opensource large language models (LLMs) on Chinese spoken-to-written style conversion. We ensured randomness in our data selection process by shuffling all the annotated texts. Subsequently, we partitioned these texts into training, development, and test sets, maintaining a distribution ratio of 8:1:1. The training set encompasses 5,789 texts, while both the development and test sets consist of 724 texts each. We used the following baseline models: Bamboo is not only is not a food for pandas, but it can also be made into a lot of furniture. Bamboo is not only a food for pandas, but it can also be made into a lot of furniture. Bamboo is not a food for pandas, but it can also be made into a lot of furniture. Figure • BART • CPT We fine-tuned BART and CPT with the training set of CS2W and tested BLOOM, BELLE, ChatGLM, and GPT3.5-turbo under the zero-and 5-shot settings. The demonstrations used for the 5-shot setting and the method of selecting them are described in detail in Appendix C. The numbers of parameters of the models, as well as the hyperparameters, are described in detail in Appendix D. We evaluated models with BLEU Table Under the zero-shot setting, ChatGLM exhibits a slight advantage over GPT3.5-turbo in terms of BLEU-3 and BLEU-4 scores, while GPT3.5-turbo excels in the remaining metrics. In contrast, under the 5-shot setting, GPT3.5-turbo emerges as the frontrunner, achieving the highest scores across all metrics. It's worth noting that, except for BELLE-7B-0.2M, all models demonstrate improved performance under the 5-shot setting when compared to their zero-shot setting performance. Table To assess the impact of spoken-to-written language conversion on downstream tasks, we employed Chinese-to-English machine translation as a representative task. We randomly selected 100 normalized text references from the test set and then manually translated them into English to serve as references for the Chinese-to-English machine translation task. We utilized OPUS-MT We conducted further experiments on the BLOOM series models to investigate the impact of the number of demonstrations in the prompts on these LLMs. Specifically, we selected BLOOM series models with parameter of 1.1B In experiments with two to three demonstrations, limited prompt diversity leads to a performance decrease. With more demonstrations and diversity, the model performance gradually increases again. In this paper, we have presented a dataset CS2W, which is the first open-source Chinese spoken-towritten conversion dataset. The dataset covers four types of conversion problems commonly occurring in Chinese spoken texts. We manually annotate the type and span for each conversion problem and provide high-quality written style normalized texts. The dataset is used as a benchmark testbed to evaluate the performance of advanced LLMs on spoken-to-written style conversion and would promote future research on this underexplored direction. For all zero-shot and few-shot experiments, we used the same prompt for all models. However, prompt selection is important for large language models. We plan to use more prompts and prompt engineering methods to conduct experiments on the curated dataset in the future. All data used in this study are freely available to the public. The raw data are from a public dataset built in previous work. We follow the policy of using these data without infringing any copyright issues. The curated dataset in this study is for academic research purposes only. All annotators are well paid according to the number of their annotations. MagicData-RAMC is an ASR transcription dataset that comprises 351 sets of spontaneous conversations in Chinese Mandarin. Each set features natural dialogues between two speakers discussing a single topic, and it includes both audio files and transcribed texts that preserve disfluencies, grammatical errors, and ASR transcription errors. We manually select sentences from the ASR transcriptions, ensuring they are self-contained in meaning but with conversion problems for annotation. In transcribed text, we do not extract the following three types of sentences: Incomplete Sentences, Sentences that require context, and Sentences that are too short. In spontaneous conversations, a speaker often breaks off abruptly or is interrupted by another speaker, resulting in many incomplete sentences in transcriptions. e.g. 我和他刚刚准备出去玩，就。 (He and I were just getting ready to hang out but.) This indicates that the speaker stops speaking at the word "but" or is interrupted by someone else. Incomplete sentences cannot be easily understood by annotators and, as a result, may not be accurately annotated. Spontaneous conversations have continuity and many sentences need to be understood in context. e.g. Speaker 1: 你这次数学考试考了多少 分？ (Speaker 1: What grade did you get on this math test? ) Speaker 2: 一 百 多 一 点 ， 比 不 上 你 。 (Speaker 2: little over a hundred, no more than you. ) Speaker 2's speech is a response to Speaker 1's question, which may not be fully comprehensible without considering Speaker 1's speech. Sentences that require context can also be challenging to annotate accurately. In spoken language, phrases such as "no problem," "yes," and "okay" are frequently employed, and these are considered too general. Therefore, we refrain from selecting sentences with fewer than 5 tokens. Second In conclusion, we select complete sentences with conversion problems whose lengths are appropriate. Automatic speech recognition (ASR) plays a vital role in a wide range of NLP application scenarios. Spoken language, which serves as a fundamental input for plenty of downstream tasks, is transcribed into written text in a spoken style. However, they often inherently contain disfluencies, grammatical errors, and colloquial words. This dataset consists of transcribed texts with conversion problems for annotation. For each sentence, the annotator needs to annotate the type and the range of conversion problems and write the corresponding written language. We further categorize the conversion problem with a two-level classification system shown in Figure The elements that make a sentence not fluent are referred to as "disfluency", which can be categorized into R-type and Filler Words based on their structures. Filler Words Filler Words, such as "uh" and "ah," have no specific meaning and are often used to indicate pauses and hesitations in the speaker's discourse. Additionally, common words like "yeah" and "okay" are also sometimes classified as filler words. e.g. 好吧，这是个，嗯，一个不错的主意。 (Well, this is, you know, a good plan.) In this sentence, the phrases "well" and "you know" lack specific meaning and should be annotated as "Filler words." R-type The standard structure of R-type disfluency encompasses three elements: the reparandum, an optional interregnum, and the associated repair. The reparandum consists of words that the speaker initially intends to discard, representing an unintended inclusion in the utterance. This section typically comprises one or more words slated for repetition or correction. The interregnum, often comprising fixed phrases like "uh" or "you know," serves as a non-lexicalized component, contributing filler words without specific meaning. Lastly, the repair phase involves correcting or repeating words from the reparandum, thereby refining the overall coherence of the utterance. e.g. 让我们，我的意思是，让我来解决这 个问题。 (Let us, I mean, let me work on the problem.) In the provided example, the sentence "Let us" functions as the reparandum, embodying the words originally unintended for inclusion. The subsequent phrase "let me" constitutes the repair, correcting the preceding reparandum. The interregnum in this instance is "I mean," a non-essential filler phrase devoid of substantive meaning. ASR Transcription Errors are occasional homophone mistakes in ASR transcriptions. CS2W is built upon the existing MagicData-RAMC dataset e.g. 这个艺术家很有菜花。 (The artist is very cauliflower. According to the intended meaning of the sentence, the correct version should be "The artist is very talented." Therefore, "cauliflower" needs to be annotated as an "ASR Transcription Error." The transcription text of spoken language often includes grammatical errors because speakers in conversations often lack careful thinking. Common grammatical errors in spoken language include Missing Words, Redundant Words, and Incorrect Word Order. Missing Words Missing Words include missing subjects, missing predicates, missing objects, missing function words, and missing modifiers. e.g. 那时我们有机会扳平比分，但是我们没 有机会。 (We had a chance to equalize, but we didn't it.) This sentence is missing a verb. The correct sentence is "We had a chance to equalize, but we didn't take it". Redundant Words Redundant Words include redundant subjects, redundant predicates, redundant objects, redundant function words, and redundant modifiers. e.g. 它们的皮毛很有光泽,可以用肉眼很难看 出来。 (Their fur is shiny and can be hardly seen with the naked eye.) The modifier in this sentence is redundant. The fur is shiny so it should be visible to the naked eye. The word "hardly" should be deleted. Incorrect Word Order Incorrect Word Order is also common in spoken transcribed texts because of the frequent inversions in spoken language. e.g. 昨天看了新买的一部电影我在电视上。 (Yesterday watched a newly purchased movie I on TV.) In Chinese, the correct sentence is "Yesterday I watched a newly purchased movie on TV." Spoken language often contains informal expressions, such as some popular Internet phrases, which are called "Colloquial Words." e.g. 这明明是你的功劳，却被同事抢走了， 你真是一个大怨种。 (This is obviously your credit, but your coworkers took it away. You're such an unlucky guy.) We need to replace all informal expressions with formal ones. In Chinese, the Internet phrase "大怨 种" is used to describe people who are aggrieved but have no way to complain. Therefore, we should replace it with the more formal expression "You're unlucky." In real spontaneous conversations, a single sentence often contains multiple conversion problems. e.g. 在国内成立野牛，这个，水牛研究中 心，有利于帮助适应人工环境。 (The establish-ment of the bison, I mean, buffalo research center in the country will help to adapt to an artificial environment.) This sentence contains both disfluency and grammatical errors. The corrected version should be: "The establishment of the buffalo research center in the country will help them adapt to an artificial environment." We built an annotation platform to accelerate our annotation progress. When using it, the annotator needs to select the conversion type and annotation range of the current sentence. Then, the annotator provides the written language corresponding to this spoken language. Next, we will present the annotation rules for different conversions. Please note that on the annotation platform, each sentence is wordsegmented into individual words, each of which can be selected to make it easier for the annotator to annotate the conversion range. Disfluency The annotator selects the Disfluency button on the annotation platform. For Filler Words, the annotator should annotate their range. For Rtype, both the reparandum and the interregnum should be annotated, but the repair does not need to be annotated. This is because the sentence can be corrected by simply deleting the reparandum and the interregnum. The annotator selects the ASR Transcription Errors button on the annotation platform and annotates the range of the ASR Transcription Errors. The annotator selects the Grammatical Errors button on the annotation platform. For Missing Words, the annotator needs to annotate the two words before and after the missing part. For Redundant Words, the annotator needs to annotate the redundant part. As for Incorrect Word Order, the entire sentence has to be annotated. The annotator selects the Colloquial Words button on the annotation platform and annotates the range of the colloquial words. Mixed Type First, the annotator selects the button for the initial conversion type and annotates the range of that conversion. Subsequently, the annotator sequentially selects the buttons for the other conversion types and annotates their respective ranges. To ensure annotation consistency and quality, we implement a two-round annotation process. In the first round, we enlist eight native Chinese speakers as part-time annotators after providing preannotation training. In the second round, we conduct manual evaluation and re-annotation, with the authors of this paper serving as senior annotators. In the event of inter-annotator disagreements, the annotator in the second round reannotates the sentence and submits the results of both rounds to the senior annotator. There are two scenarios of inter-annotator disagreements. The first scenario is when the annotator from either the first or second round makes an incorrect annotation. e.g. 当他回到车车间时，已经有了明显的变 化。 (When he returned to the ga-, garage, had changed markedly.) The sentence exhibits two conversion problems, namely Disfluency and Missing Words. The annotator in the first round accurately annotates the disfluency but overlooks the grammatical errors. In the second round, the annotator correctly annotates both conversion problems. The senior annotator will then determine the correct annotation. The second case is that the sentence can be corrected in multiple ways, which is common in grammatical errors. e.g. Source sentence: 如果人们连续看上四五 个小时的电视节目，就会感到十分疲劳。 (If people watch TV programs for four or five hours in a row, will feel very tired.) Target sentence 1: 人们如果连续看上四五个 小时的电视节目，就会感到十分疲劳。 (People in case of watching TV programs for four or five hours in a row will feel very tired.) Target sentence 2: 如果人们连续看上四五 个小时的电视节目，他们就会感到十分疲劳。 (If people watch TV programs for four or five hours in a row, they will feel very tired.) This sentence lacks a subject and has two potential solutions. First, considering the conversion as an "Incorrect Word Order," the corresponding written language is target sentence 1. Second, considering the conversion as a "Missing Word," the written language is target sentence 2. In this scenario, the senior annotator selects the solution with a smaller edit distance. If the edit distances are equal, the senior annotator opts for the first-round solution. Regarding the prompts used for zero-and fewshot settings, we tried two different prompts on BLOOM-7B under the zero-shot and 5-shot settings. Prompt 1: "下面有一个口语到书面语风格转 换任务, 请把口语修改为书面语: 口语：{源句} 书面语:" (Here's a spoken-to-written style conversion task, please rewrite the spoken language into the written language: spoken: {source sentence} written:). Prompt 2: "下面有一个语法纠错任务, 请把 错误的文本修改为正确的文本: 错误文本: {源 句} 正确文本: " (Here's a grammatical error correction task, please correct the wrong text into the right text: wrong text: {source sentence} correct text: ) Prompt 2 outperforms Prompt 1 on all metrics. We speculate that the model may struggle to comprehend the definition of the spoken-to-written language conversion task. However, grammatical error correction is a widely-used task, and the four conversion problems, except for colloquial words, can be considered either simple or complex grammatical errors, allowing it to perform well. Hence, we adopt Prompt 2 for the rest of our experiments. Under the zero-shot setting, the prompt is "下面 有一个语法纠错任务，请把错误的文本修改为 正确的文本：错误文本：{源句} 正确文本：" (Here's a grammatical error correction task, please correct the wrong text into the right text: wrong text: {source sentence} correct text:). We expect LLMs to be capable of correcting all types of conversion problems. Therefore, it is crucial to ensure the diversity of demonstrations provided to LLMs. Given that CS2W encompasses four types of conversion problems, we incrementally add demonstrations of different types as the number of demonstrations increases. The demonstrations for each conversion problem are randomly selected from the validation set. These demonstrations differ from the input sentences in the test set and contain only one conversion problem. Additionally, since CS2W is dominated by disfluency as the primary conversion problem, we specifically select two demonstrations, one for R-type and another for Filler Words, under the 5-shot setting. In Section 5.4 and 5.7, the demonstrations used in 5-shot experiments are shown in Figure While our initial intuition suggested that including a higher diversity of conversion types in demonstrations would enable LLMs to address more types of conversions and subsequently improve results, our findings have led us to reconsider this notion. To test this hypothesis, we randomly selected five demonstrations from the validation set and repeated the selection process three times, resulting in three distinct prompts used for the 5-shot setting. The specific demonstrations included in each prompt are detailed in Table We tested BELLE-7B-2M, ChatGLM-6B, and GPT3.5-turbo under the 5-shot settings with these three prompts. The results, in comparison to those presented in Section 5.4, are summarized in Table In summary, our experimental findings challenge our initial hypothesis that greater diversity in examples would consistently enhance the ability of LLMs to address a wider range of conversion types and lead to improved results. Surprisingly, it is not always the case. The prompt with the most diversity, Prompt 1, displays the weakest performance. In contrast, Prompt 2 and Prompt 3, which feature fewer types of conversions but a higher concentration of demonstrations with disfluencies, delivered more favorable results. We attribute this phenomenon to the prevalence of disfluency within the CS2W dataset. When the number of demonstrations with disfluencies surpasses a certain threshold, the overall performance tends to improve. However, it's worth noting that the dataset's distribution of conversion types may play a pivotal role in these results. In a scenario where the four conversion types were more balanced, prompts with greater diversity might have exhibited improved performance. The numbers of parameters of the models used in the experiments are shown in Table
| 707 | 2,901 | 707 |
NewsRecLib: A PyTorch-Lightning Library for Neural News Recommendation
|
NewsRecLib 1 is an open-source library based on Pytorch-Lightning and Hydra developed for training and evaluating neural news recommendation models. The foremost goals of NewsRecLib are to promote reproducible research and rigorous experimental evaluation by (i) providing a unified and highly configurable framework for exhaustive experimental studies and (ii) enabling a thorough analysis of the performance contribution of different model architecture components and training regimes. NewsRecLib is highly modular, allows specifying experiments in a single configuration file, and includes extensive logging facilities. Moreover, NewsRecLib provides outof-the-box implementations of several prominent neural models, training methods, standard evaluation benchmarks, and evaluation metrics for news recommendation.
|
Personalized news recommendation has become ubiquitous for customizing suggestions to users' interests Despite the abundance of model designs, research on neural news recommenders (NNRs) suffers from two major shortcomings: (i) a surprising amount of non-reproducible research It is thus particularly difficult to evaluate the impact of specific components in NNR architecture and training (e.g., news encoder, user modeling, training objectives) on the overall performance of the model In this work, we introduce NewsRecLib, an open source library for NNRs, to remedy these critical limitations.
|
NewsRecLib is powered by Hydra Logging and profiling. The library supports multiple standard tools (e.g., WandB Overall, NewsRecLib is designed to support the development and benchmarking of NNRs as well as the specific analysis of contributions of common components of the neural recommendation pipelines. In this paper, we discuss the building blocks of NewsRecLib and provide an overview of the readily available models. For a detailed documentation on the usage of the library, we refer to its project page. Figure NewsRecLib is highly modularized: it decouples core components to the largest extent possible. Concretely, we decouple the essential building blocks of a NNR, namely the news encoder (NE), the user encoder (UE), and the click predictor. NE is further decomposed into a configurable set of feature encoders (i.e., components that embed different aspects of the news, e.g., title, topical category or named entities). Different model components can be interchanged with corresponding sub-modules of other recommenders, ensuring freedom in choosing each building block of a model independently of the other components (i.e., by mixing the NE of "NNR 1" with the UE of "NNR 2"), in contrast to practices in existing NNR libraries, in which sub-components are tied to concrete NNR architectures that introduced them. Because of this, NewsRecLib allows for clear-cut and comprehensive analyses of impact of NNR components on their overall performance. Reproducibility strongly relies on the transparency of each step and component in the pipeline, as well as the availability of metadata regarding the factors that influence the model (e.g., hyperparameter values, training objective) and the environment in which it is trained and evaluated (e.g., library versions). Because of this, NewsRecLib leverages the Hydra Each concrete module setting is specified and retrieved automatically from a dedicated configuration file which can be accessed by all the pipeline components. A variety of callbacks supported by PyTorch Lightning (e.g., model checkpointing, early stopping, debugging) can be defined, and modified via a corresponding configuration. A single configuration file guides each experiment: the default configurations of the used modules and callbacks are hierarchically inherited and can be overridden. Experiment configurations can also be overwritten directly from the command line, removing the need to store many similar configuration files: this facilitates fast experimentation and minimizes boilerplate code. Experiments can be executed on CPU, GPU, and in a distributed fashion by specifying the type of accelerator supported in PyTorch Lightning. The integration with extensive logging capabilities (see §2.3) ensures that any modifications are persistently stored in the experiment directory, together with other log files and model checkpoints. Fig. With Hydra's pluggable architecture as its backbone, every part of the recommendation pipeline is transparent to the user. NewsRecLib records comprehensive information during training, including number of trainable model parameters and total model size, runtimes, training and validation losses. Moreover, it stores important metadata regarding hyperparameter settings, operating system, PyTorch version, environment details, and dependencies between libraries. Any profiler supported by PyTorch can be incorporated by a simple modification of the corresponding configuration file. NewsRecLib supports widely used loggers like WandB NNR performance heavily depends on model hyperparameters, making hyperparameter optimization a crucial ingredient in the empirical evaluations of NNRs. NewsRecLib supports hyperparameter tuning using the Optuna framework NewsRecLib currently encompasses two popular benchmark datasets, 13 news recommendation models, and various evaluation metrics. Datasets. We provide out-of-the-box utilities for two prominent monolingual news recommendation benchmarks: MIND Recommendation Models. NewsRecLib provides implementations for 10 general-purpose NNRs and 3 fairness-aware recommenders. To support analysis of model components, for the models that did not use PLMs in their NEs (but rather contextualized embeddings with convolutional or attention layers), we implement an additional variant with a PLM-based NE (as proposed in Training Objectives. Most NNR models are trained with point-wise classification objectives Evaluation Metrics. NewsRecLib integrates standard accuracy-based metrics, such as AUC, MRR, and nDCG@k. Additionally, we implement aspectbased diversity and aspect-based personalization defined in In the past decade, numerous frameworks for the development and comprehensive evaluation of recommender systems have been proposed to address the problem of reproducibility in the field Microsoft Recommenders In contrast to these frameworks, NewsRecLib focuses solely on the state-of-the-art neural news recommendation models, providing utilities for the most used benchmark datasets, architectures, training techniques, and evaluation metrics tailored to news recommendation. NewsRecLib unifies disparate implementations of recent neural news recommenders in a single open-source library that is built on top of mature frameworks for deep learning (PyTorch Lightning), evaluation (TorchMetrics), and configuration (Hydra). We conduct experiments with the pre-implemented recommendation models from NewsRecLib to investigate their performance when (1) trained with the original architecture (e.g., NE based on word embeddings and contextualization layer) and ( We carry out the evaluation on the MINDsmall Model AUC MRR nDCG@5 nDCG@10 Dctg@10 Dsnt@10 AUC MRR nDCG@5 nDCG@10 Dctg@10 Dsnt@10 DKN 50. 100-dimensional TransE embeddings Table In this work, we introduced NewsRecLib, a highly configurable, modular and easily extensible framework for neural news recommendation. Our library is specifically designed to foster reproducible research in recommender systems and rigorous evaluation of models -users only need to create a single configuration file for an experiment. We briefly described the underlying principles of NewsRecLib and the structure of its building blocks. The framework currently provides two standard benchmark datasets, loading and pre-processing functions, 13 neural recommendation models, different training objectives and hyperparameters optimization strategies, numerous evaluation metrics, extensive logging capabilities, and GPU support. We believe that NewsRecLib is a useful tool for the community that will (i) catalyze reproducible NNR research, (ii) foster fairer comparisons between the models, and (iii) facilitate identification of NNR components that drive their performance. While we have striven to build a comprehensive library for the design and fair evaluation of neural news recommendation models, several additional factors must be taken into account. Firstly, even though we aim to replicate the original implementations of the models to the highest degree possible, discrepancies in our code and results can arise from the usage of different frameworks, as well as scarce availability of implementation details in the source code or publications of some of the recommenders. Secondly, our library is heavily dependent on the changes and maintenance of the frameworks on which it is built, namely PyTorch Lightning (and by extension, PyTorch), Hydra, TorchMetrics, Optuna. As such, new plugins for logging (e.g., Neptune (Neptune team, 2019), Comet Lastly, neural news recommendation is a computationally expensive endeavor which requires availability of large compute resources. Although NewsRecLib technically supports execution of experiments on CPU, this would be not only highly inefficient and time-consuming, but also infeasible for large-scale datasets with hundreds of thousands of users and news. Consequently, users should ideally have access to GPUs to efficiently use our library. Users of our library should differentiate the recommendation models available in NewsRecLib from the originals. Consequently, they should explicitly credit and cite both NewsRecLib, as well as the original implementations, as specified on our GitHub page. A.1 Configuration Logging Figs. 4 and 5 illustrate an example of how the configuration of each of the pipeline's components is logged when the training process is initiated. Fig. NewsRecLib provides, to date, implementations of 10 general NNRs: • DKN • NPA • NAML • NRMS • LSTUR • TANR • CAUM • MINS • CenNewsRec • MINER
| 816 | 596 | 816 |
Searching for the X-Factor: Exploring Corpus Subjectivity for Word Embeddings
|
We explore the notion of subjectivity, and hypothesize that word embeddings learnt from input corpora of varying levels of subjectivity behave differently on natural language processing tasks such as classifying a sentence by sentiment, subjectivity, or topic. Through systematic comparative analyses, we establish this to be the case indeed. Moreover, based on the discovery of the outsized role that sentiment words play on subjectivity-sensitive tasks such as sentiment classification, we develop a novel word embedding SentiVec which is infused with sentiment information from a lexical resource, and is shown to outperform baselines on such tasks.
|
Distributional analysis methods such as Word2Vec While there have been active and ongoing research on improving word embedding methods (see Section 5), there is a relative dearth of study on the impact that an input corpus may have on the quality of the word embeddings. The previous preoccupation centers around corpus size, i.e., a larger corpus is perceived to be richer in statistical information. For instance, popular corpora include Wikipedia, Common Crawl, and Google News. We postulate that there may be variations across corpora owing to factors that affect language use. Intuitively, the many things we write (a work email, a product review, an academic publication, etc.) may each involve certain stylistic, syntactic, and lexical choices, resulting in meaningfully different distributions of word cooccurrences. Consequently, such factors may be encoded in the word embeddings, and input corpora may be differentially informative towards various NLP tasks. In this work, we are interested in the notion of subjectivity. Some NLP tasks, such as sentiment classification, revolve around subjective expressions of likes or dislikes. Others, such as topic classification, revolve around more objective elements of whether a document belongs to a topic (e.g., science, politics). Our central hypothesis is that word embeddings learnt from input corpora of contrasting levels of subjectivity perform differently when classifying sentences by sentiment, subjectivity, or topic. As the first contribution, we outline an experimental scheme to explore this hypothesis in Section 2, and conduct a series of controlled experiments in Section 3 establishing that there exists a meaningful difference between word embeddings derived from objective vs. subjective corpora. We further systematically investigate factors that could potentially explain the differences. Upon discovering from the investigation that sentiment words play a particularly important role in subjectivity-sensitive NLP tasks, such as sentiment classification, as the second contribution, in Section 4 we develop SentiVec, a novel word embedding method infused with information from lexical resources such as a sentiment lexicon. We further identify two alternative lexical objectives: Logistic SentiVec based on discriminative logistic regression, and Spherical SentiVec based on soft clustering effect of von Mises-Fisher distributions. In Section 6, the proposed word embeddings show evident improvements on sentiment classification, as compared to the base model Word2Vec and other baselines using the same lexical resource.
|
We lay out the methodology for generating word embeddings of contrasting subjectivity, whose effects are tested on several text classification tasks. As it is difficult to precisely quantify the degree of subjectivity of a corpus, we resort to generating word embeddings from two corpora that contrast sharply in subjectivity, referring to them as the Objective Corpus and the Subjective Corpus. Objective Corpus As virtually all contents are written by humans, an absolutely objective corpus (in the philosophical sense) may prove elusive. There are however exemplars where, by construction, a corpus aspires to be as objective as possible, and probably achieves that in practical terms. We postulate that one such corpus is Wikipedia. Its list of policies and guidelines Subjective Corpus By extension, one may then deem a corpus subjective if its content does not at least meet Wikipedia's neutral point of view requirement. In other words, if the content is replete with personal feelings and opinions. We posit that product reviews would be one such corpus. For instance, Amazon's Community Guideline Word Embeddings For the comparative analysis in Section 3, we employ Word2Vec (reviewed below) to generate word embeddings from each corpus. Later on in Section 4, we will propose a new word embedding method called SentiVec. For Word2Vec, we use the Skip-gram model to train distributional word embeddings on the Objective Corpus and the Subjective Corpus respectively. Skip-gram aims to find word embeddings that are useful for predicting nearby words. The objective is to maximize the context probability: where W is an input corpus and C(w) is the context of token w. The probability of context word w , given observed word w is defined via softmax: where v w and v w are corresponding embeddings and V is the corpus vocabulary. Though theoretically sound, the formulation is computationally impractical and requires tractable approximation. where σ( • ) is a sigmoid function, and { ŵi } k i=1 are negative samples. Summing up all the contextword pairs, we derive the NS Skip-gram objective: Training word embeddings with Skip-gram, we keep the same hyperparameters across all the runs: 300 dimensions for embeddings, k = 5 negative samples, and window of 5 tokens. The Objective and Subjective corpora undergo the same preprocessing, i.e., discarding short sentences (< 5 tokens) and rare words (< 10 occurrences), removing punctuation, normalizing Unicode symbols. To compare word embeddings, we need a common yardstick. It is difficult to define an inherent quality to word embeddings. Instead, we put them through several evaluation tasks that can leverage word embeddings and standardize their formulations as binary classification tasks. To boil the comparisons down to the essences of word embeddings (which is our central focus), we rely on standardized techniques so as to attribute as much of the differences as possible to the word embeddings. We use logistic regression for classification, and represent a text snippet (e.g., a sentence) in the feature space as the average of the word embeddings of tokens in the snippet (ignoring out-ofvocabulary tokens). The evaluation metric is the average accuracy from 10-fold cross validation. There are three evaluation tasks of varying degrees of hypothetical subjectivity, as outlined below. Each may involve multiple datasets. Sentiment Classification Task This task classifies a sentence into either positive or negative. We use two groups of datasets as follows. The first group consists of 24 datasets from UCSD Amazon product data The second is Cornell's sentence polarity dataset v1.0 Subjectivity Classification Task This task classifies a sentence into subjective or objective. The dataset is Cornell's subjectivity dataset v1.0 Topic Classification Task We use the 20 Newsgroups dataset We conduct a series of comparative analyses under various setups. For each, we compare the performance in the evaluation tasks when using the Objective Corpus and the Subjective Corpus. Table Controlling for Corpus Size In Setup II, we keep the number of sentences in both corpora the same, by randomly downsampling sentences in the Subjective Corpus. This procedure consequently reduces the number of types and tokens (see Table 1, Setup II, Corpus Statistics). Note that the number of tokens in the Subjective corpus is now fewer than in the Objective, the latter suffers no change. Yet, even after a dramatic reduction in size, the Subjective embeddings still outperform the Objective significantly on both datasets of the sentiment classification task (+4% on Amazon and +2.5% on RT), while showing similar performance on subjectivity and topic classifications. This bolsters the earlier observation that sentiment classification is more sensitive to subjectivity. While there is a small effect due to corpus size difference, the gap in performance between Subjective and Objective embeddings on sentiment classification is still significant and cannot be explained away by the corpus size alone. Controlling for Vocabulary While the Subjective Corpus has a much smaller vocabulary (i.e., # types), we turn a critical eye on whether its apparent advantage lies in having access to special word types that do not exist in the Objective Corpus. In Setup III, we keep the training vocabulary the same for both, removing the types that are Table Significant Words To get more insights on the difference between the Subjective and Objective corpora, we analyze the mistakes word embeddings make on the development folds. At this point we focus on the sentiment classification task and specifically on the Amazon data, which indicates the largest performance differences in the controlled experiments (see Table As words are still the main unit of information in distributional word embeddings, we extract words strongly associated with misclassified sentences. We employed log-odds ratio with informative Dirichlet prior method Table We also observe linguistic negations (i.e., not, Don't). For instance, the word most associated with the Objective-specific mistakes (excluding the Subjective misclassified sentences) is not, which suggests that perhaps Subjective word embedding accommodates better understanding of linguistic negations, which may partially explain the difference. However, our methodology as outlined in Section 2.2 permits exchangeable word order and is not intended to analyze structural interaction between words. We focus on further analysis of sentiment words, leaving linguistic negations in word embeddings for future investigation. Controlling for Sentiment Words To control for the "amount" of sentiment in the Subjective and Objective corpora, we use sentiment lexicon compiled by In summary, evidences from the series of controlled experiments support the existence of some X-factor to the Subjective embeddings, which confers superior performance in subjectivity-sensitive tasks such as sentiment classification. where L word2vec (W ; C) is the Skip-gram objective as in (4); L lex (W, L) is a lexical objective for corpus W and lexical resource L; and λ is a tradeoff parameter. Lexical resource L = {X i } n i=1 comprises of n word sets, each X i contains words of the same category. For sentiment classification, we consider positive and negative word categories. Logistic SentiVec admits lexical resource in the form of two disjoint word sets, L = {X 1 , X 2 }, X 1 ∩ X 2 = ∅. The objective is to tell apart which word set of L word w belongs to: We further tie these probabilities together, and cast the objective as a logistic regression problem: where v w is a word embedding and τ is a direction vector. Since word embeddings are generally invariant to scaling and rotation when used as downstream feature representations, τ can be chosen randomly and fixed during training. We experiment with randomly sampled unit length directions. For simplicity, we also scale embedding v w to its unit length when computing v w • τ , which now equals to cosine similarity between v w and τ . When v w is completely aligned with τ , the cosine similarity between them is 1, which maximizes P(w ∈ X 1 ) and favors words in X 1 . When v w is opposite to τ , the cosine similarity equals to -1, which maximizes P(w ∈ X 2 ) and predicts vectors from X 2 . Orthogonal vectors have cosine similarity of 0, which makes both w ∈ X 1 and w ∈ X 2 equally probable. Optimizing (6) makes the corresponding word embeddings of X 1 and X 2 gravitate to the opposite semispaces and simulates clustering effect for the words of the same category, while the Word2Vec objective prevents words from collapsing to the same directions. Optimization The objective in (6) permits simple stochastic gradient ascent optimization and can be combined with negative sampling procedure for Skip-gram in (5). The gradient for unnormalized embedding v w is solved as follows: The optimization equation for v w , when w ∈ X 2 , can be derived analogously. Spherical SentiVec extends Logistic SentiVec by dealing with any number of lexical categories, L = {X i } n i=1 . As such, the lexical objective takes on generic form: Each P (w ∈ X i ) defines embedding generating process. We assume each length-normalized v w for w of L is generated w.r.t. a mixture model of von Mises-Fisher (vMF) distributions. vMF is a probability distribution on a multidimensional sphere, characterized by parameters µ (mean direction) and κ (concentration parameter). Sampled points are concentrated around µ; the greater the κ, the closer the sampled points are to µ. We consider only unimodal vMF distributions, restricting concentration parameters to be strictly positive. Hereby, each X i ∈ L is assigned to vMF distribution parameters (µ i , κ i ) and the membership probabilities are defined as follows: where Z κ is the normalization factor. The Spherical SentiVec lexical objective forces words of every X i ∈ L to gravitate towards and concentrate around their direction mean µ i . As in Logistic SentiVec, it simulates clustering effect for the words of the same set. In comparison to the direction vector of Logistic SentiVec, mean directions of Spherical SentiVec when fixed can substantially influence word embeddings training and must be carefully selected. We optimize the mean directions along with the word embeddings using alternating procedure resembling K-means clustering algorithm. For simplicity, we keep concentration parameters tied, κ 1 = κ 2 = ... = κ n = κ, and treat κ as a hyperparameter of this algorithm. Optimization We derive optimization procedure for updating word embeddings assuming fixed direction means. Like Logistic SentiVec, Spherical SentiVec can be combined with the negative sampling procedure of Skip-gram. The gradient for unnormalized word embedding v w is solved by the following equation: Once word embedding v w (w ∈ X i ) is updated, we revise direction mean µ i w.r.t. maximum likelihood estimator: Updating the direction means in such a way ensures that the lexical objective is non-decreasing. Assuming the stochastic optimization procedure for L word2vec complies with the same nondecreasing property, the proposed alternating procedure converges. There have been considerable research on improving the quality of distributional word embeddings. Villegas et al. ( The objective of experiments is to study the efficacy of Logistic SentiVec and Spherical SentiVec word embeddings on the aforementioned text classification tasks. One natural baseline is Word2Vec, as SentiVec subsumes its context prediction objective, while further incorporating lexical category prediction. We include two other baselines that can leverage the same lexical resource but in manners different from SentiVec, namely: Retrofitting SentiVec also outperforms the two baselines that benefit from the same lexical resources. Retrofitting does not improve upon Word2Vec, with the two embeddings essentially indistinguishable (the difference is only noticeable at the second decimal point). Refining makes the word embeddings perform worse on the sentiment classification task. One possible explanation is that Refining normally requires fine-grained labeled lexicon, where the words are scored w.r.t. the sentiment scale, whereas we use sentiment lexicon of two labels (i.e., positive or negative). SentiVec accepts coarse-grained sentiment lexicons, and potentially could be extended to deal with fine-grained labels. As previously alluded to, topic and subjectivity classifications are less sensitive to the subjectivity within word embeddings than sentiment classification. One therefore would not expect much, if any, performance gain from infusion of sentiment information. However, such infusion should not subtract or harm the quality of word embeddings either. Table The diagrams are presented for Objective Embeddings Figure Figure In summary, SentiVec effectively provides an advantage for subjectivity-sensitive task such as sentiment classification, while not harming the performance of other text classification tasks. We explore the differences between objective and subjective corpora for generating word embeddings, and find that there is indeed a difference in the embeddings' classification task performances. Identifying the presence of sentiment words as one key factor for the difference, we propose a novel method SentiVec to train word embeddings that are infused with the sentiment polarity of words derived from a separate sentiment lexicon. We further identify two lexical objectives: Logistic SentiVec and Spherical SentiVec. The proposed word embeddings show improvements in sentiment classification, while maintaining their performance on subjectivity and topic classifications.
| 652 | 2,600 | 652 |
STEP: Staged Parameter-Efficient Pre-training for Large Language Models
|
Pre-training large language models faces significant memory challenges due to the large size of model weights. We propose STaged parameter-Efficient Pre-training (STEP), which combines ideas from parameter-efficient tuning and staged training. We conduct experiments on pre-training models of various sizes and demonstrate that STEP can achieve up to a 40.4% reduction in maximum memory requirement compared to vanilla pre-training while maintaining comparable performance.
|
Large Language Models (LLMs) have become a fundamental technology in artificial intelligence. One challenge we aim to address in the research on LLMs is the vast amount of computational resources needed for pre-training, e.g., LLaMA To tackle this challenge, methods for reducing computational costs during pre-training have been actively studied. For example, ReLoRA For this goal, we propose a method that combines ideas of Parameter-Efficient Tuning (PET) We explore the effectiveness of STEP in pretraining experiments by comparing the baseline (Vannila pre-training) and conventional method (ReLoRA) under the same computational cost. We demonstrate that STEP achieves up to a 40.4% reduction in maximum memory requirements compared to vanilla pre-training while maintaining comparable validation es.
|
Memory Efficient Training for LLMs Several memory-efficient training approaches have been actively developed in the literature of training LLMs Staged Training As briefly described in Section 1, our goal is to develop a method for pre-training LLMs that can achieve comparable performance at the same computational cost while reducing the maximum memory requirements during pre-training. The following four procedures are an overview of STEP and how it efficiently trains LLMs; (Procedure 1) STEP performs a vanilla pretraining on a model with a much smaller size than the target model size as an initial model. (Procedure 2) STEP expands the layers of the initial model to increase its size. (Procedure 3) STEP also introduces the PET parameters given by the parameter-efficient adaptors for the layers trained in Procedure 1. (Procedure 4) STEP continues to pre-train the parameters in layers newly added in Procedure 2 and the adaptors added in Procedure 3 while freezing those in layers trained in Procedure 1. Note that The first to fourth red right-arrow in Figure This section explains how we expand the layers in Procedure 2. Given a model with n layers, the Growth Layer Operator modifies the structure of the model's layers. We use Interpolation We further extend it by incorporating an idea of a fusing method that averages the parameters of the two layers This section provides additional information about Procedure 3, which introduces PET parameters by the adaptors. We specifically focus on the low-rank adaptation method We assume that the maximum memory requirement during the pre-training can be estimated by the size of model states, which include the parameters of the model itself, the gradients of the model parameters being trained, and the optimizer state. 2 Moreover, we assume that we use a typical Transformer model and the Adam optimizer (Kingma and Ba, 2014), which are a commonly used configuration for pre-training LLMs. Additionally, we assume that all parameters are represented as 32-bit floating-point numbers. Consequently, when the number of parameters in one layer of the Transformer is P layer and the number of layers in the model is n, the memory usage of the model state, expressed in bytes, is given by where the optimizer state of Adam consists of two parts; the gradient momentum and variance. Regarding the maximum memory requirement for STEP, let n i be the number of layers increased in the i-th stage from the i -1 stage in STEP, where n 0 = 0. Let N i represent the total number of layers in the i-th stage model, namely, N i = i k=1 n k . Moreover, E(P layer ) denotes the number of parameters for the single layer, P layer , added by PET. Then, we can estimate the maximum memory requirement for the stage i, that is, M i , as follows: where the term 4N i-1 P layer represents the number of frozen model parameters already trained in the 1 to i -1 stages, the term 16n i P layer indicates the number of newly added model parameters with optimization states added in Procedure 2 and the term 16N i-1 E(P layer ) represents the number of PET parameters with optimization states added in Procedure 3. Let L be the number of layers for the model that is finally obtained. Then, the solution of the following minimization problem can minimize the maximum memory requirement during the pre-training: This section demonstrates the effectiveness of the proposed method, STEP, through the pre-training experiments of LLMs. We investigate whether STEP can achieve a comparable validation perplexity to vanilla pre-training at the same computational cost. We also compared with We used C4 The model configuration follows an architecture based on LLaMA We evaluated the effectiveness of STEP when the Growth Layer operator is applied once during its pre-training. This means that we set K = 2 in Equation 3 for STEP. Given the number of layers L with the fixed dimension of hidden layers, we compute n 1 and n 2 that can minimize the maximum memory requirements by Equation Layers are added to the upper part of the Transformer layers. The discussion about the position where layers are added is provided in Appendix E. Table Type of Growth Layer Operators: We conducted an ablation study on Growth Layer Operators (Procedure 2) in STEP. We compared three Growth Layer operators: Stacking, Queueing, Interpolation-Copy, and Interpolation-Mean. Stacking is proposed in The results of this ablation study are shown in Table Pre-training LLM requires substantial memory, posing a challenge for research. We proposed a new training method called STEP, which enables LLM pre-training with limited memory requirements. Our experiments demonstrated that STEP achieved comparative performance to vanilla LLM pre-training while minimizing peak memory usage. Several limitations of our study should be addressed in future research. First, while we conducted experiments with up to two stages in STEP, the effectiveness of using more than three stages remains unexplored. Second, although our methods reduced memory usage, we did not observe significant enhancements in training speed. Third, although validation is the standard metric for evaluating the performance of pre-training, it is still unknown whether the models pre-trained by the proposed method can improve the downstream tasks. To investigate the downstream task, we need to fine-tune all the pre-trained models. Finally, our experiments focused on relatively smaller model sizes compared to the recent LLMs with billions of parameters, such as those with 7B or more. A About Zero Initialization In STEP, we use ReLoRA (LoRA) for PET and LLaMA as the model. When not considering Grouped Query Attention The training configurations used in the experiment are shown in Table To verify whether the mean initialization in Interpolation actually possesses the desired properties, we compare the cosine similarity between the output of the added layer ϕ new_i after Interpolation and the input to the layer ϕ i+1 before Interpolation. In this case, if the output of ϕ new_i is similar to the input of ϕ i+1 , it can be considered that ϕ new_i appropriately processes the output from ϕ i and outputs something that is easy for ϕ i+1 to process. We apply Interpolation to a 334M model with nine layers that have been trained on the C4 dataset This ablation study investigates the most effective position to add new layers when applying the Growth Layer operator using Interpolation-Mean in Procedure 2. In this ablation study, we perform layer additions on 409M →668M configurations in Table
| 473 | 805 | 473 |
Low-Quality Product Review Detection in Opinion Summarization
|
Product reviews posted at online shopping sites vary greatly in quality. This paper addresses the problem of detecting lowquality product reviews. Three types of biases in the existing evaluation standard of product reviews are discovered. To assess the quality of product reviews, a set of specifications for judging the quality of reviews is first defined. A classificationbased approach is proposed to detect the low-quality reviews. We apply the proposed approach to enhance opinion summarization in a two-stage framework. Experimental results show that the proposed approach effectively (1) discriminates lowquality reviews from high-quality ones and (2) enhances the task of opinion summarization by detecting and filtering lowquality reviews.
|
In the past few years, there has been an increasing interest in mining opinions from product reviews In this paper, we demonstrate the aforementioned biases and define a standard specification to measure the quality of product reviews. We then manually annotate a set of ground-truth with real world product review data conforming to the specification. To automatically detect low-quality product reviews, we propose a classification-based approach learned from the annotated ground-truth. The proposed approach explores three aspects of product reviews, namely informativeness, readability, and subjectiveness. We apply the proposed approach to opinion summarization, a typical opinion mining task. The proposed approach enhances the existing work in a two-stage framework, where the low-quality review detection is applied right before the summarization stage. Experimental results show that the proposed approach can discriminate low-quality reviews from high-quality ones effectively. In addition, the task of opinion summarization can be enhanced by detecting and filtering low-quality reviews. The rest of the paper is organized as follows: Section 2 introduces the related work. In Section 3, we define the quality of product reviews. In Section 4, we present our approach to detecting lowquality reviews. In Section 5, we empirically verify the effectiveness of the proposed approach and its use for opinion summarization. Section 6 summarizes our work in this paper and points out the future work.
|
The problem of evaluating helpfulness of reviews In practice, researchers in this area considered the problem as a ranking problem and solved it with regression models. In the process of model training and testing, they used the ground-truth derived from users' votes of helpfulness provided by Amazon. As we will show later in Section 3, these models all suffered from three types of voting bias. In our work, we avoid using users' votes by developing a specification on the quality of reviews and building a ground-truth according to the specification. One area of research on opinion mining from product reviews is to judge whether a review expresses a positive or a negative opinion. For example, Turney (2006) presented a simple unsupervised learning algorithm in judging reviews as "thumbs up" (recommended) or "thumbs down" (not recommended). Another area of research on opinion mining is to extract and summarize users' opinions from product reviews In this section, we will first show three biases of users' votes observed on Amazon, and then present our specification on the quality of product reviews. In our study, we use the product reviews on digital cameras crawled from Amazon as our data set. The data set consists of 23,141 reviews on 946 digital cameras. At the Amazon site, users could vote for a review with a "helpful" or "unhelpful" label. Thus, for each review there are two numbers indicating the statistics of these two labels, namely the number of "helpful" votes and that of "unhelpful" ones. Certainly, the ground-truth has the advantage of convenience. However, we identify three types of biases that make the Amazon ground-truth not always suitable for determining the quality of reviews. We describe these biases in details in the rest of this section. At the Amazon site, users tend to value others' opinions positively rather than negatively. From Figure It's a nice compact and light camera and takes great photos and videos. Only complaint (other than E18) is the limit of 30-second videos on 640x480 mode. I've got a 512MB compact flash card, I should be able to take as much footage as I have memory in one take. There also exists a bootstrapping effect of "hot" reviews at the Amazon site. Figure Publication date can influence the accumulation of users' votes. In Figure Besides these aforementioned biases, using the raw rating from readers directly also fails to provide a clear guideline for what a good review consists of. In this section, we provide such a guideline, which we name as the specification (SPEC). In the SPEC, we define four categories of review quality which represent different values of the reviews to users' purchase decision: "best review", "good review", "fair review", and "bad review". A generic description of the SPEC is as follows: A best review must be a rather complete and detailed comment on a product. It presents several aspects of a product and provides convincing opinions with enough evidence. Usually a best review could be taken as the main reference that users only need to read before making their purchase decision on a certain product. The first review in Figure A fair review contains a very brief description on a product. It does not supply detailed evaluation on the product, but only comments on some aspects of the product. For example, the third review in Figure A bad review is usually an incorrect description of a product with misleading information. It talks little about a specific product but much about some general topics (e.g. photography). For example, the last review in Figure I purchased this camera about six months ago after my Kodak Easyshare camera completely died on me. I did a little research and read only good things about this Canon camera so I decided to go with it because it was very reasonably priced The Sony DSC "P10" Digital Camera is the top pick for CSC. Running against cameras like Olympus stylus, Canon Powereshot, Sony V1, Nikon, Fuji, and More. The new release of 5.0 mega pixels has shot prices for digital cameras up to $1000+. This camera I purchased through a Private Dealer cost me $400.86. The Retail Price is Running $499.00 to $599.00. Purchase this camera from a wholesale dealer for the best price $377.00. Great Photo Even in dim light w/o a flash. The p10 is very compact. Can easily fit into any pocket. The camera can record 90 minutes of mpeg like a home movie. There are a lot of great digital cameras on the market that shoot good pictures and video. What makes the p10 the top pick is it comes with a rechargeable lithium battery. Many use AA batteries, the digital camera consumes theses AA batteries in about two hours time while the unit is on. That can add continuous expense to the camera. It's also the best resolution on the market. 6.0 megapix is out, though only a few. And the smallest that we found. Also the best price for a major brand. There is nothing wrong with the 2100 except for the very noticeable delay between pics. The camera's digital processor takes about 5 seconds after a photo is snapped to ready itself for the next one. Otherwise, the optics, the 3X optical zoom and the 2 megapixel resolution are fine for anything from Internet apps to 8" x 10" print enlarging. It is competent, not spectacular, but it gets the job done at an agreeable price point. I want to point out that you should never buy a generic battery, like the person from San Diego who reviewed the S410 on According to the SPEC defined above, we built a ground-truth from the Amazon data set. We randomly selected 100 digital cameras and 50 reviews for each camera. Totally we have 4,909 reviews since some digital cameras have fewer than 50 unique reviews. Then we hired two annotators to label the reviews with the SPEC as their guideline. As the result, we have two independent copies of annotations on 4,909 reviews, with the labels of "best", "good", "fair", and "bad". Table (1) where the "preference pair" is defined as a pair of reviews with a order. For example, a best review and a good review correspond to a preference pair with the order of "best review preferring to good review". The "all preference pairs" are collected from one of the annotations (the annotation 1 or the annotation 2) by ignoring the pairs from the same category. The "incorrect preference pairs" are the preference pairs collected from the Amazon ground-truth but not with the same order as that in the all preference pairs. The order of the preference pair collected from the Amazon ground-truth is evaluated on the basis of the percentage score as described in Section 3.1. The error rate of preference pairs based on the annotation 1 and that based on the annotation 2 are 0.448 and 0.446, respectively, averaged over 100 digital cameras. The high error rate of preference pairs demonstrates that the Amazon ground-truth diverges from the annotations (our ground-truth) significantly. To discover which kind of ground-truth is more reasonable, we ask an additional annotator (the third annotator) to compare these two kinds of ground-truth. More specifically, we randomly selected 100 preference pairs whose orders the two kinds of ground-truth don't agree on (called incorrect preference pairs in the evaluation above). As for our ground-truth, we choose the Annotation 1 in the new test. Then, the third annotator is asked to assign a preference order for each selected Note that the third annotator is blind to both our specification and the existing preference order. Last, we evaluate the two kinds of ground-truth with the new annotation. Among 100 pairs, our ground-truth agrees to the new annotation on 85 pairs while the Amazon ground-truth agrees to the new annotation on 15 pairs. To confirm the result, yet another annotator (the fourth annotator) is called to repeat the same annotation independently as the third one. And we obtain the same statistical result (85 vs. 15) although the fourth annotator does not agree with the third annotator on some pairs. In practice, we treat the reviews in the first three categories ("best", "good" and "fair") as highquality reviews and those in the "bad" category as low-quality reviews, since our goal is to identify low quality reviews that should not be considered when creating product review summaries. We employ a statistical machine learning approach to address the problem of detecting low-quality products reviews. Given a training data set 𝐷 = 𝑥 𝑖 , 𝑦 𝑖 1 𝑛 , we construct a model that can minimize error in prediction of y given x (generalization error). Here 𝑥 𝑖 ∈ 𝑋 and 𝑦 𝑖 = {𝑖𝑔 𝑞𝑢𝑎𝑙𝑖𝑡𝑦 , 𝑙𝑜𝑤 𝑞𝑢𝑎𝑙𝑖𝑡𝑦} represents a product review and a label, respectively. When applied to a new instance x, the model predicts the corresponding y and outputs the score of the prediction. In our study, we focus on differentiating lowquality product reviews from high-quality ones. Thus, we treat the task as a binary classification problem. We employ SVM (Support Vector Machines) The construction of SVM needs labeled training data (in our case, the categories are "high-quality reviews" and "low-quality reviews"). Briefly, the learning algorithm creates the "hyper plane" in (2), such that the hyper plane separates the positive and negative instances in the training data with the largest "margin". Product features (e.g., "image quality" for digital camera) in a review are good indicators of review quality. However, different product features may refer to the same meaning (e.g., "battery life" and "power"), which will bring redundancy in the study. In this paper, we formulize the problem as the "resolution of product features". Thus, the problem is reduced to how to determine the equivalence of a product feature in different forms. In We use edit distance When using contextual similarity, we split all the reviews into sentences. For each mention of a product feature, we take it as a query and search for all the relevant sentences. Then we construct a vector for the mention, by taking each unique term in the relevant sentences as a dimension of the vector. The cosine similarity between two vectors of mentions is then present to measure the contextual similarity between two mentions. To detect low-quality reviews, our proposed approach explores three aspects of product reviews, namely informativeness, subjectiveness, and readability. We denote the features employed for learning as "learning features", discriminative from the "product features" we discussed above. As for informativeness, the resolution of product features is employed when we generate the learning features as listed below. Pairs mapping to the same product feature will be treated as the same product feature, when we calculate the frequency and the number of product features. We apply the approach proposed in We also use a list of product names and a list of brand names to generate the learning features. Both lists can be collected from the Amazon site because they are relatively stable within a time interval. The learning features on the informativeness of a review are as follows. Sentence level (SL) The number of sentences in the review The average length of sentences The number of sentences with product features Word level (WL) The number of words in the review The number of products (e.g., DMC-FZ50, EX-Z1000) in the review The number of products in the title of a review The number of brand names (e.g., We make use of several features at paragraph level which indicate the underlying structure of the reviews. These features include, The number of paragraphs in the review The average length of paragraphs in the review The number of paragraph separators in the review Here, we refer to the keywords, such as "Pros" vs. "Cons" as "paragraph separators". The keywords usually appear at the beginning of paragraphs for categorizing two contrasting aspects of a product. We extract the nouns and noun phrases at the beginning of each paragraph from the 4,909 reviews and use the most frequent 30 pairs of keywords as paragraph separators. Table We also take the subjectiveness of reviews into consideration. Unlike previous work In this section, we describe our experiments with the proposed classification-based approach to lowquality review detection, and its effectiveness on the task of opinion summarization. In our proposed approach, the problem of assessing quality of reviews is formalized as a binary classification problem. We conduct experiments by taking reviews in the categories of "best", "good", and "fair" as high-quality reviews and those in the "bad" category as low-quality reviews. As for classification model, we utilize the SVMLight toolkit We take the approach that utilizes only the category of features on sentence level (SL) as the baseline, and incrementally add other categories of features on informativeness, readability and subjectiveness. We can see that both the features on word level (WL) and those on product feature level (PFL) can improve the performance of classification much. The features on readability can still increase the accuracy although the contribution is much less. The features on subjectiveness, however, make no contribution. We also conduct a more detailed analysis on each individual feature. Two categories of features on "title" and "brand name" have poor performance, which is due to the lack of information in the title and the low coverage of brand names in a review, respectively. One potential application of low-quality review detection is the opinion summarization of reviews. The process of opinion summarization of reviews with regards to a query of a product consists of the following steps 2. For each product feature, generate a positive opinion set and a negative opinion set of opinion segments, denoted as POS(𝑓) and NOS(𝑓). 3. For each product feature, aggregate the numbers of segments in POS(𝑓) andNOS(𝑓), as opinion summarization on the product feature. In this process, all the reviews contribute the same. However, different reviews do hold different authorities. A positive/negative opinion from a high-quality review should not have the same weight as that from a low-quality review. We use a two-stage approach to enhance the reliability of summarization. That is, we add a process of low-quality review detection before the summarization process, so that the summarization result is obtained based on the high-quality reviews only. We are to demonstrate how much difference the proposed two-stage approach can bring into the opinion summarization. We use the best classification model trained as described in Section 5.1 to filter low-quality reviews, and do summarization on the high-quality reviews associated to the 50 test queries. We denote the proposed approach and the old approach as "two-stage" and "one-stage", respectively. Due to the limited space, we only give a visual comparison of the two approaches on "image quality" in Figure To show the comparison on more features in a compressed space, we give the statistic ratio of change between two approaches instead. As for the evaluation measure, we define "RatioOfChange" (ROC) on a feature f as, where Rate * (f) is defined as, Table There is no standard way to evaluate the quality of opinion summarization as it is rather a subjective problem. In order to demonstrate the impact of the two-stage approach, we turn to external authoritative sources other than Amazon.com as the objective evaluation reference. We observe that CNET Figure In this paper, we studied the problem of detecting low-quality product reviews. Our contribution can be summarized in two-fold: (1) we discovered three types of biases in the ground-truth used extensively in the existing work, and proposed a specification on the quality of product reviews. The three biases that we discovered are imbalance vote bias, winner circle bias, and early bird bias. (2) Rooting on the new ground-truth (conforming to the proposed specification), we proposed a classification-based approach to low-quality product review detection, which yields better performance of opinion summarization. We hope to explore our future work in several areas, such as further consolidating the new ground-truth from different points of view and verifying the effectiveness of low-quality review detection with other applications.
| 749 | 1,506 | 749 |
TIMEDIAL: Temporal Commonsense Reasoning in Dialog
|
Everyday conversations require understanding everyday events, which in turn, requires understanding temporal commonsense concepts interwoven with those events. Despite recent progress with massive pre-trained language models (LMs) such as T5 and GPT-3, their capability of temporal reasoning in dialogs remains largely under-explored. In this paper, we present the first study to investigate pre-trained LMs for their temporal reasoning capabilities in dialogs by introducing a new task and a crowd-sourced English challenge set, TIMEDIAL. We formulate TIME-DIAL as a multiple choice cloze task with over 1.1K carefully curated dialogs. Empirical results demonstrate that even the best performing models struggle on this task compared to humans, with 23 absolute points of gap in accuracy. Furthermore, our analysis reveals that the models fail to reason about dialog context correctly; instead, they rely on shallow cues based on existing temporal patterns in context, motivating future research for modeling temporal concepts in text and robust contextual reasoning about them. The dataset is publicly available at:
|
Humans can effortlessly reason about temporal concepts of everyday events such as their duration, frequency, or relative ordering ( For instance, for the first dialog in Table Although previous works have studied temporal reasoning in natural language, they have either focused on specific time-related concepts in isolation, such as temporal ordering and relation extraction In this work, we make the first systematic study of temporal commonsense reasoning in a multi-turn dialog setting. The task involves complex reasoning that requires operations like comparison and arithmetic reasoning over temporal expressions and the need for commonsense and world knowledge. We design a new task for dialog-based temporal reasoning and present a new challenge set in English, called TIMEDIAL, to evaluate language understanding models on the task. We formulate the problem as a crowd-sourced cloze task with multiple choices based on dialogs in the DailyDialog dataset The challenge set requires the models to demonstrate understanding of the context and use temporal commonsense to make right choices. Our final challenge set consists of 1.1K carefully curated dialog instances. We then study the performance of several stateof-the-art pre-trained language models on TIME-DIAL along several dimensions including modeling paradigms (classification, mask filling, and generation), the scope of dialog contexts, in-domain vs. out-of-domain training, dependence on shallow text matching for reasoning, and the types of reasoning required. Our experiments demonstrate that offthe-shelf, pre-trained language models cannot effectively reason about temporal aspects in a dialog, even with domain-specific finetuning. Our findings indicate that large-scale pre-trained models even after fine-tuning may not be sufficient for robust temporal reasoning in dialogs, and motivate future research toward modeling temporal concepts over diverse everyday events, and contextual reasoning about them.
|
We formulate the dialog-based temporal commonsense reasoning problem as a cloze task Having a non-trivial set of options is crucial to build a challenge set and to avoid accidental spurious biases The TIMEDIAL dataset is derived from DailyDialog data Our data collection process involves two steps: (1) identifying dialogs that are rich in temporal expressions, and (2) asking human annotators to provide correct and incorrect options for cloze instances derived from these dialogs. We now describe these steps in detail. Temporal expression identification. Here, we select dialogs that are rich with temporal information, in order to focus on complex temporal reasoning that arises in natural dialogs. Temporal expressions are automatically identified with SU-Time Human annotated options. Next, we make spans in the dialogs. For a dialog, we mask out each temporal expression that contains numerals, each resulting in a cloze question that is then sent for human annotation. This resulted in 1,526 instances for annotation. For each masked span in each dialog, we obtain human annotation to derive a fixed set of correct and incorrect options given the context. Concretely, given a masked dialog and a seed correct answer (i.e., the original text) for the masked span, the annotators To ensure that the annotated incorrect options are not too trivially distinguishable by the models (as discussed in §2), we define three rules for the annotators to follow. • Rule 1: Phrase Matching. The rater should first try to pick another temporal span from the dialog context that makes syntactic/semantic sense (e.g., when the span is of the appropriate type, such as duration, for the masked span) but is still incorrect according to commonsense. • Rule 3: Open-ended. If neither of the above rules is applicable, then raters can come up with an incorrect option using their own judgment. The two incorrect options are required to differ from each other as much as possible. Rules-1&2 are designed to confuse models that rely on shallow pattern matching. Finally, to ensure the quality of the human-annotated options, we perform a subsequent round of human validation on the gathered data. The validators identify and fix issues such as duplicate options, unreasonable or obscure annotations w.r.t natural usage, or ungrammatical annotations that do not fit in the context. Table Answering different instances in the dataset requires different types of core reasoning abilities, such as comparison, arithmetic inference, or reasoning based on world knowledge or general commonsense. To facilitate fine-grained analysis, we also annotate the reasoning categories for a randomly sampled set of 100 dialogs. Though each instance can involve multiple reasoning types, we associate it with one predefined category label that indicates the primary type of reasoning it requires. Table We consider a broad set of methods and evaluate their performance on our challenge TIMEDIAL dataset. These methods vary in terms of the modeling paradigms, the scope of the dialog contexts, and training settings. In particular, they encompass the major ways pre-trained LMs are currently used in downstream tasks ( §4.1) which often outperform earlier specialized non-pretrained models. We also consider different lengths of context used in reasoning, varying by their vicinity to the masked span ( §4.2). Finally, we study different training settings, including zero-shot, in-domain, and out-of-domain training ( §4.3). We experiment across three major modeling paradigms: (i) Binary Classification, (ii) Mask Filling, and (iii) Generation. Figure In this setting, we formulate the task as a binary classification problem, i.e., we use a classifier to measure the probability of the candidate in the (masked dialog context, candidate) pair being a correct answer. Any powerful LM -e.g., BERT In an unlabeled corpus, we use the SUTime tool Output: (3) Generation to annotate temporal spans. We mask each temporal span in this corpus and use the masked text as one positive example for binary classification. To generate negative example, we randomly sample another temporal span from the dialog context and use it as a negative example for the masked temporal span. The resulting data is noisy because the randomly sampled temporal span can also logically fit in the masked span in the given context; however, we assume the likelihood of that happening is low. We leave drawing harder negative instances using heuristics to future work. We also use the mask filling approach of BERTlike mask language models (MLMs). For each dialog context and a candidate temporal span of m tokens, we replace the blank in the dialog context with m masked tokens. We then evaluate the likelihood of predicting the temporal span tokens for those masked positions, and make average across the positions. A key advantage of this method is that we can directly apply a BERT model in the zero-shot manner since the model was pretrained in the same way, as for accommodating for [MASK] fillings. Additionally, we also finetune BERT's MLM for learning task specific properties. The third method is a fully generative approach using the text-to-text paradigm of T5 We aim to study the influence of context on a model's temporal reasoning in dialog by incorporating varying scopes of dialog context based on their vicinity to the target span. Since the dialogs in TIMEDIAL are rich in temporal concepts, we want to evaluate LMs' dependence on shallow text matching vs. the ability to accurately understand the causal relations between those concepts (see Table • Full context, where the model is presented with the complete available dialog to reason on. Due to our design of challenging negatives, the full context can often confuse models that rely on shallow cues. • Local context, where we provide only with the utterances that immediately precede and follow the target utterance. • Target context, where the context is restricted to only the particular utterance that contains the masked span. For all models, we consider two common training settings, e.g., in-domain data, which is typically small, and out-of-domain training where a large amount of data is available. Table In Using the proposed TIMEDIAL challenge set, we next conduct extensive experiments and analyses on the different model variants and context settings. We use either 4x4 or 8x8 Cloud TPUs V3 pod slices Table Overall. The generation model based on T5-LARGE and finetuned on the in-domain DailyDialog data achieves the best performance. However, its 2-best accuracy (74.8) lagged far behind the human performance, demonstrating the difficulty of the TIMEDIAL challenge set. Zero-shot vs. out-of-domain vs. in-domain. When comparing the different training data setup, we observe that models with in-domain training using the DailyDialog data (e.g., LARGE-IN) consistently outperforms those trained on the large out-ofdomain Meena dataset (e.g., LARGE-OUT). Both setups outperform the zero-shot models (without any fine-tuning) (e.g., LARGE-ZERO). The results show that the large LMs still highly depend on indomain or at least dialog data to grasp and enhance their temporal reasoning ability in dialog context. Further, we see increasing performance with increasing model size, which is not unexpected given the complexity of the task. Next, we analyze the different types of errors based on different rules for negative option creation in the annotation process. In particular, the phrase matching rule picks an exact time span from the dialog context, and numeral matching picks numerals from the dialog context. Thus, models picking those incorrect options imply reliance on spurious shallow text matching features. Figure Table Table In 5 out of the 12 cases, using the local context yields equal or higher reasoning accuracy. The results show that the LMs still fall short of properly modeling the rich dialog contexts and making effective use of all information to do reasoning. Figure Temporal commonsense reasoning. Early studies related to temporal analysis define time in the context of sets and relations Some recent work has focused on building challenging benchmarks for temporal commonsense reasoning. Story Cloze Test focuses on stereotypical causal temporal and causal relations between events Commonsense reasoning with LMs. With the recent success of large pre-trained language models (LMs) We introduced TIMEDIAL, a challenge set consistting of 1.1K multiple-choice cloze questions for temporal commonsense reasoning in dialog. The dataset is carefully curated to evaluate a models' ability to do temporal commonsense/numerical reasoning over dialog context. In order to establish strong baselines and provide information on future model development, we conducted extensive experiments with state-of-the-art language models with different settings: the scope of context, weak supervision strategies, and learning objectives. While humans can easily answer these questions (97.8% accuracy), even our best model variant (T5-large with in-domain training) struggles on this challenge set (73%). Moreover, our qualitative error analyses show that these large language models often rely on shallow, spurious features (particularly text matching) when answering these questions, instead of truly doing reasoning over the context.
| 1,117 | 1,979 | 1,117 |
Building Lexicon for Sentiment Analysis from Massive Collection of HTML Documents
|
Recognizing polarity requires a list of polar words and phrases. For the purpose of building such lexicon automatically, a lot of studies have investigated (semi-) unsupervised method of learning polarity of words and phrases. In this paper, we explore to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences. The key idea is to develop the structural clues so that it achieves extremely high precision at the cost of recall. In order to compensate for the low recall, we used massive collection of HTML documents. Thus, we could prepare enough polar sentence corpus.
|
Sentiment analysis is a recent attempt to deal with evaluative aspects of text. In sentiment analysis, one fundamental problem is to recognize whether given text expresses positive or negative evaluation. Such property of text is called polarity. Recognizing polarity requires a list of polar words and phrases such as 'good', 'bad' and 'high performance' etc. For the purpose of building such lexicon automatically, a lot of studies have investigated (semi-) unsupervised approach. So far, two kinds of approaches have been proposed to this problem. One is based on a thesaurus. This method utilizes synonyms or glosses of a thesaurus in order to determine polarity of words As for the second approach, it depends on the definition of co-occurrence whether the hypothesis is appropriate or not. In Turney's work, the co-occurrence is considered as the appearance in the same window In this paper, we explore to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the ex- tracted polar sentences. An overview of the proposed method is represented in Figure The key idea is to develop the structural clues so that it achieves extremely high precision at the cost of recall. As we will see in Section 2.3, the precision was extremely high. It was around 92% even if ambiguous cases were considered as incorrect. In order to compensate for the low recall, we used massive collection of HTML documents. Thus, we could build enough polar sentence corpus. To be specific, we extracted 500,000 polar sentences from one billion HTML documents. The contribution of this paper is to empirically show the effectiveness of an approach that makes use of the strength of massive data. Nowadays, terabyte is not surprisingly large, and larger corpus would be obtained in the future. Therefore, we think this kind of research direction is important.
|
Our method begins by automatically constructing polar sentence corpus with structural clues (step 1). The basic idea is exploiting certain language and layout structures as clues to extract polar sentences. The clues were carefully chosen so that it achieves high precision. The original idea was represented in our previous paper Some polar sentences are described by using characteristic language structures. Figure What characterizes this sentence is the singly underlined phrase. In this phrase, 'riten (advantage)' is followed by postpositional particle '-ha', which is Japanese topic marker. And hence, we can recognize that something positive is the topic of the sentence. This kind of linguistic structure can be recognized by lexico-syntactic pattern. Hereafter, such words like 'riten (advantage)' are called cue words. In order to handle the language structures, we utilized lexico-syntactic patterns as illustrated below. riten-ha advantage-POST (polar) koto-desu to-POST A sub-tree that matches (polar) is extracted as polar sentence. It is obvious whether the polar sentence is positive or negative one. In case of Figure Two kinds of layout structures are utilized as clues. The first clue is the itemization. In Figure The other clue is table structure. In Figure ¯The sound is natural. ¯Music is easy to find. ¯Can enjoy creating my favorite play-lists. ¯The remote controller does not have an LCD display. ¯The body gets scratched and fingerprinted easily. ¯The battery drains quickly when using the backlight. Figure 2 To be exact, the doubly underlined part is polar clause. However, it is called polar sentence because of the consistency with polar sentences extracted by using layout structures. 7.0km/litter Mileage(highway) 9.0km/litter Plus This is a four door car, but it's so cool. The seat is ragged and the light is dark. Figure It is easy to extract polar sentences from the itemization. Such itemizations as illustrated in Figure As for table structures, two kinds of tables are considered (Figure The method was applied to one billion HTML documents. In order to get dependency tree, we used KNP In order to investigate the quality of the corpus, two human judges (judge A/B) assessed 500 polar sentences in the corpus. According to the judge A, the precision was 91.4%. 459 out of 500 polar sentences were regarded as valid ones. According to the judge B, the precision was 92.0% (460/500). The agreement between the two judges was 93.5% (Kappa value was 0.90), and thus we can conclude that the polar sentence corpus has enough quality After error analysis, we found that most of the errors are caused by the lack of context. The following is a typical example. There is much information. This sentence is categorized into positive one in the corpus, and it was regarded as invalid by both judges because the polarity of this sentence is ambiguous without context. As we described in Section 1, the hypothesis of co-occurrence based method is often inappropriate. The next step is to acquire polar phrases from the polar sentence corpus (step 2 and 3 in Figure From the corpus, candidates of polar phrases are extracted together with their counts (step 2). As is often pointed out, adjectives are often used to express evaluative content. Considering that polarity of isolate adjective is sometimes ambiguous (e.g. high), not only adjectives but also adjective phrases (noun + postpositional particle + adjective) are treated as candidates. Adjective phrases are extracted by the dependency parser. To handle negation, an adjective with negation words such as 'not' is annotated by NEGATION tag. For the sake of readability, we simply represent adjective phrases in the form of 'noun-adjective' by omiting postpositional particle, as in the Figure For each candidate, we count the frequency in positive and negative sentences separately. Intuitively, we can expect that positive phrases often appear in positive sentences, and vice versa. However, there are exceptional cases as follows. Although the price is high, its shape is beautiful. Although this sentence as a whole expresses positive evaluation and it is positive sentence, negative phrase 'price is high' appears in it. To handle this, we hypothesized that positive/negative phrases tend to appear in main clause of positive/negative sentences, and we exploited only main clauses to count the frequency. For each candidate, we determine numerical value indicating the strength of polarity, which is referred as polarity value. On the basis of this value, we select polar phrases from the candidates and add them to our lexicon (step 3). For each candidate , we can create a contingency table as follows. ´ ´ ÔÓ×µ is the frequency of in positive sentences. ´ ÔÓ×µ is that of all candidates but . ´ Ò µ and ´ Ò µ are similarly decided. From this contingency table, 's polarity value is determined. Two ideas are examined for compari-son. One is based on chi-square value and the other is based on Pointwise Mutual Information (PMI). The chi-square value is a statistical measure used to test the null hypothesis that, in our case, the probability of a candidate in positive sentences is equal to the probability in negative sentences. Given Table ¾ ´ µ ܾ´ µ ݾ´ÔÓ× Ò µ ´Ü ݵ ´Ü ݵ ¾ ´Ü ݵ Here, ´Ü ݵ is the expected value of ´Ü ݵ under the null hypothesis. Although ¾ ´ µ´ ¼µ indicates the strength of bias toward positive or negative sentences, its direction is not clear. We determined polarity value so that it is greater than zero if appears in positive sentences more frequently than in negative sentences and otherwise it is less than zero. È Î ¾´ µ ¾ ´ µ È´ Ò µ È ´ ÔÓ×µ ¾ ´ µ ÓØ ÖÛ × È´ ÔÓ×µ is 's probability in positive sentences, and È´ Ò µ is that in negative sentences. They are estimated by using Table ´ ÔÓ×µ ´ ÔÓ×µ • ´ ÔÓ×µ È´ Ò µ ´ Ò µ ´ Ò µ • ´ Ò µ PMI based polarity value Using PMI, the strength of association between and positive sentences (and negative sentences) is defined as follows PMI based polarity value is defined as their difference. This idea is the same as , it is regarded as negative phrase. Otherwise, it is regarded as neutral. Only positive and negative phrases are added to our lexicon. By changing , the trade-off between precision and recall can be adjusted. In order to avoid data sparseness problem, if both ´ ÔÓ×µ and ´ Ò µ are less than three, such candidates were ignored. As described in Section 1, there have been two approaches to (semi-) unsupervised learning of polarity. This Section introduces the two approaches and other related work. Kamps et al. built lexical network by linking synonyms provided by a thesaurus, and polarity was defined by the distance from seed words ('good' and 'bad') in the network Compared with our approach, the drawback of using thesaurus is the lack of scalability. It is difficult to handle such words that are not contained in a thesaurus (e.g. newly-coined words or colloquial words). In addition, phrases cannot be handled because the entry of usual thesaurus is not phrase but word. Another approach is based on an idea that polar phrases conveying the same polarity co-occur with each other in corpus. Ø×´Õµ means the number of hits returned by a search engine when query Õ is issued. AE Ê means NEAR operator, which enables to retrieve only such documents that contain two queries within ten words. Hatzivassiloglou and McKeown constructed lexical network and determine polarity of adjectives Popescu and Etzioni applied relaxation labeling to polarity identification Kanayama and Nasukawa used both intra-and inter-sentential co-occurrence to learn polarity of words and phrases In summary, the strength of our approach is to exploit extremely precise structural clues, and to use In some review sites, pros and cons are stated using such layout that we introduced in Section 2. Some work examined the importance of such layout Some studies addressed supervised approach to learning polarity of phrases Kobayashi et al. proposed a framework to reduce the cost of manually building lexicon A test set consisting of 405 adjective phrases were created. From the test set, we extract polar phrases by looking up our lexicon. The result was evaluated through precision and recall The test set was created in the following manner. 500 adjective phrases were randomly extracted from the Web text. Note that there is no overlap between our polar sentence corpus and this text. After removing parsing error and duplicates, 405 unique adjective phrases were obtained. Each phase was manually annotated with polarity tag (positive, negative and neutral), and we obtained 158 positive phrases, 150 negative phrases and 97 neutral phrases. In order to check the reliability of annotation, another human judge annotated the same data. The Kappa value between the two judges was 0.73, and we think the annotation is reliable. From the test set, we extracted polar phrases by looking up our lexicon. As for adjectives in the lexicon, partial match is allowed. For example, if the lexicon contains an adjective 'excellent', it matches every adjective phrase that includes 'excellent' such as 'view-excellent' etc. As a baseline, we built lexicon similarly by using polarity value of We evaluated the result of polar phrase extraction. By changing the threshold , we investigated recallprecision curve (Figure The Figures show that both of the proposed methods outperform the baselines. The best F-measure was achieved by PMI ( =1.0). Although Turney's method may be improved with minor configurations (e.g. using other seeds etc.), we think this results indicate the feasibility of the proposed method. Al-Figure Comparing the two proposed methods, PMI is always better than chi-square. Especially, chi-square suffers from low recall, because the size of lexicon is extremely small. For example, when the threshold is 60, the precision is 80% and the recall is 48% for negative phrases. On the other hand, PMI would achieve the same precision when recall is around 80% ( is between 0.5 and 1.0). Turney's method did not work well although they reported 80% accuracy in To further assess our method, we did two additional experiments. In the first experiment, to investigate the effect of data size, the same experiment was conducted using 1/n (n=1,5,10,15,20) of the entire polar sentence corpus (Table In the other experiment, the lexicon was evaluated directly so that we can examine polar words and phrases that are not in the test set. We think it is difficult to fully assess low frequency words in the previous setting. Two human judges assessed 200 unique polar words and phrases in the lexicon (PMI, =1.0). The average precision was 71.3% (Kappa value was 0.66). The precision is lower than the result in Table The Table It is important to discuss how general our frame- -2.9 -3.3 yakkaida (annoying) -11.9 -3.9 shomo-hayai (consumption-quick) -17.7 -4.4 work is. Although the lexico-syntactic patterns shown in Section 2 are specific to Japanese, we think that the idea of exploiting language structure is applicable to other languages including English. Roughly speaking, the pattern we exploited can be translated into 'the advantage/weakness of something is to ...' in English. It is worth pointing out that lexico-syntactic patterns have been widely used in English lexical acquisition In this paper, we explore to use structural clues that can extract polar sentences from Japanese HTML documents, and build lexicon from the extracted polar sentences. The key idea is to develop the structural clues so that it achieves extremely high precision at the cost of recall. In order to compensate for the low recall, we used massive collection of HTML documents. Thus, we could prepare enough polar sentence corpus. Experimental result demonstrated the feasibility of our approach.
| 651 | 1,889 | 651 |
Undirected Machine Translation with Discriminative Reinforcement Learning
|
We present a novel Undirected Machine Translation model of Hierarchical MT that is not constrained to the standard bottomup inference order. Removing the ordering constraint makes it possible to condition on top-down structure and surrounding context. This allows the introduction of a new class of contextual features that are not constrained to condition only on the bottom-up context. The model builds translation-derivations efficiently in a greedy fashion. It is trained to learn to choose jointly the best action and the best inference order. Experiments show that the decoding time is halved and forestrescoring is 6 times faster, while reaching accuracy not significantly different from state of the art.
|
Machine Translation (MT) can be addressed as a structured prediction task Statistical models usually define f as: f (x) = arg max y∈Y Score(x, y), where Score(x, y) is a function whose parameters can be learned with a specialized learning algorithm. In MT applications, it is not possible to enumerate all y ∈ Y. HMT decoding applies pruning (e.g. Cube Pruning UMT's fast decoding is made possible through even stronger pruning: the decoder chooses a single action at each step, never retracts that action, and prunes all incompatible alternatives to that action. If this extreme level of pruning was applied to the CKY-like beam-decoding used in standard HMT, translation quality would be severely degraded. This is because the bottom-up inference order imposed by CKY-like beam-decoding means that all pruning decisions must be based on a bottom-up approximation of contextual features, which leads to search errors that affect the quality of reordering and lexical-choice Removing the bottom-up inference order constraint makes it possible to condition on top-down structure and surrounding context. This undirected approach allows us to integrate contextual features such as the Language Model (LM) in a more flex-ible way. It also allows us to introduce a new class of undirected features. In particular, we introduce the Context-Free Factor (CFF) features. CFF features compute exactly and efficiently a bound on the context-free cost of a partial derivation's missing branches, thereby estimating the future cost of partial derivations. The new class of undirected features is fundamental for the success of a greedy approach to HMT, because the additional nonbottom-up context is sometimes crucial to have the necessary information to make greedy decisions. Because UMT prunes all but the single chosen action at each step, both choosing a good inference order and choosing a correct action reduce to a single choice of what action to take next. To learn this decoding policy, we propose a novel Discriminative Reinforcement Learning (DRL) framework. DRL is used to train models that construct incrementally structured output using a local discriminative function, with the goal of optimizing a global loss function. We apply DRL to learn the UMT scoring function's parameters, using the BLEU score as the global loss function. DRL learns a weight vector for a linear classifier that discriminates between decisions based on which one leads to a complete translation-derivation with a better BLEU score. Promotions/demotions of translations are performed by applying a Perceptron-style update on the sequence of decisions that produced the translation, thereby training local decisions to optimize the global BLEU score of the final translation, while keeping the efficiency and simplicity of the Perceptron Algorithm Our experiments show that UMT with DRL reduces decoding time by over half, and the time to rescore translations with the Language Model by 6 times, while reaching accuracy non-significantly different from the state of the art.
|
In this section, we present the UMT framework. For ease of presentation, and following synchronous-grammar based MT practice, we will henceforth restrict our focus to binary grammars A UMT decoder can be formulated as a function, f , that maps a source sentence, x ∈ X , into a structure defined by y = (t, h) ∈ Y, where t is the translation in the target language, and h is the synchronous tree structure generating the input sentence on the source side and its translation on the target side. Synchronous-trees are composed of instantiations of synchronous-rules, r, from a grammar, G. A UMT decoder builds synchronous-trees, h, by recursively expanding partial synchronous-trees, τ . τ includes a partial translation. Each τ is required to be a connected sub-graph of some synchronous-tree h. Thus, τ is composed of a subset of the rules from any h that generates x on the source side, such that there is a connected path between any two rules in τ . Differently from the partial structures built by a bottom-up decoder, τ does not have to cover a contiguous span on x. Formally, τ is defined by: 1) The set of synchronous-rule instantiations in τ : 2) The set of connections among the synchronousrule instantiations, C. Let c i = (r i , r j i ) be the notation to represent the connection between the i-th rule and the rule r j i . The set of connections can be expressed as: 3) The postcondition set, P , which specifies the non-terminals in τ that are available for creating new connections. Each postcondition, p i = (r x , X y ) i , indicates that the rule r x has the non-terminal X y available for connections. The index y identifies the non-terminal in the rule. In a binary grammar y can take only 3 values: 1 for the first non-terminal (the left child of the source side), 2 for the second non-terminal, and h for the head. The postcondition set can be expressed as: The set of carries, K. We define a different carry, κ i , for each non-terminal available for connections. Each carry stores the extra information required to correctly score the non-local interactions between τ and the rule that will be connected at that non-terminal. Thus |K| = |P |. Let κ i be the carry associated with the postcondition p i . The set of carries can be expressed as: Algorithm 1 gives details of the UMT decoding algorithm. The decoder takes as input the source sentence, x, the parameters of the scoring function, w, and the synchronous-grammar, G. At line 2 the partial synchronous-tree τ is initialized by setting I, C, P and K to empty sets ∅. At line 3 the queue of candidate connection-actions is initialized as Q ≡ { [r leaf , null] | r leaf is a leaf rule}, where null means that there is no postcondition specified, since the first rule does not need to connect to anything. A leaf rule r leaf is any synchronous rule with only terminals on the right-hand sides. At line 4 the main loop starts. Each iteration of the main loop will expand τ using one connection-action. The loop ends when Q is empty, implying that τ covers the full sentence and has no more missing branches or parents. The best scoring action according to the parameter vector w is popped from the queue at line 5. The scoring of connection-actions is discussed in details in Section 3.2. At line 6 the selected connection-action is used to expand τ . At line 7 the queue of candidates is updated accordingly (see lines 19-21). At line 8 the decoder it-erates the main loop, until τ is complete and is returned at line 9. Lines 10-18 describe the CreateConnection(•) procedure, that connects the partial synchronoustree τ to the selected rule r via the postcondition p i specified by the candidate-action selected in line 5. This procedure returns the resulting partial synchronous-tree: τ ≡ τ ⋖ [r, p i ] . At line 11, r is added to the rule set I. At line 12 the connection between r and r p i (the rule specified in the postcondition) is added to the set of connections C. At line 13, p i is removed from P . At line 14 the carry k i matching with p i is removed from K. At line 15 the set of carries K is updated, in order to update those carries that need to provide information about the new action. At line 16 new postconditions representing the non-terminals in r that are available for subsequent connections are added in P . At line 17 the carries associated with these new postconditions are computed and added to K. Finally at line 18 the updated partial synchronous-tree is returned. In the very first iteration, the CreateConnection(•) procedure has nothing to compute for some lines. Line 11 is not executed since the first leaf rule needs no connection and has nothing to connect to. lines 12-13 are not executed since P and K are ∅ and p i is not specified for the first action. Line 15 is not executed since there are no carries to be updated. Lines 16-17 only add the postcondition and carry relative to the leaf rule head link. The procedure used to update Q is reported in lines 19-21. At line 20 all the connection-actions involving the expansion of p i are removed from Q. These actions are the incompatible alternatives to the selected action. In the very first iteration, all actions in Q are removed because they are all incompatible with the connected-graph constraint. At line 21 new connection-actions are added to Q. These are the candidate actions proposing a connection to the available non-terminals of the selected action's new rule r. The rules used for these new candidate-actions must not be in conflict with the current structure of τ (e.g. the rule cannot generate a source side terminal that is already covered by τ ). Training a UMT model simply means training the parameter vector w that is used to choose the best scoring action during decoding. We propose a novel method to apply a kind of minimum error rate training (MERT) to w. Because each action choice must be evaluated in the context of the complete translation-derivation, we formalize this method in terms of Reinforcement Learning. We propose Discriminative Reinforcement Learning as an appropriate way to train a UMT model to maximize the BLEU score of the complete derivation. First we define DRL as a novel generic training framework. RL can be applied to any task, T , that can be formalized in terms of: 1) The set of states S 3) The transition function T : S × A s → S, that specifies the next state given a source state and performed action A policy is defined as any map π : S → A. Its value function is given by: where path(s 0 |π) ≡ s 0 , s 1 , • • • , s σ |π is the sequence of states determined by following policy π starting at state s 0 . The Q-function is the total future reward of performing action a 0 in state s 0 and then following policy π: Standard RL algorithms search for a policy that maximizes the given reward. Because we are taking a discriminative approach to learn w, we formalize our optimization task similarly to an inverse reinforcement learning problem end if 9: until convergence 10: Return(w); approach simplifies the task of learning the reward function in two respects: the learned reward function only needs to be monotonically related to the true reward function, and this property only needs to hold for the best competing alternatives. This is all we need in order to use the discriminative reward function in an optimal classifier, and this simplification makes learning easier in cases where the true reward function is too complicated to model directly. In RL, an optimal policy π * is one which, at each state s, chooses the action which maximizes the future reward Q π * (s, a). We assume that the future discriminative reward can be approximated with a linear function Qπ (s, a) in some featurevector representation φ : S × A s → R d that maps a state-action pair to a d-dimensional features vector: Qπ (s, a) = w φ(s, a) where w ∈ R d . This gives us the following policy: The set of parameters of this policy is the vector w. With this formalization, all we need to learn is a vector w such that the resulting decisions are compatible with the given information about the optimal action sequence. We propose a Perceptron-like algorithm to learn these parameters. Algorithm 2 describes the DRL meta-algorithm. The Trainer takes as input φ, the task T , and a generic set of data D describing the behaviors we want to learn. The output is the weight vector w of the learned policy that fits the data D. The algorithm consists in a single training loop that is repeated until convergence (lines 2-9). At line 3 a state, s, is sampled from S. At line 4, â is set to be the action that would be preferred by the current w-policy. At line 5 an action, a ′ , is sampled from A s such that a ′ = â. At line 6 the algorithm checks if preferring path(T (s, â), π w ) over path(T (s, a ′ ), π w ) is a correct choice according to the behaviors data D that the algorithm aims to learn. If the current w-policy contradicts D, line 7 is executed to update the weight vector to promote Φ w (s, a ′ ) and penalize Φ w (s, â), where Φ w (s, a) is the summation of the features vectors of the entire derivation path starting at (s, a) and following policy π w . This way of updating w has the effect of increasing the Q(•) value associated with all the actions in the sequence that generated the promoted structure, and reducing the Q(•) value of the actions in the sequence that generated the penalized structure 3 . We have described the DRL meta-algorithm to be as general as possible. When applied to a specific problem, more details can be specified: 1) it is possible to choose specific sampling techniques to implement lines 3 and 5; 2) the test at line 6 needs to be detailed according to the nature of T and D; 3) the update statement at line 7 can be replaced with a more sophisticated update approach. We address these issues and describe a range of alternatives as we apply DRL to UMT in Section 3.2. To apply DRL we formalize the task of translating x with UMT as T ≡ {S, {A s }, T, R, γ}: 1) The set of states S is the space of all possible UMT partial synchronous-trees, τ ; 2) The set A τ,x is the set of connection-actions that can expand τ connecting new synchronousrule instantiations matching the input sentence x on the source side; 3) The transition function T is the connection function τ ≡ τ ⋖ a formalized in Section 2 and detailed by the procedure CreateConnection(•) in Algorithm 1; 4) The true reward function R is the BLEU score. BLEU is a loss function that quantifies the difference between the reference translation and the output translation t. The BLEU score can be computed only when a terminal state is reached and a full translation is available. Thus, the rewards are all zero except at terminal states, called a Pure De-3 Preliminary experiments with updating only the features for â and a ′ produced substantially worse results. layed Reward function; 5) Considering the nature of the problem and reward function, we choose an undiscounted setting: γ = 1. Next we specify the details of the DRL algorithm. The data D consists of a set of pairs of sentences, D ≡ {(x, t * )}, where x is the source sentence and t * is the reference translation. The feature-vector representation function φ maps a pair (τ, a) to a real valued vector having any number of dimensions. Each dimension corresponds to a distinct feature function that maps: {τ } × A τ,x → R. Details of the features functions implemented for our model are given in Section 4. Each loop of the DRL algorithm analyzes a single sample (x, t * ) ∈ D. The state s is sampled from a uniform distribution over s 0 , s 1 , • • • , s σ |π . The action a ′ is sampled from a Zipfian distribution over {A τ,x -â} sorted with the Qπw (s, a) function. In this way actions with higher score have higher probability to be drawn, while actions at the bottom of the rank still have a small probability to be selected. The if at line 6 tests if the translation produced by path(T (s, a ′ ), π w ) has higher BLEU score than the one produced by path(T (s, â), π w ). For the update statement at line 7 we use the Averaged Perceptron technique Algorithm 2 can be easily adapted to implement the efficient Averaged Perceptron updates (e.g. see Section 2.1.1 of In this section we show how the features designed for bottom-up HMT can be adapted to the undirected approach, and we introduce a new feature from the class of undirected features that are made possible by the undirected approach. Local features depend only on the action rule r. These features can be used in the undirected approach without adaptation, since they are independent of the surrounding structure. For our experiments we use a standard set of local features: the probability of the source phrase given the target phrase; the lexical translation probabilities of the source words given the target words; the lexical translation probabilities of the target words given the source words; and the Word Penalty feature. Contextual features are dependent on the interaction between the action rule r and the available context. In UMT all the needed information about the available context is stored in the carry κ i . Therefore, the computation of contextual features whose carry's size is bounded (like the LM) requires constant time. The undirected adaptation of the LM feature computes the scores of the new n-grams formed by adding the terminals of the action rule r to the current partial translation τ . In the case that the action rule r is connected to τ via a child nonterminal, the carry is expressed as Where W L and W R are respectively the left and right boundary target words of the span covered by τ . This notation is analogous to the standard star notation used for the bottom-up decoder (e.g. Where W L and W R are respectively the left and right boundary target words of the surrounding context provided by τ . The boundary words stored in the carry and the terminals of the action rule are all the information needed to compute and score the new n-grams generated by the connection-action. In addition, we introduce the Context-Free Factor (CFF) features. An action rule r is connected to τ via one of r's non-terminals, X r,τ . Thus, the score of the interaction between r and the context structure attached to X r,τ can be computed exactly, while the score of the structures attached to other r nonterminals (i.e. those in postconditions) cannot be computed since these branches are missing. Each of these postcondition nonterminals has an associated CFF feature, which is an upper bound on the score of its missing branch. More precisely, it is an upper bound on the context-free component of this score. This upper bound can be exactly and efficiently computed using the Forest Rescoring Framework To efficiently compute CFF features we run the Inside-Outside algorithm with the (max, +) semiring We implement our model on top of Cdec We experiment on the NIST Chinese-English parallel corpus. The training corpus contains 239k sentence pairs with 6.9M Chinese words and 8.9M English words. The test set contains 919 sentence pairs. The hierarchical translation grammar was extracted using the Joshua toolkit Table Table To test the impact of the CFF features, we trained and tested UMT with DRL with and without these features. This resulted in an accuracy decrease of 2.3 BLEU points. Thus these features are important for the success of the greedy approach. They provide an estimate of the score of the missing branches, thus helping to avoid some actions that have a good local score but lead to final translations with low global score. To validate the results, additional experiments were executed on the French to Italian portion of the Europarl corpus v6. This portion contains 190k pairs of sentences. The first 186k sentences were used to extract the grammar and train the two models. The final tests were performed on the remaining 4k sentence pairs. With this corpus we measured a similar speed gain. HMT b30 is 2.3 times slower at decoding compared to UMT, and 6.1 times slower at rescoring, while UMT loses 1.1 BLEU points in accuracy. But again the accuracy differences are not considered significant. We measured a p-value of 0.25, which is not significant at the 0.1 level. Models sharing similar intuitions have been previously applied to other structure prediction tasks. For example, Other approaches learning a model specifically for greedy decoding have been applied with suc-cess to other less complex tasks. Similarities can be found between DRL and previous work that applies discriminative training to structured prediction: In general, we believe that greedy-discriminative solutions are promising for tasks like MT, where there is not a single correct solution: normally there are many correct ways to translate the same sentence, and for each correct translation there are many different derivation-trees generating that translation, and each correct derivation tree can be built greedily following different inference orders. Therefore, the set of correct decoding paths is a reasonable portion of UMT's search space, giving a well-designed greedy algorithm a chance to find a good translation even without beam search. In order to directly evaluate the impact of our proposed decoding strategy, in this paper the only novel features that we consider are the CFF features. But to take full advantage of the power of discriminative training and the lower decoding complexity, it would be possible to vastly increase the number of features. The UMT's undirected nature allows the integration of non-bottom-up contextual features, which cannot be used by standard HMT and PbMT. And the use of a historybased model allows features from an arbitrarily wide context, since the model does not need to be factorized. Exploring the impact of this advantage is left for future work. The main contribution of this work is the proposal of a new MT model that offers an accuracy/complexity balance that was previously unavailable among the choices of hierarchical models. We have presented the first Undirected framework for MT. This model combines advantages given by the use of hierarchical synchronousgrammars with a more efficient decoding algorithm. UMT's nature allows us to design novel undirected features that better approximate contextual features (such as the LM), and to introduce a new class of undirected features that cannot be used by standard bottom-up decoders. Furthermore, we generalize the training algorithm into a generic Discriminative Reinforcement Learning meta-algorithm that can be applied to any structured prediction task.
| 712 | 3,052 | 712 |
Frame Semantics for Stance Classification
|
Determining the stance expressed by an author from a post written for a two-sided debate in an online debate forum is a relatively new problem in opinion mining. We extend a state-of-the-art learningbased approach to debate stance classification by (1) inducing lexico-syntactic patterns based on syntactic dependencies and semantic frames that aim to capture the meaning of a sentence and provide a generalized representation of it; and (2) improving the classification of a test post via a novel way of exploiting the information in other test posts with the same stance. Empirical results on four datasets demonstrate the effectiveness of our extensions.
|
Given a post written for a two-sided topic in an online debate forum (e.g., "Should abortion be allowed?"), the task of debate stance classification involves determining which of the two sides (i.e., for or against) its author is taking. For example, a stance classification system should determine that the author of the following post is anti-abortion. Post 1: Abortion has been legal for decades and no one seems to have a problem with it. That's ridiculous! There are millions of people in the world who would love to have children but can't. Previous approaches to debate stance classification have focused on three debate settings, namely congressional floor debates Our goal in this paper is to improve the state of the art in stance classification of online debates, focusing in particular on ideological debates. Specifically, we present two extensions, one linguistic and the other extra-linguistic, to the state-of-the-art supervised learning approach to this task proposed by We evaluate our approach to stance classification of ideological debates on datasets collected for four domains from online debate forums. Experimental results demonstrate the effectiveness of our approach: it outperforms an improved version of Anand et al.'s approach by 2.6-7.0 accuracy points on the four domains. The rest of the paper is organized as follows. We first present our datasets in Section 2. Section 3 describes our two learning-based baseline systems for stance classification. Sections 4 and 5 discuss our two extensions. Finally, we show evaluation results in Section 6 and present conclusions in Section 7.
|
For our experiments, we collect debate posts from four popular domains, Abortion (ABO), Gay Rights (GAY), Obama (OBA), and Marijuana (MAR). Each post should receive one of two domain labels, for or against, depending on whether the author of the post supports or opposes abortion, gay rights, Obama, or the legalization of marijuana. To see how we obtain these domain labels, let us first describe the data collection process in more detail. We collect our debate posts for the four domains from an online debate forum 1 . In each domain, there are several two-sided debates. Each debate has a subject (e.g., "Abortion should be banned") for which a number of posts were written by different authors. Each post is manually tagged with its author's stance (i.e., yes or no) on the debate subject. Since the label of each post represents the subject stance but not the domain stance, we need to automatically convert the former to the latter. For example, for the subject "Abortion should be banned", the subject stance yes implies that the author opposes abortion, and hence the domain label for the corresponding label should be against. We construct one dataset for each domain. Statistics of these datasets are shown in Table We employ as baselines two stance classification systems, Our first baseline, Anand et al.'s approach, is a supervised method that trains a stance classifier 1 Related work on stance classification of congressional debates has found that enforcing author constraints (ACs) can improve classification performance (e.g., We first use the learned stance classifier to classify the test posts as in the first baseline, and then postprocess the labels of the test posts. Specifically, we sum up the confidence values Our first extension to To motivate this extension, let us take a closer look at Anand et al.'s attempt to generalize using syntactic dependencies. Note that any approach that aims to generalize using syntactic dependencies suffers from several weaknesses. First, the semantic relationship between the pair of lexical items involved in each of these features is not encoded. This means that the resulting features do not adequately capture the meaning of the underlying sentence. Second, replacing a word with its POS tag is a syntactic, not semantic, generalization, and doing so further abstracts the resulting feature from the meaning of the underlying sentence. Above all, while the resulting features are intended to improve generalizations, they can provide very limited generalizations. To see why, consider two semantically similar sentences "I hate arrogant people" and "I dislike arrogant people". Ideally, any features that intend to provide a generalized representation of these sentences should be able to encode the fact that they are semantically similar. However, Anand et al.'s features would fail to do so because they cannot capture the fact that "hate" and "dislike" are semantically similar. In the rest of this section we describe how we generate a semantic generalization of a sentence to capture its meaning. Our approach to semantic generalization involves (1) inducing from the training data a set of patterns that aim to provide a semantic generalization of the sentences in the training posts and (2) using them in combination with the baseline systems to classify a test post. Below we describe these two steps in detail. This step is composed of two sub-steps. For each domain, we extract a list of topics. We define a topic as a word sequence that (1) starts with zero or more adjectives and ends with one or more nouns and (2) appears in at least five posts from the domain. Using this method, for example, we can extract "abortion", "partial-birth abortion", "birth control", etc., as the topics for Abortion. Given a sentence, we create patterns to capture its information using syntactic dependencies and semantic frames. (1) Some people hate guns. (2) Some people do not like guns. Subject-Frame-Object (SFO) patterns. We create a set of SFO patterns for a transitive verb if (1) it is a frame target Next, note that since the subject of hate is the target of the frame People and its object is a topic, we need to create patterns in a similar manner, resulting in Patterns 3 and 4. Note that People in these two patterns (with 'P' capitalized) is the name of the frame People, not the word people appearing in the sentence. To provide better generalization, we create a simplified version of each SFO pattern by replacing the frame name representing subject/object with the value DC. This results in Patterns 5-8. For sentence (2), we can generate patterns in a similar manner, resulting in Patterns 9-16. For example, Pattern 9 contains the element NEG, which encodes the fact that the verb like is negated. Pattern 10 deserves discussion. Since the positive sentiment-bearing verb like is negated, the sentiment value of Pattern 10 is -, which encodes the fact that not like has a negative sentiment. The negation value of Pattern 10 is POS rather than NEG, reflecting the fact that not like does not appear in a negative context. In other words, the sentiment value needs to be flipped if the verb is negated, and so may the negation value. It is worth noting that Patterns 2 and 10 are identical, which provides suggestive evidence that sentences (1) and ( Dependency-Frame (DF) patterns. We create a set of DF patterns for a dependency relation d if (1) both arguments of d are frame targets or (2) the head is a frame target and the dependent is a topic. For example, in the dependency relation dobj(hate,guns), both hate and guns are frame targets, as discussed above, and guns is a topic, so a set of DF patterns (Patterns 17-20 in Table (3) the frame name of the head is EF; (4) the frame name of the dependent is Weapon; (5) the head is not negated; and (6) we don't care about the sentiment of the head. Pattern 18 is the same as Pattern 17, except that it takes into account the sentiment value of the verb. Patterns 19 and 20 replaces the frame name of the dependent with the topic name, which is guns. The negation and sentiment values are computed in the same way as those in the SFO patterns. Frame-Element-Topic (FET) patterns. We create one FET pattern for every (v,f e) pair in a sentence where v is a verb and a frame target, and f e is a topic and a frame element of v's frame. In this step, we will use the patterns learned in Step 1 in combination with the baseline systems to classify a test post. A simple way to combine the learned patterns with the baseline systems would be to augment the feature set they employ with the learned patterns. One potential weakness of this method is that the impact of these patterns could be undermined by the fact that they are significantly outnumbered by the baseline features, particularly the n-gram features. For this reason, we decided to train another stance classifier, which we will refer to as the semantics-based classifier, c s . Like the baseline stance classifier c b , (1) c s is trained using SVM light , (2) each training instance for c s corresponds to a training post, and (3) its class label is the stance the post expresses. Unlike c b , however, the features employed by c s are created from the learned patterns. Specifically, from each pattern we create one binary feature whose value is 1 if and only if the corresponding pattern is applicable to the training post under consideration. A natural question, then, is: how can we combine the decisions made by c b and c s ? To answer this question, we applied both classifiers to the de- At first glance, we should just abandon c s because of its consistently poorer performance. However, since the two classifiers are trained on disjoint feature sets (one is lexico-syntactic and the other semantic), we hypothesize that the mistakes they made on the development set could be complementary. To confirm this hypothesis, we compute the percentage of posts in the development set that are correctly classified by one but not the other. Results of this experiment are shown in Table Given these results, we hypothesize that performance could be improved by combining the predictions made by c b and c s . Since c b consistently outperforms c s on all datasets, we use c s to make a prediction if and only if (1) c b cannot predict confidently and ( To classify a debate post p in the test set, we have so far exploited only the information extracted from p itself. However, it is conceivable that we can improve the classification of p by exploiting the information extracted from other test posts that have the same stance as p. This is the goal of our second extension. To see why doing so can improve the classification of p, we make a simple observation: some posts are easier to classify than the others. Typically, posts containing expressions that are strong indicators of the stance label are easier to classify than those that do not. As an example, consider the following posts: Post 2: I don't think abortion should be illegal. Post 3: What will you do if a woman's life is in danger while she's pregnant? Do you still want to sacrifice her life simply because the fetus is alive? It should be fairly easy for a human to see that the authors of both posts support abortion. However, Post 2 is arguably easier to classify than Post 3: Post 2 has an easy-to-determine stance, whereas Post 3 has a couple of rhetorical questions that may be difficult for a machine to understand. Hence, we might be able to improve the classification of Post 3 by exploiting information from other posts that have the same stance as itself (which in this case would be Post 2). In practice, however, we are not given the information of which posts have the same stance. In the two subsections below, we discuss two simple methods of determining whether two posts are likely to have the same stance. The first method, which we will refer to as M 1 , is fairly straightforward: we posit that two posts are likely to have the same stance if they are written by the same author. Given a test post p to be classified, we can use this method to identify a subset of p's same-stance posts. For convenience, we denote this set as SameStancePosts(p). The question, then, is: how can we exploit information in SameStancePosts(p) to improve the classification of p? One way would be to combine the content of the posts in SameStancePosts(p) with that of p (i.e., by taking the union of all the binaryvalued feature vectors), and use the class value of the combined post as the class value of p. However, rather than simply combining all the posts to form one big post, we generalize this idea by (1) generating all possible combinations of posts in SameStancePosts(p); (2) for each such combination, combine it with p; (3) classify each combination obtained in (2) using the SVM classifier; (4) sum the confidence values of all the combinations; and (5) use the signed value as the class value of p. Note that if SameStancePosts(p) contains n posts, the number of possible combinations is n i=0 n i . For efficiency reasons, we allow each combination to contain at most 10 posts. At first glance, it seems that the combination method described in the previous paragraph is an alternative implementation of ACs. (Recall that ACs are inter-post constraints that ensure that two posts written by the same author for the same domain should receive the same label.) Nevertheless, there are two major differences between our combination method and ACs. First, in ACs, the same-author posts can only interact via the confidence values assigned to them. On the other hand, in our proposal, the same-author posts interact via Feature Definition SameDebate whether authors posted in same debate SameThread whether authors posted in same thread Replied whether one author replied to the other Table feature sharing. In other words, in ACs, the sameauthor posts interact after they are classified by the stance classifier, whereas in our proposal, the interaction occurs before the posts are classified. Second, in ACs, all the same-author posts receive the same stance label. On the other hand, this is not necessarily the case in our proposal, because two same-author posts can be classified using different combinations. In other words, ACs and our combination method are not the same. In fact, they can be used in conjunction with each other. Using M 1 to identify same-stance posts has a potential weakness. If an author has composed a small number of posts, then the number of combinations that can be generated will be small. In the extreme case, if an author has composed just one post p, then no combinations will be generated using M 1 . To enable p to benefit from our idea of exploiting same-stance posts, we propose another method to identify same-stance posts, M 2 , which is a generalization of M 1 . In M 2 , we posit that two posts are likely to have the same stance if they are written by the same author or by similar-minded authors. Given test post p, we can compute SameStancePosts(p) using the definition of M 2 , and apply the same 5-step combination method described in the previous subsection to SameStancePosts(p) to classify p. The remaining question is: given an author, a, in the test set, how do we compute his set of similar-minded authors, A similar ? To do this, we train a binary author-agreement classifier on the training set to generate A similar for a. Specifically, each training instance corresponds to a pair of authors in the training set having one of two class labels, agree (i.e., authors have the same stance) and disagree (i.e., authors have opposing stances). We represent each instance with two types of features. Features of the first type are obtained by taking the difference of the feature vectors corresponding to the two authors under consideration, where the feature vector of an author is obtained by taking the union of the feature vectors corresponding to all of the posts written by her. Taking the difference would allow the learner to focus on those features whose values differ in the feature vectors. For the second type of features, we use author interaction information encoded as three binary features (see Table Results are expressed in terms of accuracy obtained via 5-fold cross validation, where accuracy is the percentage of test instances correctly classified. Since all experiments require the use of development data for parameter tuning, we use three folds for model training, one fold for development, and one fold for testing in each fold experiment. Results are shown in Table Next, we incorporate our first extension, pattern induction, into the better of the two baselines (i.e., the second baseline). Results of combining c b and c s to classify the test posts (together with the ACs) are shown in row 3 of Table Before incorporating our second extension, let us recall our earlier hypothesis that combining c b and c s using our method would be better than training just one classifier that combines the features used by c b and c s . The reason behind our hypothesis was that simply combining the feature sets would undermine the impact of pattern-based features because they would be significantly outnumbered by the features in c b . To confirm this hypothesis, we showed in row 4 of Table As we can see, this classifier (referred to as c bs in the table) together with the ACs performs significantly worse than the c b +c s +AC system (row 3) on all datasets. In fact, the c b +AC system (row 2) outperforms the c bs +AC system on OBA, but they are statistically indistinguishable on the remaining datasets. These results suggest that combining the pattern-based features with the baseline features into one feature set renders the former ineffective. Finally, we incorporate our second extension, the one that involves generating combinations of test posts written by the same author (M 1 ) and by both the same author and similar-minded authors (M 2 ). Results of these experiments are shown in rows 5-6 of Table Overall, our two extensions yield a stance classification system that significantly outperforms the better baseline on all four datasets, with an average improvement of 6.4 accuracy points. Given the better performance of the combination-based systems, a natural question is: can we further improve performance by applying our combination methods to generate artificial posts and use them as additional training instances? To answer this question, we apply both M 1 and M 2 to generate additional training instances, using a random selection of same-stance authors in place of M 2 's k-nearest neighbor method. However, neither method yields an improvement in performance over the method on which it is based. We speculate that since all the posts in the training combinations are already present in the training set as individual posts, they are more likely to be farther away from the hyperplane than the individual posts, meaning that they are less likely to be support vectors. This in turn implies that they are less likely to affect classification performance. To gain additional insights into our approach, we performed a qualitative analysis of the errors produced by our best-performing system below. Failure to accumulate decisions from several clues. Authors often express their stance using a group of sentences where the latter sentence(s) indicate the actual stance and the initial sentence(s) may give a false impression about the author's stance. Consider Post 1 (see Section 1) and Post 4. Post 4: I agree abortion creates stress and pain. I agree it kills a potential life. That does not mean it is right to ban abortion. In Post 1, the author is anti-abortion, whereas in Post 4, the author is pro-abortion. However, the first sentence in Post 1 gives a misleading clue about the author's stance, and so do the first two sentences in Post 4. Since all the systems discussed in the paper operate on one sentence at a time, they are all prone to such errors. One way to address this problem could be to determine how adjacent sentences are related to each other via the use of discourse relations. Presence of materials irrelevant to stance. Because of the informal style of writing, we often find long posts with one or two sentences indicating the actual stance of the author. The rest of such posts often include descriptions of an author's personal experience, comments or questions directed to other authors etc. Such long posts are frequently misclassified for all four domains. Consider the following example. Post 5: Marijuana should at least be decriminalized. Driving stoned, however, is something totally different and should definitely be a crime. Also, weed can't kill you, unlike cigarettes and alcohol. In my opinion cigarettes should definitely be ille-gal, but they're so ingrained into our culture that I doubt that is going to happen any time soon. In this post, the author supports the legalization of marijuana. However, the only useful hints about her stance are "marijuana should at least be decriminalized" and "weed can't kill you". The rest of the post is not helpful for stance classification. Convoluted posts appearing later in long post sequences. As a post sequence gets longer, authors tend to focus on specific aspects of a debate and consequently, it becomes more difficult to classify their stances, even with the context-based features (features taken from the immediately preceding post) proposed by Anand et al. Consider the following post sequence, where only the first post (P1) and the nth post (Pn) are shown due to space limitations. [P1: Anti-Obama] Obama is a pro-abortionist. Killing babies is wrong so stop doing it. The new health reform bill is not good. There are some good things but more worse than good. You could have just passed some laws instead of making a whole bill. • • • [Pn: Pro-Obama] Killing fetuses isn't wrong. Besides, we could use those fetuses for stem cell research. As we can see, the author of P1 does not support Obama because of his pro-abortion views. In Pn, a pro-Obama author explains why she thinks abortion is not wrong. However, without the context from P1 that Obama is pro-abortion, it is not easy for a machine to classify Pn correctly. This problem is more serious in ABO and GAY than in the other domains as the average length of a post sequence in these two domains is larger. We examined the under-studied task of stance classification of ideological debates. Employing our two extensions yields a system that outperforms an improved version of Anand et al.'s approach by 2.6-7.0 accuracy points. In particular, while existing approaches to debate stance classification have primarily employed lexico-syntactic features, to our knowledge this is the first attempt to employ FrameNet for this task to induce features that aim to capture the meaning and provide semantic generalizations of a sentence. In addition, our method for identifying and exploiting same-stance posts during the inference procedure provides further gains when used on top of our FrameNet extension.
| 657 | 1,614 | 657 |
Facilitating Multi-turn Emotional Support Conversation with Positive Emotion Elicitation: A Reinforcement Learning Approach
|
Emotional support conversation (ESC) aims to provide emotional support (ES) to improve one's mental state. Existing works stay at fitting grounded responses and responding strategies (e.g., question), which ignore the effect on ES and lack explicit goals to guide emotional positive transition. To this end, we introduce a new paradigm to formalize multi-turn ESC as a process of positive emotion elicitation. Addressing this task requires finely adjusting the elicitation intensity in ES as the conversation progresses while maintaining conversational goals like coherence. In this paper, we propose SUPPORTER, a mixture-of-expert-based reinforcement learning model, and well design ES and dialogue coherence rewards to guide policy's learning for responding. Experiments verify the superiority of SUPPORTER in achieving positive emotion elicitation during responding while maintaining conversational goals including coherence.
|
Emotional support (ES) aims to reassure a person to recover from emotional distress and improve one's mental state To achieve this goal, a typical practice is modeling empathy, which aims to perceive and understand the situation and feelings of others Yeah! I don't even know what is going to happen with our final. That is really upsetting and stressful. Maybe you can talk to your parents or a close friend about this, they are also frustrated due to the same experience. Yes, you are right! I am not the only one affected by the pandemic. Come on! I believe that you will definitely be able to overcome the difficulties and find a solution! Yeah, definitely! Thanks for your support, I feel better now. interaction. (2) Lack of awareness of emotional elicitation. Only emanating emotional resonance fails to help users jump out of negative mental states. Although To this end, we introduce multi-turn ESC with positive emotion elicitation, a new paradigm aims to progressively empathize and elicit users to reach a better mental state through multi-turn conversation. Addressing this task is challenging (an example is in Figure In this paper, we propose SUPPORTER Our contributions are summarized as follows: (1) We introduce a new paradigm by carefully dissecting the challenges of formalizing multi-turn ESC as a process of positive emotion elicitation. (2) We propose SUPPORTER, an MoE-based RL model with carefully constructed ES and dialogue coherence rewards, elicits positive emotion during responding while maintaining dialogue coherence. (3) Extensive experiments show the superiority of SUPPORTER with automatic, interactive human, and novel ES and dialogue coherence evaluations.
|
Empathetic Conversation To construct a warm dialogue system, a milestone is to endow it with empathy Emotional Support Conversation To remedy above deficiencies, To free users from emotional distress and advance the conversation towards an optimistic state, positive emotion elicitation is an intuitive solution 3 Preliminaries The purpose of generative processes to enhance elicitation induces two attendant issues: First, with-out proper controls may lead to greedily pursuing the goals of elicitation while discarding the contextual coherence, e.g., "Come on!" with strong elicitation as a response in the context of the user continuing to express negative emotions. Second, whether the response meets the user's expectations needs feedback from the user's future utterance. It means maintaining coherence with future dialogue is also crucial. In this work, we construct contextual and future dialogue coherence rewards to guide the model's learning of bi-coherent expressions and perform the automatic and interactive human evaluation of conversational goals including coherence. In Figure We use the updated state to generate response and further optimize the policy by measuring how well the response reaches the goal of ES and dialogue coherence with the well-designed parallel rewards. As a key component of SUPPORTER, we first introduce the structure of multi-task mixture-of-expert. Dialogue Encoder Following Emotion Experts To track possible transitions of user's emotional states, emotion experts are associated with contextual and future user emotion predictions. We extract M fine-grained emotional reactions for each utterance in the corpus, which are inferred from COMET We project the [CLS] representations h X,pos and h X,neg of positive and negative experts to predict positive and negative emotion, respectively: which is supervised by the positive and negative emotions collected in the e * pos and e * neg sets of the user's last utterance in the dialogue context using cross-entropy loss: (3) Note that an utterance may be inferred to the emotions with different polarities due to cognitive differences For future emotion experts, we adopt the above method to get L f tr-emo pos and L f tr-emo neg losses and train them to predict the positive and negative emotions of the user's future utterance (i.e., next turn utterance). In this way, emotion experts can learn various emotion-level features by L emo loss: Keyword Experts To meet the need for dialogue coherence, keyword experts are associated with keyword predictions that act on maintaining coherence with contextual and future utterances. Here, a bidirectional emotion keyword graph G is constructed, which is also used in coherence rewards designing (a construction example is in Appendix A). We extract the salient keywords of each utterance in the corpus as vertices using a rule-based approach Contextual keyword experts are transformed similarly to emotion experts, and their [CLS] representations h ctx-kws X,pos and h ctx-kws X,neg can be obtained from positive and negative keyword experts H ctx-kws X,pos and H ctx-kws X,neg , respectively. We infer the one-hop neighbors of contextual keywords from the "forward-positive" and "forward-negative" relations respectively in G to enhance the perception of the target keywords in the golden response. Specifically, we use attention ) where E ctx-kws pos and E ctx-kws neg are positive and negative neighbor embedding matrices that share parameters with the dialogue encoder. We then concatenate e ctx-kws pos and e ctx-kws neg with H ctx-kws X,pos and H ctx-kws X,neg respectively at the token level, and use an MLP layer to fuse them to obtain keywordenhanced experts H ctx-kws X,pos-kws and H ctx-kws X,neg-kws : Further, we take the positive and negative key-words in the golden response as supervision to optimize the L ctx-kws pos and L ctx-kws neg losses adopting cross-entropy (this process can refer to above emotion prediction task). Similarly, multi-hop reasoning on G, i.e., "forward → forward → backwardpositive" and "forward → forward → backwardnegative" (clarified in Appendix A), is performed to obtain keywords coherent with the future utterance. Taking the positive and negative keywords in future utterance as the prediction target, the keyword-enhanced future keyword experts can be optimized by L f tr-kws pos and L f tr-kws neg losses. In this way, keyword experts can learn various expression-level features by L kws loss: Multi-task Training To make the experts retain the primitive semantics without hindering their respective diversity, we give them a minor constraint. Specifically, we average the representations of emotion and keyword experts to get h X,exp , and make it close to sequence representation h X by optimizing the MSE loss with a minor hyperparameter α: where d h is the dimension of h X . Then, we jointly train the multi-task MoE by optimizing L exp loss: We use the standard reinforcement learning framework (Sutton and Barto, 2018) as the backbone. State We concatenate the dialogue context and the extracted keywords as the initial state s 1 ∈ S, i.e., s 1 = {C, C kws } (we omit the subscript t of dialogue context C t for simplicity). At each step, the prompt token sequence E generated by the policy determined expert (i.e., action) triggers an update of the state. We record the observed state where η(•) is an ELU activation function with a dropout layer, ⊙ is the hadamard product, ϕ(•) is the softmax function. A k is a binarized vector for pruning the action space, and we set it as a full-one vector due to the small number of experts. Rewards To guide policy learning, we reward the decision made at each step by measuring how well the response generated from updated state s k+1 provides ES and maintains dialogue coherence. (1) Conversation-level ES Reward: aims to dynamically adjust the elicitation intensity of positive emotion as the conversation progresses defined as: Here, f ES (•) measures the positive emotion level of an utterance using the emotion classification model developed by (2) Turn-level ES Reward: aims to capture the feedback of user's next turn emotion defined as: Here, P ED tES measures the relative positive emotion distance between the generated response y and the user's future (i.e., next turn) utterance c f . We encourage P ED tES to get smaller with the approaching of current turn T to M T , i.e., supervising smooth elicitation in the latter stage and improving tolerance to emotional fluctuations. (3) Contextual Dialogue Coherence Reward: aims to constrain generated response y to maintain coherence with context C by measuring their coherence at keyword-level and sentence-level. First, we reconstruct a dataset (4) Future Dialogue Coherence Reward: aims to introduce the consideration of coherence with the user's future utterance c f . Similarly, we reconstruct a dataset where N f,kws is the number of keywords in y kws that have a backward relation with keywords c f kws of c f in G. (5) Total reward. The total reward is r = w cES * r cES +w tES * r tES +w cDC * r cDC +w f DC * r f DC . Info. # We set K-step iterations, and the goal of agent learning is to maximize the expected cumulative reward: , where θ is the learned parameter and γ is the discount coefficient. The agent is optimized by L agent loss and its policy gradient is defined as: (13) where G is the discounted cumulative reward from the initial state to the terminal state. Finally, we take the hidden state H S,K+1 of the state s K+1 to generate the response, where the decoder is optimized by L gen loss: Joint Training Our model is finally jointly trained by optimizing L joint loss: 5 Experiments Dataset Our experiments are conducted on the widely used ESConv We implement all models with Pytorch, and all pretrained models (i.e., BlenderBot, Bart) use small versions. We set the number of steps K = 2 and reward weights w cES = w cDC = 0.1, w tES = w f DC = 1.0 (selected using a grid-search approach with two values {0.1, 1.0} for each hyperparameter). We extract M = 10 emotional reactions for each utterance. The maximum number of conversation turn M T is set to 10. The discount factor γ is 0.99, the hyperparameter α is 1e-5, and the batch size is 16. We use Adam optimizer (Kingma and Ba, 2015) with an initial learning rate of 2e-5 and a linear warmup of 120 steps for training on a GPU-V100 machine. The warm start stage is trained for 5 epochs, and the joint training stage is set to 3 epochs. The decoding settings are consistent with We adopt Perplexity (PPL), Bleu (B-n) and Distinct (D-n) to evaluate the general generation quality and diversity of the models. To measure how well the generated responses achieve goals, we define (1) ES scores containing conversation-level (cES) and turn-level (tES), i.e., r cES and r tES , measure the elicitation intensity of positive emotion involving conversation progress and the perceived intensity to the user's next turn emotion; (2) Dialogue coherence scores containing contextual (cDC) and future (fDC), i.e., r cDC and r f DC , measure the coherence with the context and the user's future utterance. The former reaches comparable or even worse results than the baselines, and the latter greedily achieves the goal of maximizing the rewards resulting in low dialogue quality. We recruited three crowdsourcing workers and exposed them to 100 negative situations randomly sampled from the test set. They were asked to engage in multi-turn conversation with the models to simulate the process of seeking ES and to choose the better one (Win) from a model pair by considering five aspects, respectively: (1) Fluency: which bot's response is more fluent and understandable? (2) Informativeness: which bot's response is more diverse and specific, and contains more information? (3) Coherence: which bot's response is more coherent with context in a multi-turn conversation? (4) Supportiveness: which bot provides more effective ES, i.e., is more likely to elicit users to change their emotions from negative to positive? (5) Overall: generally, which bot is more preferred? As in Table To analyze the quality of the experts, we show the specificity of the experts learned by SUPPORTER. As shown in Figure To further explore the adjustability of elicitation intensity of positive emotion in multi-turn conversation, we analyze the trend of positive emotion distance with the dialogue As shown in Figure We further analyze the impact of the number of iteration steps K. In Table In this paper, we introduce a new paradigm to formalize multi-turn ESC as a process of positive emotion elicitation and propose an MoE-based rein-forcement learning model SUPPORTER with welldesigned ES and dialogue coherence rewards. Extensive experiments verify the superiority of our model in providing effective ES for positive emotion elicitation while maintaining conversational goals including coherence. Our work will facilitate future work to develop ESC with positive emotion elicitation for improving the users' mental state. We discuss three limitations of this work as follows. The first one is the instability of reinforcement learning. Reward-driven policy learning is an essential advantage of this work because it is better equipped with the positive emotion-driven process of ESC than existing works and can model flexible ESC expression beyond the training data. However, this flexibility also suffers from instability, which calls for additional knowledge or strategies to refine the learning process. The second one is the need for further reference to psychological theory. An advantage of our work is to learn posterior ESC patterns integrating the dialogue context and future feedback in the form of rewards. However, there is still other valuable prior knowledge to be referred from psychology studies, e.g., the CBT (cognitive-behavioral therapy) methods. This kind of prior knowledge can be used as additional knowledge to refine the learning process as mentioned in the first limitation. The third one is that the reward design can be further optimized. The ideal case is to construct a high-quality dataset with human-feedback labels for training reward model (e.g., the constructed example of ChatGPT). At the same time, the larger parameter of the reward model, the more conducive it is to learn a robust policy and avoid it overfitting to the reward function. However, such optimizations need a trade-off with cost. In this paper, the ESConv dataset used in our experiments is a publicly-available benchmark for emotional support conversation, which does not contain sensitive and personal information as well as unethical language. Our work builds on this dataset to study positive emotion elicitation to improve the user's mental state. Therefore, we focus on constructing a dialogue system to provide emotional support from families and friends in the daily scenarios limited by this dataset rather than profes-sional psychological counseling or psychological treatment. For risky non-daily scenarios such as self-harm or suicide-related conversations, we do not claim that the dialogue system we built has a treatment or improvement effect on them. Additionally, we also ensure the anonymity of our interactive human evaluation. We believe our work meets ACL's Code of Ethics. A construction example of the bidirectional emotion keyword graph G is in Figure One-hop Reasoning on Graph G For the contextual keyword "close", its one-hop neighbor reasoned by the "forward-positive" relation is "understand", and the one reasoned by the "forwardnegative" relation is "frustrated". Further, the one-hop neighbors reasoned by the "forward" relation are the union of the one-hop neighbors of the above two relations, i.e., "understand" and "frustrated". For the keyword "frustrated" of the response, it cannot reason the one-hop neighbor using the "backward-positive" relation. Therefore, its one-hop neighbors reasoned by the "backward" relation are the same as the one-hop neighbors reasoned by the "backward-negative" relation, i.e., "close", "warning", and "pandemic". Multi-hop Reasoning on Graph G Taking the "forward → forward → backward-positive" multihop reasoning as an example, using the "forward" relationship for the contextual keywords to perform one-hop reasoning can obtain the set of neighbors that contain the keywords of the response, which we regard as the extended keyword set of the response determined by the context. Using the keywords in this set as a starting point to perform the second-hop reasoning by the "forward" relation can result in the expanded keyword set of the user's future utterance (i.e., the user's next turn utterance) determined by the response. Further, similarly, the third-hop reasoning is performed using the "backward-positive" relation to determine the extended positive keywords set of the response coherent to the future utterance. In Table Situation#1 I have been laid out for not performing. But the actual situation is the customer was satisfied my service. But the expectation of the management was not to do that. (3 rd -5 th turn) User: I was attending a customer. He was having a grievance which I sorted out. But the management did not like that. User: I do love my kids, but they are expecting some toys and stuff from Santa. Bot: Would it be possible to look at some more home made stuff for this year that will still be fun and meaningful? User: I can not craft. If I made them something it would be so bad. MoEL I am so sorry to hear that. I am sorry you are having to go through this. MIME I understand how you feel. I have been there too. I know how hard it is to the next year. I am sure there are some options out there. BlenderBot-Joint Have you considered doing crafts with your kids? I think that would be a great idea! That way you can feel like your kids are being cared for and you can really enjoy seeing them happy. Ground-Truth I have seen some people make small ornaments filled with candies and treats and other small goodies. So even if all of the gifts feel small, they can feel like they are getting a lot of them. And you could even get them to have fun decorating and painting the ornaments! and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Limited by the space. Crowdsourcing workers are from Amazon Mechanical Turk. D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank. D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank. D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank. Oh I am really sorry to hear that, did you have the same job for a long time? That sounds very difficult to deal with. Have you considered getting some more craft supplies? Maybe something that you can use to make a list of things you would like to get your kids for Christmas? SUPPORTER I know how hard it can be to make things. I know it is a terrible feeling to feel bad about yourself, but it is okay to be proud of yourself for being able to make them. Ground-Truth I know it can be really hard to feel like you can not make anything worthwhile, but I am sure there's something you could come up with that could still put a smile on your kids' faces. Context (6 th -7 th turn) User: Do you have any craft ideas? Bot: Maybe you could look at involving your kids in the process. The feeling of doing something together as a family can make it way more fun. It could be a shared experience your family can look back on fondly later on. User: That is a good idea. I would like to have it done before Christmas. How about some wooden toys or a cardboard box fort? MoEL I am so glad to hear that. I hope that you can do something fun! MIME I think you can do it! I have been through this, and you are not alone? BlenderBot-Joint That sounds like a great idea! Maybe even something for them to look forward to after Christmas. I know I love the idea of making toys and toys! Bart-Joint Yes, I think you should try to get it done before Christmas, but maybe even after that. You can see how hard it is to find a good time to do something.
| 928 | 1,694 | 928 |
Simultaneous Translation with Flexible Policy via Restricted Imitation Learning
|
Simultaneous translation is widely useful but remains one of the most difficult tasks in NLP. Previous work either uses fixed-latency policies, or train a complicated two-staged model using reinforcement learning. We propose a much simpler single model that adds a "delay" token to the target vocabulary, and design a restricted dynamic oracle to greatly simplify training. Experiments on Chinese↔English simultaneous translation show that our work leads to flexible policies that achieve better BLEU scores and lower latencies compared to both fixed and RL-learned policies.
|
Simultaneous translation, which translates sentences before they are finished, is useful in many scenarios such as international conferences, summits, and negotiations. However, it is widely considered one of the most challenging tasks in NLP, and one of the holy grails of AI Simultaneous translation is previously studied as a part of real-time speech recognition system (WRITE) using the pretrained base model. This R/W model is trained by reinforcement learning (RL) method without updating the base model. We aim to combine the merits of both efforts, that is, we design a single model end-toend trained from scratch to perform simultaneous translation, as with
|
Let x = (x 1 , . . . , x n ) be a sequence of words. For an integer 0 ≤ i ≤ n, we denote the sequence consisting of the first consecutive i -1 words in x by x <i = (x 1 , . . . , x i-1 ). We say such a sequence x <i is a prefix of the sequence x, and define s x if sequence s is a prefix of x. Conventional Machine Translation Given a sequence x from the source language, the conventional machine translation model predicts the probability distribution of the next target word y j at the j-th step, conditioned on the full source sequence x and previously generated target words y <j , that is p(y j | x, y <j ). The probability of the whole sequence y generated by the model will be To train such a model, we can maximize the probability of ground-truth target sequence conditioned on the corresponding source sequence in a parallel dataset D, which is equivalent to minimize the following loss: (1) In this work, we use Transformer Prefix-to-Prefix Framework Previous work As a simple example in this framework, To obtain a flexible and adaptive policy, we need our model to be able to take both READ and WRITE actions. Conventional translation model already has the ability to write target words, so we introduce a "delay" token ε in target vocabulary to enable our model to apply the READ action. Formally, for the target vocabulary V , we define an extended vocabulary (2) Each word in this set can be an action, which is applied with a transition function δ on a sequence pair (s, t) for a given source sequence x where s x. We assume ε cannot be applied with the sequence pair (s, t) if s = x, then we have the transition function δ as follows, where s • x represents concatenating a sequence s and a word x. Based on this transition function, our model can do simultaneous translation as follows. Given the currently available source sequence, our model continues predicting next target word until it predicts a delay token. Then it will read a new source word, and continue prediction. Since we use Transformer model, the whole available source sequence needs to be encoded again when reading in a new source word, but the predicted target sequence will not be changed. Note that the predicted delay tokens do not provide any semantic information, but may introduce some noise in attention layer during the translation process. So we propose to remove those delay token in the attention layers except for the current input one. However, this removal may reduce the explicit latency information which will affect the predictions of the model since the model cannot observe previous output delay tokens. Therefore, to provide this information explicitly, we embed the number of previous delay tokens to a vector and add this to the sum of the word embedding and position embedding as the input of the decoder. We first introduce a restricted dynamic oracle Restricted Dynamic Oracle Given a pair of full sequences (x, y) in data, the input state of our restricted dynamic oracle will be a pair of prefixes (s, t) where s x, t y and (s, t) = (x, y). The whole action set is V + defined in the last section. The objective of our dynamic oracle is to obtain the full sequence pair (x, y) and maintain a reasonably low latency. For a prefix pair (s, t), the difference of the lengths of the two prefixes can be used to measure the latency of translation. So we would like to bound this difference as a latency constraint. This idea can be illustrated in the prefix grid (see Figure By this definition, we know that this oracle can always find an action sequence to obtain (x, y). When the input state does not satisfy any latency constraint, then this dynamic oracle will provide only one action, applying which will improve the length difference. Note that this dynamic oracle is restricted in the sense that it is only defined on the prefix pair instead of any sequence pair. And since we only want to obtain the exact sequence from data, this oracle can only choose the next groundtruth target word other than ε . In many cases, the assumption |x| = |y| does not hold. To overcome this limitation, we can utilize the length ratio γ = |x|/|y| to modify the length difference: d = |s| -γ|t|, and use this new difference d in our dynamic oracle. Although we cannot obtain this ratio during testing time, we may use the averaged length ratio obtained from training data Training with Restricted Dynamic Oracle We apply imitation learning to train our translation model, using the proposed dynamic oracle as the expert policy. Recall that the prediction of our model depends on the whole generated prefix including ε (as the input contains the embedding of the number of ε ), which is also an action sequence. If an action sequence a is obtained from our oracle, then applying this sequence will result in a prefix pair, say s a and t a , of x and y. Let p(a | s a , t a ) be the probability of choosing action a given the prefix pair obtained by applying action sequence a. Then the averaged probability of choosing the oracle actions conditioned on the action sequence a will be f (a, π x,y,α,β ) = a∈π x,y,α,β (sa,ta) . To train a model to learn from the dynamic oracle, we can sample from our oracle to obtain a set, say S(x, y), of action sequences for a sentence pair (x, y). The loss function for each sampled sequence a ∈ S(x, y) will be For a parallel text D, the training loss is Directly optimizing the above loss may require too much computation resource since for each pair of (x, y), the size of S(x, y) (i.e. the number of different action sequences) can be exponentially large. To reduce the computation cost, we propose to use two special action sequences as our sample set so that our model can learn to do translation within the two latency constraints. Recall that the latency constraints of our dynamic oracle π x,y,α,β are defined by two bounds: α and β. For each bound, there is a unique action sequence, which corresponds to a path in the prefix grid, such that following it can generate the most number of prefix pairs that make this bound tight. Let a α (x,y) (a β (x,y) ) be such an action sequence for (x, y) and α (β). We replace S(x, y) with {a α (x,y) , a β (x,y) }, then the above loss for dataset D becomes . This is the loss we use in our training process. Note that there are some steps where our oracle will return two actions, so for such steps we will have a multi-label classification problem where labels are the actions from our oracle. In such cases, Sigmoid function for each action is more appropriate than the Softmax function for the actions will not compete each other We observed that the model trained on the two special action sequences occasionally violates the latency constraints and visits states outside of the designated band in prefix grid. To avoid such case, we force the model to choose actions such that it will always satisfy the latency constraints. That is, if the model reaches the aggressive bound, it must choose a target word other than ε with highest score, even if ε has higher score; if the model reaches the conservative bound, it can only choose ε at that step. We also apply a temperature constant e t to the score of ε , which can implicitly control the latency of our model without retraining it. This improves the flexibility of our trained model so that it can be used in different scenarios with different latency requirements. To investigate the empirical performance of our proposed method, we conduct experiments on NIST corpus for Chinese-English. We use NIST 06 (616 sentence pairs) as our development set and NIST 08 (691 sentence pairs) as our testing set. We apply tokenization and byte-pair encoding (BPE) We use BLEU We tried three different pairs for α and β: (1, 5), (3, 5) and (3, 7), and summarize the results on testing sets in Figure Figure We provide an example from the development set of Chinese-to-English translation in Table To analyze the effects of proposed techniques on the performance, we also provide an ablation study on those techniques for our model trained with α = 3 and β = 5 in Chineseto-English translation. The results are given in Table 4, and show that all the techniques are important to the final performance and using Sigmoid function is critical to learn adaptive policy. Table We have presented a simple model that includes a delay token in the target vocabulary such that the model can apply both READ and WRITE actions during translation process without a explicit policy model. We also designed a restricted dynamic oracle for the simultaneous translation problem and provided a local training method utilizing this dynamic oracle. The model trained with this method can learn a flexible policy for simultaneous translation and achieve better translation quality and lower latency compared to previous methods.
| 575 | 666 | 575 |
Minimum Translation Modeling with Recurrent Neural Networks
|
We introduce recurrent neural networkbased Minimum Translation Unit (MTU) models which make predictions based on an unbounded history of previous bilingual contexts. Traditional back-off n-gram models suffer under the sparse nature of MTUs which makes estimation of highorder sequence models challenging. We tackle the sparsity problem by modeling MTUs both as bags-of-words and as a sequence of individual source and target words. Our best results improve the output of a phrase-based statistical machine translation system trained on WMT 2012 French-English data by up to 1.5 BLEU, and we outperform the traditional n-gram based MTU approach by up to 0.8 BLEU.
|
Classical phrase-based translation models rely heavily on the language model and the reordering model to capture dependencies between phrases. Sequence models over Minimum Translation Units (MTUs) have been shown to complement both syntax-based However, the estimation of higher-order n-gram models becomes increasingly difficult due to data sparsity issues associated with large n-grams, even when training on over one hundred billion words Recently, several feed-forward neural networkbased models have achieved impressive improvements over traditional back-off n-gram models in language modeling However, feed-forward networks do not directly address the limited context issue either, since predictions are based on a fixed-size context, similar to back-off n-gram models. We therefore focus in this paper on recurrent neural network architectures, which address the limited context issue by basing predictions on an unbounded history of previous events which allows to capture long-span dependencies. Recurrent architectures have recently advanced the state of the art in language modeling Specifically, we introduce two recurrent neural network-based MTU models to address the is- sues regarding data sparsity and limited context sizes by leveraging continuous representations and the unbounded history of the recurrent architecture. Our first approach frames the problem as a sequence modeling task over minimal units ( §3). The second model improves over the first by modeling an MTU as a bag-of-words, thereby allowing us to learn representations over sub-structures of minimal units that are shared across MTUs ( §4). Our models significantly outperform the traditional back-off n-gram based approach and we show that they act complementary to a very strong recurrent neural network-based language model based solely on target words ( §5). 2 Minimum Translation Units We can straightforwardly estimate an n-gram model over MTUs to estimate the probability of a sentence pair using standard back-off techniques commonly employed in language modeling. For example, a trigram model in target left-to-right order factors the sentence pair in Figure If we would like to model larger contexts, then we quickly run into data sparsity issues. To illustrate this point, consider the parameter growth of an n-gram model which is driven by the vocabulary size |V | and the n-gram order n: O(|V | n ). Clearly, the exact estimation of higher-order n-gram probabilities becomes more difficult with large n, leading to the estimation of events with increasingly sparse statistics, or having to rely on statistics from lower-order events with backoff models, which is less desirable. Even wordbased language models rarely ventured so far much beyond 5-gram statistics as demonstrated by Neural network-based sequence models tackle the data sparsity problem by learning continuous word representations, that group similar words together in continuous space. For example, the distributional representations induced by recurrent neural networks have been found to have interesting syntactic and semantic regularities
|
The first model we introduce is based on the recurrent neural network language model of The model is factored into an input layer, a hidden layer with recurrent connections, and an output layer (Figure The state of the hidden layer is determined by the input layer and the hidden layer configuration of the previous time step h t-1 . The weights of the connections between the layers are summarized in a number of matrices: U represents weights from the input layer to the hidden layer, and W represents connections from the previous hidden layer to the current hidden layer. Matrix V contains weights between the current hidden layer and the output layer. The hidden and output layers are computed via a series of matrix-vector products and nonlinearities: where k exp {z k } are sigmoid and softmax functions, respectively. Additionally, the network is interpolated with a maximum entropy model of sparse n-gram features over input MTUs Naïve computation of the probability distribution over the next MTU is very expensive for large vocabularies, such as commonly encountered for MTU models (Table minimal unit conditioned on the class: This requires the number of classes and MTUs to be evenly balanced, i.e., each class contains exactly as many minimal units as there are classes. Figure The previous model treats MTUs as atomic symbols which leads to large vocabularies requiring large parameter sets and expensive inference. However, similar MTUs may share the same words, or words which are related in continuous space. The atomic MTU model does not exploit this since it cannot access the internal structure of a minimal unit. The approach we pursue next is to break MTUs into individual source and target words The structure of the bag-of-words MTU RNN models is shown in Figure However, there are a number of computational issues with this model: First, we cannot efficiently factor the word layer w t into classes such as for the atomic MTU RNN model because we require all its activations to compute the MTU output layer y t . This reduces the best case complexity of computing the word layer from O( |V |) back to linear in the number of source and target words |V |. In practice this results in between 200-1000 more activations that need to be computed, depending on the word vocabulary size. Second, turning the MTU output layer into a convolutional layer is not enough to sufficiently reduce the computational effort to compute the output activations since the number of connections between the word and MTU layers is very imbalanced. This is because frequent words, such as function words, are part of many MTUs and therefore have a very high out-degree, e.g., the neuron representing "the" has over 82K outgoing edges. On the other hand, infrequent words, have a very low out-degree. This imbalance makes it hard to efficiently compute activations and error gradients, even on a GPU, since some neurons require substantially more work than others. The simplified model computes the probability of the next MTU m t+1 as a product of individual word probabilities: where we predict a sequence of source and target words a 1 , . . . , a u ∈ m t+1 with a class-structured output layer, similar to the atomic model ( §3). Training still uses a cross entropy criterion and back propagation through time, however, error vectors are computed on a per-word basis, instead of a per-MTU basis. Direct connections between the input and output layers are based on source and target words which is less sparse than basing direct features on entire MTUs such as for the original bag-of-words model. Overall, the simplified model retains the bag-ofwords input representation of the original model, while permitting the efficient factorization of the word-output layer into classes. We evaluate the effectiveness of both the atomic MTU RNN model ( §3) and the simplified bag-ofwords MTU RNN model ( §4) in an n-best rescoring setting, comparing against a trigram back-off MTU model as well as the phrasal decoder 1-best output which we denote as the baseline. Baselines. We experiment with an in-house phrase-based system similar to Moses The 1-best output by the phrase-based decoder is the baseline accuracy. As a second baseline we experiment with a trigram back-off MTU model trained on all extracted MTUs, denoted as n-gram MTU. The trigram MTU model is estimated with the same modified Kneser-Ney framework as the target side language model. All MTU models are trained in target left-to-right MTU order which performed well in initial experiments. Evaluation. We test our approach on two different data sets. First, we train a German to English system based on the data of the WMT 2006 shared task Second, we experiment with a French to English system based on 102M words of training data from the WMT 2012 campaign. The majority of the training data set is parliamentary proceedings except for about 5m words which are newswire; all MTU models are trained on the newswire subset since we found similar accuracy to using all data in initial experiments. We evaluate on four newswire domain test sets from 2008, 2010 and 2011 as well as the 2010 system combination test set containing between 2034 to 3003 sentences. Log-linear weights are estimated on the 2009 data set com-prising 2525 sentences. We evaluate all systems in a single reference BLEU setting. Rescoring Setup. We rescore the 1000-best output of the baseline phrase-based decoder by either the trigram back-off MTU model or the RNN models. The baseline accuracy is obtained by choosing the 1-best decoder output. We reestimate the log-linear weights for rescoring by running a further iteration of MERT with the additional feature values; we initialize the rescoring feature weight to zero and try 20 random restarts. At test time we use the new set of log-linear weights to rescore the test set n-best list. Neural Network Setup. We trained the recurrent neural network models on between 88% and 93% of each data set and used the remainder as validation data. The vocabulary of the atomic MTU RNN model is comprised of all MTU types which were observed more than once in the training data. We first report the decoder 1-best output as the first baseline and then rescore our two data sets (Table How do our models compare to other neural network approaches that rely only on target side information? To answer this question we compare to the strong language model of Mikolov (2012; RNNLM) which has recently improved the stateof-the-art in language modeling perplexity. The results (Table Can our models act complementary to a strong RNN language model? Our Our approach of modeling Minimum Translation Units is very much in line with recent work on ngram-based translation models (Crego and Yvon, 2010), and more recently, continuous space-based translation models Other work on applying recurrent neural networks to machine translation Minimum Translation Unit models based on recurrent neural networks lead to substantial gains over their classical n-gram back-off models. We introduced two models of which the best improves accuracy by up to 1.5 BLEU over the 1-best decoder output, and by 0.8 BLEU over a trigram MTU model in an n-best rescoring setting. Our experiments have shown that representing MTUs as bags-of-words leads to better accuracy since this exploits similarities in the internal structure of Minimum Translation Units, which is not possible when modeling them as atomic symbols. We have also shown that our models are complementary to a very strong RNN language model In future work, we would like to make the initial version of the bag-of-words model computationally more tractable using a better GPU implementation. This model combines the efficient bag-ofwords input representation with the ability to predict MTUs as single units while explicitly modeling the constituent words in an intermediate layer.
| 662 | 3,108 | 662 |
Exploiting Position and Contextual Word Embeddings for Keyphrase Extraction from Scientific Papers
|
Keyphrases associated with research papers provide an effective way to find useful information in the large and growing scholarly digital collections. In this paper, we present KPRank, an unsupervised graph-based algorithm for keyphrase extraction that exploits both positional information and contextual word embeddings into a biased PageRank. Our experimental results on five benchmark datasets show that KPRank that uses contextual word embeddings with additional position signal outperforms previous approaches and strong baselines for this task.
|
Keyphrase extraction is the task of automatically extracting a small set of descriptive words or phrases that can accurately summarize the topics discussed in a document Most of the previous approaches to keyphrase extraction are either supervised or unsupervised. While supervised approaches perform generally better Inspired by the Transformer models
|
In this section, we describe our unsupervised graphbased algorithm called KPRank, that exploits both position information of the words in a document along with contextual word embeddings for computing a biased PageRank score for each candidate word. Our approach consists of three steps: (1) candidate word selection and word graph construction; (2) word scoring by biased PageRank; and (3) candidate phrase formation. For a target doucment D, we first apply a partof-speech filter Preliminaries. PageRank where O(n j ) = n k ∈Adj(n j ) w jk and Adj(n j ) is the set of all adjacent nodes of node n j ∈ N . p i is defined below. In order to prevent the PageRank from getting stuck in cycles or dead ends, a dumping factor α was added to Eq. ( be the probability distribution of randomly jumping to any node in the graph. For an unbiased PageRank, this is a uniform distribution, with p i = 1 |N | , for all i from 1 to |N |. For a biased PageRank, this probability distribution is not uniform, but rather the nodes in the graph are visited preferentially, with some nodes being visited more often than others, depending on the p i value for node n i (Haveliwala, 2003). Key2Vec is an example of (topic) biased PageRank for keyphrase extraction that computes p i for node n i using the cosine similarity between the embedding of word/phrase corresponding to node n i and a theme vector for the entire document, which corresponds to the aggregated word/phrase embeddings from the document's title KPRank. In our proposed approach, we calculate p i for node n i using two types of scores: theme (or topic) score and positional score. We multiply both scores to assign a final weight to node n i before running the biased PageRank algorithm. Both scores and their calculation are explained below. To calculate the theme score (ts i ) for node n i ∈ N , we first calculate a theme vector (T D ) for document D. A theme vector is obtained by averaging SciBERT 10 , whereas for a word occurring on position 100, its ps i score is 1 100 . The intuition behind this weighting scheme is to give higher weight to words appearing in the beginning of a document since in scientific writing, authors tend to use keyphrases very early in the document (even from the title) The biased PageRank scores for each node n i are finally calculated by iteratively applying Eq. ( In our experiments, the PageRank scores are updated until the difference between two consecutive iterations is ≤ 0.001 or for 100 iterations. Table phrases. We consider phrases with the regular expression (adjective)*(noun)+, of length up to four words, to generate candidate phrases. We used stemmed version of each word using Porter stemmer. We use POS tagger from Python's NLTK toolkit. The score for each candidate phrase is calculated by summing up the scores of its individual words For evaluation, we use five datasets, which we describe below. We use the combination of controlled (author assigned) and uncontrolled (reader assigned) keyphrases as gold-standard phrases. We used uncontrolled keyphrases when available. Table Inspec Krapivin NUS (Nguyen and ACM contains 30,000 papers published in ACM conferences with a train and test split consisting of 10, 000 and 20, 000 papers, respectively. For each dataset we use its test set for evaluation. Evaluation metrics. To evaluate the performance of different methods, we use micro avg. F1-score. We report the performance for the top 5 and 10 candidate phrases returned by different methods as in The effect of position, contextual embeddings, and the comparison with previous works. To see the effect of positional information, we compare the performance of KPRank that uses contextual SciBERT (SB) embeddings along with positional information (denoted as KPRank(SB)) with that of its counterpart that does not use positional information (denoted as KPRank(SB-POS)). Moreover, to see the effect of contextual embeddings, we compare the performance of SciBERT-based KPRank (KPRank(SB)) with that of KPRank that uses FastText non-contextual word embeddings Tables 2 shows these comparisons on SemEval, Inspec, Krapivin, NUS, and ACM. It can be seen from the table that adding position information shows much higher improvement in the performance of KPRank, i.e. KPRank(SB) substantially outperforms KPRank(SB-POS). Moreover, KPRank(SB) outperforms KPRank(FastText) on all the datasets except for Krapivin. Importantly, KPRank(SB) outperforms most baseline methods, including Key2Vec (by a large margin) e.g., on Se-mEval, KPRank(SB) achieves an F1@5 of 22.51% as compared with 17.54% achieved by Key2Vec. We can also notice from Table Figure Drawing from an information processing perspective, this paper examines how information technology (IT) has been a catalyst in the development of new forms of organizational structures. [...] to the present environmental instability that now characterizes many industries. Specifically, the authors suggest that advances in IT have enabled managers to adapt existing forms and create new models for organizational design that better fit requirements of an unstable environment. Anecdotal example. To see the quality of predicted phrases by the KPRank(SB), we randomly selected a paper from the Inspec dataset and evaluated the KPRank(SB) on it. We manually inspected the top-10 predictions by the KPRank(SB) and contrasted them with the gold-standard keyphrases. The title, abstract, gold-standard keyphrases and top-10 predicted keyphrases for this paper are shown in Figure We can also see that KPRank(SB) did not predict gold-standard phrase "environmental instabily." A closer inspection of the document and both types of scores (theme score and positional score) assigned by KPRank(SB) to both constituent words of the gold-standard phrase that was not ranked in top-10 predictions revealed that these constituent words have lower values of theme score and they both appear only once in the document. Hence, the Pagerank algorithm will not boost these words. Inspecting other errors, we found that KPRank can fail to predict phrases that contain words that are less frequent in the document and their word embeddings are far from the theme vector. In this paper, we proposed a novel unsupervised graph-based algorithm, named KPRank, which incorporates both positional appearances of the words along with contextual word embeddings for computing a biased PageRank score for each candidate word. Our experimental results on five datasets show that incorporating position information into our biased KPRank model yields better performance compared with a KPRank that does not use the position information, and SciBERTbased KPRank usually outperforms FastText-based KPRank on this task. Moreover, KPRank outperforms strong baseline methods. In the future, it would be interesting to explore KPRank on other domains, such as Biology, and Social Science.
| 550 | 352 | 550 |
kNN-BOX: A Unified Framework for Nearest Neighbor Generation
|
Augmenting the base neural model with a token-level symbolic datastore is a novel generation paradigm and has achieved promising results in machine translation (MT). In this paper, we introduce a unified framework kNN-BOX, which enables quick development and visualization for this novel paradigm. kNN-BOX decomposes the datastore-augmentation approach into three modules: datastore, retriever and combiner, thus putting diverse kNN generation methods into a unified way. Currently, kNN-BOX has provided implementation of seven popular kNN-MT variants, covering research from performance enhancement to efficiency optimization. It is easy for users to reproduce these existing work or customize their own models. Besides, users can interact with their kNN generation systems with kNN-BOX to better understand the underlying inference process in a visualized way. In experiment section, we apply kNN-BOX for machine translation and three other seq2seq generation tasks (text simplification, paraphrase generation and question generation). Experiment results show that augmenting the base neural model with kNN-BOX can bring large performance improvement in all these tasks. The code and document of kNN-
|
Equipping the base neural model with a symbolic datastore is a novel paradigm for enhancing generation quality. However, we notice that these kNN generation methods are implemented with diverse codebases, e.g., Fairseq In this paper, we introduce a unified framework kNN-BOX for nearest neighbor generation, which supports quick development and visualization anal-ysis. Our framework decomposes the datastoreaugmentation approach into three modules: DATA-STORE, RETRIEVER and COMBINER, thus putting diverse kNN generation methods into a unified way (Figure Another useful feature of kNN-BOX is supporting visualized interactive analysis. Via our provided web service, users can interact with their kNN model and observe its inference process, e.g. the content and distribution of its retrieval results (Figure Experiment results on machine translation datasets show that kNN-BOX is a reliable platform for model reproduction and development. In addition, we apply kNN-BOX for three other seq2seq tasks, i.e., text simplification, paraphrase generation and question generation. Experiment results show that augmenting the base neural model with kNN-BOX is also beneficial in these tasks, showing the great potential of nearest neighbor generation and the wide usage of our kNN-BOX toolkit. At the time of writing, we are happy to see that kNN-BOX has been used as the backbone of this year's ACL paper 2 Background: kNN-MT Before introducing kNN-BOX, we recap kNN-MT approach in this section. Generally, kNN-MT framework aims at memorizing translation knowledge in parallel corpus C into a datastore D and use it to augment the NMT model M during inference. Memorizing Knowledge into Datastore To extract translation knowledge, translation pair (X , Y) is fed into M for teacher-forcing decoding. At time step t, the continuous representation of the translation context (X , Y <t ), i.e. the hidden state h t from the last decoder layer, is taken as key: and the target token y t is taken as value. Each key-value pair explicitly memorizes the translation knowledge: generating the value token at the decoder hidden state key. With a single forward pass over the entire corpus, the full datastore D can be constructed: Generating with Memorized Knowledge The constructed datastore is then combined with the base NMT model as an augmentation memory. During inference, the NMT model retrieves related knowledge from the datastore to adjust its own translation prediction. Specifically, the NMT model uses the contextualized representation of the test translation context (X , Y <t ) to query the datastore for nearest neighbor representations and the corresponding target tokens N k = {(h j , y j )} k j=1 . The retrieved entries are then converted to a distribution over the vocabulary: (2) where s measures the similarity between h t and h j : Here, d denotes L 2 -square distance and T is the temperature. In the end, the output distribution of the NMT model and symbolic datastore are interpolated with the weight λ: Recent Advances in kNN-MT To make kNN-MT more effective, efficient and explainable, various methods have been devised. (2023a) explore the interpretability issue in kNN-MT and builds a light and more explainable datastore according to the capability of the NMT model. 3 Unified Framework: kNN-BOX This section describes how we design and implement kNN-BOX, and introduce how users run kNN-BOX for developing kNN generation models and interacting with the deployed model visually.
|
We develop kNN-BOX based on the widely-used generation framework Fairseq, making it easy to apply kNN-BOX for other generation tasks. The overall workflow of kNN-BOX is illustrated in Figure • DATASTORE: saving generation knowledge as key-values pairs (Equation • RETRIEVER: retrieving nearest neighbors from the datastore during inference. • COMBINER: converting retrieval results to a distribution (Equation With this design, diverse kNN models can be implemented in a unified way. For a specific kNN variant, it usually makes a modification on one of the three modules, compared to vanilla kNN generation model. Therefore, users can customize the corresponding module and quickly develop a kNN generation model. Supporting visual interactive analysis is another useful feature of kNN-MT. By saving intermediate computation results, we enable kNN-BOX to visualize the inference process. We hope this feature will help users to better understand their own model. Developing New Models kNN-BOX is designed not only for reproducing existing work, but also for developing new models on new tasks. For each module, users can pick one of its implementation from kNN-BOX or customize their own version, and combine three modules together to build a new kNN generation model. In this process, only few lines of codes needs to be added, which can save users a lot of time. More importantly, this implementation fashion enables users to easily build a fused model, e.g., combining the most explainable datastore (PLACDATSTORE) with the strongest combiner (ROBUSTCOMBINER). To perform generation tasks other than machine translation, users only need to switch the training corpus to build a task-specific datastore. Visualizing Generalization Process By running our provided script to launch a web page (shown in Figure input window and tune generation hyperparameters in the upper-left panel. The generated results, both detokenized and tokenized, will then be displayed. Taking kNN-MT as an example, after clicking a word in the translation, users can see the translation probability given by both NMT model and kNN-MT model. Moreover, detailed information of the retrieved datastore entries will be displayed in the bottom panel. By selecting on a certain nearest neighbor point, users can see the corresponding value token, translation context and query-key distance. Overall, the visualization page can help user to interact with their kNN generation model and explore its inner working process. To evaluate the effectiveness of kNN-BOX, we conduct experiments on machine translation and three other seq2seq tasks. Dataset For machine translation, we adopt four German-English OPUS datasets Base Neural Model On OPUS dataset, we follow previous kNN-MT work and use the winner model of WMT'19 De-En news translation task We use BLEU score calculated by sacrebleu kNN-BOX can help user to quickly augment the base NMT model with kNN methods. By running our provided shell scripts, users can quickly reproduce existing kNN-MT models. Table In experiment section, we show that kNN-BOX can not only be applied for enhancing neural machine translation model, but also for enhancing neural generation model in other seq2seq tasks. In the future, we will keep update this toolkit to provide implementation of more retrieve-andgenerate methods and optimize the framework to make it more user-friendly, and explore the possibility to apply kNN-BOX for more generation tasks. We discuss two potential limitations of our kNN-BOX toolkit below: • Inference Latency: The nearest neighbor retrieval system queries the datastore at each timestep, which introduces inference latency. • Datastore reusability: The datastore is constructed using a specific model, which limits its reusability. This means that the datastore cannot be seamlessly integrated or utilized with other models.
| 1,202 | 3,497 | 1,202 |
Automatic Acquisition of Two-Level Morphological Rules
|
We describe and experimentally evaluate a complete method for the automatic acquisition of two-level rules for morphological analyzers/generators. The input to the system is sets of source-target word pairs, where the target is an inflected form of the source. There are two phases in the acquisition process: (1) segmentation of the target into morphemes and (2) determination of the optimal two-level rule set with minimal discerning contexts. In phase one, a minimal acyclic finite state automaton (AFSA) is constructed from string edit sequences of the input pairs. Segmentaiion of the words into morphemes is achieved through viewing the AFSA as a directed acyclic graph (DAG) and applying heuristics using properties of the DAG as well as the elementary edit operations. For phase two, the determination of the optimal rule set is made possible with a novel representation of rule contexts, with morpheme boundaries added, in a new DAG. We introduce the notion of a delimiter edge. Delimiter edges are used to select the correct twolevel rule type as well as to extract minimal discerning rule contexts from the DAG. Results are presented for English adjectives, Xhosa noun locatives and Afrikaans noun plurals.
|
Computational systems based on the two-level model of morphology Up to now, these two components had to be coded largely by hand, since no automated method existed to acquire a set of two-level rules for input sourcetarget word pairs. To hand-code a 100% correct rule set from word pairs becomes almost impossible when a few hundred pairs are involved. Furthermore, there is no guarantee that such a hand coded lexicon does not contain redundant rules or rules with too large contexts. The usual approach is rather to construct general rules from small subsets of the input pairs. However, these general rules usually allow overrecognition and overgeneration -even on the subsets from which they were inferred. Simons The contribution of this paper is to present a complete method for the automatic acquisition of an op-timal set of two-level rules (i.e. the second component above) for source-target word pairs. It is assumed that the target word is formed from the source through the addition of a prefix and/or a suffix 1. Furthermore, we show how a partial acquisition of the morphotactic description (component one) results as a by-product of the rule-acquisition process. For example, the morphotactic description of the target word in the input pair The right-hand side of this morphotactic description is then mapped on the left-hand side,
|
happy+er happi 0 er For this example the two-level rule y:i ¢~ p:pcan be derived. These processes are described in detail in the rest of the paper: Section 2 provides an overview of the two-level rule formalism, Section 3 describes the acquisition of morphotactics through segmentation and Section 4 presents the method for computing the optimal two-level rules. Section 5 evaluates the experimental results and Section 6 summarizes. Two-level rules view a word as having a lezical and a surface representation, with a correspondence between them Lexical: h appy + e r Surface: h appi 0 e r Each pair of lexical and surface characters is called a feasible pair. A feasible pair can be written as lezicabcharac~er:surface-charac~er. Such a pair is called a default pair when the lexicai character and surface character are identical (e.g. h:h). When the lexical and surface character differ, it is called a special pair (e.g. y:i). The null character (0) may appear as either a lexical character (as in +:0) or a surface character, but not as both. 1Non-linear operations (such as infixation) are not considered here, since the basic two-level model deals with it in a round-about way. We can note that extensions to the basic two-level model have been proposed to handle non-linear morphology Two-level rules have the following syntax The morphotactics of the input words are acquired by (1) computing the string edit difference between each source-target pair and (2) merging the edit sequences as a minimal acyclic finite state automaton. The automaton, viewed as a DAG, is used to segment the target word into its constituent morphemes. A string edit sequence is a sequence of elementary operations which change a source string into a target string In general, the prefix-root boundary is just the reverse of the root-suffix boundary, i.e. INSERT-DELETE-NOCHANGE, with the DELETE operation being optional. The heuristic resulting from this observation is a bias giving highest precedence to INSERT operations, followed by DELETE and NOCHANGE, in the first half of the edit sequence. In the second half, the precedence is reversed. A single source-target edit sequence may contain spurious INSERTs which are not considered to form part of a morpheme. For example, the O:i insertion in Example 7 should not contribute to the suffixer to form -ier, since -ier is an allomorph of -er. To combat these spurious INSERTs, all the edit sequences for a set of source-target words are merged as follows: A minimal acyclic finite state automaton (AFSA) is constructed which accepts all and only the edit sequences as input strings. This AFSA is then viewed as a DAG, with the elementary edit operations as edge labels. For each edge a count is kept of the number of different edit sequences which pass through it. A path segment in the DAG consisting of one or more INSERT operations having a similar count, is then considered to be associated with a morpheme in the target word. The O:e O:r INSERT sequence associated with the -er suffix appears more times than the O:i O:e O:r INSERT sequence associated with the -ier suffix, even in a small set of adjectively-related source-target pairs. This means that there is a rise in the edge counts from O:i to O:e (indicating a root-suffix boundary), while O:e and O:r have similar frequency counts. For prefixes a fall in the edge frequency count of an INSERT sequence indicates a prefix-root boundary. To extract the morphemes of each target word, every path through the DAG is followed and only the target-side of the elementary operations serving as edge labels, are written out. The null characters (0) on the target-side of DELETEs are ignored while the target-side of Phase one can segment only one layer of affix additions at a time. However, once the morpheme boundary markers (+) have been inserted, phase two should be able to acquire the correct two-level rules for an arbitrary number of affix additions: prefizl +prefiz2+. . . +roo~+suffizl +suffiz2+ .... To acquire the optimal rules, we first determine the full length lexical-sufface representation of each word pair. This representation is required for writing two-level rules (Section 2). The morphotactic descriptions from the previous section provide sourcetarget input pairs from which new string edit sequences are computed: The right-hand side of the morphotactic description is used as the source and the left-hand side as the target string. For instance, Example 8 is written as: To formulate a two-level rule for the source-target pair happy-unhappier, we need a correspondence pair (CP) and a rule type (op), as well as a left context (LC) and a right context (RC) (see Section 2). Rules need only be coded for special pairs, i.e. IN-SERTs, DELETEs or REPLACEs. The only special pair in the above example is y:i, which will be the CP of the rule. Now the question arises as to how large the context of this rule must be? It should be large enough to uniquely specify the positions in the lexical-surface input stream where the rule is applied. On the other hand, the context should not be too large, resulting in an overspecified context which prohibits the application of the rule to unseen, but similar, words. Thus to make a rule as general as possible, its context (LC and RC) should be as short as possible s . By inspecting the edit sequence in Example 10, we see that y changes into i when y is preceded by a p:p, which serves as our first attempt at a (left) context for y:i. Two questions must be asked to determine the correct rule type to be used y:i ¢~ p:p _ 2Our assumption is that such a minimal cost mapping will lead to an optimal rule set. In most (if not all) of the examples seen, a minimal mapping was also intuitively acceptable. sit abstractions (e.g. sets such as V denoting vowels) over the regular pairs are introduced, it will not be so simple to determine what is "a more general context". However, current implementations require abstractions to be explicitly instantiated during the compilation process ( However, this soon becomes a tedious and error prone task when the number of source-target pairs increases, due to the complex interplay of rules and their contexts. It is important to acquire the minimal discerning context for each rule. This ensures that the rules are as general as possible (to work on unseen words as well) and prevents rule conflicts. Recall that one need only code rules for the special pairs. Thus it is necessary to determine a rule type with associated minimal discerning context for each occurrence of a special pair in the final edit sequences. This is done by comparing all the possible contiguous 4 contexts of a special pair against all the possible contexts of all the other feasible pairs. To enable the computational comparison of the growing left and right contexts around a feasible pair, we developed a "mixed-context" representation. We call the particular feasible pair for which a mixed-context is to be constructed, a marker pair (MP), to distinguish it from the feasible pairs in its context. The mixedcontext representation is created by writing the first feasible pair to the left of the marker pair, then the first right-context pair, then the second left-context pair and so forth: [ 13] LC1, RC1, LC2, RC2, LC3, RC3, ..., MP The marker pair at the end serves as a label. Special symbols indicate the start (SOS) and end (EOS) of an edit sequence. If, say, the right-context ofa MP is shorter than the left-context, an out-of-bounds symbol (OOB) is used to maintain the mixed-context format. For example the mixed-context of y:i in the edit sequence in Example 10, is represented as: [ 14] p:p, +:0, p:p, e:e, a:a, r:r, h:h, EOS, +:0, OOB, n:n, OOB, u:u, SOS, OOB, y:i The common prefixes of the mixed-contexts are merged by constructing a minimal AFSA which accepts all and only these mixed-context sequences. 4A two-level rule requires a contiguous context. The transitions (or edges, when viewed as a DAG) of the AFSA are labeled with the feasible pairs and special symbols in the mixed-context sequence. There is only one final state for this minimal AFSA. Note that all and only the terminal edges leading to this final state will be labeled with the marker pairs, since they appear at the end of the mixed-context sequences. More than one terminal edge may be labeled with the same marker pair. All the possible (mixed) contexts of a specific marker pair can be recovered by following every path from the root to the terminal edges labeled with that marker pair. If a path is traversed only up to an intermediate edge, a shortened context surrounding the marker pair can be extracted. We will call such an intermediate edge a delimiter edge, since it delimits a shortened context. For example, traversing the mixed context path of y:i in Example 14 up to e:e would result in the (unmixed) shortened context: [ 25] p:p p:p _ +:0 e:e From the shortened context we can write a two-level rule y:i op p:p p:p _ ÷:0 e:e which is more general than a rule using the full context: [ 27] y:i op SOS u:u n:n h:h a:a p:p p:p _ +:0 e:e r:r For each marker pair in the DAG which is also a special pair, we want to find those delimiter edges which produce the shortest contexts providing a true answer to at least one of the two rule type decision questions given above. The mixed-context prefix-merged AFSA, viewed as a DAG, allow us to rephrase the two questions in order to find answers in a procedural way: Question 1 Traverse all the paths from the root to the terminal edges labeled with the marker pair L:S. Is there an edge el in the DAG which all these paths have in common? If so, then question one is true for the environment E constructed from the shortened mixed-contexts associated with the path prefixes delimited by el. Consider the terminal edges which same L-component as the marker pair L:S and which are reachable from a common edge e2 in the DAG. Do all of these terminal edges also have the same S-component as the marker pair? If so, then question two is true for the environment E constructed from the shortened mixed-contexts associated with the path prefixes delimited by e2. For each marker pair, we traverse the DAG and mark the delimiter edges nearest to the root which allow a true answer to either question one, question two or both (i.e. el = e2). This means that each path from the root to a terminal edge can have at most three marked delimiter edges: One delimiting a context for a ~ rule, one delimiting a context for a rule and one delimiting a context for a ~ rule. The marker pair used to answer the two questions, serves as the correspondence part (Section 2) of the rule. To continue with Example 14, let us assume that the DAG edge labeled with e:e is the closest edge to the root which answers true only to question one. Then the ~ rule is indicated: y:i ¢* a:a p:p p:p _ ÷:0 e:e r:r The reasons for this preference are that the ¢~ rule • provides a more precise statement about the applicable environment of the rule and it • seems to be preferred in systems designed by linguistic experts. Furthermore, from inspecting examples, a delimiter edge indicating a ~ rule generally delimits the shortest contexts, followed by the delimiter for ¢~ and the delimiter for ~. The shorter the selected context, the more generally applicable is the rule. We therefore select only one rule per path, in the following preference order: (1) ¢~, (2) ~ and (3) ~. Note that any of the six possible precedence orders would provide an accurate analysis and generation of the pairs used for learning. However, our suggested precedence seems to strike the best balance between over-or underrecognition and over-or undergeneration when the rules would be applied to unseen pairs. The mixed-context representation has one obvious drawback: If an optimal rule has only a left or only a right context, it cannot be acquired. To solve this problem, two additional minimal AFSAs are constructed: One containing only the left context information for all the marker pairs and one containing only the right context information. The same process is then followed as with the mixed contexts. The final set of rules is selected from the output of all three the AFSAs: For each special pair 1. we select any of the ¢~ rules with the shortest contexts of which the special pair is the lefthand side, or 2. if no ¢~ rules were found, we select the shortest and ~ rules for each occurrence of the special pair. They are then merged into a single ¢~ rule with disjuneted contexts. The rule set learned is complete since all possible combinations of marker pairs, rule types and contexts are considered by traversing all three DAGs. Furthermore, the rules in the set have the shortest possible contexts, since, for a given DAG, there is only one delimiter edge closest to the root for each path, marker pair and rule type combination. Our process works correctly for examples given in Note that this set contains ambiguity: The locative of ingubo is either engubeni or engutyeni. Our process must learn the necessary two-level rules to map ingubo to engubeni and engutyeni, as well as to map both engubeni and engutyeni in the other direction, i.e. to ingubo. Similarly, indlu and ikhaya each have two different locative forms. Furthermore, the two source words inkosi and iinkosi (the plural of inkosi) differ only by a prefixed i, but they have different locative forms. This small difference between source words provides an indication of the sensitivity required of the acquisition process to provide the necessary discerning information to a two-level morphological processor. At the same time, our process needs to cope with possibly radical modifications between source and target words. Consider the mapping between ubuchopho and its locative ebucotsheni. Here, the only segments which stay the same from the source to the target word, are the three letters -buc-, the letter -o-(the deletion of the first -h-is correct) and the second -h-. The target words are correctly segmented during phase one as: level rule compiler KGEN (developed by Nathan Miles) was used to compile the acquired rules into the state tables required by PC-KIMMO. Both PC-KIMMO and KGEN are available from the Summer Institute of Linguistics. [ 23] The ~ and ~ rules of a special pair can be merged into a single ~=~ rule. For example the four rules above for the special pair q-:O can be merged into In this way the 24 rules are reduced to a set of 16 rules which contain only a single ¢~ rule for each special pair. This merged set of 16 two-level rules analyze and generate the input word pairs 100% correctly. The next step was to show the feasibility of automatically acquiring a minimal rule set for a wide coverage parser. To get hundreds or even thousands of input pairs, we implemented routines to extract the lemmas ("head words") and their inflected forms from a machine-readable dictionary. In this way we extracted 3935 Afrikaans noun-plural pairs which served as the input to our process. Afrikaans plurals are almost always derived with the addition of a suffix (mostly -e or -s) to the singular form. Different sound changes may occur during this process. For example 6, gemination, which indicates the shortening of a preceding vowel, occurs frequently (e.g. hat ---* katte), as well as consonant-insertion (e.g. has ---* haste) and elision (ampseed --~ ampsede). Several sound changes may occur in the same word. For example, elision, consonant replacement and gemination occurs in loof ---* lowwe. Afrikaans (a Germanic language) has borrowed a few words from Latin. Some of these words have two plural forms, which introduces ambiguity in the word mappings: One plural is formed with a Latin suffix (-a) (e.g. emetikum --~ emetika) and one with an indigenous suffix (-s) (emetih.m emetih ms). Allomorphs occur as well, for example -ens is an allomorph of the suffix -s in bed + s ---, beddens. During phase one, all but eleven (0.3%) of the 3935 input word pairs were segmented correctly. To facilitate the evaluation of phase two, we define a simple rule as a rule which has an environment consisting of a single context. This is in contrast with an environment consisting of two or more contexts disjuncted together. Phase two acquired 531 simple rules for 44 special pairs. Of these 531 simple rules, 500 are ~ rules, nineteen are ¢~ rules and twelve are ~ rules. The average length of the simple rule contexts is 4.2 feasible pairs. Compare this with the nAil the examples comes from the 3935 input word pairs.
| 1,217 | 1,347 | 1,217 |
Compressing and Debiasing Vision-Language Pre-Trained Models for Visual Question Answering
|
Despite the excellent performance of visionlanguage pre-trained models (VLPs) on conventional VQA task, they still suffer from two problems: First, VLPs tend to rely on language biases in datasets and fail to generalize to outof-distribution (OOD) data. Second, they are inefficient in terms of memory footprint and computation. Although promising progress has been made in both problems, most existing works tackle them independently. To facilitate the application of VLP to VQA tasks, it is imperative to jointly study VLP compression and OOD robustness, which, however, has not yet been explored. This paper investigates whether a VLP can be compressed and debiased simultaneously by searching sparse and robust subnetworks. To this end, we systematically study the design of a training and compression pipeline to search the subnetworks, as well as the assignment of sparsity to different modality-specific modules. Our experiments involve 3 VLPs, 2 compression methods, 4 training methods, 2 datasets and a range of sparsity levels. Our results show that there indeed exist sparse and robust subnetworks, which are competitive with the debiased full VLP and clearly outperform the debiasing SoTAs with fewer parameters on OOD datasets VQA-CP v2 and VQA-VS. 1
|
Visual Question Answering (VQA) Although large-scale vision-language pretrained models (VLPs) achieve further improvements in the in-distribution (ID) VQA benchmark The dataset-bias problem in VQA is well studied by numerous debiasing methods based on conventional small-scale models This paper jointly studies the compression and debiasing problems of VLP for the VQA task. To this end, we combine the existing debiasing and pruning methods to establish a training and compression pipeline, and conduct extensive experiments with the pre-trained lxmert, which is the most popular VLP in VQA, under different OOD settings. We show that there exist sparse lxmert subnetworks that are more robust than the full model, which suggests that the goal of OOD robustness and computational efficiency can be achieved simultaneously. We also present a comprehensive study on the design of the training and compression pipeline, as well as the assignment of sparsity to different model modules, to identify subnetworks with better OOD generalization. Our findings highlight the importance of 1) Employing a two-stage training and compression pipeline and integrating the debiasing objective throughout the entire process. 2) If there are two debiasing methods working well with the full model, training the full model with the relatively poor-performing one and compressing it with the better one. 3) Assigning modality-specific sparsity to different modules of VLP. Our main contributions are as follows: (1) We present the first (to our knowledge) systematic study on sparsity and OOD robustness for VLPs. (2) Our empirical studies on the training and compression pipeline and sparsity assignment can serve as a valuable guideline for the future design of VLP subnetwork searching methods. (3) We obtain subnetworks that outperform existing debiasing So-TAs in terms of the trade-off between accuracy and model size on OOD datasets VQA-CP v2 and VQA-VS (see Fig.
|
This section takes lxmert as an example to introduce how we extract subnetworks. Lxmert contains an embedding layer, a visual fc layer, a pooler layer, a VQA-specific classifier and a stack of Transformer layers, which involve three encoders: lan-guage encoder (L enc ), object relationship encoder (R enc ) and cross-modality encoder (C enc ). We adopt unstructured pruning to obtain a compressed version (i.e., a subnetwork) of the original VLPs. Specifically, given a VLP f (θ) with parameters θ, we apply a binary pruning mask m ∈ {0, 1} |θ| to the model parameters, which gives rise to f (m⊙θ), where ⊙ is the element-wise product. The parameters to be pruned are: where W emb , W vis-fc and W plr are the weights of embedding layer, vision fc layer and pool layer, θ Lenc ∪ θ Renc ∪ θ Xenc are the parameters of Transformer layers. More details of lxmert can be found in App. B.1. Another model visualBERT We consider two representative pruning methods, i.e., magnitude-based pruning Magnitude-based Pruning approximates the importance of model parameters based on their absolute values and eliminates the less important ones. We adopt the basic version of magnitude-based pruning, i.e., one-shot magnitude pruning (OMP). OMP can optionally be combined with further finetuning of the pruned subnetwork to recover the performance drop. Mask Training directly optimizes the binary pruning mask m towards the given objectives. Specifically, each weight matrix W ∈ R d i ×do is associated with two mask matrices, namely a binary mask m ∈ {0, 1} d i ×do and a real-valued mask m ∈ R d i ×do . In the forward propagation, m is computed from m through binarization: where ϕ is the threshold. Then, the original weight matrix W is replaced with a pruned one m ⊙ W. When it comes to backward propagation, we follow We initialize m according to the magnitudes of the pre-trained weights of lxmert. This strategy is shown to be more effective than random initialization for pre-trained language models where α ≥ 1 is a hyper-parameter. At initialization, we set the threshold ϕ = 0.01 (any other value with the same order of magnitude should also be fine). To ensure that the subnetwork satisfies the given sparsity, ϕ is re-computed every t m training steps. The deabising methods in VQA usually contain a main model and a biased model. The biased model, which learns the language bias, is used to measure the training samples' bias degree and adjust the training loss for the main model. We experiment with SoTAs debiasing methods, i.e., LMH where δ denotes the sigmoid function. Learned-Mixin +H (LMH) adds a biased model to learn biases during training, as follows: where p b and p m are the predicted distribution of biased model and main model, respectively. g(h) determines how much to trust the learned biases, based on lxmert's last hidden representation h. Following RUBi adopts a training strategy similar to LMH to regularize the main model's probability, and uses standard cross-entropy as the training loss: LPF measures the bias degree as α k = p b [a k ] to regularize the loss of the main model: where the γ is a tunable hype-parameter. Given the pre-trained lxmert f (θ pt ), our goal is to find a subnetwork f (m ⊙ θ f t ) that satisfies a target sparsity level s and maximizes the OOD performance: ) where E OOD denotes OOD evaluation, ∥∥ 0 is the L 0 norm and |θ pr | is the total number of parameters in θ pr . This goal is achieved by searching the optimal m and θ f t in model training and compression. Eq. 9 only specifies the overall sparsity. In this work, we also explore a finer-grained control over sparsity, which allocates different sparsity to different modules of lxmert, given that the overall sparsity is satisfied. Concretely, we consider three modules from different modalities, i.e., the language module, the visual module and the cross-modality module. The constraint in the optimization problem is then rewritten as If not otherwise specified, we set the sparsity of every weight matrix to target sparsity. For example, if s = 70% and there is no modality-specific constraint, then all weight matrices are at 70% (uniform sparsity). If s L = 50%, then all weight matrices in θ Lan are at 50% sparsity, while s R and s X could be different (modality-specific sparsity). In this section, we study the proper design of the training and compression pipeline, under the basic framework described in Sec. 2.5. Here we focus on the mask training compression method, as it has been shown to generally outperform OMP with further fine-tuning. Our main observations can be described from three perspectives: First, it is recommended to introduce the debiasing loss across Stage1, Stage2 and (if applicable) Stage3. The reason is three-fold: 1) As shown by Fig. Second, Stage3 is unnecessary if it adopts the same training objective as Stage2. Comparing the blue and red (or cyan) bars in Fig. Third, it is recommended to use different debiasing methods in the two stages and leave the better one to Stage2. As shown in Fig. In this section, we mainly investigate three questions: (1) How does compression affect lxmert's OOD generalization ability? (2) How to design the training and pruning pipeline to achieve a good sparsity-performance trade-off? (3) How to assign sparsity to different modality-specific modules? We conduct experiments on the OOD benchmarks VQA-CP v2 Subnetworks from BCE Fine-tuned lxmert. We compress the BCE fine-tuned lxmert using OMP and mask training and introduce either L bce or L lmh in the pruning (for mask training) or further fine-tuning process (for OMP). The results are shown in the upper row of Fig. 3) "mask train" outperforms "OMP" in general, which suggests that directly optimizing the subnetwork structure is more effective than debiasing a compressed subnetwork by further fine-tuning. Subnetworks from lxmert Fine-tuned with Debiasing Methods. From the lower row of Fig. Pruning Each Single Modality-specific Module. Since lxmert uses different modules to encode the multi-modal data, it is intuitive to hypothesize that different modules of lxmert may capture the language bias to different extents. To validate this hypothesis, we compress the language, visual and cross-modality modules, respectively. As presented 4 "lpf+lmh" denotes "lxmert(lpf) + mask train(lmh)" by Fig. When the full model is lxmert(bce) (the orange and cyan lines), compressing the language or crossmodality module has a positive effect on the OOD performance, and the accuracy generally improves as sparsity increases from 10% to 90%. By contrast, compressing the visual module results in inferior results than compressing the other two modules, even if the number of remaining parameters is larger (note that the visual module has a smaller number of parameters than the other two modules). These results suggest that, for the biased lxmert(bce), the language and cross-modality modules capture more training set bias than the visual module, which supports the above hypothesis. In terms of "lxmert(lmh) + mask train(lmh)" (the red line), although compression does not lead to performance improvement like compressing lxmert(bce), the results also demonstrate that the language and cross-modality modules are more compressible than the visual module. Searching for Appropriate Modality-specific Sparsity. Motivated by the above findings, we search for appropriate modality-specific sparsity by performing mask training with a variety of sparsity configurations (see App. C.4) for the three modules while keeping the overall sparsity the same. As we can see in Fig. Fig. VQA-CP v2 is widely used in the literature of debiasing VQA systems. However, it only considers the question-type-based bias. To account for other potential biases, VQA-VS constructs several types of OOD test sets according to different shortcuts (e.g., keyword and key object). As a result, VQA-VS is more challenging and allows us to analyze the results on different biases. In this section, we search sparse and robust lxmert subnetworks in VQA-VS based on the major findings obtained from VQA-CP v2. The Effect of Compression. Fig. 3) The "lmh+lmh" 5 subnetworks improve over full lxmert (lmh) on both ID and OOD test sets, across a wide range of sparsity levels, suggesting that lxmert can also be simultaneously compressed and debiased on VQA-VS. The Effect of Modality-specific Sparsity. Fig. 10 also shows that compressing different modalityspecific modules has different effect on VQA-VS, as in VQA-CP v2. The language module is the most compressible while compressing the visual module results in the sharpest performance decline. To compare modality-specific sparsity and uniform sparsity, we directly inherit the sparsity configuration selected in Sec. 3.4 on VQA-CP v2. Fig. In this section, we will compare the best training and compression solutions identified in the previous sections with the current SoTA debiasing methods. Tab. 1 shows the results on VQA-CP v2. We find that: The accuracy of our methods (10% lxmert and 30% lxmert) beats the previous non-VLP debi-5 Since most debiasing methods (e.g., LPF and RUBi) fail on VQA-VS (see Tab.2), we only use LMH in VQA-VS. However, combining LMH and other effective debiasing methods in different stages may further outperform "lmh+lmh", as found in VQA-CP v2. We leave it for future work. asing SoTAs with 1.55% and 5.79%, respectively, with fewer or similar amounts of parameters, establishing new state-of-the-arts. 2) Our methods (30% lxmert and 50% lxmert) outperform the debiased full lxmert, even with much fewer parameters. 3) Full lxmert(lpf) and full lxmert(lmh) are good at different question types, which can partly explain why combining them in different stages produces more robust subnetworks. We also add experiments on a more recent VLP mPLUG Results on VQA-VS are presented in Tab. 2. We can observe that: 1) Our methods "bce+bce" 10% lxmert and "lmh+lmh" 30% lxmert outperform all the non-VLP debiasing methods in both ID and OOD settings, with similar or fewer parameters. 2) Except for LMH, other debiasing methods underperform BCE in OOD-mean. LMH improves the OOD accuracy at the cost of ID accuracy decline. 3) The "lmh+lmh" subnetworks (even with 50% remaining parameters) obviously improve the ID performance of lxmert (lmh) and retain comparable OOD performance. 4) Compared with "bce+bce", the OOD advantage of "lmh+lmh" outweighs its ID disadvantage at 50% to 90% parameters. With fewer remaining parameters, the overall performance of "bce+bce" is superior. To facilitate the application of VLP-based VQA systems, this paper presents the first joint study on the compression and debiasing problems of VLP for the VQA task. Through extensive experiments with three VLPs (i.e., lxmert, visual-BERT and mPLUG), we analyze the impact of compression on the OOD generalization ability. We present a comprehensive study on the design of the training and compression pipeline for a good sparsity-performance trade-off, and provide some valuable findings about the assignment of sparsity to different modality-specific modules. The compressed lxmert subnetworks in this paper outperform the SoTA debiasing methods with fewer or similar model parameter counts. Although we have empirically verified that the adoption of modality-specific sparsity is beneficial for the search for more robust subnetworks, our work still does not provide a solution on how to determine the optimal sparsity assignment effectively and efficiently. We invite follow-up studies to further address it in future work. A.1 Overcoming Dataset Bias in VQA Most VQA systems heavily rely on the information of the question to predict answers no matter the content of the given image. That is they learned the language biases in datasets. They are not robust and always perform poor in the OOD setting where the language biases they learned in training set are invalid for test set. To promote the development of models that overcome such problem, VQA-CP v2 Recently, VLPs Model compression techniques for Transformerbased pre-trained models are well developed (mainly around BERT), including pruning B.1 lxmert Architecture and Subnetworks For lxmert, the embedding layer and visual fc layer map language-modality input (token sequences obtained by WordPiece tokenizer) and visionmodality input (36 object features obtained by Faster R-CNN The attention modules have four kinds of weight matrices, i.e., the query, key and value matrices W Q,K,V ∈ R d model ×d model , and the output matrix We adopt unstructured pruning to obtain a compressed version (i.e., a subnetwork) of the original VLPs. Specifically, given a VLP f (θ) with Figure where W emb , W vis-fc and W plr are the weights of embedding layer, vision fc layer and pool layer, θ Lenc ∪ θ Renc ∪ θ Xenc are the parameters of Transformer layers: (12) where CX, CL and CR are the language selfattention, visual self-attention and cross-attention modules respectively. Similar to lxmert, visualBERT is composed of an embedding layer, a visual projection layer, a pooler layer, a stack of Transformer layers. Differently, visualBERT's Transformer layers only involve a single encoder (V enc ). The parameters of visual-BERT to be pruned are: where W emb and W plr are the weights of embedding layer and pool layer, θ Venc are the parameters of Transformer layers: where V = 12. LMH takes a step further based on Produce of Experts (PoE) where p b is the predicted distribution of biased model, and indicates the bias degree of the sample. In this way, when a sample is heavily biased, that is, p b is large, the main model will not output a large p m for it during training. Following To selectively adjust the main model's behavior, LMH adds a learn function g to explicitly deter- mask train(bce)" (full lines) at all sparsity levels. As the sparsity increases, the gaps widen. This shows the initialization strategy we adopt is more effective than random initialization. Subnetworks at 90% Sparsity From Fig. For the overall target sparsity of 50% and 70%, we adopt the following procedure to search the comfortable zone for the modality-specific sparsity: First, we traverse [10%, 30%, 50%, 70%, 90%] (i.e., step size of 20%) to assign modality-specific sparsity for any two modules, and compute the modality-specific sparsity for the remaining one Second, we use the same method to traverse the reduced range with a smaller step size of 5%. In this way, we can determine the most comfortable zone for the modality-specific sparsity. Similarly, when the overall target sparsity is 90%, we directly traverse 80% ∼ 98% with a step size of 2% to search the most comfortable zone of the modality-specific sparsity. D More Experiments on VQA-VS D.1 Performance on varying OOD test sets of VQA-VS The Effect of Compression without Debiasing For simplicity, we categorize the nine OOD test sets into 3 categories of different modalities, i.e., language-based (OOD-lang), visual-based (OODvis) and cross-modality (OOD-crsM) ones. We report the average accuracy of each category, as well as the IID accuracy and the average accuracy of all OOD test sets (OOD-mean) in Fig. The lower part of Fig.
| 1,263 | 1,953 | 1,263 |
ZhuJiu: A Multi-dimensional, Multi-faceted Chinese Benchmark for Large Language Models
|
The unprecedented performance of large language models (LLMs) requires comprehensive and accurate evaluation. We argue that for LLMs evaluation, benchmarks need to be comprehensive and systematic. To this end, we propose the ZhuJiu benchmark, which has the following strengths: (1) Multi-dimensional ability coverage: We comprehensively evaluate LLMs across 7 ability dimensions covering 51 tasks. Especially, we also propose a new benchmark that focuses on knowledge ability of LLMs. (2) Multi-faceted evaluation methods collaboration: We use 3 different yet complementary evaluation methods to comprehensively evaluate LLMs, which can ensure the authority and accuracy of the evaluation results. (3) Comprehensive Chinese benchmark: ZhuJiu is the pioneering benchmark that fully assesses LLMs in Chinese, while also providing equally robust evaluation abilities in English. (4) Avoiding potential data leakage: To avoid data leakage, we construct evaluation data specifically for 37 tasks. We evaluate 9 current mainstream LLMs and conduct an in-depth discussion and analysis of their results. The ZhuJiu benchmark and open-participation leaderboard are publicly released at
|
With the continuous development of large language models (LLMs), the emergence of GPT4 (Ope-nAI, 2023) is enough to trigger a new wave of technology. Various types of LLMs have recently been rapidly developing, such as Llama2 2 †Corresponding author conduct comprehensive and objective evaluations of LLMs to fully understand their strengths and limitations. Specifically, on the one hand, for applicators, they need to understand the overall performance of LLMs or the advantages of LLMs in a specific aspect. Constructing comprehensive and authoritative benchmarks can help applicators significantly improve the efficiency of using LLMs. On the other hand, for developers, the improvement direction of LLMs requires accurate evaluation results as guidance. An objective and fair benchmark can help them carry out relevant research work on LLMs more targetedly. To this end, scholars conduct extensive research on evaluations for LLMs and construct some superior benchmarks. Normally, the evaluation for LLMs includes two aspects: ability evaluation and evaluation method. Although traditional benchmarks such as GLUE Despite these successful efforts for LLMs' evaluations, existing studies still suffer from several limitations: (1) Current benchmarks tend to focus on evaluating LLMs on a single dimension of their abilities, which can not provide a comprehensive evaluation of LLMs. (2) Most benchmarks only use a single evaluation method, which may not provide an accurate evaluation of all the abilities of LLMs. For example, while HELM In this paper, we propose the ZhuJiu Benchmark to solve above mentioned problems, which can fill the gap in the development of a comprehensive benchmark for evaluating LLMs in Chinese. The advantages of the ZhuJiu are as follows: (1) Multi-dimensional ability coverage: we evaluate LLMs from 7 ability dimensions, including knowledge, Chinese-specific, language, reasoning, refusal, safety and robustness abilities, covering 51 datasets to provide a comprehensive performance assessment. In addition, we also proposed a new paradigm for evaluating the knowledge ability. (2) Multi-faceted evaluation methods coordination: we use Metrics Evaluation, Scoring Evaluation, and Comparative Evaluation for comprehensively evaluating LLMs to ensure authoritative and accurate evaluation results. (3) Comprehensive Chinese benchmark: ZhuJiu is the pioneering Chinese benchmark that can comprehensively evaluate LLMs, while allowing equivalent assessment in English. (4) Avoiding potential data leakage: in addition to collecting 14 commonly used datasets, we construct 37 datasets for the evaluation of LLMs, ensuring maximum avoidance of data leakage and evaluation fairness. The overall evaluation process is shown in Figure We also release an online evaluation platform that supports multiple functions including visualizations of evaluation results, participating in model arena and submission of evaluation model, etc. Moreover, we evaluate 9 publicly available LLMs, including ChatGLM In summary, the contributions of this paper are as follows: • We propose ZhuJiu, the first Chinese benchmark that covers multi-dimensions of ability and employs multi-faceted evaluation methods in collaboration. Meanwhile in the ZhuJiu we construct a novel benchmark for evaluating knowledge ability and 37 evaluation datasets to prevent data leakage issues. • We release an online evaluation platform that enables users to evaluate LLMs. We will continue to improve the platform, and update the evaluation leaderboard. • Using the ZhuJiu benchmark, we evaluate 9 current LLMs, to comprehensively and deeply explore their abilities, providing valuable insights to inform future LLM development.
|
As stated above, the ZhuJiu benchmark uses 3 evaluation methods to assess the abilities across seven dimensions of LLMs. This section provides a detailed introduction to the ZhuJiu benchmark covering the evaluation methods, datasets, and ability dimensions. We also detail the specific scoring rules in Appendix A. The evaluation framework is shown in Figure Unlike previous works that only use a single evaluation method Metrics Evaluation is an indispensable component in LLM assessment, providing objective results The abilities demonstrated by Comparative evaluation is the most intuitive evaluation method. In this paper, we drew inspiration from the work of Chatbot Arena For a benchmark, the most crucial part is undoubtedly its data source and data quality. In ZhuJiu, our evaluation data comes from two parts. On the one hand, we use 14 currently popular LLMs evaluation datasets. On the other hand, considering the serious issue of data leakage when solely using public datasets for LLMs evaluation, which could compromise the fairness of evaluation results, we constructed 37 evaluation datasets based on Chat-GPT (OpenAI, 2022). To ensure the generality of ZhuJiu, we evaluate LLMs using 14 publicly available datasets, which are essential due to their high quality and ability to accurately evaluate the performance of LLMs in certain aspects. To address the issue of data leakage in LLMs evaluation, we are inspired by PandaLM To better understand the processes of data construction and evaluation in a more intuitive way, we take Scoring Evaluation as an example to demonstrate the process, as shown in Figure With the help of the aforementioned evaluation methods and datasets, we can assess the abilities of LLMs in 7 aspects. We will provide a detailed introduction to the specific evaluation methods and details in this section. To comprehensively evaluate the knowledge abilities of LLMs, we conduct the evaluation from four perspectives: world knowledge, commonsense knowledge, linguistic knowledge, and concept. For each evaluation perspective, we select the appropriate properties of accuracy, robustness, completeness, and timeliness to construct evaluation datasets for evaluating LLMs. Detailed descriptions of these four properties are provided in Appendix B, using a detailed framework shown in Figure For world knowledge, on the one hand, we utilize the GAOKAO-bench For commonsense knowledge, we select commonsense triplets as the basic data and construct evaluation datasets based on the evaluation properties of accuracy and robustness. We then use Scoring Evaluation to evaluate LLMs. For linguistic knowledge, we use Chinese FrameNet (CFN) For concept, we manually select common entity words as the original data and evaluate the accuracy and robustness of LLMs concepts with Scoring Evaluation. Following SuperCLUE (Liang Xu and others from SuperCLUE team, 2023), and conventional Chinese evaluations, the Chinese-specific ability evaluation aims to use corpora with Chinese unique characteristics as the original data to form evaluation data. These corpora include ChID We conduct a comprehensive evaluation of LLMs' language ability from both aspects of language understanding and language generation. For evaluating LLMs' language understanding ability, we choose to evaluate them on the tasks of reading comprehension and coreference resolution. We find that using existing datasets could achieve good evaluation results, and the datasets we use included C3 As the evaluation of LLMs' reasoning ability is less affected by data leakage Regarding the refusal ability, we can understand it like this: To know what you know and to know what you do not know, that is true knowledge. For constructing datasets of refusal ability, we drew inspiration from the categories of Known-Unknown Questions proposed in For the evaluation of safety ability, we follow Sun et al., 2023a's classification of safety ability and further summarize and categorize them. We derive a total of 9 evaluation tasks from 6 perspectives, including Insult, Human Health (Physical harm and Mental health), Social Topic (Unfairness discrimination and Ethics morality), Serious Risk (Criminal Activity and Unsafe Instruction Topic), Goal Hijacking and Role play instruction. Subsequently, we employ the Scoring Evaluation to assess LLMs. Traditional robustness evaluation primarily focuses on assessing the impact of adding perturbations of varying granularity to the text on the performance of the model We develop an online platform to provide a range of services for the community as follows: Visualizations of evaluation results We publish the rankings of all model evaluations on the platform, including specific scores for each ability and evaluation method, and the rankings will be updated continuously as the evaluations progress. Participating in Model Arena We launch a oneon-one model arena feature on our platform, where everyone can support the LLMs they believe perform better based on their own judgment. Please refer to Figure Submission of Evaluation Model We also encourage everyone to actively participate in our evaluations and join the leaderboard. On our platform, we allow users to submit applications for evaluation. To facilitate the utilization and advancement of LLMs, the primary emphasis of ZhuJiu's inaugural evaluation phase is directed towards opensource LLMs with a parameter magnitude of approximately 10 billion, including: ChatGLM-6B We report the overall performance in Table ??, and show more detailed assessment results in our platform. From the results, we can obtain some intriguing findings: (1) Model-Performance is Limited by Model-Size: Based on the results in table ??, it becomes evident that models with a parameter size of around 10 billion still exhibit significant limitations in overall performance compared to (2) Lower Limit Sets Upper Limit: The analysis reveals that Phoenix (3) Knowledge is Power: In ZhuJiu, our primary focus lies in the knowledge ability of LLMs, as the pivotal task at hand is to ensure LLMs acquire accurate knowledge and effectively harness their acquired knowledge. However, in this version, the majority of LLMs exhibit subpar performance in terms of knowledge capacity, making the ZhuJiu benchmark exceptionally challenging. The results reveal that Chat-GLM2 In this work, we present ZhuJiu, the pioneering multi-dimensional ability coverage, multi-faceted evaluation methods collaboration Chinese benchmark. ZhuJiu is capable of using 3 evaluation methods to comprehensively evaluate LLMs across 7 ability dimensions, using 51 datasets. Additionally, we independently construct 37 evaluation datasets to maximize the avoidance of data leakage issues in LLM evaluation. We also focus on expanding the evaluation of knowledge ability, providing a new framework for assessing LLMs' knowledge ability. Finally, we provide a comprehensive and continuously updated evaluation platform with multiple functions and in the first season of ZhuJiu, we evaluate 9 open-source LLMs. In the future, we plan to (1) continuously construct high-quality evaluation datasets to enrich ZhuJiu, (2) further perfect the assessment of knowledge ability and develop new evaluation methods for Chinese characteristic ability, (3) further perfect the platform's functionality and update the platform's information. We will comprehensively evaluate the model from seven ability dimensions and 3 assessment methods to ensure the thoroughness and authority of the evaluation results. Specifically, the comprehensive evaluation process can be broken down into three steps. Step 1 For each ability dimension score A, we will take the average of LLM's scores d = [d 1 , . . . , d n ] on each dataset as LLM's score for that ability dimension: Step 2 For each evaluation method score E, LLM's score is the average of its scores A = [A 1 , . . . , A m ] for each ability dimension: Step 3 LLM's scores E = [E 1 , E 2 , E 3 ] for each evaluation method are standardized and then averaged to obtain LLM's final score on ZhuJiu: In the evaluation process of knowledge ability, we mainly evaluate from the properties of accuracy, robustness, completeness and timeliness. For each property, we will randomly generate one hundred sets of evaluation data for evaluation. Here we Need to explain the specific indicators of each evaluation • Accuracy: Evaluate whether the content of the model's reply is correct through Exact Match (EM) and ChatGPT (OpenAI, 2022), and calculate the accuracy rate in the 100 questions answered correctly by the model. • Robustness: We use the same set of data to use ChatGPT to randomly generate five different ways of asking questions, and then score according to whether the model is stable in replying to different questions generate by the same set of data. The principle of scoring is that the more stable the content of the reply, the higher the score. • Completeness: Only for the evaluation of world knowledge, scoring is based on the proportion of standard answers cover in the model's reply content. For example, according to the calculation of a question with a full score of 10, for the data "(中国四大 " "(The Four Great Inventions of ancient China-include-gunpowder, compass, papermaking, printing)" generate the evaluation question "中国的四大发明包括哪 些?" "What are the Four Great Inventions of ancient China?", if the model answers "火药,指南针,造纸术,印刷术" "gunpowder, compass, papermaking, printing", it will get a full score of 10, and answer "火药,指南 针,造纸术,瓷器" "gunpowder, compass, papermaking, china" has a correct rate of 75 percent and a score of 7.5. • Timeliness: It is only aim at the evaluation of world knowledge, and specifically evaluates the update degree of LLMs knowledge, similar to accuracy, and evaluates whether the answer of the model is correct or not according to EM and ChatGPT. In the scoring evaluation method, we use GPT-4 to score the answers of the model being tested. The evaluation content covers 37 testing tasks corresponding to 7 capabilities, and the evaluation datasets are all generated by GPT and manually reviewed to prevent data leakage. For each evaluation task, there are more than three nearly characteristic evaluation indicators. Please act as a language model evaluator, you need to assess a language model's poetry writing ability. Your assessment results need to consider whether the article meets the poetry format requirements, its relevance to the question, the depth, and creativity of the answer. Your assessment results need to provide an explanation for the quality of the poetry, please try to remain objective. After the explanation, provide a score from 1-10 for each perspective. Please act as a language model evaluator, you need to assess a language model's ability to output text in a specified formal format. Your assessment results need to consider whether the text conforms to the format requirements of the corresponding scene, whether it conforms to the role identity, whether it is logical, and whether the text is fluent. Your assessment results need to provide an explanation for the text, please try to remain objective. After the explanation, provide a score from 1-10 for each perspective. [Question]prompt[Start of response]response[End of response] Your output format needs to strictly follow the JSON format. The output JSON dictionary includes two keys: 'Explanation' and 'Score'. The value of 'Explanation' is in string format. The value of 'Score' is a nested dictionary, containing the following keys: 'Correctness of Format', 'Match with Role Identity', 'Logic', 'Text Fluency'. You only need to output the JSON assessment result.
| 1,176 | 3,720 | 1,176 |
CRUSH4SQL: Collective Retrieval Using Schema Hallucination For Text2SQL
|
Existing Text-to-SQL generators require the entire schema to be encoded with the user text. This is expensive or impractical for large databases with tens of thousands of columns. Standard dense retrieval techniques are inadequate for schema subsetting of a large structured database, where the correct semantics of retrieval demands that we rank sets of schema elements rather than individual elements. In response, we propose a two-stage process for effective coverage during retrieval. First, we instruct an LLM to hallucinate a minimal DB schema deemed adequate to answer the query. We use the hallucinated schema to retrieve a subset of the actual schema, by composing the results from multiple dense retrievals. Remarkably, hallucination -generally considered a nuisance -turns out to be actually useful as a bridging mechanism. Since no existing benchmarks exist for schema subsetting on large databases, we introduce three benchmarks. Two semi-synthetic datasets are derived from the union of schemas in two wellknown datasets, SPIDER and BIRD, resulting in 4502 and 798 schema elements respectively. A real-life benchmark called SocialDB is sourced from an actual large data warehouse comprising 17844 schema elements. We show that our method 1 leads to significantly higher recall than SOTA retrieval-based augmentation methods.
|
State-of-the-art language model based Text-to-SQL generators provide impressive accuracies on well known benchmarks, but they require the entire DB schema as input, along with the user question text In contrast, real-life datasets may have thousands of tables with hundreds of columns per table. E.g., a real-life data warehouse of data about various social indicators of a country comprises of more than 17.8 thousand columns! For such large schema, we cannot afford to include the entire schema in the prompt preceding each query; only a high-recall subset of the schema can be attached to each question. LLM-as-a-service usually charges for each token exchanged between client and server, so we want the subset to be as small as possible while ensuring high recall. Even for in-house (L)LMs or other Text-to-SQL methods, admitting extraneous schema elements as candidates for use in the generated SQL reduces its accuracy Retrieving a subset of a corpus of passages to augment the LLM prompt has become an emerging area We argue (and later demonstrate) that Text-to-SQL needs a more circumspect approach to jointly leverage the strengths of LLMs and dense retrieval. Consider a question "What is the change in female school enrollment and GDP in Cameroon between 2010 and 2020?", to be answered from a large database like World bank data 2 . An effective Text-to-SQL system needs to realize that GDP and female school enrollment are two key 'atoms' in the question and match these to tables Development indicators and Education statistics respectively. Additionally, it needs to generalize Cameroon to Country, and 2010,2020 to year to match correct columns in these tables. This requires generalization and phrase-level matching via LLMs (with all the world knowledge they incorporate); pre-deep-NLP and 'hard' segmentation techniques Our setting thus requires us to retrieve, score and select sets of schema elements that collectively cover or explain the whole query. This bears some superficial similarity with multi-hop question answering (QA). Closer scrutiny reveals that, in multihop QA benchmarks such as HotPotQA, each question comes with only 10 passages, out of which 8 are 'distractors' and two need to be selected to extract the answer. The best-performing systems Our contributions: In this paper, we propose a new method called CRUSH 3 that leverages LLM hallucination (generally considered a nuisance) in conjunction with dense retrieval, to identify a small, high-recall subset of schema elements for a downstream Text-to-SQL stage. CRUSH first uses fewshot prompting of an LLM to hallucinate a minimal schema that can be used to answer the given query. In the example above, the LLM might hallucinate a schema that includes tables like • Indicators(name, country, year) and • Education enrollment data(type, country, year, value) The hallucinated schema contains strings that are significantly closer to the gold table names mentioned earlier. We use the hallucinated schema elements to define a collection of index probes for fast index-based retrieval of the actual schema elements in the DB. Finally, CRUSH approximately solves a novel combinatorial subset selection objective to determine a high-recall, small-sized schema subset. The objective includes special terms to 3 Collective Retrieval Using Schema Hallucination maximize coverage of distinct elements of the hallucinated schema while rewarding connectivity of the selected subset in the schema graph. Our second contribution involves the creation of three novel benchmarks for the task of retrieval augmentation related to Text-to-SQL conversion on a large schema. We developed two semi-synthetic benchmarks, encompassing 4502 and 768 columns respectively, by taking a union of all databases from the well-known SPIDER benchmark, as well as the relatively recent BIRD benchmark. The third benchmark is sourced from a production data warehouse and features 17.8 thousand columns. This serves to mitigate a critical limitation in existing Text-to-SQL benchmarks, which have much smaller schema. Beyond its large scale, our third benchmark introduces additional challenges such as a significantly higher overlap in column names, and challenging lexical gaps (when compared to SPIDER), between schema mentions in the question and the actual schema name. Using these benchmarks, we present an extensive empirical comparison between CRUSH and existing methods. We show consistent gains in recall of gold schema elements, which translates to increased accuracy of Text-to-SQL generation. The results of our analysis provide valuable insights into the weaknesses of the existing singleembedding or token-level representations.
|
We are given a large database schema D consisting of a set of tables T , with each table t ∈ T comprising of a set of columns c ∈ C(t). We will use d to denote a schema element ('document'), which can be either a table or a column. A schema element d has a textual name or description S(d). The text associated with a column t.c is written as the concatenation S(t).S(c) as shown in Figure Apart from the database, the input includes a natural language question x. Question x is associated with a (possibly unknown) correct ('gold') SQL query q(x). The gold SQL query q(x) mentions a subset R(q(x)) of the schema elements from D. Almost always, |R(q(x))| ≪ |D|. Our goal is to retrieve from D, a (small) subset R(x) ⊂ D that includes R(q(x)), i.e., R(q(x)) ⊆ R(x). The question x will be concatenated with the schema subset R(x) and input to a Text-to-SQL model to convert the question into an SQL query. There are multiple reasons to minimize |R(x)|. Hallucinate the minimal schema of a relational database that can be used to answer the natural language question. Here are some Examples: Example 1: Question: Show the season, the player, and the name of the team that the player belongs to. Tables (3) The performance of even in-house (L)LMs or other Text-to-SQL systems degrade when extraneous schema elements are presented as possible candidates for inclusion in the SQL query. In this paper we focus on the task of efficiently retrieving R(x) from D, while maximizing the recall of gold schema elements in R(q(x)). We assume that the schema D is indexed in a pre-processing step. Each table t(c 1 , . . .) in the database D is exploded into the form t.c for each column, where '.' is a separator character. Each column now acts as a 'document' d in an information retrieval system, except that we will retrieve sets of documents. Each document is sent into a pretrained transformer, SGPT (Muennighoff, 2022), or the LLM service to get an embedding d. Successfully matching a user question x to a relevant schema R(q(x)) could be quite non-trivial, since the schema element names or descriptions are often not directly mentioned in the question. For example, consider the question x = Count the number of members in the Bootup Baltimore Club older than 18. In order to match this question prop-erly to gold schema D, containing such schema elements as Age, Club description, and Club members, etc., we need to perform multiple types of lexical and syntactic reasoning on the question and schema: • extract segments like older than 18 and Bootup Baltimore Club, • generalize the first segment to a likely schema element Age, and, • for the second segment, instead of attempting to match strings like Bootup Baltimore verbatim, match to a schema element called Club description. To bridge such large lexical gap between the tokens of x and the gold schema R(q(x)), we designed a two-phase approach. In the first phase, we transform x into an intermediate form K(x) comprising of multiple generalized segments leveraging an LLM. In the second phase, we use K(x) to retrieve from the DB schema D suitable schema element subset R(x), by approximately optimizing a combinatorial objective that collectively maximizes coverage of all elements of K(x). We describe these two phases next. Our goal here is to infer from the question text x, a set K(x) of intermediate search strings, which, when used to probe a suitable index over the client's DB schema, will retrieve R(x). After unsuccessful attempts with query decomposition and token-level retrieval methods, we proposed to harness an LLM to aid with this task. Initially, we attempted to use the LLM to extract variable name mentions or perform conventional query decomposition like in We use state-of-the-art LLMs with few-shot prompting to hallucinate such a schema. We employ GPT-3 (text-davinci-003) with a fixed prompt comprising of six in-context examples as shown in the first half of Table We experimented with a few other prompt types, and we will present a comparison in Section 5.4 of our empirical evaluation. In this stage our goal is to retrieve a subset R(x) from D so that collectively R(x) is closest to the halluncinated schema K in the context of x. First, we retrieve a candidate set C(x) using K as probes on the indexed embeddings in D, and then we collectively match C(x) to K(x). K(x) consists of a set of hallucinated tables with their hallucinated columns, each of the form t(c 1 , . . .). These are written out as a set of "t.c 1 " column names prefixed with table names. Henceforth, we regard K(x) as a set of such hallucinated texts {k}. Each k ∈ K(x) is converted into an embedding vector k for retrieving 'real' schema elements, via the following steps: 1: form concatenation "x t.c" 2: apply a pretrained transformer, SGPT (Muennighoff, 2022) 3: get per-token contextual embedding vectors 4: average-pool per-token embeddings into k (Through experiments, we will show that using the form "x t.c" improves recall, compared to not prefixing x.) At this stage, K(x) has been converted into a bag of vectors {k}. We perform a nearest neighbor search on D using each key vector k, and retain some number of top matches per probe k based on cosine similarity of their embeddings. This gives us the candidate set C(x) of schema elements from D. See Figure Cosine baseline: A baseline method may simply return C(x). However, C(x) thus collected was observed to be biased toward generic, uninformative schema names such as Name and Identifier that appear across many tables. This hurts coverage. Therefore, we design a more careful optimization around C(x), described next. We extract from the candidate set C(x) a manageable subset R(x) with size |R(x)| ≤ B for some size budget B, that provides coverage to all parts of K, and also to reward connectivity in the schema graph on the retrieved subset. Recall that a large budget not only results in a higher expense to use an LLM-as-a-service, but may also provide a more challenging generation task for the downstream Text-to-SQL module. Entropy-guided similarity: Instead of just cosine as the distance between a k ∈ K and d ∈ D, we refine the similarity to score match of rarer columns higher. Consider some k ∈ K(x) that has good matches with many schema elements d ∈ C(x). In information retrieval, this is analogous to a query word with low inverse document frequency (IDF) in the multinomial simplex ∆ |C(x)-1| . This multinomial distribution has an entropy, which we will denote by H(k). If H(k) is large, that means k has no sharp preference for any schema element in C(x), so its impact on the perceived similarity cos(k, d) should be dialed down, inspired by TFIDF vector space model from information retrieval. We achieve this effect via the score where σ(•) is the standard sigmoid shifted by the average entropy H defined as the average H(k) over all k ∈ K. Coverage score: We will assume that the hallucinated schema elements are all informative, so we would like to 'cover' all of them using R(x). We score coverage of a k by R(x) using a soft maximization function defined as The first part of our collective objective is Connections between 'documents': Schema elements are not isolated, but related to each other using the rich structure of the schema graph smx(e(d, d ′ ) : The function smx is chosen instead of directly summing e(d, d ′ ) for all pairs in R(x) to prevent quadratically growing rewards for large subgraphs. Overall objective: Combining the two desiderata, we get our overall objective as with a balancing hyperparameter ♣. It is possible to express the above optimization as a mixed integer linear program. In practice, we find it expeditious to use a simple greedy heuristic. Also, we fix ♣ = 1 in all experiments. Given a question x, we obtain the embedding x of x, i.e., x = M (x), and then retrieve the K nearest neighbors of x in D as R(x) = K-NN(x, {d : d ∈ D}). This method has the limitation that the top-K retrieved elements may be skewed towards capturing similarity with only a subset of the gold schema R(q(x)). The RESDSQL LLMs for ranking: Sun et al. ( Other uses of LLMs in scoring and ranking items are in recommender systems LLMs for question decompostion and retrieval: Decomposing complex questions has been of interest for some years now # LLM prompt: Hallucinate a minimal schema of a relational database that can be used to answer the natural language question. Here are some examples: x Count the number of members in the Bootup Baltimore Club older than 18. K Club(Name, id, description, location), member_of_club(club id, student id), Student(id, age) x What are the names of all stations with a latitude smaller than 37.5? K Station(Name, Latitude) x Show the season, the player, and the name of the team that players belong to. K Match_season(season, team, player), Team(name, team identifier) x Find the first name and age of the students who are playing both Football and Lacrosse. K SportsInfo(sportname, student id), Student(age, first name, student id) x What are the names of tourist attractions reachable by bus or is at address 254 Ottilie Junction? K Locations(address, location id), Tourist_attractions(how to get there, location id, name) x Give the name of the highest paid instructor. K Instructor(Name, Salary) Hallucinated K generated by LLM given input x x What are the names of properties that are either houses or apartments with more than 1 room? K Property(name, type, number of rooms) x Which employee received the most awards in evaluations? Give me the employee name. K Employee(name, employee id), Evaluations(employee id, awards) x What is the document name and template id with description with the letter 'w' in it? K Document(name, description, template id) x What semester ids had both Masters and Bachelors students enrolled? K Semester(id, start date, end date), Enrollment(semester id, student id, degree), Student(id, name) In this section we compare CRUSH with existing methods for schema subsetting. We also present a detailed ablation on the various design options for CRUSH. We test on the following two benchmarks that we designed, because of the absence of any preexisting large-schema benchmark. SpiderUnion: This is a semi-synthetic benchmark derived from SPIDER Our evaluation metric, given a retrieved set R(x), is recall, defined as |R(q(x))∩R(x)| |R(q(x))| . For each question in the test set, since the gold SQL is available, we have access to the gold retrieval set R(q(x)) ⊂ D. We measure recall only over column names, since if a column name is selected in R(x), the table name is always implicitly selected. BirdUnion: Following the same approach as with SpiderUnion, we created BirdUnion from BIRD SocialDB: We created this benchmark from a real-life data warehouse, which collates statistics on various social, economic, and health indicators from a large country, featuring diverse geographical and temporal granularities. The complete database schema is publicly accessible, though we withhold the URL during the anonymity period. The warehouse holds approximately 1046 tables and a total of 18,685 columns. Each table and column carries descriptive names. From administrators of the Web-site, we obtained 77 questions along with the gold tables, which contain the answers. Some examples of questions and schema names can be found in Table Single DPR (SGPT): This is the popular Dense Passage Retrieval (DPR) baseline where we use the SGPT LM (Muennighoff, 2022) to embed x into a single embedding vector x and retrieve from D based on cosine similarity with a d ∈ D. Single DPR (OpenAI): As above, except we use OpenAI's Similarity and Search Embeddings API (text-embedding-ada-002) as the LM. Token-level Embedding (ColBERT): Instead of searching with a single embedding we perform token-decomposed retrieval for finer-grained interaction In Table • We see a significant boost in the recall by CRUSH, particularly at low to medium budget levels. For example, on SpiderUnion at a budget of ten columns, we recall 83% of gold columns whereas the best existing method only gets up to 77%. On the BirdUnion dataset, we recall 76% with CRUSH, while the best alternative method reaches only 56%. On the more challenging SocialDB dataset we recall 58% of gold tables whereas the best alternative method gets 49%. • Token-level methods are worse than even Single DPR-based methods. • Embedding of OpenAI is slightly better than than of SGPT. In Table Impact of improved recall on Text-to-SQL generation accuracy: We use the state-of-art RES-DSQL For the BirdUnion dataset, the RESDSQL system could not handle the larger schema at budget 100, but we expect a similar trend. Before we arrived at the schema hallucination approach, we experimented with other prompts motivated by the techniques in the question decomposition literature. In Table (1) Variables: that seeks to identify key isolated variables mentioned in x, and (2) Relations: that relates the variables to a subject, roughly analogous to a table name. We present a comparison of different prompt types in Table In CRUSH we used six examples of hallucinated schema in the prompt for in-context learning. We reduce that number to two and four and present the results in Table In Tables Budget-→ r @ 3 r @ 5 r @ 10 r @ 20 r @ 30 r @ 50 r @ 100 Budget-→ r @ 3 r @ 5 r @ 10 r @ 20 r @ 30 r @ 50 r @ 100 SpiderUnion Single DPR (OpenAI) 0. shot in-context examples. Low standard deviation suggests a high level of robustness across different samples. In Table CRUSH includes a number of careful design choices. In Table • During retrieval, we obtain the embeddings of a k ∈ K jointly with x. In contrast, if we independently embed k, the recall drops significantly. • After retrieval, the overall objective of collective selection (Eq 6) incorporates three key ideas: entropy guided similarity, edge scores, and coverage of hallucinated schema elements. We study the impact of each. We remove the entropy discounting in Eqn. (2), and observe a drop in recall at low budget levels. When we remove the edge scores, we also see a mild drop. • To study the impact of coverage, we replace the soft-max function smx() with a simple summation so that for each selected d ∈ D, the reward is just the sum of similarity to each k ∈ K. We find that the recall suffers. A coverage encouraging objective is important to make sure that the selected items are not over-represented by matches to a few k ∈ K. While LLMs incorporate vast world knowledge and corpus statistics, they may be unfamiliar with Budget r @ 3 r @ 5 r @ 10 r @ 20 r @ 30 r @ 50 r @ 100 CRUSH 0. (possibly private) client DB schemas, which can be very large, rendering impractical or expensive any attempt to upload the full schema in-context along with questions for Text-to-SQL applications. Remarkably, we find a workable middle ground by allowing the LLM to hallucinate a schema from the question and limited in-context examples with no reference to the client schema. Then we formulate a novel collective optimization to map the hallucinated schema to real DB schema elements. The resulting real schema subset that is retrieved has a small size, yet high recall This schema subset can be readily uploaded to (L)LM-based Text-to-SQL methods. The reduced space of client DB schema elements also improves the accuracy of generated SQL for state-of-the-art Text-to-SQL implementa-tions. Dataset Budget r @ 3 r @ 5 r @ 10 r @ 20 r @ 30 EM Mean 0.33 0. Budget r @ 3 r @ 5 r @ 10 r @ 20 r @ 30 EM Mean 0.34 0. 8 Acknowledgement
| 1,338 | 4,703 | 1,338 |
Deep Neural Representations for Multiword Expressions Detection
|
Effective methods for multiword expressions detection are important for many technologies related to Natural Language Processing. Most contemporary methods are based on the sequence labeling scheme applied to an annotated corpus, while traditional methods use statistical measures. In our approach, we want to integrate the concepts of those two approaches. We present a novel weakly supervised multiword expressions extraction method which focuses on their behaviour in various contexts. Our method uses a lexicon of English multiword lexical units acquired from The Oxford Dictionary of English as a reference knowledge base and leverages neural language modelling with deep learning architectures. In our approach, we do not need a corpus annotated specifically for the task. The only required components are: a lexicon of multiword units, a large corpus, and a general contextual embeddings model. We propose a method for building a silver dataset by spotting multiword expression occurrences and acquiring statistical collocations as negative samples. Sample representation has been inspired by representations used in Natural Language Inference and relation recognition. Very good results (F1=0.8) were obtained with CNN network applied to individual occurrences followed by weighted voting used to combine results from the whole corpus. The proposed method can be quite easily applied to other languages.
|
Multiword expressions (henceforth MWEs) have been studied for decades, defined in different ways in literature with different denotations of this term, e.g. see the overview in Several MWE characteristics or identifying properties have been postulated, e.g.: arbitrariness, institutionalisation, limited semantic variability (especially non-compositionality and nonsubstitutability), domain specificity, and limited syntactic variability Extraction of MWEs and their description in a semantic lexicon (e.g. as a reference resource) is important for many NLP applications like semantic indexing, knowledge graph extraction, vector models, topic modelling etc. Due to the specific properties of MWEs as whole units, their automated description by the distributional semantics method, e.g. embeddings, is not guaranteed, especially in the case of MWEs of lower frequency. Traditionally, MWEs extraction is preceded by finding collocations (frequent word combinations) by statistical or heuristic association measures and filtering them by syntactic patterns. However, in this way mainly the frequency-related aspect is covered. The peculiar behaviour of MWEs as a language unit may be observed in linguistic contexts, and methods based on the well-known sequence labelling scheme try to do that. They explore MWE specific behaviour of as a language expressions across text contexts, where the contexts are represented by contextual embeddings (neural language models). However, such approaches require a lot of hard manual work on text annotation. In addition, due to the corpus size limitation, most potential MWEs are observed only in a few, if not singular uses, while a lexicon element by a definition is a ready-to-use unit to be included in different contexts and, as such, should be studied. Thus, we want to fully explore the expected MWE characteristic aspects, including frequency, and to reduce the amount of manual work required. MWE annotated corpora are very rare and small, e.g. PARSEME
|
Initially statistical association measures calculated on the basis of word co-occurrence statistics in corpora were used for discovering and ranking collocations as potential MWEs Recently, attention was shifted to MWE extraction perceived as a sequence labelling problem, e.g. To counteract this, a model based on LSTM, the long short-term memory networks and CRF is proposed MWEs can be also represented as subgraphs enriched with morphological features However, such sequence labeling approaches focus on word positions and orders in sentences, and seem to pay less attention to the semantic incompatibility of MWEs or semantic relations between their components. Furthermore, sequence labeling methods do not emphasize the semantic diversity of MWE occurrence contexts. Thus, they overlook one of the most characteristic MWE factors: components of a potential MWE co-occur together regardless of the context. It allows us to distinguish a lexicalised MWE from a mere collocation or even a term strictly related to one domain. To the best of our knowledge, the concept of using deep neural contextual embeddings to describe the semantics of the MWEs components and the semantic relations between them in a detection task has not been sufficiently studied, yet. Moreover, due to the sparsity of the MWEs occurrences in the corpus, the corpus annotation process is very time consuming and can lead to many errors and low inter-annotator agree-ment. For this reason, we propose a lexicon-based corpus annotation method. We assume that the vast majority of MWEs are monosemous, automatically extract the sentences containing the MWE occurrences, and treat all sentences including a given MWE (as a word sequence) as representing the same multiword lexical unit. The conducted analysis of the existing resources has shown that it is difficult to find a large annotated dataset for the multiword expressions detection task. PARSEME shared task and multilingual corpus Concerning language expressions that are not MWEs, i.e. negative samples from the ML perspective, they are not listed or mentioned in the dictionaries. Having a corpus annotated with MWE occurrences we could extract expressions that are not as negative samples. However genuine MWEs are more frequent or statistically specific. Thus, 'normal' language expressions would be too obviously different. Instead, we noticed that statistical association measures produce very long ranking lists of collocations. Further down the ranking, MWE occurrences are quickly dwindling away. In addition, we are interested only in specific structural types of collocations that match structural types of MWEs acquired from a dictionary. To generate the list of incorrect MWEs, we selected three popular association measures We then used the three resulting lists to search for sentences including collocation or MWE occurrences in the BNC corpus. The searched expressions were simply recognised by comparing lemma sequences. Some recognition error may appear, but the potential error ration seems to be very small (single percents). If multiple MWE/collocation lemma sequences were detected among the sentence lemmas, then their occurrences were considered as separate training samples (positive or negative), see Alg. 1. In order to evaluate our method of detecting sentences containing MWEs, we extracted 4 randomly selected samples containing 100 found sentences each. A linguist conducted the analysis and found that 99% of the sentences contained correct MWE occurrences. The analysis was performed only on sentences corresponding to positive samples -MWEs from the dictionary, but similar results can be expected for collocations from the lists. Our work resulted in the creation of three datasets of MWE and collocation occurrences, named on the basis of the sources of knowledge: • ODE-PMI dataset -dataset containing occurrences of correct MWEs from the ODE dictionary and the incorrect ones obtained via the PMI measure, • ODE-Dice dataset -dataset containing oc-currences of correct MWEs from the ODE dictionary and the incorrect ones obtained via the Dice measure, • ODE-Chi2 dataset -dataset containing occurrences of correct MWEs from the ODE dictionary and the incorrect ones obtained via the Chi2 measure. Algorithm 1 Procedure of obtaining sentences (s) containing MWEs from the corpus (C) by comparing sentence word lemmas (l i ∈ [l 0 , l 1 , . . . , l n ]) to the list (M ) of lemmatised MWEs (m j ∈ [m 0 , m 1 , . . . , m k ]) for l i ∈ s do 4: for m j ∈ M do 5: if l i ∈ m j then 6: sentence_list.insert(s) The proposed representation has been inspired by the ones often used in the Natural Language Inference domain and also in the task of semantic relations extraction The obvious target is the potential compositionality of an expressions: MWE or non-lexicalised collocation. In the case of compositional expressions we expect to see some kind of inclusion relation. However, we assumed that contextual embeddings allow us to go beyond focusing only on semantic compositionality, e.g. some syntactic idiosyncrasy should be also visible in relation between contextual embeddings of the whole expression and its component. Moreover, in order to minimise the effect of accidental properties of some specific context we try to collect representations of the same expressions (MWEs and collocations) across as many contexts as possible. The obtaining of contextual MWE embeddings is described in Eq. 1. An MWE embedding ( ---→ m sent ) in the sentence context (sent) is an average of the WordPiece subtoken (s ∈ S msent ) vectors ( -→ v s ) related to the MWE components. ---→ In the next step, the MWE occurrence was replaced subsequently with each of its components in order to obtain their contextual embeddings ( --→ c sent ) by averaging the corresponding subtoken vectors representations ( -→ ν s ) related to the substituted components (S csent ), see Eq. 2. --→ The final baseline embedding ( -→ B ) of a training sample related to a sentence (sent) containing MWE (m) and one of its components (c) is described in Eq. 3. (3) Our element-wise difference vector based representation Diff-Emb ( -→ D), described in Eq. 5 leverages the absolute difference between non-contextual component embeddings ( -→ w 1 --→ w 2 ) obtained via the skipgram model from the fastText library - We also decided to consider the relevance of Hadamard product vectors, which we included in our Prod-Emb representation ( -→ P ), explained in Eq. 7. It consists of the Hadamard product of non-contextual fastText component embeddings ( -→ w 1 ⊙ -→ w 2 ) and the averaged vector of Hadamard products between the component (c ∈ m) embeddings and MWE (m) embedding (avg_prod m,sent ) in the context of the sentence (sent) described in Eq. 6 In order to combine the difference-based and product-based approaches we developed the Mean-Emb representation ( -→ M ), explained in Eq. 8. It consists of the averaged difference vector ( ----------→ avg_dif f m,sent ) and the averaged Hadamard product vector ( ----------→ avg_prod m,sent ) described in Eq. 4 and 6 respectively. For all conducted experiments we selected a singletask binary classification, where the classifier aims to predict the correct label out of 2 possible ones (lexicalised vs non-lexicalised) for the expression represented by one of the vector representations: baseline, Diff-Emb, Prod-Emb or Mean-Emb. In the process of generating the contextual embeddings we used the XLM-RoBERTa Tab. 1 shows the evaluation results for each representation on the ODE-PMI dataset. Each value is averaged over ten folds. The Mean-Emb representation combining both the knowledge based on the difference vector and the Hadamard product vector achieved the best results. The performance of the CNN model trained on all representations and evaluated on the ODE-Dice dataset is shown in Tab. 2. The best performance can be observed for the Mean-Emb model. Each of the developed representations achieved better results than the baseline vector representation. The evaluation results for the classifier trained on each representation and evaluated on the ODE-Chi2 dataset are shown in Tab. The idea of silver dataset enables transformation of any corpus into a dataset for MWE extraction, only if a limited lexicon of MWE examples is provided as a starting point -a kind of seed lexicon to be expanded. We can leverage a MWE annotated corpus, too, in the same way as a lexicon to extract the initial list of MWEs, but a large non-annotated corpus stays the basis. Several linguistic resources can be also merged, any MWE annotated text, as well as lexicons. Time-consuming and expensive corpus annotation is avoided. Moreover, it seems to be easier to maintain high quality lexicon than corpus annotation, e.g. due to potential errors and discrepancies between single annotations. A lexicon can be edited by several linguists, and metrics such as inter-annotator agreement can be easily calculated. What is more, such a transformation of lexiconbased knowledge into a dataset enables the use of deep neural network models that require large number of training samples. This is one of the reasons why our CNN method, pre-trained on contextual embeddings with weighted voting, applied to MWE recognition achieved several times better results than methods based on contextual embeddings and recurrent neural in the PARSEME shared task in general Our approach may be applied to texts in different languages, both to obtain multilingual collections and to apply transfer learning to facilitate the knowledge about MWEs in one language to MWE recognition in another language. This may be particularly relevant for low-resource languages, and it definitely a direction for further research. Another advantage of the proposed method is faster training and prediction in comparison to sequence labeling methods. In our case, the model gets the full sample representation only once before prediction. This shortens the inference time. Our vector representations support MWEs longer than two words. In the case of multiword units containing three and more words, the difference and product vectors calculated between two MWE components can be replaced with the vector obtained via the same operation, but averaged over all MWE component pairs. The obtained results show that non-lexicalised representations, i.e. those that do not include vectors for components and the whole expression 2 perform better independently of the kind of a measure used to extract collocations. All representations except the baseline are built from differences and 2 A contextual vector of the whole expression somehow includes a picture of the particular expression and its lexemes. products of vectors, not the vectors itself. Thus they are more focused on representing relations between a potential MWE and its components. It is worth to be emphasised that lexical split was also implemented in order to prevent the models to remember concrete words instead of learning patterns for behaviour of proper MWEs. There are no large differences between results for different measure, but, with some caution, we can observe that results obtained with PMI are slightly better, while in the case of PMI the measure is naturally is filtered by 0 threshold and produces potentially more interesting collocations, thus harder to be distinguished from the proper MWEs. Our three representations allowed classifier to achieve significantly better results in comparison to the baseline approach focused on the component and MWE embedding. The context provided additional information on the MWE semantics, which improved the model performance. This is related to the noncompositional nature of the MWEs, which meaning cannot be inferred from their component meanings. Our approach based on difference and product vectors forced the models significantly reduced the training time. It may be more important in practice, when the training time and inference time are more important than the quality of prediction. On the other hand, the method based on contextual embeddings allows transforming any set of texts with the use of dictionary knowledge into an annotated corpus containing occurrences of the MWEs and their components. The model, by examining the semantic differences between the component and the entire expression, takes into account the variability of the context, which should allow for the extraction of the MWE meaning following the assumption of its monosemous character. In future work, we want to use our methods to generate corpora in other languages, which will be later used to train models in the multilingual MWEs detection task and to explore the transfer learning mechanism in a language-independent MWE detection.
| 1,411 | 1,998 | 1,411 |
ICE: Idiom and Collocation Extractor for Research and Education
|
Collocation and idiom extraction are wellknown challenges with many potential applications in Natural Language Processing (NLP). Our experimental, open-source software system, called ICE, is a python package for flexibly extracting collocations and idioms, currently in English. It also has a competitive POS tagger that can be used alone or as part of collocation/idiom extraction. ICE is available free of cost for research and educational uses in two user-friendly formats. This paper gives an overview of ICE and its performance, and briefly describes the research underlying the extraction algorithms.
|
Idioms and collocations are special types of phrases in many languages. An idiom is a phrase whose meaning cannot be obtained compositionally, i.e., by combining the meanings of the words that compose it. Collocations are phrases in which there is a semantic association between the component words and some restrictions on which words can be replaced and which cannot. In short, collocations are arbitrarily restricted lexeme combinations such as look into and fully aware. Many scientists from diverse fields have worked on the challenging tasks of automated collocation and idiom extraction, e.g., see To tackle this void, we introduce a feature-rich system called ICE (short for Idiom and Collocation Extractor), which has two versions: one is flexible and pipelined seamlessly for research purposes as a component of a larger system such as a question answering system, and the second as a web-based tool for educational purposes. ICE has a modular architecture and also includes a POS tagger, which can be used alone or as part of collocation or idiom extraction. An experiment with the CoNLL dataset shows that ICE's POS tagger is competitive against the Stanford POS tagger. For ease of use in research, we provide ICE as a python package. For collocation extraction, ICE uses the IR models and techniques introduced by For idiom extraction, ICE uses the semanticsbased methods introduced by Thus, ICE includes extraction methods for idioms and collocations that are state-of-the-art. Other tools exist for collocation extraction, e.g., see
|
As ICE's algorithms are based on Bing search, users must provide a valid user id for the Bing API. ICE receives a list of sentences as an input and outputs a list of all collocations and idioms. It first splits the input sentences using NLTK sentence tokenizer, then generates n-grams and part of speech tags. ICE's n-gram generator takes care of punctuation marks and has been shown to be better than NSP's n-gram generator. Finally, the output n-grams are given to the collocation and idiom detection algorithms. Collocation and idiom extraction has been done by the algorithm given by For collocations, ICE pipelines the first and second methods, then pipelines them with the third or the fourth method (both options are available in the code). These methods are connected sequentially. This means that if something is considered as a collocation in one component, it will be added to the list of collocations and will not be given to the next component (yes/no arrows in the diagram). Table The Ngram Extractor receives all sentences and generates n-grams ranging from bigrams up to 8grams. It uses NLTK sentence and word tokenizers for generating tokens. Then, it combines the generated tokens together taking care of punctuation to generate the n-grams. Dictionary Check uses WordNet The next component is Online Dictionary. It searches online dictionaries to see if the n-gram exists in any of them or not. It uses Bing Search API Web Search and Substitution is the next component in the pipeline. This method uses Bing Search API to obtain hit counts for a phrase query. Then each word in the n-gram will be replaced by 5 random words (one at the time), and the hit counts are obtained. At the end, we will have a list of hit counts. These values will be used to differentiate between collocations and non-collocations. The last component in the pipeline of collocation extraction is Web Search and Independence. The idea of this method is to check whether the probability of a phrase exceeds the probability that we would expect if the words are independent. It uses hit counts in order to estimate the probabilities. These probabilities are used to differentiate between collocations and non-collocations. When running the collocation extraction function, one of the components should be selected out of the third and fourth ones. The Idiom Extractor diagram is relatively simpler. Given the input n-gram, it creates n + 1 sets. The first contains (stemmed) words in the meaning of the phrase. The next n sets contain stemmed word in the meaning of each word in the n-gram. Then it applies the set difference operator to n pairs containing the first set and each of the n sets. The Or subsystem considers a phrase as an idiom if at least one word survives one of the subtractions (union of difference sets should be non-empty). For the And, at least one word has to exist that survived every subtraction (intersection of difference sets should be non-empty) Performance. ICE outperforms both Text-NSP and MWEToolkit. On the gold-standard dataset, ICE's F1-score was 40.40%, MWE-Toolkit's F1-score was 18.31%, and Text-NSP had 18%. We also compared our idiom extraction with AMALGr method Sample Code. Below is the sample code for using ICE's collocation extraction as part of a bigger system. For idiom extraction you can use IdiomExtractor class instead of collocationExtractor. with all keys on the line."] >>from ICE import CollocationExtractor >>extractor = CollocationExtractor. with_collocation_pipeline( "T1" , bing_key = "Temp", pos_check = False) >> print(extractor. get_collocations_of_length( input, length = 3)) >> ["on the line"] Educational Uses. ICE also has a web-based interface for demonstration and educational purposes. A user can type in a sentence into an input field and get a list of the idioms or collocations in the sentence. A screen-shot of the web-based interface is shown in Figure ICE is a tool for extracting idioms and collocations, but it also has functions for part of speech
| 606 | 1,548 | 606 |
Importance-Aware Data Augmentation for Document-Level Neural Machine Translation
|
Document-level neural machine translation (DOCNMT) aims to generate translations that are both coherent and cohesive, in contrast to its sentence-level counterpart. However, due to its longer input length and limited availability of training data, DOCNMT often faces the challenge of data sparsity. To overcome this issue, we propose a novel Importance-Aware Data Augmentation (IADA) algorithm for DOCNMT that augments the training data based on token importance information estimated by the norm of hidden states and training gradients. We conduct comprehensive experiments on three widely-used DOCNMT benchmarks. Our empirical results show that our proposed IADA outperforms strong DOCNMT baselines as well as several data augmentation approaches, with statistical significance on both sentence-level and document-level BLEU.
|
Document-level Neural Machine Translation (DOCNMT) has achieved significant progress in recent years, as evidenced by notable studies To overcome this issue, the data augmentation (DA) technology In this work, we combine IADA with two pop-ular data augmentation methods, word dropout
|
Document-Level NMT In recent years, numerous approaches have been proposed for documentlevel neural machine translation (DOCNMT). One early model, proposed by Data Augmentation Data augmentation (DA) approaches for NMT are commonly categorized into two classes, word replacement and back/forward translation. Ours Our novel Important-Aware Data Augmentation (IADA) method effectively encourages the DOCNMT model to leverage the contextual information. Our empirical results conform that IADA is compatible with the classical DA approaches, such as DOCBT and DOCFT. In this section, we introduce the task of DOCNMT in Section 3.1, our proposed IADA framework in Section 3.2, our token importance measures in Section 3.3, and our training objective in Section 3.4. The standard sentence-level NMT (SENTNMT) model ignores surrounding context information, whose probability of translation is defined as: where x x x i and y y y i are the i-th source and target training sentence, y i,t denotes the t-th token in y y y i and |•| indicates the sequence length. Different from SENTNMT, DOCNMT has the access to both current sentence and context sentences for translation. Given a document pair {X X X i , Y Y Y i }, we define X X X i = {C C C x x x i , x x x i } and Y Y Y i = {C C C y y y i , y y y i }, where x x x i and y y y i are the current sentence pair, and C C C x x x i and C C C y y y i are their corresponding context. The translation probability of y y y i in DOCNMT is: Because of paralysis, my grandmother's legs have stopped working. Today, she had another attack. Aufgrund von Lähmung haben die Beine meiner Großmutter aufgehört zu arbeiten. Heute hatte sie einen weiteren Anfall. Figure 2: An illustrative example of IADA. Strikethrough indicates perturbation. The "sie" is semantically connected to "she", "grandmother", and "Großmutter". IADA is inclined to mask "she" in the current sentence and other less-important words in the context. Tokens in blue are similarly affected by IADA. Existing DOCNMT models only demonstrate limited usage of the context Importance-Aware Data Augmentation (IADA) is built on top of this goal. Specifically, IADA first perturbs the important tokens in the current sentence to be translated, which encourages the DOCNMT models to recover those information using the document context. IADA then perturbs the less important tokens in the context, highlighting the useful contextual information. Note that these two steps can be performed simultaneously. As shown in Figure To implement this design, IADA perturbs the original document pair and obtain X Accordingly, the translation probability of a DOCNMT model with IADA is: IADA uses a token-specific replacement probability p i,t to determine the tokens to be replaced in these sentences. For example, the token x i,t in the source document X X X i is replaced: where Ω(•) could be an arbitrary replacement strategy. IADA can be incorporated with various existing replacement strategies. In this paper, we show the effectiveness of two versions of IADA, IADA DROP (with word dropout) and IADA REPL (with word replacement). Token-Specific Replacement Probability As discussed above, in IADA, the important tokens in the context should be assigned lower replacement probabilities, while the important tokens in the current sentence should be assigned higher replacement probabilities. Therefore, for the token x i,t in the source document X X X i , we define its corresponding p i,t as: where p ctx and p cur are the initial replacement probabilities for the context and current sentence respectively, and σ(•) is the sigmoid function whose output can be interpreted as a probability. Importance Normalization To properly control the spread of token importance scores, we propose to normalize the token importance score ψ(x i,t ) across all tokens in X X X i as: where ) is the original token importance score. α is the hyper-parameter that controls the spread of token importance scores. We also apply this normalization process to ψ(y i,t ) in the target documents. In this section, we discuss how IADA determines the word importance score ϕ(x i,t ) for the DOC-NMT training instances. Schakel and The meaning of a word is dynamic according to its surrounding context. Thus, we propose to use the norm of topmost layer hidden states h h h x i,t from encoder, which incorporates the context-aware information Likewise, given a target document Y Y Y i , we obtain importance score ϕ TNORM (y i,t ): where We use hidden states given by the topmost point-wise feedforward networks in the encoder or decoder to compute the TNORM, before the layer normalization Norm of Gradients (GNORM) TNORM is context-aware but ignores the source-target alignment information, as ϕ TNORM (x i,t ) in Equation We obtain the importance score ϕ GNORM (x i,t ): where L(X X X i , Y Y Y i , θ θ θ) is the loss function with the input of X X X i and Y Y Y i seeking for the optimal parameters θ θ θ. The identical process can be directly applied to y i,t . Note that the gradient g g g x i,t or g g g y i,t in this process is not used for updating θ θ θ. As described in Equation original loss, see Equation 13 agreement loss, see Equation As defined in Equation 2, the conventional training objective of the DOCNMT models for a document pair {X X X i , Y Y Y i }, namely the original loss, can be defined as: The perturb loss is defined in the same way for { X X X i , Ỹ Y Y i }. Furthermore, given the equivalence between the perturbed and original samples, we introduce an extra agreement loss, namely Jensen-Shannon divergence: where 4 Experiments We evaluate IADA against various competitive baselines from two categories, the DOCNMT baselines and the data augmentation baselines. with a number of data augmentation approaches: • Word Dropout (WORDDROP) Word dropout Datasets In our experiments, we evaluated the performance of our model on three English-German translation datasets: the small-scale benchmarks TED Evaluation We evaluate the translation quality using sentence-level SacreBLEU Inference We translate test examples in their original order, beginning with the first sentence independent of context. Previous translations serve as the context for the current translation. Hyperparameters All the approaches in this works, including IADA and baselines, are trained from scratch with the identical hyperparameters. The model is randomly initialized and optimized with Adam (Kingma and Ba, 2015) with β 1 = 0.9, β 2 = 0.98 and the learning rate α = 5×10 -4 . The model is trained with the batch size of 32K tokens for both datasets and the dropout rate p = 0.3. The batch size of 32K tokens is achieved by using the batch size of 4096 tokens and updating the model for every 8 batches. The learning rate schedule is the same as described in The model architecture for all the approaches in this work is Transformer-base We present the main results in Table Comparison with other approaches Our IADA DROP and IADA REPL models surpass other DOCNMT models in performance without requiring additional neural modules or incurring computational overhead. Moreover, IADA models also outperform other competitive DA approaches on both s-BLEU and d-BLEU. They exhibit substantial performance gains on all three benchmarks, demonstrating their effectiveness in training DOCNMT models for both low-resource and high-resource settings. In contrast, other DA approaches only exhibit marginal improvements on the large benchmark Europarl. ingly, we observe that WORDREPL-based methods (IADA REPL and WORDREPL) slightly outperform the WORDDROP-based methods (IADA DROP and WORDDROP). We hypothesize that WORDREPLbased methods generate more diverse synthetic data by replacing selected tokens with distinct random tokens, compared with replaceing selected tokens with ⟨MASK⟩. Lastly, we also observe that GNORM outperforms TNORM, confirming our hypothesis in Section 3.3. In this section, we conduct ablation studies to show the effectiveness of IADA components based on IADA REPL on the TED benchmark. Perturbation Strategy Our proposed perturbation strategy's effectiveness is demonstrated by enumerating all possible strategies for token importance measures in Table information in the current sentence. Results consistently indicate that all other perturbation strategies are suboptimal compared to our strategy. This success is attributed to the design of IADA, which encourages DOCNMT models to leverage contextual information. Token Importance Measures To demonstrate the effectiveness of our proposed importance measures, we replace ψ(•) in Equation Importance Normalization We examine the impact of importance normalization (Equation We analyze IADA from various aspects in this section, including contextual awareness, robustness, compatibility with DOCBT/DOCFT, simulated low-resource scenario, and linguistic analysis. Contextual Awareness In our analysis, we evaluate the contextual awareness of DOCNMT models using a contrastive test set. We focus on the accuracy of different anaphoric pronoun types (Table Simulated Low-Resource Scenario We also examine the usefulness of IADA in low-resource training scenarios. We vary the size of the training data (D trn ) for TED from 20% (around 40K) to 100% (around 200K). The performance gap (∆ {•} ) compared to the DOC2DOC model is shown in Figure In our experiment, we test the effectiveness of IADA in mitigating negative impacts of irrelevant and dis- ruptive context. We randomly replace two out of three sentences in the gold context of training instances with sentences from other documents. Results on TED (Table We analyze perturbed tokens from WORDREPL and IADA REPL with GNORM using linguistic analysis, focusing on five significant Part-Of-Speech (POS) tags. The results (Figure In this paper, we present IADA, a new method for generating high-quality syntactic data for DOC-NMT. By leveraging token importance, IADA augments existing training data by perturbing important tokens in the current sentences while keeping those less important ones unchanged. This encourages DOCNMT models to effectively utilize contextual information. We propose TNORM and GNORM to measure token importance. We also introduce the agreement loss to prevent the training samples from being overly hard to learn after perturbation. Results demonstrate that IADA outperforms competitive DOCNMT approaches as well as several data augmentation methods. Our analysis reveals that IADA enhances DOC-NMT models' contextual awareness, robustness, and is compatible with DOCBT and DOCFT techniques. IADA also shows significant benefits in low-resourced settings. Linguistic analysis validates the effectiveness of IADA in identifying important tokens. Overall, our findings highlight the efficacy of IADA in improving syntactic data generation for DOCNMT. Comparing with standard optimization techniques, our proposed IADA with the TNORM and GNORM requires additional forward and backward computation. For each training step, IADA with TNORM requires one additional forward pass, and IADA with GNORM requires one additional forward and backward pass. Note that IADA is only applied to the training stage and has no impact on the DOC-NMT inference.
| 827 | 283 | 827 |
Manually Constructed Context-Free Grammar For Myanmar Syllable Structure
|
Myanmar language and script are unique and complex. Up to our knowledge, considerable amount of work has not yet been done in describing Myanmar script using formal language theory. This paper presents manually constructed context free grammar (CFG) with "111" productions to describe the Myanmar Syllable Structure. We make our CFG in conformity with the properties of LL(1) grammar so that we can apply conventional parsing technique called predictive top-down parsing to identify Myanmar syllables. We present Myanmar syllable structure according to orthographic rules. We also discuss the preprocessing step called contraction for vowels and consonant conjuncts. We make LL (1) grammar in which "1" does not mean exactly one character of lookahead for parsing because of the above mentioned contracted forms. We use five basic sub syllabic elements to construct CFG and found that all possible syllable combinations in Myanmar Orthography can be parsed correctly using the proposed grammar.
|
Formal Language Theory is a common way to represent grammatical structures of natural languages and programming languages. The origin of grammar hierarchy is the pioneering work of Noam
|
Myanmar is a syllabic script and also one of the languages which have complex orthographic structures. Myanmar words are formed by collection of syllables and each syllable may contain up to seven different sub syllabic elements. Again, each component group has its own members having specific order. Basically, Myanmar script has 33 consonants, 8 vowels (free standing and attached) The following subsection will give the details of each sub syllabic group. Each Myanmar consonant has default vowel sound and itself works as a syllable. In writing formal grammar for a Myanmar syllable, there are some cases where two or more Myanmar characters combine each other and the resulting combined forms are also used in Myanmar traditional writing system though they are not coded directly in the Myanmar Unicode chart. Such combinations of vowel and medials are described in detail below. Two or more Myanmar attached vowels are combined and formed new three members { , , } in the vowel set. Description Vowel The above mentioned combinations of characters are considered as one vowel or medial in constructing the grammar. The complete sets of elements for vowels and meidals used in writing grammar are depicted in the table below. 3 , , , "Table As mentioned in the earlier sections, we choose only 5 basic sub syllabic components namely consonants (C), medial (M), attached vowels (V), vowel killer (K) and diacritics (D) to describe Myanmar syllable. As our intended use for syllabification is for sorting, we omit stand-alone vowels and digits in describing Myanmar syllable structure. Further, according to the sorting order of Myanmar Orthography, standalone vowels are sorted as the syllable using the above 5 sub syllabic elements having the same pronunciation. For example, stand-alone vowel "ဣ" is sorted as consonant "အ" and attached vowel " " combination as "အ ". 3 Sorting order of Medials and attached vowels in Myanmar Orthography In Myanmar language, a syllable with only one consonant can be taken as one syllable because Myanmar script is Abugida which means all letters have inherent vowel. And, consonants can be followed by vowels, consonant, vowel killer and medials in different combinations. One special feature is that if there are two consonants in a given syllable, the second consonant must be followed by vowel killer (K). We The combinations among five basic sub syllabic components can also be described using Finite State Automaton. We also find that Myanmar orthographic syllable structure can be described in regular grammar. "Figure In the above FSA, an interesting point is that only one consonant can be a syllable because Myanmar consonants have default vowel sounds. That is why, state 2 can be a final state. For instance, a Myanmar Word "မ န မ" (means "Woman" in English) has two syllables. In the first syllable "မ န ", the sub syllabic elements are Consonant(မ) + Vowel( ) +Consonant(န)+ Vowel Killer( )+Diacritics( ). The second syllable has only one consonant "မ". Context free (CF) grammar refers to the grammar rules of languages which are formulated independently of any context. A CF-grammar is defined by: 1. A finite terminal vocabulary V T . 2. A finite auxiliary vocabulary V A . 3. An axiom SV A . 4. A finite number of context-free rules P of the form A where AV A and {V A U V T }* (M. The following figure is a part of parse table made from the productions of the proposed LL(1) grammar. In the above table, the topmost row represents terminal symbols whereas the leftmost column represents the non terminal symbols. The entries in the table are productions to apply for each pair of non terminal and terminal. An example of Myanmar syllable having 4 different sub syllabic elements is parsed using proposed grammar and the above parse table. The parsing steps show proper working of the proposed grammar and the detail of parsing a syllable is as follows. "Table This study shows the powerfulness of Chomsky`s context free grammar as it can apply not only to describe the sentence structure but also the syllable structure of an Asian script, Myanmar. Though the number of productions in the proposed grammar for Myanmar syllable is large, the syntactic structure of a Myanmar syllable is correctly recognized and the grammar is not ambiguous. Further, in parsing Myanmar syllable, it is necessary to do preprocessing called contraction for input sequences of vowels and consonant conjuncts or medials to meet the requirements of traditional writing systems. However, because of these contracted forms, single lookahead symbol in our proposed LL(1) grammar does not refer exactly to one character and it may be a combination of two or more characters in parsing Myanmar syllable. Myanmar script is syllabic as well as aggulutinative script. Every Myanmar word or sentence is composed of series of individual syllables. Thus, it is critical to have efficient way of recognizing syllables in conformity with the rules of Myanmar traditional writing system. Our intended research is the automatic syllabification of Myanmar polysyllabic words using formal language theory. One option to do is to modify our current CFG to recognize consecutive syllables as a first step. We found that if the current CFG is changed for sequence of syllables, the grammar can be no longer LL(1). Then, we need to use one of the statistical methods, for example, probabilistic CFG, to choose correct productions or best parse for finding syllable boundaries. Again, it is necessary to calculate the probability values for each production based on the frequency of occurrence of a syllable in a dictionary we referred or using TreeBank. We need Myanmar corpus or a tree bank which contains evidence for rule expansions for syllable structure and such a resource does not yet exist for Myanmar. And also, the time and cost for constructing a corpus by ourselves came into consideration. Another approach is to construct finite state transducer for automatic syllabification of Myanmar words. If we choose this approach, we firstly need to construct regular grammar to recognize Myanmar syllables. We already have Myanmar syllable structure in regular grammar. However, for finite state syllabification using weights, there is a lack of resource for training database. We still have many language specific issues to be addressed for implementing Myanmar script using CFG or FSA. As a first issue, our current grammar is based on five basic sub-syllabic elements and thus developing the grammar which can handle all seven Myanmar sub syllabic elements will be future study. Our current grammar is based on the code point values of the input syllables or words. Then, as a second issue, we need to consider about different presentations or code point values of same character. Moreover, we have special writing traditions for some characters, for example, such as consonant stacking eg. ဗု ဒ္ဓ (Buddha), မန ္တလေး (Mandalay, second capital of Myanmar), consonant repetition eg. က (University), kinzi eg. အင်္ ဂ (Cement), loan words eg. ဘတ် (စ် ) (bus). To represent such complex forms in a computer system, we use invisible Virama sign (U+1039). Therefore, it is necessary to construct the productions which have conformity with the stored character code sequence of Myanmar Language.
| 994 | 185 | 994 |
Are Missing Links Predictable? An Inferential Benchmark for Knowledge Graph Completion
|
We present InferWiki, a Knowledge Graph Completion (KGC) dataset that improves upon existing benchmarks in inferential ability, assumptions, and patterns. First, each testing sample is predictable with supportive data in the training set. To ensure it, we propose to utilize rule-guided train/test generation, instead of conventional random split. Second, InferWiki initiates the evaluation following the open-world assumption and improves the inferential difficulty of the closed-world assumption, by providing manually annotated negative and unknown triples. Third, we include various inference patterns (e.g., reasoning path length and types) for comprehensive evaluation. In experiments, we curate two settings of InferWiki varying in sizes and structures, and apply the construction process on CoDEx as comparative datasets. The results and empirical analyses demonstrate the necessity and high-quality of InferWiki. Nevertheless, the performance gap among various inferential assumptions and patterns presents the difficulty and inspires future research direction. Our datasets can be found in
|
Knowledge Graph Completion (KGC) aims to predict missing links in KG by inferring new knowledge from existing ones. Attributed to its reasoning ability, KGC models are crucial in alleviating the KG's incompleteness issue and benefiting many downstream applications, such as recommendation In this paper, we re-think the task of KGC and construct a new benchmark dubbed InferWiki that highlights three fundamental objectives: Test triples should be inferential: this is the essential requirement of KGC. Each test triple should have supportive samples in the train set. However, we observe two major issues of current KGC datasets: unpredictable and meaningless test triples, which may hinder evaluating and advancing stateof-the-arts. As shown in Table Test triples may be inferred positive, negative, or unknown. Following open-world assumption: what is not observed in KG is not necessar- ily false, but unknown Inference has various patterns. Concentrating on limited patterns in evaluation may bring in severe bias. Domain-specific datasets Kinship To this end, we curated an Inferential KGC dataset extracted from Wikidata and establish the benchmark with two settings of varying in sizes and structures: InferWiki64k and InferWiki16k. Instead of random split, we mine rules via Any-BURL Our contributions can be summarized as follows: • We summarize three principles of KGC: inferential ability, assumptions and patterns, and construct a rule-guided dataset. • We highlight the importance of negatives and unknowns, and initiate open-world evaluation. • We conduct extensive experiments to establish the benchmark. The results and deep analyses verify the necessity and challenge of Infer-Wiki, providing insights for future research.
|
We can roughly classify current KGC datasets into two groups: inferential and non-inferential datasets. The first group is usually manually curated to ensure each testing sample can be inferred from training data through reasoning paths, while they only focus on specific relations, such as Families The limited scale and inference patterns make them not challenging. HOLE The second group of datasets are automatically derived from public KGs and randomly split positive triples into train/test, leading to a risk of testing samples non-inferential from training data. Popular datasets include FB15k-237 In fact, inference is an important ability for intelligence. Various fields study how inference is done in practice, ranging from logic to cognitive psychology. Inference helps people make reliable predictions, which is also an expected ability for AI models. Indeed, once deployed, a model may have to make a prediction when there is no evidence in the training set. But, instead of an unreliable guess, we highlight the ability to know unknown, a.k.a. open-world assumption. Therefore, we aim to curate an large-scale inferential benchmark InferWiki including various inference patterns and testing samples (i.e., positive, negative, and unknown), for better evaluation. We list the statistics in Table We describe our dataset construction that comprises four steps: data preprocessing, rule mining, ruleguided train/test generation, and inferred test labeling. We then give a detailed analysis. More and more studies utilize Wikidata 1 as a knowledge resource due to its high quality and large quantity. We utilize the September 2019 English dump in experiments. Data preprocessing aims to define relation vocabulary and extract two sets of triples from Wikidata: a large one for rule mining T r and a relatively small one for dataset generation T d . The reason for using two sets is to avoid the leakage of rules. In other words, some frequent rules on the large set may be very few on the small set. The different distributions shall avoid that rule mining methods will easily achieve high performance. Besides, more triples can improve the quality of mined rules. In contrast, the relatively small set is enough for efficient KGC training and evaluation. In specific, we first extract all triples that consist of two entity items and one relation with English labels. We then remove the repeated triples and obtain 40,199,175 triples with 7,734,841 entities and 1,170 different relation types. Considering rule mining efficiency, we reduce the relation vocabulary by (1) manually filtering out meaningless relations, such as movie ID or film rating, (2) removing relations of InstanceOf and subClassOf following existing benchmarks 1 Since developing advanced rule mining models is not the focus of this paper and several mature tools are available online, such as AMIE+ Given a set of triples (i.e., the large set T r ), this step aims to automatically learn rules F = {(f p , λ p )} P p=1 , where f p denotes a horn rule, e.g., spouse(x, y) ∧ father(x, z) ⇒ mother(y, z), and λ p ∈ [0, 1] denotes the confidence of f p . For each rule f p , the left side of ⇒ is called the premise, and the right side is called the conclusion, where the conclusion contains a single atom and the premise is a conjunction of several atoms in the Horn rule scheme. We can ground specific entities to replace x, y, z in f p , which shall denote an inferential relationship between premise and conclusion triples. For example, given spouse(LeBron James, Savannah Brinson) and father(LeBron James, Bronny James), we may infer a new triple mother(Savannah Brinson, Bronny James). Of course, not all of the mined rules are reasonable. To alleviate the negative impacts of unreasonable rules, we rely on more data (a large set of triples) and keep high-confidence rules only. Particularly, we follow the suggested configuration of AnyBURL. We run it for 500 seconds to ensure that all triples can be traversed at least once and obtain 251,317 rules, where 168,996 out of them whose confidence meets λ p > 0.1 have been selected as the rule set to guide dataset construction. Different from existing benchmarks, InferWiki provides inferential testing triples with supportive data in the training set. Moreover, it aims to include as many inference patterns as possible and these patterns are better evenly distributed to avoid biased evaluation. Thus, this step has four objectives: ruleguided split, path extension, negative supplement, and inference pattern balance. Rule-guided Split grounds the mined rules F on triples T d to obtain premise triples and corresponding conclusion triples. All premise triples form a training set, and all conclusion triples form a test set. Thus, they are naturally guaranteed to be inferential. For correctness, all of premise triples must exist in the given triple set T d , while conclusion triples are not necessarily in T d and may be generated for further annotation (i.e., Section 3.4). For example, given a rule spouse(x, y) ∧ father(x, z) ⇒ mother(y, z), we traverse all of the given triples and find entities LeBron James, Savannah Brinson, and Bronny James that meet the premise. We then add the premise triples spouse(LeBron James, Savannah Brinson) and father(LeBron James, Bronny James) into the training set, and generate the conclusion triple mother(Savannah Brinson, Bronny James) for testing, no matter it is given or not. Path Extension aims to increase the inference path patterns by (1) adding more reasoning paths for the same testing triple, and (2) elongating paths by replacing those premise triples that have reasoning paths. For example, we replace father(LeBron James, Bronny James) with two triples that can infer it: father(LeBron James, Bryce James) and brother(Bronny James, Bryce James). The original path is then extended by one hop. Correspondingly, we define the confidence of extended paths as the multiplication of all involved rules. Longer paths will challenge long-distance reasoning ability. Negative Supplement is to generate negative triples if we cannot annotate the same number of negatives with positive triples. Otherwise, we will face an imbalance issue. Following conventions, we randomly corrupt the head or tail entities in a positive triple with the following constraints: (1) the relation of the positive triple is exclusive, e.g., placeOfBirth, if the ratio from head to tail entities is smaller than a threshold (we choose 1.2 heuristically in experiments); otherwise, the corrupted negative triple may be actually positive, leading to false negative errors. (2) We choose positive triples from the test set for corruption to improve the difficulty -the model has to correctly infer the corresponding positive triple from training data, then classify the corrupted triple as negative through the confliction. Particularly, for non-exclusive relation types, most of their corrupted results should be unknown following open-world assumption. The inferred test set covers such cases, which will be discussed in Section 3.4. Inference Pattern Balance aims to balance various inference patterns, including path length, relation types, and relation patterns Different from existing datasets, InferWiki aims to include positive, negative, and unknown testing triples, to evaluate the model under two types of assumptions: open-world assumption and closedworld assumption. The main difference between them is whether unknown triples are regarded as negatives. That is, the open-world evaluation is a three-class classification problem (i.e., positive, negative, and unknown). The closed-world evaluation targets only positive and negative triples, and we can simply relabel unknown triples as negatives without changing the test set. So far, we have two test sets: one is generated via rule guidance, and the other contains the supplemented negatives. This section aims to label the generated triples. First, we automatically label the triples with positive if they exist in Wikidata. Then, we manually annotate the remaining 4,053 triples. The annotation guideline can be found in Appendix B. Note that all of the unknowns are factually incorrect but not inferential. To assess the quality of annotations, we verify a random selection of 300 test triples (100 for each label). The annotators agree with our labels 84.3% of the time. We further investigate the disagreements by relabeling 100 samples. 85% of the time, humans prefer an unknown, while automatic labeling tends to assign them with positive or negative labels. This suggests the inferential difference between humans and machines -the capacity of knowing unknown. Finally, we remove the entities that are not in any of the grounded paths and their triples. We randomly select half of the test set as valid. This forms InferWiki64k. We further extract a dense subset InferWiki16k by filtering out the positive triples whose confidence is smaller than 0.6. Correspondingly, negative/unknown triples are reduced to keep balance. The statistics is listed in Table The answer is yes but not exactly. As shown in Table Next, we analyze the relation patterns and path length distribution through comparisons with existing KGC datasets. Due to the different construction pipelines, existing datasets are difficult to offer quantitative statistics. We thus apply our pipeline on CoDEx Further analysis of relation, entity and neighbor distributions can be found in Appendix D&E. Although we carefully design the construction of in-ferWiki, there are still two types of limitations: rule biases and dataset errors, that can to be addressed along with the development of KG techniques in the future. In terms of rule biases, AnyBURL may be over-estimated due to its role in the construction. Although we utilize two triple sets to avoid rule leakage, their overlap may still bring unfair performance gain to AnyBURL. We consider synthesize several rule mining results to improve InferWiki in the next version. In terms of dataset errors, first, to balance positive and negative triples in the larger InferWiki64k, we follow conventions to randomly sample a portion of negatives. These negatives may be unknown if following open-world assumption. We manually assess the randomly sampled negatives and find a 15.7% error rate. Therefore, we conduct open-world experiments on the smaller InferWiki16k, all of whose testing negatives are verified by humans. The second type of errors is due to unreasonable rules for dataset split, which is caused by prediction errors of existing rule mining models. However, there is no suitable evaluation in this field to provide quantitative analysis. Our ongoing work aims to develop an automatic evaluation for path rationality to improve the mining quality, and thus facilitate our inferential pipeline. We benchmark performance on InferWiki for the tasks: (1) Link Prediction, the task of predicting the missing head/tail entity for a given query triple (?, r, t) or (h, r, ?). Models are encouraged to rank correct entities higher than others in the vocabulary. We adopt the filtering setting (Bordes et al., 2013) that excludes those entities, if the predicted triples have been seen in the train set. Mean reciprocal rank (MRR) and hits@k are standard metrics for evaluation. (2) Triple Classification aims to predict a label for each given triple (h, r, t). The label following open-world assumption is trinary y ∈ {-1, 0, 1} and becomes binary y ∈ {-1, 1} when adopting closed-world assumption -all 0-label triples are re-labeled with -1, since our unknown triples are factually negative yet non-inferential from training data. Since KGC models output real-value scores for triples, we classify scores into labels by choosing one or two thresholds per relation type on valid. Accuracy, precision, recall, and F1 are measurements. For comprehensive comparison, we choose three types of representative models as baselines: (1) Knowledge Graph Embedding models, including TransE Table Figure Since most baselines fail in judging unknown as negative, we now investigate them following open-world assumption to see their ability in recog- nizing unknown triples. Table Table InferWiki16k CoDEx-m-infer MRR Hit@1 Hit@10 MRR Hit@1 Hit@10 MRR Hit@1 Hit@10 dataset construction, although we utilize two sets of triples to minimize rule leakage. Actually, inference of rules may be more important than we thought to improve the reliability and interpretability of knowledge-driven models. This also motivates us to incorporate rule knowledge into KGC training for advanced reasoning ability Figure We present the Hit@1 tail prediction on Infer-Wiki64k regarding relation patterns in Table We investigate the impacts of rule-based train/test generatation by comparing CoDEx-m-infer with Hier Comp Others TransE .000 . Table We illustrate the most frequent relation types and their distribution of InferWiki64k and Infer-Wiki16k in Figure We highlighted three principles for KGC datasets: inferential ability, assumptions, and patterns, and contribute a large-scale dataset InferWiki. We established a benchmark with three types of seven KGC models on two tasks of triple classification and link prediction. The results present a detailed analysis regarding various inference patterns, which demonstrates the necessity of an inferential guarantee for better evaluation and the difficulty of new open-world triple classification. In the future, we are interested in cross-KGs inference and transfer The second group of datasets are automatically derived from public KGs and randomly split positive triples into train/valid/test, leading to a risk of testing samples non-inferential from training data. FB13 We provide the following annotation guidelines for annotators to label inferred triples in Section 3.4. Task This is a two-step annotations. First, you must annotate each triple with the label y ∈ {1, -1}, where 1 denotes that the triple is correct and -1 denotes that the triple is incorrect. You can find the answer from anywhere you want, such as commonsense, Wikipedia, and professional websites. If you cannot find any evidence to support the statement, you shall choose label -1. Second, you must annotate each incorrect triple with the label ŷ ∈ {0, -1}, where 0 denotes that you do not know the answer. Now, you can find the answer from our provided triples. If you cannot find any evidence to support the statement, you shall choose label 0. Examples Here are some examples judged using three types of knowledge sources. • Commonsense: • Professional websites: To annotate the triple (Bahrain-Merida 2019, hasPart, Carlos Betancur), you may search the person in professional websites, such as • Wikipedia: Given the triples (Tōkaidō Shinkansen, connectsWith, Osaka Higashi Line) and (Tōkaidō Shinkansen, con-nectsWith, San'yō Main Line), you can find related station information from the page of Tōkaidō Shinkansen. You can find that Osaka Higashi Line shares a transfer station with Tōkaidō Shinkansen, thus label it with 1. And, San'yō Main Line doesn't show up in the page, you may label it with -1. InferWiki is able to analyze relation patterns for each path, including symmetry, inversion, hierarchy, and composition, where detailed explanations and examples are listed in Table We illustrate the most frequent relation types and their distribution of InferWiki64k and Infer-Wiki16k in Figure Figure Our experiments are run on the server with the following configurations: OS of Ubuntu 16.04.6 LTS, CPU of Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, and GPU of GeForce RTX 2080 Ti. We use OpenKE
| 1,099 | 1,741 | 1,099 |
A Multimodal Framework to Detect Target Aware Aggression in Memes
|
Internet memes have gained immense traction as a medium for individuals to convey emotions, thoughts, and perspectives on social media. While memes often serve as sources of humor and entertainment, they can also propagate offensive, incendiary, or harmful content, deliberately targeting specific individuals or communities. Identifying such memes is challenging because of their satirical and cryptic characteristics. Most contemporary research on memes' detrimental facets is skewed towards highresource languages, often sidelining the unique challenges tied to low-resource languages, such as Bengali. To facilitate this research in lowresource languages, this paper presents a novel dataset MIMOSA (MultIMOdal aggreSsion dAtaset) in Bengali. MIMOSA encompasses 4,848 annotated memes across five aggression target categories: Political, Gender, Religious, Others, and non-aggressive. We also propose MAF (Multimodal Attentive Fusion), a simple yet effective approach that uses multimodal context to detect the aggression targets. MAF captures the selective modality-specific features of the input meme and jointly evaluates them with individual modality features. Experiments on MIMOSA exhibit that the proposed method outperforms several state-of-theart rivaling approaches. Our code and data are available at
|
Recently, the rise of social media has given prominence to a distinct multimodal phenomenon known as meme, a composition of an image coupled with concise textual content. While memes are often humorous, they can propagate hate, offense, and aggression by incorporating political or cultural elements. Such undesired memes pose a significant threat to social harmony, as they can potentially harm individuals or specific groups based on their political philosophy, sexual orientation, religious beliefs, and more. As memes have become crucial in influencing social interactions, there has been a notable rise in research focused on meme analysis. This research includes analyzing the emotions Bengali memes have gained significant traction recently, reaching a broad audience and influencing public opinion while promoting negativity and violence. Detecting objectionable Bengali memes is currently in the developing stage due to the limited availability of tools such as OCR. Nonetheless, two works With the motivation mentioned above, we develop a novel corpus of Bengali memes encompassing various levels of aggression. On the technical front, prior studies reveal that state-of-the-art multimodal systems, effective in many visual-linguistic tasks, struggle with meme analysis. Memes rely heavily on context and often lack a clear connection between visual and textual elements. Moreover, memes contain much noise, making them distinct from other, more structured multimodal data. To tackle these issues, we develop a multimodal attentive fusion-based model to identify the targets of aggression within these memes. Our significant contributions are as follows. • We develop a novel multimodal aggression dataset MIMOSA consisting of 4,848 Bengali memes labeled with four aggression (Political, Gendered, Religious, and Others) and one non-aggressive class. • We propose MAF, a simple yet effective multimodal fusion approach that utilizes the attentive multimodal representation of the input meme and the individual modality-specific features to learn the subtle aggression elements better. • Finally, we perform extensive experiments on MIMOSA and show that MAF outperforms eleven state-of-the-art unimodal and multimodal baselines in terms of all the evaluation measures.
|
This section demonstrates the previous studies that have already been conducted on objectionable content (i.e., hate, offense, and aggression) detection based on unimodal and multimodal content. Unimodal Based Objectionable Content Detection: Most research on objectionable content detection (OCD) focused on analyzing textual data. Over the years, the topic has become a prominent research issue among researchers of different languages (2020) and Various methods were employed over the years for hate, aggression, and offense detection. Earlier studies used machine learning In contrast to only text and image-based OCD, several works have been accomplished considering the multimodal information in recent years. Differences with existing researches: While there has been significant progress in multimodal hate speech and offensive content detection, a notable gap exists in the research landscape regarding multimodal aggression detection, especially in lowresource languages (i.e., Bengali). Our investigation revealed that only two works Per our exploration, no benchmark dataset is explicitly developed for identifying aggression and its targets from the multimodal data. To fill this void, we developed MIMOSA: a novel target-aware multimodal aggressive memes dataset in Bengali. To create MIMOSA, we followed the guidelines provided by the Following existing works on aggression detection Aggressive memes can be offensive or hateful, but not all offensive or hateful memes represent aggression. Offensive content We have collected memes from various social media platforms and online sources to create the dataset. To ensure representativeness and reduce biases to a particular source, we collected data from diverse sources (e.g., Facebook, Instagram, Pinterest, and different Bengali Blogs). Figure A set of keywords such as "Bengali Memes," "Bengali Funny Memes," "Bengali Offensive Memes," "Bengali Aggressive Memes," "Bengali Troll Memes," "Bengali Political Memes," "Bengali Political Troll Memes," "Bengali Feminism Troll Memes," "Bengali Islam Troll Memes," "Bengali Hinduism Troll Memes," and "Bengali Celebrity Troll Memes" were used to search the memes. We used neutral keywords not explicitly tied to specific aggression themes to reduce biases to any specific category. Despite our best efforts, the dataset may have inherent biases, a common challenge in the development process. We collected the memes only from public domains, social media pages, and groups to avoid copyright infringement. Through this search process, 4,980 memes were collected from March 2022 to February 2023. During the data accumulation period, we have discarded memes that fall under the following categories: (i) memes that have information from only one modality (either visual or textual), (ii) memes that contain cartoons (as AI systems often face difficulty to process them), and (iii) memes that are visibly unclear (blurred). Figure A.2 illustrates some filtered samples. We discarded 132 memes based on the above criterion and finished with a total of 4,848 memes. Afterward, we extract the meme caption using an OCR 1 . However, we manually checked the extracted captions to correct any missing words and spelling as OCR in Bengali is not well-established. Finally, the memes and their associated captions are forwarded to the annotators to start the annotation process. MIMOSA was manually labeled into five categories: four aggression targets categories (political aggression (PAg), religious aggression (RAg), gendered aggression (GAg), others (Oth)) and a non-1 After reviewing existing works on aggression detection The annotators were asked to adhere to the class definitions to ensure labeling consistency. Initially, the annotators were asked to determine whether the meme was aggressive or non-aggressive based on the class definition. If an aggressive meme is discovered, they were instructed to further categorize it into one of the specific aggression targets. The annotators were also asked to provide reasoning for annotation decisions, which the expert will use as a reference in cases of disagreement. Initially, the annotators were trained with a small set of memes before being given a more extensive set to annotate independently. The training assisted in familiarizing the annotators with the task and ensuring consistency in their decisions. Three annotators (computer science undergraduates) each performed manual annotation, and the labels were verified by an expert (a professor with more than 20 years of research experience in NLP). More details of the annotators and the annotation process are provided in the Appendix B. To assess annotation quality, we used inter-annotator agreement metrics like Cohen's kappa coefficient For model training and evaluation, the dataset is divided into train (70%), validation (15%), and test (15%) sets. This section describes the proposed multimodal framework for target-aware aggression identification. The system takes memes and their corresponding caption as input. We employed state-of-the-art models to encode the memes' visual and textual information. Afterward, we use an attentive fusion mechanism to create a multimodal representation by selectively focusing on the encoded visual and textual features. Figure To encode the visual information of the memes, we use the image encoder of a pre-trained visual- To make a multimodal representation, the obtained visual and textual vector representations are fused using a multi-head self-attention (MSA) block This section discusses the baselines and their performance comparison with the proposed method (MAF). We will also illustrate the proposed approach's superiority by examining the errors. To experiment with MIMOSA, we developed several state-of-the-art computational models, including unimodal visual models, unimodal textual models, and multimodal models pre-trained on both modalities. We use two primary metrics for the evaluation: weighted f 1 -score (WF1) and macro-averaged mean absolute error (MMAE) To validate the performance of the proposed multimodal framework, we develop several models considering unimodal information (only visual or textual) and multimodal information (visual and textual). For the unimodal visual-only models, we employed three well-known architectures: ResNet50 • Early Fusion: We combine the intermediate feature representations of ViT and the Bangla-BERT model for the early fusion approach. • Late Fusion: The softmax prediction scores of the ViT and Bangla-BERT models are utilized to construct the late fusion model. • CLIP: It is a multimodal model trained on noisy image-text pair using contrastive learning • BLIP: BLIP (Bootstrapping Language-Image Pre-training) • ALBEF: ALBEF (Align Before Fuse) In the case of the CLIP and BLIP models, we extract the visual and textual embedding representations by fine-tuning them on the developed dataset. Afterward, we combined both representations and trained them on top of a softmax layer. Table Classwise Models Performance: To see the performance across different aggression target classes, we further investigate the classification reports (shown in Figure This outcome further demonstrates the efficacy of the proposed method in identifying the targets of aggressive memes. The results showed that the proposed MAF is superior in identifying the targets of aggressive memes more accurately compared to the only visual and textual approach. However, to examine the mistakes of the proposed method, we perform a detailed error analysis using quantitative and qualitative ways. We also consider the best visual and textual models for better demonstration. Quantitative Analysis: To perform quantitative analysis, we use the confusion metrics of the models shown in Figure This paper presented a novel multimodal dataset, MIMOSA, consisting of 4,848 memes, for detecting the targets of Bengali aggressive memes into five classes. This research also proposed a multimodal deep neural network MAF for the down-stream task. Experiments on MIMOSA demonstrated the efficacy of MAF outperformed eleven state-of-art unimodal and multimodal baselines. We plan to extend the dataset for more domains and languages. The future aim is to investigate the proposed model's performance on other datasets to enhance its generalization capabilities. Though the proposed method (MAF) demonstrates superior performance, there still exist some constraints. First, it is likely that in some cases, the MAF may focus on irrelevant parts of the visual and textual features during attentive fusion. For example, suppose the dataset contains misleading captions or irrelevant textual information. In that case, the attention mechanism might align with those parts of the image that are visually unrelated, leading to biased representations and thus providing suboptimal results. Second, upon analyzing the misclassified memes, we observed that the proposed MAF struggled with memes that contained subtle or sarcastic content. Addressing the challenge of mitigating bias and obtaining accurate annotations is a pivotal concern when labeling a dataset
| 1,314 | 2,277 | 1,314 |
BRIO: Bringing Order to Abstractive Summarization
|
Abstractive summarization models are commonly trained using maximum likelihood estimation, which assumes a deterministic (onepoint) target distribution in which an ideal model will assign all the probability mass to the reference summary. This assumption may lead to performance degradation during inference, where the model needs to compare several system-generated (candidate) summaries that have deviated from the reference summary. To address this problem, we propose a novel training paradigm which assumes a non-deterministic distribution so that different candidate summaries are assigned probability mass according to their quality. Our method achieves a new state-of-the-art result on the CNN/DailyMail (47.78 ROUGE-1) and XSum (49.07 ROUGE-1) datasets. Further analysis also shows that our model can estimate probabilities of candidate summaries that are more correlated with their level of quality. 1
|
Neural methods for abstractive summarization
|
To understand whether existing models can accurately perform such relative comparisons, we conducted a preliminary study on pre-trained BART We introduce a training paradigm which requires the abstractive model to be able to be accurate with respect to predicting the tokens in the reference summaries and coordinated with respect to Figure the candidate summaries. In other words, we give the abstractive model a dual role: as a generation model, it generates the output summaries in an autoregressive way; as an evaluation model, it can be used to score the quality of candidate summaries by estimating a probability distribution over candidate outputs. The generation model is trained using the standard MLE loss, but to train the evaluation model we introduce a contrastive loss Our main contribution is to change the target distribution of abstractive models from a one-point deterministic distribution assumed by MLE training to a non-deterministic distribution in which candidate summaries are also assigned probability mass according to their quality. The new SOTA performance on The goal of abstractive summarization is to create a function g that takes a source document D and generates an appropriate summary S S ← g(D) (1) Training Objective Neural abstractive summarization models aim to learn a neural model g that results in good summaries. Maximum likelihood estimation (MLE) is the standard training algorithm. It aims to maximize the likelihood of the reference summary S * , i.e., where θ denotes the parameters of g and p g θ denotes the probability distribution entailed by these parameters. The summation is over the training set and {D (i) , S * (i) } is the i-th training sample. For a specific sample {D (i) , S * (i) }, Eq. 2 is equivalent to minimizing the sum of negative loglikelihoods of the tokens {s in the reference summary S * whose length is l, which is the cross-entropy loss: where S * <j denotes the partial reference sequence {s * 0 , • • • , s * j-1 } and s * 0 is a pre-defined start token. p true is a one-hot distribution under the standard MLE framework: In practice, label smoothing where N is the size of the dictionary. Inference and Exposure Bias During inference, the abstractive model g is used to generate the candidate summary in an autoregressive manner. It is intractable to enumerate all the possible candidate outputs, so in practice methods such as beam search are used to reduce the search space. One important step in search is estimating the probability of the next word s t given the previous predicted sequence S <t : Comparing Eq. 6 with Eq. 3, the major difference is that during inference the model makes new predictions based on its own previous predictions S <t instead of the reference S * <t . As a result, even if the generation model g achieves very high accuracy w.r.t. Eq. 3, once S <t starts to deviate from S * , there is the risk that the performance of g will significantly degrade. This problem has been identified as the exposure bias Eq. 6 implies that the abstractive model g should be able to assign higher estimated probability to the better candidate summary during inference. However, this intuition is not directly captured in the standard MLE objective used in training -a model obtaining zero MLE loss would assign zero probability to any candidate summary different from the reference. This is obviously improper for any task where multiple reasonable generations may exist We next describe precisely how we encourage coordination through contrastive learning. The candidate quality measure M can be defined in many ways. In this work we define it as the ROUGE where S i and S j are two different candidate summaries and ROUGE(S i , S * ) > ROUGE(S j , S * ), ∀i, j, i < j. λ ij is the margin multiplied by the difference in rank between the candidates, i.e., where α is the length penalty hyperparameter. This loss gives the abstractive model a dual purpose, first as a reference-free evaluation model, which can be used in a two-stage summarization pipeline, where it is used to score the candidates generated by a pre-trained generation model and select the final output from them. However, since the autoregressive generation depends on both the token-level prediction accuracy and sequencelevel coordination, the model fine-tuned with the contrastive loss alone can no longer be used as a generation model. Multi-task Fine-tuning Following where γ is the weight of the contrastive loss. We note that the contrastive and the cross-entropy loss can effectively complement each other -since the contrastive loss is defined on the sequence level, the token-level cross-entropy loss serves as a normalization to ensure that the model could assign balanced probability mass across the whole sequence. Training Methods of Seq2Seq Models In order to align the training objective and evaluation metric, structured losses have been used for the Seq2Seq model training. Among them, marginbased losses One main challenge of directly optimizing a Seq2Seq model with quality scores of the output is that the discrete sampling process makes the loss non-differentiable. To circumvent this problem, reinforcement learning has been used to reformulate the conditional text generation tasks Datasets We mainly use three datasets in our experiments (statistics in Appendix A). CNNDM 4 (Hermann et al., 2015) is a large scale news dataset. Following The results are shown in Tab 2. For CNNDM and NYT we use BART as the backbone model while for XSum we use the pre-trained PEGASUS model as our base model since it achieves better performance than BART. We have the following observations: (1) BRIO-Ctr outperforms SimCLS, its counterpart as an evaluation model in a two-stage summarization framework. Specifically, both BRIO-Ctr and SimCLS are used to score the candidate summaries generated by a Seq2Seq abstractive model (BART). The final outputs are selected based on those scores. We attribute BRIO-Ctr's superior performance to its use of the same model architecture (BART) for both candidate generation and scoring, while SimCLS uses RoBERTa as the evaluation model. As a result, BRIO-Ctr maximizes the parameter sharing between the two stages, and preserves the power of the Seq2Seq model pre-trained on the same dataset. (2) BRIO-Mul is able to establish the new stare-of-the-art performance on CNNDM. Notably, the previous state-of-the-art model, GSum, takes additional guidance as input and needs a separate encoder to encode the guidance information, while BRIO-Mul uses the same parameterization of BART. Compared to other methods (ConSum, SeqCo, GOLD) that aim to improve upon BART, BRIO-Mul performs much better, showing the effectiveness of our training method. (3) Since on XSum we use PEGASUS instead of BART as the base model, the result shows that our method is not restricted to the specific choice of the base model. We further perform some in-depth analyses from diverse perspectives on the CNNDM dataset to gain more insights into our proposed method. Generation-Finetuning as a Loop Since the fine-tuned model (BRIO-Mul) is still able to gen- Increasing the Beam Width While theoretically a larger beam width (i.e. the number of candidates maintained during beam search) would allow more candidates to be considered and therefore increase the upper bound of the performance, in practice model performance may be lower if the beam width is too large. The reason for this phenomenon is closely related to the low sequence-level coordination of the generator. Specifically, increasing the beam width may introduce candidates with lower quality The results in Tab. 6 show that (1) Our model can significantly improve the model performance when either ROUGE or BERTScore is used as the target evaluation metric for ordering candidate summaries. This suggests that it is possible to use our method to optimize any specific target metric, making our method an alternative to reinforcement learning or minimum risk training. (2) Our model that is trained on one evaluation metric (e.g. BERTScore) also achieves improvement on another metric (e.g. ROUGE) compared with the baseline model, which indicates that the improvement made by our model is not from exploiting the potential weaknesses of individual metrics. Besides, this result also demonstrates a non-trivial degree of agreement between ROUGE and BERTScore. Novel n-grams We compare the ratio of novel n-grams in reference, BRIO-Mul's, and BART's summaries. As Tab. 7 shows, our model is more "abstractive" compared to BART, although reference summaries still contain more novel n-grams. This is likely due to the fact that our model is optimized at the sequence-level, allowing more freedom for paraphrasing and compression. We further investigate the relation of the "abstractiveness" and model performance by com- paring our model (BRIO-Mul) with the baseline model (BART) on different buckets of test examples grouped by the "novelty" of the reference summaries, where D and S * are the source document and reference summary respectively, G D and G S * are the sets of bigrams in D and S * , 1 is the indicator function. The results in Fig. (2) our model achieves larger improvement over the baseline model. Rank Correlation We computed the rank correlation between the estimated probabilities of the candidate summaries calculated by the generators and the quality scores of the candidate summaries. We use Eq. 9 to calculate the estimated probabilities Calibration requires that a model's confidence on its predictions is equal to the accuracy of these predictions We investigate this relation from the opposite direction by evaluating whether our model (BRIO-Mul), which is trained to have better sequencelevel performance, would also be more calibrated at the token-level compared with the baseline models that are trained using MLE and label smoothing. We follow previous work by using the Expected Calibration Error where the samples are grouped into M equal-width buckets by confidence (conf), B m denotes the m-th bucket, and n is the total number of samples. Following The results in Tab. 9 show that BRIO-Mul is better calibrated compared to BART, suggesting that our method helps to improve the token-level calibration by explicitly encouraging the model to have more accurate sequence-level probability estimations. The reliability graph is shown in Fig. The training paradigm proposed in this paper may be extended to any Seq2Seq model. However, it can be a non-trivial overhead to generate the candidate summaries using large neural models on the entire training set. On the other hand, recent work BRIO-Mul chelsea beat manchester city 3-1 in the youth cup final at the etihad stadium. tammy abraham scored twice in the first half to give chelsea the lead. dominic solanke scored late on to seal the win for the home side. Reference alejandro valverde won ahead of julian alaphilippe and michael albasini. chris froome finished 123rd after a crash during the final 12 kilometres. team sky's sports director gabriel rasch praised froome for finishing. rasch said froome was 'banged up' but expects to ride tour de romandie. BART movistar rider alejandro valverde won fleche wallonne on wednesday. team sky's chris froome fell in the final 12km but finished the race. philippe gilbert pulled out of the race after a bad crash 50km from the end. click here for more cycling news. 2021; Tab. 10 presents an interesting pattern we observed when comparing the results of BRIO-Mul and BART, which demonstrates that our method helps the abstractive model to filter out noise patterns in the original data. Specifically, some of the reference summaries (331/11490) in CNNDM contains the phrase "click here", pointing to a hyperlink, and 103 source documents also contain this phrase. BART picked up this pattern, and generates this phrase in 96 output summaries. On the contrary, our model learns to ignore this noise pattern and never generated it across the whole test set, likely because it identified that generated candidates with this pattern rarely achieve a high ROUGE score, and downweighted the probability accordingly. In this work, we presented a new training paradigm that assigns candidate outputs probability mass according to their quality using contrastive learning. While our method has achieved significant improvement on abstractive summarization, we note several directions for the future work to explore. First, since our method makes no assumptions specifically about the summarization task, it can be extended to other conditional text generation tasks such as machine translation. Second, it is possible to apply our method in a reinforcement learning setting, where the candidate summaries are dynamically generated. Finally, in experiments we only used diverse beam search to generate the candidate summaries, but it is likely that other candidate generation methods could yield further improvements.
| 911 | 44 | 911 |
Joint Learning of Representations for Web-tables, Entities and Types using Graph Convolutional Network
|
Existing approaches for table annotation with entities and types either capture the structure of table using graphical models, or learn embeddings of table entries without accounting for the complete syntactic structure. We propose TabGCN, which uses Graph Convolutional Networks to capture the complete structure of tables, knowledge graph and the training annotations, and jointly learns embeddings for table elements as well as the entities and types. To account for knowledge incompleteness, TabGCN's embeddings can be used to discover new entities and types. Using experiments on 5 benchmark datasets, we show that TabGCN significantly outperforms multiple state-of-the-art baselines for table annotation, while showing promising performance on downstream table-related applications.
|
Table data abounds in webpages and organizational documents. Annotation of table entries, such as columns, cells and rows, using available background knowledge (e.g. Yago, DBPedia, Freebase, etc.), such as knowledge of entities and their types, helps in better understanding and semantic interpretation of such tabular data. The challenge, however, is that such web tables do not adhere to any standard format, schema or convention Initial research on table annotation In this work, we propose the TabGCN model that uses a Graph Convolutional Network (GCN) We demonstrate these capabilities of TabGCN using experiments on 5 benchmark web table datasets comparing against 5 existing models. We show that WebGCN significantly improves performance for entity and type annotation. For the other tasks, we show that the same embeddings show impressive performance. No existing model can perform all of these tasks. Our contributions are as follows: (a) We propose a model called TabGCN based on the GCN architecture that captures the complete syntactic structure of tables as well as the knowledge representation, and learns embeddings of tables, rows, columns and cells, as well as entities and types jointly and in an end-to-end fashion. (b) TabGCN addresses incompleteness in the knowledge representation by discovering new entities and types. (c) TabGCN significantly outperforms 5 existing approaches in 5 different benchmark datasets for the task of table annotation. (d) The trained embeddings show impressive performance in downstream tasks such as identifying similar tables and assignment of rows to appropriate tables.
|
Existing literature on table annotation considers two different types of tables. In general, web tables In terms of approaches, one category of work considers graphical models to capture table structure and performs joint inference for entity and type classification The second category focuses on embeddings for tables For the problem of extending the knowledge graph from tables, 3 Problem: Table Tables: We are given a set of tables S. Fig. Entities and Types: We assume background knowledge of entities and entity types (or simply, types). Let T denote the set of types, and E the set of entities. Each entity E is associated with a type T (E). For each entity E, there is also an entity description or lemma L(E). In Fig. We assume that tables contain information about entities E and types T . Specifically, each column of each table corresponds to a single type, each cell corresponds to a specific entity of that type. In our example table, cell x 11 is annotated with entity E123, and column C 1 is annotated with type T12:Person. Let T (C k j ) denote the type associated with column C k j , and E(x k ij ) the entity associated with the cell x k ij . Let A e be the set of all entity annotations of cells, and A t that of all type annotations of columns. In the semi-supervised table annotation task, we are given the entire set of tables S but only a subset A o e ⊂ A e of the entity annotations, and a subset A o t ⊂ A t of the type annotations are observed. The task is to annotate the unannotated cells and columns of the tables, using the observed annotations as training data. Let T o denote the set of unique types seen in A o t , and E o the set of unique entities seen in A o e . In the incomplete knowledge setting, T o ⊂ T , indicating that all the types are not seen in the training annotations. Similarly, all the entities are also not seen in training: E o ⊂ E. Now, the task for the unannotated cells and columns is three-fold. The first is to decide whether these correspond to observed entities A o e and observed types A o t . We call this novelty classification. Next, the non-novel table columns need to be annotated with observed types T o , and the non-novel table cells with observed entities E o . We call these type detection and entity detection respectively. Finally, the columns corresponding to novel types need to be grouped according to distinct novel types, and the cells corresponding to novel entities need to be grouped according to distinct novel entities. We call these type discovery and entity discovery respectively. We first present at a high level the network architecture of our model, which we call TabGCN. The core components TabGCN are (I) a Graph Convolutional Network (GCN), which captures the various syntactic relationships between table and knowledge elements, and then jointly learns embeddings for these via the convolution operation. The GCN embeddings of table elements contain information about both types and entities. These are fed into two parallel components: (II) the Type Classification component, and (III) the Entity Classification component. The Type Classification component first projects the GCN embedding of table columns to a type space using a type-projection matrix, and then uses a soft-max layer to classify this type embedding according to observed types. Similarly, the Entity Classification component first projects the GCN embeddings of table cells to an entity space using an entity-projection matrix, and then uses a soft-max layer to classify this entity embedding according to observed entities. Fig. Graph Convolutional Network: Graph Convolutional Networks (GCN) (1) u = x u . For classification, a linear classification layer is added on top of final GCN layer. Function f () is a non-linear activation for which we used ReLU. Recall that edges in a GCN serve to bring the embeddings of their end nodes closer, the extent being determined by their weight. With this intuition we create the edges R of different types reflecting the underlying semantics of tables and the annotations. These are table edges R t , knowledge edges R k , annotation edges R a , and lexical similarity edges R l . Table edges capture the semantics of web tables, which do not have special anchor columns. These are of four categories: a cell-column edge between a cell node x k ij and its corresponding column node C k j ; a cell-row edge between a cell node x k ij and its corresponding row node R k i ; a column-table edge between a column node C k j and its corresponding table node S k ; and a row-table edge between a row node R k i and its corresponding table node S k . Knowledge edges connect each entity node E o with its corresponding type node T (E o ). Annotation edges are of two categories: an entity annotation edge for each entity annotation in A o e between a cell node x k ij and its labeled entity node E(x k ij ); and a type annotation edge for each type annotation in A o t between a column node C k j and its labeled type node T (C k j ). Lexical similarity edges are added between pairs of cells in the same or different tables whose lexical similarity, computed using character-based Jaccard All edges are bi-directional. Each of the 8 edge categories above has its own parameters (W l ) for each layer in our GCN. Self loops are added for nodes associated with textual input, specifically, cells and entities with lemmas. For the input representation of such nodes, we use the pre-trained word embeddings for each of their constituent tokens, and take their mean. For this paper we used GloVe Type Classification: The final (K th ) GCN layer generates an embedding h for each node v. This contains information about both types and entities. For type classification of a column node c, we first get its type embedding h t (c) by projecting h (K) c to a type space using a type projection matrix P t : h t (c) = P t h (K) c . Then we get the probability distribution g t (c) over known types T o for the type embedding h t (c) using a soft-max layer with weight matrix θ t : g t (c) = σ(h t (c); θ t ). The type projection matrix P t and the sigmoid weight matrix θ t form the parameters for this component. We follow a similar approach for entity classification of a cell node x. We first project its GCN embedding h (K) x to an entity space using an entity projection matrix e . The entity embedding h e (x) is passed through a soft-max layer with weight matrix θ e to get the probability distribution g e (x) over known entities E o . The entity projection matrix P e and the sigmoid weight matrix θ e form the parameters for this component. The parameters for all three components are trained end-to-end using available entity annotations for cells and type annotations for columns. Specifically, we consider the type predictions g t (c) for columns and minimize classification loss with the observed entity labels T (c). We similarly minimize classification loss between entity predictions g e (x) and observed entity labels E(x) for cells. We consider a weighted combination of the entity prediction loss and the type prediction loss. We have used cross-entropy as the classification loss function, and Adam (Kingma and Ba, 2015) for optimization. Training the network in Sec.4 generates estimates of the model parameters as well as embeddings for all table and knowledge elements. In this section, we describe the use of these parameters and embeddings for the tasks defined in Sec.3. We use h(v) for the GCN embedding of a node v instead of h (K) v for brevity. Novelty Classification: To decide whether an unannotated column c corresponds to any of the known types in T o (novel type classification), we make use of their type embeddings. Column c corresponds to a new type if its typespace embedding h t (c) = P t h(c) is 'far away' from the type space embedding h t (T ) = P t h(T ) for all T ∈ T o . More specifically, we use δ(max T ∈T o cos(P t h(c), P t h(T )) ≤ t ), where t is the novel type threshold, and δ() is the Kronecker delta function. A similar approach is followed for deciding if an unannotated cell x corresponds to any of the known entities in E o (novel entity classification) by using the entity embeddings. Specifically, we use δ(max E∈E o cos(P e h(x), P e h(E)) ≤ e ), where e is the novel entity threshold. Type and Entity Detection: Columns and cells determined to be non-novel need to be classified according to known types and entities respectively. This can be done using forward propagation in the trained network on the embeddings of the corresponding nodes. The type prediction g t (c) of a column c is obtained as g t (c) = σ(P t h(c); θ t ). Similarly, the entity prediction g e (x) of a cell x is obtained as g t (x) = σ(P e h(x); θ e ). Type and Entity Discovery: On the other hand, columns and cells determined to be novel need to be grouped according to distinct new types and entities respectively. This is done by clustering their projections in the appropriate space. Specifically, for type discovery, we take the type embeddings h t (c) of all novel columns c and cluster these. Similarly, for entity discovery, we cluster the entity embeddings h e (x) of all novel cells x. The clustering algorithm needs to automatically determine the number of clusters in both cases. In this paper, we have used Silhouette clustering Training annotations are only provided for cells and columns. But embeddings are available for the rows and tables as well after training, and these can be used for different down-stream application tasks. Since these do not involve type or entity spaces, we directly use their GCN embeddings for these tasks. As examples, we define two such tasks here. Table In this section, we first present experimental results for table annotation, and then for down-stream table-related tasks. We compare our proposed model TabGCN with appropriate state-of-the-art baselines. We have uploaded our source code as supplementary material for reproducibility. Wiki M Web M T2Dv2 Limaye Hits@1 MRR Hits@1 MRR Hits@1 MRR Hits@1 MRR Hits@1 MRR Detection Results: We first present results for entity and type detection, addressed by most of the baselines. For this task, all models predict entity labels for all cells and type labels for all columns that are unannotated. However, evaluation is only for those cells and columns whose true entity and type labels are contained in the observed entities E o and observed types T o respectively. For evaluation, as in earlier table annotation papers Tab. 1 shows detailed results for all models across datasets. We can see that TabGCN signifi- cantly outperforms all baselines on all datasets for both detection tasks. The only exception is for Wiki Manual. The graphical model based approaches with handcrafted potential functions outperform the representation learning approaches, possibly on account of the smallness of the dataset. Among the embedding based approaches, TabGCN performs the best. Novelty Classification Results: In our second experiment, we consider novelty classification. This is an unsupervised task, where a model makes a binary decision for each unannotated column (novel type classification) and for each unannotated cell (novel entity classification). Since the decision depends on the thresholds for type ( t ) and entity ( e ), we plot F1 score on the y-axis against the corresponding threshold on the x-axis. Of the baselines, only PGM can address this task, but outputting a NONE label for the type or entity. Fig. between the assigned cluster labels (Y) and the true entity or type labels (C), where I(, ) denotes mutual information and H() denotes entropy. In Tab. 2, we see that TabGCN performs consistently above 80% for entity and type discovery across datasets, significantly outperforming ColNet and TaBERT. Ablation Study: Next, we analyze the performance of TabGCN, using multiple ablations of the full model. -K leaves out the knowledge nodes and their incident edges from the GCN graph during training. -E focuses only on types by removing all entity nodes and entity-related edges (type-entity and cell-entity annotation) from the GCN graph. It is trained only using type loss. Note that it cannot perform tasks associated with entities, specifically entity detection, novel entity classification and novel entity discovery. Similarly, -T focuses only on entities by removing all type nodes and type-related edges (type-entity and column-type annotation) from the GCN graph. It is trained only using entity loss, and cannot perform tasks associated with types. Finally, -L removes the lexical similarity edges from the GCN graph. The results are recorded in Table . 5. We can see that all the components of the architecture contribute to performance improvements. The improvement is statistically significant (using the Wilson Interval with α = 0.05) for all ablations other than -K. While -K performs comparably here, its performance drops significantly for novelty classification, as can be seen in Fig. Tasks: Finally, we include some results for the table and row related inference tasks defined at the end of Sec.5. This is to demonstrate how the learnt embeddings can benefit potential down-stream tasks. Note that TabGCN directly outputs table embeddings. Of the baselines, Tab2Vec and TaBERT output row embeddings. For these models, we create table embeddings by averaging the corresponding row embeddings. In Tab. 6, we record performance for table clustering. TabGCN again significantly outperforms both baselines for both datasets. We finally consider row-to-table assignment. In this task, one randomly selected row is removed from every table during training. These rows then need to be classified according to their parent table. Since the models output a ranking of tables for each row, we evaluate using two ranking related measures. Hits@1 measures the fraction of rows with the correct table at rank 1. Mean Reciprocal rank (MRR) is the mean of the reciprocal rank of the correct table over all rows, and its perfect value is 1.0. In Tab. 3, we again see that TabGCN performs the best across datasets. In summary, we have demonstrated the usefulness of learning embeddings of table elements and knowledge elements jointly using both entity and type losses in an end-to-end fashion for type and entity annotation on 5 benchmark datasets. In addition, we have demonstrated how the learned embeddings can be useful for downstream table-related tasks. In all cases, TabGCN has significantly outperformed multiple state-of-the-art baselines using probabilistic graphical models as well as other neural approaches. We have proposed a model for that jointly learns representations of tables, rows, columns and cell, as well as entities and types by the complete syntactic structure of all tables, the relevant entities and types and the available annotations using the Graph Convolutional Network. As a result, TabGCN unifies the benefits of probabilistic graphical model based approaches and embedding based approaches for table annotation. Using these embeddings, TabGCN significantly outperforms existing approaches for table annotation, as well as entity and type discovery.
| 788 | 1,624 | 788 |
Compositional Representation of Morphologically-Rich Input for Neural Machine Translation
|
Neural machine translation (NMT) models are typically trained with fixed-size input and output vocabularies, which creates an important bottleneck on their accuracy and generalization capability. As a solution, various studies proposed segmenting words into sub-word units and performing translation at the sub-lexical level. However, statistical word segmentation methods have recently shown to be prone to morphological errors, which can lead to inaccurate translations. In this paper, we propose to overcome this problem by replacing the source-language embedding layer of NMT with a bi-directional recurrent neural network that generates compositional representations of the input at any desired level of granularity. We test our approach in a low-resource setting with five languages from different morphological typologies, and under different composition assumptions. By training NMT to compose word representations from character trigrams, our approach consistently outperforms (from 1.71 to 2.48 BLEU points) NMT learning embeddings of statistically generated sub-word units.
|
An important problem in neural machine translation (NMT) is translating infrequent or unseen words. The reasons are twofold: the necessity of observing many examples of a word until its input representation (embedding) becomes reliable, and the computational requirement of limiting the input and output vocabularies to few tens of thousands of words. These requirements eventually lead to coverage issues when dealing with low-resource and/or morphologically-rich languages, due to their high lexical sparseness. To cope with this well-known problem, several approaches have been proposed redefining the model vocabulary in terms of interior orthographic units compounding the words, ranging from character ngrams In this paper, we propose to improve the quality of input (source language) representations of rare words in NMT by augmenting its embedding layer with a bi-directional recurrent neural network (bi-RNN), which can learn compositional input representations at different levels of granularity. Compositional word embeddings have recently been applied in language modeling and obtained successful results
|
In this paper, we use the NMT model of The model is trained by maximizing the loglikelihood of a parallel training set via stochastic gradient descent The inputs of the network are one-hot vectors, which are binary vectors with a single bit set to 1 to identify a specific word in the vocabulary. Each one-hot vector is then mapped to an embedding, a distributed representation of the word in a lower dimension but a more dense continuous space. From this input, a representation of the whole input sequence is learned using a bi-RNN, the encoder, which maps x into m dense sentence vectors corresponding to its hidden states. Next, another RNN, the decoder, predicts each target token y i by sampling from a distribution computed from the previous target token y i-1 , the previous decoder hidden state, and the context vector. The latter is a linear combination of the encoder hidden states, whose weights are dynamically computed by a feed-forward neural network called attention model Both the source and target vocabulary sizes play an important role in terms of defining the complex-ity of the model. In a standard architecture, like ours, the source and target embedding matrices actually account for the vast majority of the network parameters. The vocabulary size also plays an important role when translating from and to lowresource and morphologically-rich languages, due to the sparseness of the lexical distribution. Therefore, a conventional approach has now become to compose both the source and target vocabularies of sub-word units generated through statistical segmentation methods In this paper, we propose to perform NMT from input representations learned by composing smaller symbols, such as character n-grams In our model (Figure Given a bi-RNN with a forward (f ) and backward (b) layer, the input representation w of a token of t characters is computed from the hidden states h f t and h 0 b , i.e. the final outputs of the forward and backward RNNs, as follows: where W f and W b are weight matrices associated to each RNN and b is a bias vector We test our approach along with statistical word segmentation based open vocabulary NMT methods in an evaluation benchmark simulating a lowresource translation setting pairing English (En) with five languages from different language families and morphological typologies: Arabic (Ar), Czech (Cs), German (De), Italian (It) and Turk-ish (TR). The characteristics of each language are given in Table Typology Complexity The simple NMT model constitutes the baseline in our study and performs translation directly at the level of sub-word units, which can be of four different types: characters, character trigrams, BPE sub-word units, and LMVR sub-word units. The compositional model, on the other hand, performs NMT with input representations composed from sub-lexical vocabulary units. In our study, we evaluate representations composed from character trigrams, BPE, and LMVR units. In order to choose the segmentation method to apply on the English side (the output of NMT decoder), we compare BPE and LMVR sub-word units by carrying out an evaluation on the official data sets of Morpho Challenge 2010 The compositional bi-RNN layer is implemented in The performance of NMT models in translating each language using different vocabulary units and encoder input representations can be seen in Table Using our compositional model improves the quality of input representations for each type of vocabulary unit, nevertheless, the best performance is obtained by using character trigrams as input symbols and words as input representations. The higher quality of these input representations compared to those obtained from subword units generated with LMVR suggest that our compositional model can learn morphology better than LMVR, which was found to provide comparable performance to morphological analyzers in Turkish to English NMT and anyway , we migrate , circle and mix so much that (Compositional Model) there 's no longer the isolation necessary to become evolutionary . and by the way , we immigrate and circulate and intermix so much that you can 't any longer have the isolation that is necessary for evolution to take place . ama aslında bu resim tamamen , farklı yerlerin fotograf@@ larının (Simple Model) birles ¸tir@@ il@@ mesiyle meydana geldi In this paper, we addressed the problem of translating infrequent words in NMT and proposed to solve it by replacing the conventional sub-word embeddings with input representations compositionally learned from character n-grams using a bi-RNN. Our approach showed significant and consistent improvements over a variety of languages, making it a competitive solution for NMT of lowresource and morphologically-rich languages. In the future, we plan to optimize our implementation and to test its scalability on larger data sets.
| 1,084 | 1,116 | 1,084 |
Neural Topic Modeling with Bidirectional Adversarial Training
|
Recent years have witnessed a surge of interests of using neural topic models for automatic topic extraction from text, since they avoid the complicated mathematical derivations for model inference as in traditional topic models such as Latent Dirichlet Allocation (LDA). However, these models either typically assume improper prior (e.g. Gaussian or Logistic Normal) over latent topic space or could not infer topic distribution for a given document. To address these limitations, we propose a neural topic modeling approach, called Bidirectional Adversarial Topic (BAT) model, which represents the first attempt of applying bidirectional adversarial training for neural topic modeling. The proposed BAT builds a twoway projection between the document-topic distribution and the document-word distribution. It uses a generator to capture the semantic patterns from texts and an encoder for topic inference. Furthermore, to incorporate word relatedness information, the Bidirectional Adversarial Topic model with Gaussian (Gaussian-BAT) is extended from BAT. To verify the effectiveness of BAT and Gaussian-BAT, three benchmark corpora are used in our experiments. The experimental results show that BAT and Gaussian-BAT obtain more coherent topics, outperforming several competitive baselines. Moreover, when performing text clustering based on the extracted topics, our models outperform all the baselines, with more significant improvements achieved by Gaussian-BAT where an increase of near 6% is observed in accuracy.
|
Topic models have been extensively explored in the Natural Language Processing (NLP) community for unsupervised knowledge discovery. Latent Dirichlet Allocation (LDA) One possible way in addressing this limitation is through neural topic models which employ blackbox inference mechanism with neural networks. Inspired by variational autoencoder (VAE) Although To address these limitations, we model topics with Dirichlet prior and propose a novel Bidirectional Adversarial Topic model (BAT) based on bidirectional adversarial training. The proposed BAT employs a generator network to learn the projection function from randomly-sampled documenttopic distribution to document-word distribution. Moreover, an encoder network is used to learn the inverse projection, transforming a document-word distribution into a document-topic distribution. Different from traditional models that often resort to analytic approximations, BAT employs a discriminator which aims to discriminate between real distribution pair and fake distribution pair, thereby helps the networks (generator and encoder) to learn the two-way projections better. During the adversarial training phase, the supervision signal provided by the discriminator will guide the generator to construct a more realistic document and thus better capture the semantic patterns in text. Meanwhile, the encoder network is also guided to generate a more reasonable topic distribution conditioned on specific document-word distributions. Finally, to incorporate the word relatedness information captured by word embeddings, we extend the BAT by modeling each topic with a multivariate Gaussian in the generator and propose the Bidirectional Adversarial Topic model with Gaussian (Gaussian-BAT). The main contributions of the paper are: • We propose a novel Bidirectional Adversarial Topic (BAT) model, which is, to our best knowledge, the first attempt of using bidirectional adversarial training in neural topic modeling; • We extend BAT to incorporate the word re-latedness information into the modeling process and propose the Bidirectional Adversarial Topic model with Gaussian (Gaussian-BAT); • Experimental results on three public datasets show that BAT and Gaussian-BAT outperform the state-of-the-art approaches in terms of topic coherence measures. The effectiveness of BAT and Gaussian-BAT is further verified in text clustering.
|
Our work is related to two lines of research, which are adversarial training and neural topic modeling. Adversarial training, first employed in Generative Adversarial Network (GAN) Adversarial training has also been extensively used for text generation. For example, Seq-GAN To overcome the challenging exact inference of topic models based on directed graph, a replicated softmax model (RSM), based on the Restricted Boltzmann Machines was proposed in Our proposed neural topic models are based on bidirectional adversarial training The encoder learns a mapping function to transform document-word distribution to document-topic distribution. As shown in the top-left panel of Figure More concretely, for each document d in text corpus, E takes the document representation d r as input, where d r is the representation weighted by TF-IDF, and it is calculated by: where n i,d denotes the number of i-th word appeared in document d, |C| represents the number of documents in the corpus, and |C i | means the number of documents that contain i-th word in the corpus. Thus, each document could be represented as a V -dimensional multinomial distribution and the i-th dimension denotes the semantic consistency between i-th word and the document. With d r as input, E firstly projects it into an S-dimensional semantic space through the representation layer as follows: where W e s ∈ R S×V and b e s are weight matrix and bias term of the representation layer, h e s is the state vector normalized by batch normalization BN(•), leak denotes the parameter of LeakyReLU activation and o e s represents the output of representation layer. Then, the encoder transforms o e s into a Kdimensional topic space based on the equation below: where W e t ∈ R K×S is the weight matrix of topic distribution layer, b e t represents the bias term, θ r denotes the corresponding topic distribution of the input d r and the k-th (k ∈ {1, 2, ..., K}) dimension θ k r represents the proportion of k-th topic in document d. The generator G is shown in the bottom-left panel of Figure As pointed out in where α is the K-dimensional hyper-parameter of Dirichlet prior, K is the topic number that should be set in BAT, θ k f ∈ [0, 1], follows the constrain that K k=1 θ k f = 1, represents the proportion of the k-th topic in the document, and normalization term . To learn the transformation from documenttopic distribution to document-word distribution, G firstly projects θ f into an S-dimensional representation space based on equations: where W g s ∈ R S×K is weight matrix of the representation layer, b g s represents bias term, h g s is the state vector normalized by batch normalization, Eq. 6 represents the LeakyReLU activation parameterized with leak, and o g s is the output of the representation layer. Then, to project o g s into word distribution d f , a subnet contains a linear layer and a softmax layer is used and the transformation follows: where W g w ∈ R V ×S and b g w are weight matrix and bias of word distribution layer, d f is the word distribution correspond to θ f . For each v ∈ {1, 2, ..., V }, the v-th dimension d v f is the probability of the v-th word in fake document d f . The discriminator D is constituted by three layers (a V + K-dimensional joint distribution layer, an S-dimensional representation layer and an output layer) as shown in the right panel of Figure In BAT, the generator models topics based on the bag-of-words assumption as in most other neural topic models. To incorporate the word relatedness information captured in word embeddings where e v means the word embedding of v-th word, V is the vocabulary size, |Σ k | = det Σ k is the determinant of covariance matrix Σ k , D e is the dimension of word embeddings, p( e v |topic = k) is the probability calculated by density, and φ k is the normalized word distribution of k-th topic. With randomly sampled topic distribution θ f and the calculated topic-word distributions { φ 1 , φ 2 , ..., φ K }, the fake word distribution d f corresponding to θ f can be obtained by: where θ k is the topic proportion of the k-th topic. Then, θ f and d f are concatenated to form the fake distribution pair p f as shown in Figure In Figure To measure the distance between P r and P f , we use the Wasserstein-distance as the optimization objective, since it was shown to be more effective compared to Jensen-Shannon divergence where D(•) represents the output signal of the discriminator. A higher value denotes that the discriminator is more prone to consider the input as a real distribution pair and vice versa. In addition, we use weight clipping which was proposed to ensure the Lipschitz continuity end for 11: 12: end for 14: 15: , ω e , p a ) 16: end while The training procedure of BAT and Gaussian-BAT is given in Algorithm. 1. Here, c is the clipping parameter, n d represents the number of discriminator iterations per generator iteration, m is the batch size, α 1 is the learning rate, β 1 and β 2 are hyper-parameters of Adam (Kingma and Ba, 2014), and p a represents {α 1 , β 1 , β 2 }. In our experiments, we set the n d = 5, m = 64, α 1 = 1e-4, c = 0.01, β 1 = 0.5 and β 2 = 0.999. After model training, learned G and E will build a two-way projection between document-topic distribution and document-word distribution. Thus, G and E could be used for topic generation and cluster inference. To generate the word distribution of each topic, we use ts (k) , a K-dimensional vector, as the onehot encoding of the k-th topic. For example, ts 2 = [0, 1, 0, 0, 0, 0] T in a six topic setting. And the word distribution of the k-th topic is obtained by: Likewise, given the document representation d r , topic distribution θ r obtained by BAT/Gaussian-BAT could be used for cluster inference based on: where c r denotes the inferred cluster of d r . In this section, we first present the experimental setup which includes the datasets used and the baselines, followed by the experimental results. We evaluate BAT and Gaussian-BAT on three datasets for topic extraction and text clustering, 20Newsgroups Grolier is built from Grolier Multimedia Encycopedia, which covers almost all the fields in the world. NYTimes is a collection of news articles published between 1987 and 2007, and contains a wide range of topics, such as sports, politics, education, etc. We use the full datasets of 20Newsgroups 1 and Grolier 2 . For the NYTimes dataset, we randomly select 100,000 articles and remove the low frequency words. The final statistics are shown in We choose the following models as baselines: LDA ProdLDA Topic models are typically evaluated with the likelihood of held-out documents and topic coherence. However, conducted on the datasets with five topic number settings (3) The incorpora-tion of word embeddings in Gaussian-BAT helps generating more coherent topics. We also compare the average topic coherence values (all topics taken into account) numerically to show the effectiveness of proposed BAT and Gaussian-BAT. The results of numerical topic coherence comparison are listed in Table Moreover, to explore how topic coherence varies with different topic numbers, we also provide the comparison of average topic coherence vs. different topic number on 20newsgroups, Grolier and NYTimes (all topics taken into account). The detailed comparison is shown in Figure We further compare our proposed models with baselines on text clustering. Due to the lack of document label information in Grolier and NYTimes, we only use 20Newsgroups dataset in our experiments. The topic number is set to 20 (ground-truth categories) and the performance is evaluated by accuracy (ACC): where N t is the number of documents in the test set, ind(•) is the indicator function, l i is the groundtruth label of i-th document, c i is the category assignment, and map ranges over all possible oneto-one mappings between labels and clusters. The optimal map function can be obtained by the Kuhn-Munkres algorithm In this paper, we have explored the use of bidirectional adversarial training in neural topic models and proposed two novel approaches: the Bidirectional Adversarial Topic (BAT) model and the Bidirectional Adversarial Topic model with Gaussian (Gaussian-BAT). BAT models topics with the Dirichlet prior and builds a two-way transformation between document-topic distribution and document-word distribution via bidirectional adversarial training. Gaussian-BAT extends from BAT by incorporating word embeddings into the modeling process, thereby naturally considers the word relatedness information captured in word embeddings. The experimental comparison on three widely used benchmark text corpus with the existing neural topic models shows that BAT and Gaussian-BAT achieve improved topic coherence results. In the future, we would like to devise a nonparametric neural topic model based on adversarial training. Besides, developing correlated topic modelsis another promising direction.
| 1,522 | 2,388 | 1,522 |
Learning part-of-speech taggers with inter-annotator agreement loss
|
In natural language processing (NLP) annotation projects, we use inter-annotator agreement measures and annotation guidelines to ensure consistent annotations. However, annotation guidelines often make linguistically debatable and even somewhat arbitrary decisions, and interannotator agreement is often less than perfect. While annotation projects usually specify how to deal with linguistically debatable phenomena, annotator disagreements typically still stem from these "hard" cases. This indicates that some errors are more debatable than others. In this paper, we use small samples of doublyannotated part-of-speech (POS) data for Twitter to estimate annotation reliability and show how those metrics of likely interannotator agreement can be implemented in the loss functions of POS taggers. We find that these cost-sensitive algorithms perform better across annotation projects and, more surprisingly, even on data annotated according to the same guidelines. Finally, we show that POS tagging models sensitive to inter-annotator agreement perform better on the downstream task of chunking.
|
POS-annotated corpora and treebanks are collections of sentences analyzed by linguists according to some linguistic theory. The specific choice of linguistic theory has dramatic effects on downstream performance in NLP tasks that rely on syntactic features (1) @GaryMurphyDCU of @DemMattersIRL will take part in a panel discussion on October 10th re the aftermath of #seanref . . . While linguists will agree that in is a preposition, and panel discussion a compound noun, they are likely to disagree whether will is heading the main verb take or vice versa. Even at a more basic level of analysis, it is not completely clear how to assign POS tags to each word in this sentence: is part a particle or a noun; is 10th a numeral or a noun? Some linguistic controversies may be resolved by changing the vocabulary of linguistic theory, e.g., by leaving out numerals or introducing ad hoc parts of speech, e.g. for English to For these and other reasons, our annotators (even when they are trained linguists) often disagree on how to analyze sentences. The strategy in most previous work in NLP has been to monitor and later resolve disagreements, so that the final labels are assumed to be reliable when used as input to machine learning models.
|
Instead of glossing over those annotation disagreements, we consider what happens if we embrace the uncertainty exhibited by human annotators when learning predictive models from the annotated data. To achieve this, we incorporate the uncertainty exhibited by annotators in the training of our model. We measure inter-annotator agreement on small samples of data, then incorporate this in the loss function of a structured learner to reflect the confidence we can put in the annotations. This provides us with cost-sensitive online learning algorithms for inducing models from annotated data that take inter-annotator agreement into consideration. Specifically, we use online structured perceptron with drop-out, which has previously been applied to POS tagging and is known to be robust across samples and domains The vast majority of POS-annotated resources across languages contain mostly newswire text. Some annotated Twitter data sets do exist for English. In our experiments, we use the RITTER splits provided by The three annotation efforts The data sets differ in how they analyze many of the linguistically hard cases. Consider, for example, the analysis of will you come out to in GIM-PEL and RITTER (Figure Other differences include the analysis of abbreviations (PRT in GIMPEL; X in RITTER and FOS-TER), colon (X in GIMPEL; punctuation in RIT-TER and FOSTER), and emoticons, which can take multiple parts of speech in GIMPEL, but are always X in RITTER, while they are absent in FOS-TER. GIMPEL-TRAIN and RITTER-TRAIN are also internally inconsistent. See the bottom of Figure Since the mapping to universal tags could potentially introduce errors, we also annotated a data set directly using universal tags. We randomly selected 200 tweets collected over the span of one day, and had three annotators tag this set. We split the data in such a way that each annotator had 100 tweets: two annotators had disjoint sets, the third overlapped 50 items with each of the two others. In this way, we obtained an initial set of 100 doubly-annotated tweets. The annotators were not provided with annotation guidelines. After the first round of annotations, we achieved a raw agreement of 0.9, a Cohen's κ of 0.87, and a Krippendorff's α of 0.87. We did one pass over the data to adjudicate the cases where annotators disagreed, . . . or where they had flagged their choice as debatable. The final data set (lowlands.test), referred below to as INHOUSE, contained 3,064 tokens (200 tweets) and is publicly available at 3 Computing agreement scores The F 1-score relates to precision and recall in the usual way, i.e, as the harmonic mean between those two measure. In more detail, given two annotators A 1 and A 2 , we say the precision of A 1 relative to A 2 with respect to POS tag T in some data set X, denoted Prec T (A 1 (X), A 2 (X)), is the number of tokens both A 1 and A 2 predict to be T over the number of times A 1 predicts a token to be T . Similarly, we define the recall with respect to some tag T , i.e., Rec T (A 1 (X), A 2 (X)), as the number of tokens both A 1 and A 2 predict to be T over the number of times A 2 predicts a token to be T . The only difference with respect to standard precision and recall is that the gold standard is replaced by a second annotator, A 2 . Note that Prec T (A 1 (X), A 2 (X)) = Rec T (A 2 (X), A 1 (X)). It follows from all of the above that the F 1-score is symmetrical, i.e., The inter-annotator F 1-scores over the 12 POS tags in the universal tagset are presented in Figure We compute tag confusion probabilities from a confusion matrix over POS tags like the one in Figure We experiment with both agreement scores (F 1 and confusion matrix probabilities) to augment the loss function in our learner. The next section describes this modification in detail. We briefly introduce the cost-sensitive perceptron classifier. Consider the weighted perceptron loss on our ith example x i , y i (with learning rate α = 1), L w ( x i , y i ): In a non-cost-sensitive classifier, the weight function γ(y j , y 8: w * + = w 9: end for 10: end for 11: return w * / = (N × I) two cost-sensitive systems proposed only differ in how we formulate γ(•, •). In one model, the loss is weighted by the inter-annotator F 1 of the gold tag in question. This boils down to where X is the small sample of held-out data used to estimate inter-annotator agreement. Note that in this formulation, the predicted label is not taken into consideration. The second model is slightly more expressive and takes both the gold and predicted tags into account. It basically weights the loss by how likely the gold and predicted tag are to be mistaken for each other, i.e., (the inverse of) their confusion probability: In both loss functions, a lower gamma value means that the tags are more likely to be confused by a pair of annotators. In this case, the update is smaller. In contrast, the learner incurs greater loss when easy tags are confused. It is straight-forward to extend these costsensitive loss functions to the structured perceptron In our main experiments, we use structured perceptron Our results are presented in Figure The interesting results are the evaluations on out-of-sample evaluation data sets (FOSTER and IN-HOUSE) -lower part of Figure Discriminative learning typically benefits from regularization to prevent overfitting. The simplest is the averaged perceptron, but various other methods have been suggested in the literature. We use structured perceptron with drop-out, but results are relatively robust across other regularization methods. Drop-out works by randomly dropping a fraction of the active features in each iteration, thus preventing overfitting. Table We have seen that our POS tagging model improves over the baseline model on three out-ofsample test sets. The question remains whether training a POS tagger that takes inter-annotator agreement scores into consideration is also effective on downstream tasks. Therefore, we evaluate our best model, the CM-weighted learner, in two downstream tasks: shallow parsing-also known as chunking-and named entity recognition (NER). For the downstream evaluation, we used the baseline and CM models trained over 13 epochs, as they performed best on FOSTER-DEV (cf. Figure The set of features for chunking include information from tokens and POS tags, following We train the chunker on Twitter data The results for chunking are given in Table 2. They show that using the POS tagging model (CM) trained to be more sensitive to interannotator agreement improves performance over the baseline (BL) for the downstream task of chunking. Overall chunking F1 score improves. More importantly, we report on individual scores for NP, VP and PP chunks, where we see consistent improvements for NPs and VPs (since both nouns and verbs have high inter-annotator agreement), while results on PP are mixed. This is to be expected, since PP phrases involve adpositionals (ADP) that are often confused with particles (PRT), cf. Figure In the previous section, we saw positive effects of cost-sensitive POS tagging for chunking, and here we evaluate it on another downstream task, NER. For the named entity recognition setup, we use commonly used features, in particular features for word tokens, orthographic features like the presence of hyphens, digits, single quotes, upper/lowercase, 3 character prefix and suffix information. Moreover, we add Brown word cluster features that use 2,4,6,8,..,16 bitstring prefixes estimated from a large Twitter corpus Cost-sensitive learning takes costs, such as misclassification cost, into consideration. That is, each instance that is not classified correctly during the learning process may contribute differently to the overall error. The work most related to ours is the recent study of Our approach is also related to the literature on regularization, since our cost-sensitive loss functions are aimed at preventing over-fitting to low-confidence annotations. There is a large literature related to the issue of learning from annotator bias. In a similar vein, In general, it is important to analyze the data and check for label biases, as a machine learner is greatly affected by annotator noise that is not ran-dom but systematic Regarding measurements for agreements, several scores have been suggested in the literature. Apart from the simple agreement measure, which records how often annotators choose the same value for an item, there are several statistics that qualify this measure by adjusting for other factors, such as Cohen's κ In NLP, we use a variety of measures to assess and control annotator disagreement to produce homogenous final annotations. This masks the fact that some annotations are more reliable than others, and which is thus not reflected in learned predictors. We incorporate the annotator uncertainty on certain labels by measuring annotator agreement and use it in the modified loss function of a structured perceptron. We show that this approach works well independent of regularization, both on in-sample and out-of-sample data. Moreover, when evaluating the models trained with our loss function on downstream tasks, we observe improvements on two different tasks. Our results suggest that we need to pay more attention to annotator confidence when training predictors.
| 1,097 | 1,243 | 1,097 |
Learning to Write with Cooperative Discriminators
|
Despite their local fluency, long-form text generated from RNNs is often generic, repetitive, and even self-contradictory. We propose a unified learning framework that collectively addresses all the above issues by composing a committee of discriminators that can guide a base RNN generator towards more globally coherent generations. More concretely, discriminators each specialize in a different principle of communication, such as Grice's maxims, and are collectively combined with the base RNN generator through a composite decoding objective. Human evaluation demonstrates that text generated by our model is preferred over that of baselines by a large margin, significantly enhancing the overall coherence, style, and information of the generations.
|
Language models based on Recurrent Neural Networks (RNNs) have brought substantial advancements across a wide range of language tasks We propose a unified learning framework that can address several challenges of long-form text generation by composing a committee of discriminators each specializing in a different principle of communication. Starting with an RNN language model, our framework learns to construct a more powerful generator by training a number of discriminative models that can collectively address limitations of the base RNN generator, and Context: The two guards thudded into the door on the other side and began pounding on it, shouting furiously. Scious raised the large bunch of keys then placed them in a large pocket in his coat. "Come, we have to go," he whispered and moved up the corridor, the guards still hammering on the door. Jared looked around. The table next to the room they had been in was bare. LM: There was a small table set into the center of the table, and a table with two chairs and a table. A chair sat next to the couch. A large man stood in the doorway of the room, his arms crossed over his chest. "What 's up?" He asked.
|
Empty. It looked like some kind of prison. There were no windows, no doors, no windows. No sign of anyone in the house. No one had ever been here before. Reference: Only a flagon of wine and a couple of mugs plus a crude dice game. Light flickered in the stone corridor from lanterns on the walls. The place was dank and gloomy, moss in scattered patches on the floor and walls. The corridor ended just beyond the door to their former prison. No one else was about. Figure then learns how to weigh these discriminators to form the final decoding objective. These "cooperative" discriminators complement each other and the base language model to form a stronger, more global decoding objective. The design of our discriminators are inspired by Grice's maxims Empirical results ( §5) demonstrate that our learning framework is highly effective in converting a generic RNN language model into a substantially stronger generator. Human evaluation confirms that language generated by our model is preferred over that of competitive baselines by a large margin in two distinct domains, and significantly enhances the overall coherence, style, and information content of the generated text. Automatic evaluation shows that our system is both less repetitive and more diverse than baselines. RNN language models learn the conditional probability P (x t |x 1 , ..., x t-1 ) of generating the next word x t given all previous words. This conditional probability learned by RNNs often assigns higher probability to repetitive, overly generic sentences, as shown in Figure Several methods in the literature attempt to address these issues. Overly simple and generic generation can be improved by length-normalizing the sentence probability We propose a unified framework to address all these related challenges of long-form text generation by learning to construct a better decoding objective, generalizing over various existing modifications to the decoding objective. We propose a general learning framework for conditional language generation of a sequence y given a fixed context x. The decoding objective for generation takes the general form where every s k is a scoring function. The proposed objective combines the RNN language model probability P lm ( §3.1) with a set of additional scores s k (x, y) produced by discriminatively trained communication models ( §3.2), which are weighted with learned mixture coefficients λ k ( §3.3). When the scores s k are log probabilities, this corresponds to a Product of Experts (PoE) model Generation is performed using beam search ( §3.4), scoring incomplete candidate generations y 1:i at each time step i. The RNN language model decomposes into per-word probabilities via the chain rule. However, in order to allow for more expressivity over long range context we do not require the discriminative model scores to factorize over the elements of y, addressing a key limitation of RNNs. More specifically, we use an estimated score s k (x, y 1:i ) that can be computed for any prefix of y = y 1:n to approximate the objective during beam search, such that s k (x, y 1:n ) = s k (x, y). To ensure that the training method matches this approximation as closely as possible, scorers are trained to discriminate prefixes of the same length (chosen from a predetermined set of prefix lengths), rather than complete continuations, except for the entailment module as described in §3.2 Entailment Model. The prefix scores are re-estimated at each time-step, rather than accumulated over beam search. The RNN language model treats the context x and the continuation y as a single sequence s: (2) We introduce a set of discriminators, each of which encodes an aspect of proper writing that RNNs usually fail to capture. In all models the first layer embeds each word w into a 300-dimensional vector e(w) initialized with GloVe This model addresses the maxim of Quantity by biasing the generator to avoid repetitions. The goal of the repetition discriminator is to learn to distinguish between RNN-generated and gold continuations by exploiting our empirical observation that repetitions are more common in completions generated by RNN language models. However, we do not want to completely eliminate repetition, as words do recur in English. In order to model natural levels of repetition, a score d i is computed for each position in the continuation y based on pairwise cosine similarity between word embeddings within a fixed window of the previous k words, where The score of the continuation is then defined as where RNN rep (d) is the final state of a unidirectional RNN ran over the similarity scores d = d 1 . . . d n and w r is a learned vector. The model is trained to maximize the ranking log likelihood which corresponds to the probability of the gold ending y g receiving a higher score than the ending sampled from the RNN language model. Judging textual quality can be related to the natural language inference (NLI) task of recognizing textual entailment We train a classifier that takes two sentences a and b as input and predicts the relation between them as either contradiction, entailment or neutral. We use the neutral class probability of the sentence pair as discriminator score, in order to discourage both contradiction and entailment. As entailment classifier we use the decomposable attention model In contrast to our other communication models, this classifier cannot be applied directly to the full context and continuation sequences it is scoring. Instead every completed sentence in the continuation should be scored against all preceding sentences in both the context and continuation. Let t(a, b) be the log probability of the neutral class. Let S(y) be the set of complete sentences in y, S last (y) the last complete sentence, and S init (y) the sentences before the last complete sentence. We compute the entailment score of S last (y) against all preceding sentences in x and y, and use the score of the sentence-pair for which we have the least confidence in a neutral classification: (6) Intuitively, we only use complete sentences because the ending of a sentence can easily flip entailment. As a result, we carry over entailment score of the last complete sentence in a generation until the end of the next sentence, in order to maintain the presence of the entailment score in the objective. Note that we check that the current sentence is not directly entailed or contradicted by a previous sentence and not the reverse. The relevance model encodes the maxim of Relation by predicting whether the content of a candidate continuation is relevant to the given context. We train the model to distinguish between true continuations and random continuations sampled from other (human-written) endings in the corpus, conditioned on the given context. First both the context and continuation sequences are passed through a convolutional layer, followed by maxpooling to obtain vector representations of the sequences: The goal of maxpooling is to obtain a vector representing the most important semantic information in each dimension. The scoring function is then defined as where element-wise multiplication of the context and continuation vectors will amplify similarities. We optimize the ranking log likelihood (10) where y g is the gold ending and y r is a randomly sampled ending. In practice RNNs generate text that exhibit much less lexical diversity than their training data. To counter this effect we introduce a simple discriminator based on observed lexical distributions which captures writing style as expressed through word choice. This classifier therefore encodes aspects of the maxim of Manner. The scoring function is defined as The model is trained with a ranking loss using negative examples sampled from the language model, similar to Equation 5. Once all the communication models have been trained, we learn the combined decoding objective. In particular we learn the weight coefficients λ k in equation 1 to linearly combine the scoring functions, using a discriminative loss where A is the inference algorithm for beam search decoding. The weight coefficients are thus optimized to minimize the difference between the scores assigned to the gold continuation and the continuation predicted by the current model. Mixture weights are learned online: Each successive generation is performed based on the current values of λ, and a step of gradient descent is then performed based on the prediction. This has the effect that the objective function changes Due to the limitations of greedy decoding and the fact that our scoring functions do not decompose across time steps, we perform generation with a beam search procedure, shown in Algorithm 1. The naive approach would be to perform beam search based only on the language model, and then rescore the k best candidate completions with our full model. We found that this approach leads to limited diversity in the beam and therefore cannot exploit the strengths of the full model. Instead we score the current hypotheses in the beam with the full decoding objective: First, each hypothesis is expanded by selecting the k highest scoring next words according to the language model (we use beam size k = 10). Then k sequences are sampled from the k 2 candidates according to the (softmax normalized) distribution over the candidate scores given by the full decoding objective. Sampling is performed in order to increase diversity, using a temperature of 1.8, which was tuned by comparing the coherence of continuations on the validation set. At each step, the discriminator scores are recomputed for all candidates, with the exception of the entailment score, which is only recomputed for hypotheses which end with a sentence terminating symbol. We terminate beam search when the termination score, the maximum possible score achievable by terminating generation at the current position, is smaller than the current best score. We use two English corpora for evaluation. The first is the TripAdvisor corpus ADAPTIVELM Our first baseline is the same Adaptive Softmax CACHELM As another LM baseline we include a continuous cache language model SEQ2SEQ As our evaluation can be framed as sequence-to-sequence transduction, we compare against a seq2seq model directly trained to predict 5 sentence continuations from 5 sentences of context, using the OpenNMT attention-based seq2seq implementation SEQGAN Finally, as our use of discriminators is related to Generative Adversarial Networks (GANs), we use SeqGAN We pose the evaluation of our model as the task of generating an appropriate continuation given an initial context. In our open-ended generation setting the continuation is not required to be a specific length, so we require our models and baselines to generate 5-sentence continuations, consistent with the way the discriminator and seq2seq baseline datasets are constructed. Previous work has reported that automatic mea-sures such as BLEU Endings generated by our model and the baselines are compared against the reference endings in the original text. Results are given in Table For open-ended generation tasks such as our own, human evaluation has been found to be the only reliable measure We thoroughly enjoyed staying at this hotel. We travelled with our kids aged 4 and 6. The rooms were clean, spacious and well maintained. The beds were great and all of us slept well. The hotel provided vouchers towards breakfast at the adjoining restaurant. L2W We didn't have breakfast, so can't comment on that. The concierge was very helpful, and gave us directions, etc. Overall, we would definitely recommend this hotel to others. We would stay here again in a heartbeat. If you're looking for a great location, and good value, I'd recommend this hotel. ADAPTIVELM The breakfast buffet was very good, and the staff was very friendly and helpful, and it was a great way to start our day! I would definitely stay here again, and would recommend it to my friends, and would stay there again if I ever go back to Montreal. The location was great, and the staff was friendly and helpful. I would definitely stay here again!! CACHELM The staff was very friendly and helpful. The hotel is located in a great location. It is right in the middle of the gaslamp district. I would recommend this hotel to anyone. I would definitely stay here again. The staff were very friendly and helpful. The hotel is in a great location. The staff were very friendly and helpful. The hotel is in a great location. The staff were very friendly and helpful. We had a breakfast at Shula's & a delicious breakfast. The staff was very helpful and helpful. The breakfast was great as well. The staff was very helpful and friendly. We had a great service and the food was excellent. The restaurant was great and we used the vouchers towards whatever breakfast we ordered. The hotel had amazing grounds with a putting golf course that was fun for everyone. The pool was fantastic and we lucked out with great weather. We spent many hours in the pool, lounging, playing shuffleboard and snacking from the attached bar. The happy hour was great perk. The human evaluation is performed on 100 examples selected from the test set of each corpus, for every pair of generators that are compared. We present the examples to workers on Amazon Mechanical Turk, using three annotators for each example. The results are given in Table The absolute performance of all the evaluated systems on BLEU and Meteor is quite low (Table Under the crowd-sourced evaluation (Table For the TripAdvisor corpus, L2W is ranked higher than the baselines on overall quality, as well as on most individual metrics, with the exception that it fails to improve on contradiction and clarity over the ADAPTIVELM (which is again the most competitive baseline). Our model's strongest improvements over the baselines are on repetition and relevance. To investigate the effect of individual discriminators on the overall performance, we report the results of ablations of our model in Table The repetition module gives the largest boost by far, consistent with the intuition that many of the deficiencies of RNN as a text generator lie in semantic repetition. The entailment module (which was intended to reduce contradiction) is the weakest, which we hypothesize is the combination of (a) mismatch between training and test data (since the entailment module was trained on SNLI and MultiNLI) and (b) the lack of smoothness in the entailment scorer, whose score could only be updated upon the completion of a sentence. Surprisingly, L2W is even preferred over the original continuation of the initial text on BookCorpus. Qualitative analysis shows that L2W's continuation is often a straightforward continuation of the original text while the true continuation is more surprising and contains complex references to earlier parts of the book. While many of the issues of automatic metrics L2W generations are more topical and stylistically coherent with the context than the baselines. Table The CACHELM begins by generating the same over-used sentence and only produce short, generic sentences throughout. Seq2Seq simply repeats sentences that occur often in the training set, repeating one sentence three times and another twice. This indicates that the encoded context is essentially being ignored as the model fails to align the context and continuation. The SEQGAN system is more detailed, e.g. mentioning a specific location "Shula's" as would be expected given its highly diverse vocabulary (as seen in Table The L2W models do not fix every degenerate characteristic of RNNs. The TripAdvisor L2W generation consists of meaningful but mostly disconnected sentences, whereas human text tends to build on previous sentences, as in the reference continuation. Furthermore, while L2W re-peats itself less than any of our baselines, it still paraphrases itself, albeit more subtly: "we would definitely recommend this hotel to others." compared to "I'd recommend this hotel." This example also exposes a more fine-grained issue: L2W switches from using "we" to using "I" midgeneration. Such subtle distinctions are hard to capture during beam re-ranking and none of our models address the linguistic issues of this subtlety. Alternative Decoding Objectives A number of papers have proposed alternative decoding objectives for generation Pragmatic Communication Models Models for pragmatic reasoning about communicative goals such as Grice's maxims have been proposed in the context of referring expression generation Generative Adversarial Networks GANs However, backpropagating through discrete sequences and the inherent instability of the training objective Generation with Long-term Context Several prior works studied paragraph generation using sequence-to-sequence models for image captions We proposed a unified learning framework for the generation of long, coherent texts, which overcomes some of the common limitations of RNNs as text generation models. Our framework learns a decoding objective suitable for generation through a learned combination of sub-models that capture linguistically-motivated qualities of good writing. Human evaluation shows that the quality of the text produced by our model exceeds that of competitive baselines by a large margin.
| 755 | 1,169 | 755 |
Transformer-based Entity Typing in Knowledge Graphs
|
We investigate the knowledge graph entity typing task which aims at inferring plausible entity types. In this paper, we propose a novel Transformer-based Entity Typing (TET) approach, effectively encoding the content of neighbors of an entity. More precisely, TET is composed of three different mechanisms: a local transformer allowing to infer missing types of an entity by independently encoding the information provided by each of its neighbors; a global transformer aggregating the information of all neighbors of an entity into a single long sequence to reason about more complex entity types; and a context transformer integrating neighbors content based on their contribution to the type inference through information exchange between neighbor pairs. Furthermore, TET uses information about class membership of types to semantically strengthen the representation of an entity. Experiments on two real-world datasets demonstrate the superior performance of TET compared to the state-of-the-art.
|
A knowledge graph (KG) In light of this, it has been recently investigated the Knowledge Graph Entity Typing (KGET) task, aiming at inferring missing entity types in a KG. Most existing approaches to KGET use methods based on either embeddings or graph convolutional networks (GCN). Despite the huge progress these methods have made, there are still some important challenges to be solved. On the one hand, most embedding-based models The main objective of this paper is to introduce a transformer-based approach to KGET that addresses the highlighted challenges. The transformer architecture We use this class membership information for replacing the 'generic' relation has_type with a more fine-grained relation that captures to which class a type belongs to, enriching the semantic content of connections between entities and types. To sum up, our contributions are: • We propose a novel transformer-based framework for inferring missing entity types in KGs, encoding knowledge about entity neighbors from three different perspectives. • We use class membership of types to replace the single has_type relation with class-membership relations providing fine-grained semantic information. • We conduct empirical and ablation experiments on two real-world datasets, demonstrating the superiority of TET over existing SoTA models. Data, code, and an extended version with appendix are available at
|
The knowledge graph completion (KGC) task is usually concerned with predicting the missing head or tail entities of a triple. KGET can thus be seen as a specialization of KGC. Existing KGET methods can be classified in embedding-and GNC-based. Embedding-based Methods. ETE ConnectE GCN-based Methods. Graph Convolutional Networks (GCNs) have proven effective on modeling graph structures A common problem with these methods is that they follow a simple single-layer attention formulation, restricting the information transfer between unconnected neighbors of an entity. Transformer-based Methods. To the best of our knowledge, there are no transformer-based approaches to KGET. However, two transformerbased frameworks for the KGC task have been already proposed: CoKE In this section, we describe the architecture of our TET model (cf. Figure In this paper, a knowledge graph , where E is a set of entities, C is a set of entity types, R is a set of relation types, and T is a set of triples. Triples in T are either relation assertions (h, r, t) where h, t ∈ E are respectively the head and tail entities of the triple, and r ∈ R is the edge of the triple connecting head and tail; or entity type assertions (e, has_type, c), where e ∈ E, c ∈ C, and has_type is the instance-of relation. For e ∈ E, the relational neighbors of e is the set {(r, f ) | (e, r, f ) ∈ T }. The type neighbors of e are defined as {(has_type, c) | (e, has_type, c) ∈ T }. We will simply say neighbors of e when we refer to the relational and type neighbors of e. The goal of this paper is to address KGET task which aims at inferring missing types from C in entity type assertions. In this section, we introduce the local, global and context transformer-based modeling components of our TET model. Before defining these components, we start by discussing an important observation. A key observation is that in a KG all type assertions are uniformly defined using the relation has_type. Local Transformer [CLS] The local transformer mechanism is suitable for types that can be inferred by looking at simple structures, and for which independently considering neighbors is thus enough. However, inferring 'complex' types requires to capture the interaction between different neighbors of an entity. For instance, if we would like to infer that the entity Birmingham_City_L.F.C. has type Women's_football_clubs_in_England, we need to simultaneously consider different sources of information to support this, such as the type neighbor (has_type, Association_football_clubs) and relational neighbor (isLocatedIn, England) of Birm-ingham_City_L.F.C., and that (playsFor, Birming-ham_City_L.F.C.) and (hasGender, female) are relational neighbors of the entity Darla_Hood. To this aim, we introduce a global transformer module capturing the interaction between type-class and relational neighbors by comprehensively representing them as the input of a transformer as follows: 1. For a target entity e, we define the set Γ ′ as done in Section 3.2.2. Further, let Ξ = {(r 1 , f 1 ), . . . , (r m , f m )} denote the set of all relational neighbors of e. 2. We uniformly represent Γ ′ and Ξ as a single sequence 3. For each element in the sequence G, we assign randomly initialized word and position embeddings, and input it into a transformer. The output embedding of [CLS] is denoted G cls ∈ R d×1 . Similar to Equation (1), we define the prediction score S glo ∈ R L×1 as WRelu([G cls ]) + b. For an entity e, the local, global, and context transformers may generate multiple entity typing inference results. To address this, we adopt an exponentially weighted pooling method to aggregate prediction results S e = pool({S loc 0 , S loc 1 , ..., S loc m+n-1 , S glo , S ctx }) S e ∈ R L represents the relevance score between e and its types, and n (m) is the number of type-class (relational) neighbors of e respectively. For simplicity, we will omit the identifiers (loc,glo,ctx). We unify the numerical order of the output results of the local, global, and context transformers as follows: S e =pool({S 0 , S 1 , ..., S m+n-1 , S m+n , S m+n+1 }) We further apply a sigmoid function to S e , denoted as s e = σ(S e ), to map the scores between 0 and 1, where the higher the value of s e,k of s e , the more likely is e to have type k. The main intuition behind the local component is that the neighbors of an entity might help to determine its types, and that the contribution of each neighbor is different. For instance, if the entity Liverpool has the relational neighbor (places_lived, Daniel Craig), it is plausible to infer Liverpool has type /location/citytown. On the other hand, the neighbor (sports_team, Liverpool F.C.) may help to infer that it has type /sports/sports_team_location. To encode type-class neighbors (r class , c), similar to the input representations of BERT after the local transformer representation. The local transformer mainly pays attention to a single existing neighbor at a time in the inference process, reducing the interference between unrelated types. We perform a non-linear activation on neighbors, and then perform a linear layer operation to unify the dimension to the number of types, the final local transformer score S loc ∈ R L×(m+n) is defined as: An important observation is that the number of relations available vary from one KG to another. For instance, the YAGO43kET KG has substantially fewer relations than the FB15kET KG (cf. the dataset statistics in the Experiments Section), making the discrimination among relations in relational triples harder. To tackle this problem, for the YAGO43kET KG, we semantically enrich the representation of relations by using the type-class membership information. Specifically, for a relational neighbor (r, f ) of an entity, we use the types of f belonging to a certain class to enhance the relation r in the sequence ([CLS], r, f ) using the following steps: 1. Let Γ = {(has_type, c 1 ), (has_type, c 2 ), . . . , (has_type, c ℓ )} be the set of all type neighbors of f . We replace Γ with the set Γ ′ of corresponding type-class neighbors: {(r class 1 , c 1 ), (r class 2 , c 2 ), . . . , ((r class ℓ , c ℓ )}, i.e., representing that c i is a member of class i . 2. Based on r and Γ ′ , we construct a sequence P = (r, r class 1 , c 1 , r class 2 , c 2 , . . . , r class ℓ , c ℓ ). For each element p i of P , we assign randomly initialized word and position embeddings to capture sequence order. We then apply a transformer to capture the interaction between tokens. The output token embeddings are denoted as [p 0 , p 1 , . . . , p ℓ ]. 3. For the output token embeddings, we use three different operations to obtain the final representation of relation r: average, maximum, and minimum. For the YAGO43kET KG, we replace the word embedding r in sequence Q with P avg = ℓ i=0 p i , P max = Max(p i ), or P min = Min(p i ). For complex types, the global transformer uniformly serializes the information about the neighbors of the target entity. However, the neighbors of the target entity are pairs, and this structural information might be useful for inference. For instance, to infer that the entity Barack Obama has type 20th-century American writers, we need to consider different aspects of its relational neighbors, e.g., the neighbor (bornIn, Chicago) focuses on the birthplace, while the neighbor (write, A Promised Land) is concerned with possible careers. The global transformer serialization of pairs as a sequence may lead to two problems: First, serializing neighbors disregards the structure of the graph. Second, the importance of each element in the sequence is the same, and even elements that are not relevant for the inference will exchange information, e.g., bornIn and A Promised Land in the example above. To realize a differentiated aggregation between different neighbor pairs while preserving the graph structure, we use a context transformer module as in To train a model with positive sample score s e,k (representing that (e, has_type, k) exists in a KG) and negative sample score s ′ e,k (representing that (e, has_type, k) does not exist in KG), usually binary cross-entropy (BCE) is used as the loss function. However, there may exist a serious false negative problem, i.e., some (e, has_type, k) are valid, but they are missing in existing KGs. To overcome this problem, false-negative aware loss functions (FNA) have been proposed For the positive score s e,k and negative score s ′ e,k , the SFNA loss is defined as follows: 4 Experiments In this section, we discuss the evaluation of TET relative to twelve baselines on a wide array of entity typing benchmarks. We first describe datasets and baseline models (Sec. 4.1). Then we discuss the experimental results (Sec. 4.2). Finally, we present ablation study experiments (Sec. 4.3). Datasets. We evaluate our proposed TET model on two real-world knowledge graphs: FB15k (Bordes et al., 2013) and YAGO43k Baselines. We compare TET with twelve stateof-the-art entity typing methods, and their variants. We consider the embedding-based models ETE Table To verify the impact of each TET model component on the performance, we conduct ablation studies on FB15kET and YAGO43kET. In particular we look at the effect of: a) different transformer modules, achieves almost the same result as when the context one is also incorporated. We believe that in datasets with a more complex structure the context transformer could play a more prominent role, we leave this line of research as future work. Effect of Neighbor Content. We observe that the impact of relational neighbors is greater than that of type-class neighbors. Indeed, removing relational neighbors leads to a substantial performance degradation in YAGO43kET. When both of them are available, type-class neighbors might help relational ones to distinguish between relevant and irrelevant types for an inference. Effect of Dropping Rates. In real life KGs, many entities have sparse relations with other entities. In particular, they have few relational neighbors but a large number of types, so for their inference we lack structural relational information. Indeed, in YAGO43kET about 4.73% of its entities have five times more types than relational neighbors We also conduct an ablation study on the effect of using different number of hops. In principle multi-hop information could provide richer structural knowledge, increasing the discrimination of relational neighbors. Indeed, a positive effect of multi-hop information has been witnessed in several approaches to KGC. However, our experimental results show that the noise introduced by intermediate entities is more dominant than the additional knowledge n-hop entities and relations provide. Intuitively, for KGC multi-hop information makes a difference as it exploits the topological structure of the KG (i.e. how entities are related). However, in the input KG, types are not related between them and as our experiments show, one can not lift the topological structure at the entitylevel to the type one, explaining why there is no gain from considering multi-hop information. It would interesting to confirm this observation by using GCNs, which more naturally capture multi-hop information. In this paper, we propose a novel transformer-based model for KGET which utilizes contextual information of entities to infer missing types for KGs with minimal schema information. TET has three modules allowing to encode local and global neighborhood information from different perspectives. We also enhance the representation of entities by using class membership knowledge of types. We experimentally showed the benefits of our model. Our TET model currently suffers from two limitations. From the methodological viewpoint, a transformer mechanism introduces more parameters than embedding-based methods, bringing some computational burden and memory overhead, but they are tolerable. Also, there exist other important tasks related to types, e.g. fine-grained entity typing, aiming at classifying entity mentions into fine-grained semantic labels. TET is currently not appropriate for this kind of tasks. conduct an ablation study in which on FB15kET we randomly remove 25%, 50%, 75%, and 90% of the relation types. The results in Table
| 1,000 | 1,397 | 1,000 |
CAPSTONE: Curriculum Sampling for Dense Retrieval with Document Expansion
|
The dual-encoder has become the de facto architecture for dense retrieval. Typically, it computes the latent representations of the query and document independently, thus failing to fully capture the interactions between the query and document. To alleviate this, recent research has focused on obtaining query-informed document representations. During training, it expands the document with a real query, but during inference, it replaces the real query with a generated one. This inconsistency between training and inference causes the dense retrieval model to prioritize query information while disregarding the document when computing the document representation. Consequently, it performs even worse than the vanilla dense retrieval model because its performance heavily relies on the relevance between the generated queries and the real query. In this paper, we propose a curriculum sampling strategy that utilizes pseudo queries during training and progressively enhances the relevance between the generated query and the real query. By doing so, the retrieval model learns to extend its attention from the document alone to both the document and query, resulting in high-quality queryinformed document representations. Experimental results on both in-domain and out-ofdomain datasets demonstrate that our approach outperforms previous dense retrieval models.
|
Text retrieval aims to find the relevant documents for a given query from a large collection of documents, playing an indispensable role in opendomain question answering to compute relevance between the query and document. Recently, there has been a surge of research interest in neural network-based dense retrieval In recent years, the dual-encoder architecture has been a standard workhorse for dense retrieval. One major disadvantage of this architecture is that it can only partially extract the interactions between the query and document, since it encodes them separately. By comparison, the cross-encoder architecture can effectively capture the deep correlation between them by taking the concatenation of the query and document as input. By directly concatenating the query and document, the cross-encoder gains an advantage in capturing interactions, but also loses the advantage of pre-computing document representations during inference. Therefore, cross-encoder cannot wholly replace dual-encoder. To enhance the retrieval models' ability to capture interactions between queries and documents while maintaining retrieval efficiency, previous work mainly focuses on generating query-informed document representations. One approach, known as late interaction Compared with late interaction, dual-crossencoder (i.e., dense retrieval with document expansion) gets the query-related document representation without sacrificing the retrieval efficiency at inference. However, there exists a discrepancy between training and inference in the current dualcross-encoder retriever. Specifically, during training, the document is expanded using a real query, whereas during inference, the document is enriched with a generated query. This discrepancy causes the learned retriever overly focus on the query, yet neglect the document, when computing the document representation. During inference, if the generated query q′ significantly differs from the user-input query q, the query-related document representation will be misled by q′, thus degrading the performance. That is why the dual-cross-encoder even underperforms the vanilla dual-encoder. To address this issue, In this paper, we propose CAPSTONE, a curriculum sampling for dense retrieval with document expansion, to bridge the gap between training and inference for dual-cross-encoder. Our motivation is to expect the dual-cross-encoder retrieval model can utilize both the document d and pseudo query q′ to compute the query-informed document representation. To achieve this, we train the dual-cross-encoder retriever by gradually increasing the relevance of the pseudo query q′ to the gold query q. Specifically, at the early training stage, a pseudo query q′ irrelevant to q is selected, causing the retriever to solely rely on the document. As we progress to the late training stage, a highly related pseudo query q′ is chosen, allowing the retriever to learn to augment the document representation with the pseudo query. By doing so, we alleviate the discrepancy between training and inference. During inference, if the user-input query q is similar to the pseudo query q′, then q′ will contribute more to making the target document d be retrieved. Otherwise, the retrieval model will mainly rely on the relevance between q and d. To summarize, the main contributions of this paper are as follows: (1) We propose a curriculum learning approach to bridge the gap between train-ing and inference for dense retrieval with document expansion, further improving the query-informed document representation
|
In this section, we introduce the definition of text retrieval and three architectures for dense retrieval. Task Description. Text retrieval is meant to find the most relevant M documents D + = {d + 1 , d + 2 , . . . , d + M } for the given query q from a large corpus D = {d 1 , d 2 , . . . , d N } with N documents (M ≪ N ). Architecture. Dual-encoder (DE) is the typical dense retrieval architecture. As shown in Figure (1) Cross-Encoder Architecture. Since DE models q and d separately, it is not good at capturing the relevance between them. To capture the interactions between them, cross-encoder (CE) directly takes the concatenation of q and d as input, as shown in Figure where '+' is the concatenation operation. Dual-Cross-Encoder Architecture. Although CE can extract more fine-grained relevance between q and d, it is not suitable for retrieval. To combine the advantages of DE and CE, (3) In Section 3.1, we first identify the discrepancy between training and inference of DCE. Then, we introduce how to bridge the gap with curriculum learning in Section 3.2. Finally, we will show our proposed inference method in Section 3.3. The training objective of DE is to learn dense representations of queries and documents to maximize the similarity score between the query and positive document. The training loss is defined as follows: In-batch Negatives where D -is the set of negative documents, containing hard negatives and in-batch negatives. During training, DCE expands the positive document d and hard negatives with the gold query q (e.g., replacing d with d + q). For ease of understanding, we show how to construct positive, hard negatives and in-batch negatives in the left of Figure At inference, DCE enriches the document with the generated query q ′ rather than the user-input query q. When q and q ′ are different types of queries for the target document d, q or E q (q) will be far from q ′ or E d (q ′ ). As the retrieval model We first compute the relevance score M (q, q ′ i ) between the generated query q ′ i and the gold query q, and then sort the generated queries in ascending order according to their relevance scores to q. Next, we divide them into K groups. At the i-th training iteration, we randomly sample one generated query from the i-th group, and then use the sampled query to update the dual-cross encoder. overly relies on the query part when computing the document representation, E d (d + q ′ ) will be close to E d (q ′ ). Based on these, we can easily infer that E d (d + q ′ ) is far away from E q (q), making it more difficult to retrieve the target document d. As a result, the dense retriever with document expansion even underperforms its counterpart without using document expansion. We attribute the performance degradation to the discrepancy between training and inference. To bridge the gap between training and inference, we propose to expand the document with its generated query during training. We show the training process of our proposed method in the right of Figure Figure both the document and the generated query during training, so that the learned document representation contains the information of both parts. To fulfill this goal, we further propose the curriculum sampling. To be concrete, at the early training stage, the selected q′ has low correlations with q, forcing the retriever to use the document. As the training goes on, we select q′ with gradually increased relevance to q, encouraging the retriever to use the query. Driven by this motivation, we first generate some queries for each document d. Then, we compute the relevance score between the generated query q ′ i and the gold query q with an automatic evaluation metric M (i.e., ROUGE-L Before retrieving, we first generate S queries for each document with the generator. Next, we concatenate a document with one generated query and compute its latent representation with the document encoder. Then, we will get S different representations for each document. Corpus Expansion. Following Computing the Typical Representation of Different Views. To avoid increasing the retrieval latency, we can retain a single representation for each document. The easiest way is to set S to 1. However, we found that the retrieval performance is positively correlated to the number of views, S. This is because a larger value of S increases the probability of the user-input query being relevant to one of the generated queries. Therefore, setting S to 1 will degrade the retrieval performance compared with expanding the corpus. Given this, we propose to compute the typical representation for a document by averaging all S different views. This approach allows us to maintain only one representation for each document. As shown in Figure Datasets. We conduct experiments on three passage retrieval datasets: MS-MARCO passage ranking Evaluation Metrics. Following previous work, we use MRR@10, Recall@50, and Recall@1000 to evaluate the retrieval performance on MS-MARCO, where MRR@10 is the most important metric. We resort to nDCG@10 for TREC DL Implementation Details. We set the maximum document length to be 144 tokens, and the maximum query length to be 32 tokens. All models are optimized using the AdamW optimizer The query generator is built upon the seq2seq architecture During training, we implement our proposed curriculum sampling with all 80 queries and divide them into K groups, but at inference, we resort to the first 10 (i.e., S = 10) and first 5 (i.e., S = 5) queries to compute the typical document representation for MS-MARCO and BEIR benchmark, respectively. At the first training stage, we set K to 3, and at the second training stage, we set K to 4. TREC DL 19 TREC DL 20 MRR@10 R@50 R@1000 nDCG@10 nDCG@10 Sparse retrieval BM25 In-domain Performance. Table Zero-shot Performance. To test the out-ofdomain generalization capabilities, we first use the T5-base model To analyze our method more directly, experiments in this section are based on the vanilla DPR initialized with the ERNIE-2.0-base model In this paper, we append only one query to a document both at training and inference. As stated in 3.3, corpus expansion will enlarge the original corpus to S times during inference. In contrast, document expansion for sparse retrieval (2) document expansion; (3) asymmetric expansion, which appends one query to a document at training, but appends S queries to a document at inference. As shown in Figure Effect of Query Selection Strategies. To demonstrate the advantage of our proposed curriculum sampling, we expand the positive and negative doc- Relevance to query q Figure 6: Illustration of document expansion and corpus expansion at inference. For simplicity, only the positive document d i of the query q is shown. uments with different query selection strategies: (1) gold, meaning using the gold query (i.e., strategy used by the vanilla DCE From Figure (3) If the document is expanded with a weakly related (bottom-1) or random query at training, DPR with both settings will outperform DPR w/o query when S = 1. These two strategies enable DPR to mine query information without sacrificing document information. However, their performances are not strongly positively correlated with S like the top-1 strategies, indicating it does not fully use the query information. (4) DPR with the curriculum strategy outperforms DPR w/o query when S = 1, and its performance linearly increases with S, verifying this strategy enables DPR to utilize both the document and query to model the query-informed document representation. Effect of the Typical Representation. Although corpus expansion makes the retrieval performance of our proposed curriculum sampling improve with the increase of S (see the solid red line in Figure Comparison of Methods for Computing the Typical Representation. We consider three different pooling methods to compute the typical document representation: taking the average/max/median pooling of different document views. From Table Multi-stage Retrieval Performance. To further verify the effectiveness of our proposed approach, we apply our proposed approach to DPR and co-Condenser. From Table 5 Related Work In recent years, with the development of largescale pre-trained language models, such as BERT Dense retrieval models typically depend on extensive supervised data, comprising pairs of queries and positive documents. To address the challenge of limited training data, Query expansion enriches the query with various heuristically discovered relevant contexts. For example, GAR Document expansion augments the document with generated queries, which the document might answer. Compared with query expansion, document expansion can be conducted prior to indexing without incurring extra retrieval latency. Document expansion has shown its effectiveness on sparse retrieval models This work proposes the curriculum sampling for dense retrieval with document expansion, which enables dense retrieval models to learn much better query-related document representations. In addition, we propose to compute the typical representation of different document views, which balances inference efficiency and effectiveness. Our experimental results on the in-and out-of-domain datasets verify the effectiveness of the curriculum sampling and typical representation. There are two possible limitations of this work. The first limitation is that we need to generate synthetic queries for each document in the retrieval corpus, which is very time-consuming. Luckily, this process does not bring extra delay to retrieval. In addition, limited by sufficient computational resources, we only verify the effectiveness of our method on vanilla DPR and coCondenser. In the future, we plan to apply our approach to other dense retrieval models and verify the effectiveness of our method on these models.
| 1,366 | 3,568 | 1,366 |
Iterative Constrained Clustering for Subjectivity Word Sense Disambiguation
|
Subjectivity word sense disambiguation (SWSD) is a supervised and applicationspecific word sense disambiguation task disambiguating between subjective and objective senses of a word. Not surprisingly, SWSD suffers from the knowledge acquisition bottleneck. In this work, we use a "cluster and label" strategy to generate labeled data for SWSD semiautomatically. We define a new algorithm called Iterative Constrained Clustering (ICC) to improve the clustering purity and, as a result, the quality of the generated data. Our experiments show that the SWSD classifiers trained on the ICC generated data by requiring only 59% of the labels can achieve the same performance as the classifiers trained on the full dataset.
|
Subjectivity lexicons (e.g., (1) He was attacked by Milosevic for attempting to carve out a new party from the Socialists. Knowing that "attack" is a subjectivity clue with negative polarity will help a system recognize the negative sentiment in the sentence. But for (2), the same information is simply misleading, because the clue is used with an objective meaning. (2) A new treatment based on training T-cells to attack cancerous cells ... Any opinion analysis system which relies on a subjectivity lexicon will be misled by subjectivity clues used with objective senses (false hits). In Although SWSD is a promising tool, it suffers from the knowledge acquisition bottleneck. SWSD is defined as a supervised task, and follows a targeted approach common in the WSD literature for performance reasons. This means, for each target clue, a different classifier is trained requiring separate training data for each target clue. It is expensive and time-consuming to obtain annotated datasets to train SWSD classifiers limiting scalability. As a countermeasure, in Although non-expert annotations are cheap and fast, they still incur some cost. In this work, we aim to reduce the human annotation effort needed to generate the same amount of subjectivity sense tagged data by using a "cluster and label" strategy. We hypothesize that we can obtain large sets of labeled data by labelling clusters of instances of a target word instead of single instances. The main contribution of this work is a novel constrained clustering algorithm called Iterative Constrained Clustering (ICC) utilizing an active constraint selection strategy. A secondary contribution is a mixed word representation that is a combination of previously proposed context representations. We show that a "cluster and label" strategy relying on these two proposed components generates training data of good purity. The resulting data has sufficient purity to train reliable SWSD classifiers. SWSD classifiers trained on only 59% of the data achieve the same performance as classifiers trained on 100% of the data, resulting in a significant reduction in the annotation effort. Our results take SWSD another step closer to large scale application.
|
Our approach is inspired by a method lexicographers commonly employ to create sense inventories, where they create inventories based on evidence found in corpora. They use concordance information to mine frequent usage patterns. Training data for an SWSD classifier consists of instances of the target word tagged as having a subjective sense (S) or an objective sense (O) (subjectivity sense tagged data). We train a different SWSD classifier for each target word as in The induced clusters represent different usage patterns of a word. Thus, we build more than two clusters, even though SWSD is a binary task. This implies that two different instances of a word can both be subjective, but end up in different clusters, if they are different usages of the word. Since we are labelling clusters as a whole, we will introduce noise in the labeled data. Thus, in developing the clustering process, we need to minimize that noise and find as pure clusters as possible. The first step is to define the context representation of the instances. This is addressed in Section 3. Then, we turn in Section 4.2 to the clustering process itself. To evaluate our "cluster and label" strategy, we use two gold standard subjectivity sense tagged datasets. There has been much work on context representations of words for various NLP tasks. Clustering word instances in order to discriminate senses of a word is called Word Sense Discrimination. Context representations for this task rely on two main types of models: distributional semantic models (DSM) and feature-based models. We test both methods in this paper, using the same semantic space. The space is built from a corpus consisting of 120 million tokens. The rows of the space correspond to word forms and the columns correspond to word lemmas present in the corpus. We adopt the parameters for our semantic space from Purandere and Pedersen is the prominent representative of feature-based models. In this section, we evaluate context representations for the context clustering task on the subjectivity sense tagged data, senSWSD. The evaluation is done separately for each word. We use the same clustering algorithm for all context representations: agglomerative hierarchical clustering with average linkage criteria. In all our experiments throughout the paper, we fix the cluster size to 7 as it is done in Row 1 of Table When we look at the context clustering results for single words separately, we observe that the performance of different representations vary. There is not a single winner among all words. Thus, perhaps choosing one single representation for all the words is not optimal. Having that in mind, we try merging the dsm mul and local features representations. We leave out dsm add representation, since both dsm mul and dsm add rely on the same type of semantic information (i.e., a DSM). We hypothesize that the two representations, one relying on a semantic space and the other relying on local WSD features, may complement each other. To merge the representations, we concatenate the two feature vectors into one. First, however, we normalize each vector to unit length, since the individual vectors have different scales and would have unequal contribution, otherwise. We call this mixed representation mix rep. In Table We now turn to the clustering process. In a "cluster and label" strategy, in order to be able to label clusters, we need to annotate some of the instances in each cluster. Then, we can accept the majority label found in a cluster as its label. Thus, some manual labelling is required, preferably a small amount. We propose to provide this small amount of annotated data prior to clustering, and then perform semi-supervised clustering. This way the provided labels will guide the clustering algorithm to generate the clusters that are more suitable for our end task, namely clusters where subjective and objective instances are grouped together. Constrained clustering Constraints can be obtained from domain knowledge or from available instance labels. In our work, constraints are generated from instance labels. Each instance pair with opposing labels is considered to be cannot-linked. There are two general strategies to incorporate constraints into clustering. The first is to adapt the similarity between instances Our proposed constrained clustering method relies on some ideas from Although previous work report on average sub-stantial improvement in the clustering purity, Our proposed algorithm is closely related to (1) the agglomerative hierarchical clustering with complete-linkage and (2) the mechanism to impose cannot-link constraints described in Section 4.1. For our algorithm, we implement a second mechanism for imposing constraints proposed by ( We believe that learning a new distance metric will result in more reliable distance estimates between all instances. For this purpose, we learn a Mahalanobis distance function following the method described in The reason we choose the metric learning method As mentioned before, the choice of the set of constraints affects the quality of the end clustering. In this work, we define a novel method to choose informative instances, which we believe will have maximum impact on the end cluster quality, when they are labeled and used to generate constraints for our task. We use an iterative approach. Each iteration consists of three steps: (1) generating clusters by the process described in Section 4.2.1 imposing available constraints, (2) choosing the most informative instance, considering the cluster boundaries, and acquiring its label, (3) extending the available constraints with the ones we generate from the newly labeled instance. We consider an instance to be informative if there is a high probability that the knowledge of its label may change the cluster boundaries. The more probable that change is, the more informative is the instance. The basic idea is that if an instance is in a cluster holding instances of type a and it is close to another cluster holding instances of type b, that instance is most likely misclustered. Thus, it should be queried. Our hypothesis is that, in each iteration, the algorithm will choose the most problematic -informative -instance that will end up changing cluster boundaries. This will result in each iteration in a more reliable distance metric, which in return will provide more reliable estimates of problematic instances in future iterations. The imposed con- To define informativeness, we define a scoring function, which is used to score each data point on its goodness. The lower the score, the more likely it is that the instance is mis-clustered. Choosing the data point with the lowest score will likely change clustering borders in the next iteration. Our scoring function is based on the silhouette coefficient, a popular unsupervised cluster validation metric to measure goodness The silh const coefficient has favourable properties. First, it scores members that are close to a cluster with an opposing label lower than the members that are close to a cluster with the same label. According to our definition, these members are more informative. Figure The active selection strategy proposed by We call our proposed algorithm Iterative Constrained Clustering (ICC). In our final implementation, ICC starts by simply clustering the instances without any constraints. The algorithm queries the label of the prototypical memberthe member closest to the cluster center -of each cluster. Then, the described iterations begin. Algorithm 1 contains the complete ICC algorithm. Note that line 6 is equivalent to the algorithm of This section gives details on experiments to evaluate the purity of the semi-automatically generated subjectivity sense tagged data by our "cluster and label" strategy. We carry out detailed analysis to quantify the effect of the proposed active selection strategy and of metric learning on the purity of the generated data. We compare our active selection strategy to random selection and also to For comparison, we also evaluate the performance of annotations Now that we have a tool to generate training data for SWSD, we want to evaluate it on the actual SWSD task. We want to see if the obtained purity is enough to create reliable SWSD classifiers. For this purpose, we test ICC on mturkSWSD dataset. For each word in our dataset, we conduct 10fold cross-validation experiments. ICC is applied to training folds to label instances semiautomatically. We train SWSD classifiers on the generated training fold labels and test the classifiers on the corresponding test fold. We distinguish between queried instances and propagated labels. The queried instances are weighted as 1 and the instances with propagated labels are weighted by their silh const score, since that measure gives the goodness of an instance. The score is defined between -1 and 1. This score is normalized between 0 and 1, before it is used as a weight. SVM classifiers from the Weka package We implement two baselines. The first is simple random sampling and the second is uncertainty sampling, which is an active learning (AL) method. We use "simple margin" selection as described in It is surprising to see that uncertainty sampling overall does not do better than random sampling. We believe that it might be because of sampling bias. During AL, as more and more labels are obtained, the training set quickly diverges from the underlying data distribution. We see that our method reduces the annotation amount by 36% in comparison to uncertainty sampling and by 41% in comparison to random sampling to reach the performance of the SWSD system trained on fully annotated data. One related line of research is constrained clustering also known as semi-supervised clustering There have been only two previous works selecting constraints for constrained clustering actively Active Learning (AL) In this paper, we explore a "cluster and label" strategy to reduce the human annotation effort needed to generate subjectivity sense-tagged data. In order to keep the noise in the semiautomatically labeled data minimal, we investigate different feature space types and evaluate their expressiveness. More importantly, we define a new algorithm called iterative constrained clustering (ICC) with an active constraint selection strategy. We show that we can obtain a fairly reliable labeled data when we utilize ICC. We show that the active selection strategy we propose outperforms a previous approach by To our knowledge, this work is the first application of constrained clustering to a hard NLP problem. We showcase the power of constrained clustering. We hope that the same "cluster and label" strategy will be applicable to Word Sense Disambiguation. This will be part of our future work. This material is based in part upon work supported by National Science Foundation awards #0917170 and #0916046.
| 717 | 2,213 | 717 |
ASPECTNEWS: Aspect-Oriented Summarization of News Documents
|
Generic summaries try to cover an entire document and query-based summaries try to answer document-specific questions. But real users' needs often fall in between these extremes and correspond to aspects, high-level topics discussed among similar types of documents. In this paper, we collect a dataset of realistic aspect-oriented summaries, ASPECT-NEWS, which covers different subtopics about articles in news sub-domains. We annotate data across two domains of articles, earthquakes and fraud investigations, where each article is annotated with two distinct summaries focusing on different aspects for each domain.
|
Recent progress in text summarization In this work, we present a new dataset for evaluating single-document aspect-oriented extractive summarization which we call ASPECTNEWS. We derive subsets of examples from CNN/Daily Mail following certain topics, namely earthquakes and fraud reports. These domains are special in that the articles within them have several aspects which are repeatedly mentioned across articles and form coherent topics, e.g., impact on human lives of an earthquake. We ask annotators to select sentences relevant to such information needs, which correspond to imagined use cases. Interannotator agreement on full summaries is low due to the inherent subjectivity of the task, so rather than coming up with a consensus summary, we instead primarily evaluate against soft labels based on the fraction of annotators selecting a given sentence. To benchmark performance on this dataset, we build a system that can summarize a document conditioned on certain aspect-level keywords without assuming annotated training data for those aspects. Since there are no large-scale supervised training sets suitable for this purpose, we explore methods to generate aspect-oriented training data from generic summaries. We compare these with past approaches Our experiments on our ASPECTNEWS dataset
|
Figure 1: Examples of an earthquake-related article paired with extractive summaries from the CNN/DM dataset. "Generic" represents the selection of a general purpose summarization model. "Geo(graphy)" (colored in green) and "Recovery" (colored in orange) indicate our aspects of interest for the summary. We highlight aspect-relevant phrases in the document. and the SPACE dataset Relatively little recent work has focused on aspectoriented summarization. One line of research focuses on summarization of documents with respect to specific queries Methods Historically, most work on queryfocused summarization has addressed the multidocument setting. Datasets There are several differences between ASPECTNEWS and other existing aspect-oriented summarization datasets. Firstly, ASPECTNEWS focuses on single-document summarization, while similar aspect-oriented datasets such as the SPACE dataset of reviews The TAC 2010/2011 summarization datasets We begin by considering our target application: users who have specific information needs that they want to be satisfied. This consideration broadly falls under the category of purpose factors defined by Jones (1998) and should be accounted for in the summarization process. Our data collection process involves the following steps: (1) Identifying clusters of articles in our target domains from a large corpus of news summaries. (2) Manually specifying multiple user intents per target domain, representing the aspect of the summarization process. (3) Crowdsourcing annotation of extractive summaries in these domains based on the user intents. We draw our datasets from the English-language CNN/Daily Mail summarization dataset To retrieve our examples from these two domains, we first encode each article in CNN/DM corpus C with a text encoder E. We adopt the Universal Sentence Encoder We found this procedure to be more robust than simple keyword matching for retrieving articles with coherent aspects; for example, keyword matching for "earthquakes" resulted in returning articles primarily about tsunamis due to the imbalanced data distribution. With these two domains, we examine our dataset to derive aspects that simulate realistic information needs of users. Table Finally, to construct actual extractive summaries for evaluation in these domains, we presented the user intents to annotators on Amazon Mechanical Turk. An annotator is shown a description of intent from Table Each article was truncated to 10 sentences for ease of annotation. This assumption was reasonable for the two domains we considered, and the truncation approach has been used in In Table We also compare the overlap between aspectoriented annotation and generic extractive oracle derived from reference summaries from CNN/DM. In Table Our aspect-oriented data collection works well to create labeled evaluation data, but it is difficult to scale to produce a large training set. Identifying suitable domains and specifying user intents requires significant human effort, and collecting real test cases at scale would require a more involved user study. We build an aspect-oriented model without goldlabeled aspect-oriented training data. We do this by generating keywords for each article in CNN/DM, and training the model to learn the relationship between these keywords and a summary. Our system follows broadly similar principles to He et al. We present a scheme to generate keywords for each document from the original dataset. CNN/DM consists of pairs (D, S) of a document D and associated summary S. We aim to augment these to form (D, K, S ) triples with keywords K and a possibly modified summary S . Our mixed augmentation technique requires training the model on both (D, S) and (D, K, S ) for a given document. We now describe the steps to create this data. Keyword Extraction For each document in CNN/DM, we calculate the most important tokens in that document according to their TF-IDF ranking with respect to the entire corpus. Of these tokens, we select the ones that are present in the reference summary. This process selects tokens that are more likely to be consequential in affecting the output summary. To instill stronger dependence on the keywords, we made two modifications to this process. First, we modified the reference summary by concatenating the keywords with the reference summary before computing the extractive oracle summary. This concatenation makes the oracle extraction more likely to select sentences containing the keywords, though modifying the reference summary requires maintaining a balance between the influence of keywords and of the original gold summary. Second, we use BERTScore We can vary the parameter k controlling the number of times we append the keywords to the reference summary in order to generate the oracle extractive summary. We experiment with different level of intensity and show the result in Table (CNN) -A 7.2-magnitude earthquake has struck south of the Mariana Islands, according to the U.S. Geological Survey. We explore a variant of training where we include training data with multiple variants of each original document from the dataset. Each document in the original dataset is mapped to two training samples, (1) a document without keywords and an unmodified oracle extractive summary, (2) a document with keywords and an oracle extractive summary using our modification procedure. Our model is trained to predict a summary S from a document-keywords pair (D, K). Following BERT-SUM We refer to this model, trained on our BERTScore references with the mixed training scheme, as AOSUMM. We evaluate our model on the dataset, comparing performance on aspect-oriented summarization to several baselines. We additionally experiment on the SPACE multi-document dataset On ASPECTNEWS, we evaluate our model against the annotations using using F 1 score and ROUGE scores. It is impossible to achieve 100 F 1 on this task due to inherent disagreement between annotators. One downside of F 1 is that the model may be penalized even when the predicted sentence is very similar to the annotation, for this reason we also calculate ROUGE-1, -2, and -L scores On the SPACE corpus, we primarily focus on comparisons to quantized transformer (QT) Heuristic and QA Baselines KEYWORD takes the keywords described in Table We also compare our AOSUMM model against text summarization models, and query-focused models from previous work (retrained or off-the-shelf). (i) BERTSUM is a bert-base-cased extractive summarization model fine-tuned on CNN/DM The experimental results on AS-PECTNEWS are shown in Table The results on all the aspects of the SPACE dataset are shown in Table Keyword Sensitivity We evaluate the sensitivity of the model to different keywords. There is some overlap between the summaries returned by different keyword sets, as shown by the Jaccard similarity: some sentences may fit under both GEO and RECV, or both PEN and NATURE. Table The quake struck about 375 kilometers (233 miles) westsouthwest of Hagatna, Guam, and 445 kilometers (276 miles) west-southwest of Rota, Northern Mariana Islands. CTRLSUM GEO NEW: The location of the quake is in the province of Yucatán. NEW: There are no immediate reports of casualties or damage. The quake is centered about 375 kilometers (233 miles) west-southwest of Hagatna, Guam. The U.S. Geological Survey says it was a 7.2-magnitude quake. The earthquake is centered in the Yucatan province of Mexico. The country's geography is similar to that of the U.N. region. CTRLSUM RECV NEW: The death toll from the quake is not immediately known. The U.S. Geological Survey reports a 7.2-magnitude quake. The Mariana Islands sit about three-quarters of the way from Hawaii to the Philippines. "There is a survivor. There is an injury. There will be an aid.recovery. process," the U.N. secretary-general says. The quake is centered about 375 kilometers (233 miles) west-southwest of Hagatna, Guam. Table adding new information. Although such behavior may possibly perform well on automated metrics, it does not serve the purpose of query-focused summarization. Extractive summaries Table The Pacific Tsunami Warning Center did not issue a tsunami warning after the quake, which struck at 7:19 a.m. Saturday (5:19 p.m. ET Friday). R "We wouldn't expect any kind of significant tsunami for this event," said the center's director, Charles McCreery, noting that the quake's magnitude was relatively low to provoke one. R There were no immediate reports of casualties or damage, emergency management officials said. In this paper, we present a new dataset for aspectoriented summarization of news articles called AS-PECTNEWS. Unlike query-focused summarization datasets which are often driven by document specific facts or knowledge, this aspect-oriented task is designed to mimic common user intents in domain-specific settings. We present a keywordcontrollable system trained on synthetic data and show that it can perform well on ASPECTNEWS without training on the target domains, performing
| 618 | 1,305 | 618 |
Twitter Homophily: Network Based Prediction of User's Occupation
|
In this paper, we investigate the importance of social network information compared to content information in the prediction of a Twitter user's occupational class. We show that the content information of a user's tweets, the profile descriptions of a user's follower/following community, and the user's social network provide useful information for classifying a user's occupational group. In our study, we extend an existing dataset for this problem, and we achieve significantly better performance by using social network homophily that has not been fully exploited in previous work. In our analysis, we found that by using the graph convolutional network to exploit social homophily, we can achieve competitive performance on this dataset with just a small fraction of the training data. * Equal Contribution; work performed while both authors were visiting Singapore University of Technology and Design (SUTD).
|
Twitter ( Profiling users can enhance service quality and improve product recommendation, and hence is a widely studied problem. User occupational class prediction is an important component of user profiling and a sub-task of user demographic feature prediction. Existing approaches to predicting Twitter users' demographic attributes explore, select, and combine various features generated from text and network to achieve the best predictive performances in respective classification tasks As for user occupational class prediction, Preot ¸iuc-Pietro et al. ( In this paper, we analyze the usefulness of a user's network information over the user's tweets for predicting its occupational group. We extend the existing dataset for occupation classification (Preot ¸iuc-Pietro et al. (
|
A Graph Convolutional Network (GCN) (1) where X is the feature matrix for all the nodes with X (0) being the initial feature input of size d nodes × d f eatures , A is the adjacency matrix of dimension d nodes × d nodes , D is the degree matrix of A + λI, λ is a hyperparameter controlling the weight of a node against its neighbourhood, and W (l) and b (l) are trainable weights and bias for the l-th layer, respectively. In each layer of GCN, a node aggregates its direct neighbours' features according to  and linearly transforms the representation using W and b. A nonlinear activation function σ (e.g., ReLu) is then applied. The number of layers of GCN decides the number of hops away that the neighbours' features will be smoothed over for each node. 3 Experimental Setup We base our work on a publicly available Twitter dataset that maps 5,191 users to 9 major occupational classes To explore the role of network information in occupational class prediction, we extend the above dataset by crawling follower/following IDs (henceforth referred to as follow IDs) for each main ID (IDs corresponding to main users). For the crawled follow IDs, we further crawl their Bio descriptions. We refer to the extended dataset as ED. ED contains 4,557 main users with both followers and followings information. The remaining Twitter accounts could not be scrapped because of various reasons such as account suspension and protected tweets. Table Thus, we decide to filter the graph by keeping the follow IDs with more than 10 connections to the main IDs. All connections between main IDs are retained. The filtering step results in 29 main IDs losing all their connections. For all such isolated main IDs, we retrieve all its follow IDs having at least one other main ID connection. After all these operations, we are able to construct an un-weighted graph in which all the main IDs are connected. The filtered graph contains 34,630 unique users (including 4,557 main IDs) and 586,303 edges. Although the main users are not collected to be connected to each otheronly 2,550 main IDs have at least one direct connection to another main ID, we find that they often share common follow IDs which allows us to retrieve their social representations. To compare with previous works, we also construct a partial network dataset that contains only following IDs of all the 4,557 main IDs. We refer to this partial dataset as PD. PD adheres to the same network construction methodology as ED. We divide the dataset into training, development, and test sets using stratified split with the splitting ratio of 80%, 10%, and 10%. All the experimental results are reported on the same test set. The split information and the processed dataset ED can be found together with code on github: Node Embeddings: To encode user-user social relationship of main IDs with the follow network, we learn latent representations of all IDs (node embedding) which can be easily exploited for the prediction task. The embeddings are learned by forming node sequences using Deep Walk Based on the network processing strategy used in Next, we performed 10 random walks starting from each main ID, alternating between main ID and followers/followings with a walk length of 80. For each node, the walk sequence is used to generate embeddings using a similar approach to word2vec Text Features: To have a valid comparison with existing approaches, we construct two sets of text features: (1) bag-of-clusters (Preot ¸iuc-Pietro et al., 2015): we assign each word that appears in each main ID's concatenated tweets document to its corresponding word cluster, where the word clusters are obtained by applying spectral clustering ure We also test another setting where we do not use the Bio information: we keep the feature as a matrix of one-hot encoded vectors corresponding to all 34,630 IDs. For all the experiments, we set λ to 1 in Equation As shown in Table We first conduct experiments on our collected ED dataset with 4,557 main users using existing methods. The better accuracy among existing methods is given by the concatenated bag-ofclusters and Deepwalk embeddings: 55.0%. Next, we investigate the performance of bagof-words features from main ID tweets and follow Bios using logistic regression (LR) and support vector machines (SVM). From the experiments on tweets, we find that using the bag-ofwords features achieve comparable performance to using the bag-of-clusters features. Thus we opt for the bag-of-words representation in subsequent experiments. The optimized model using Bio text features outperforms using tweet content. It can be inferred that the Bio descriptions of follow accounts provide more useful information compared to tweets. The reason could be the higher noise in tweets, while people are comparatively more careful while writing their Bios. The next set of results uses follow network features. Based on SVM. We achieve higher accuracy (55.3%) as compared to tweets BOW (54.6%). However, the model is less effective than using follow Bio BOW. Combining both node representations and follow Bio BOW features further boosts the accuracy to 57.5%. To analyze the importance of Bios in conjunction with social network information, we exploit graph convolutional networks. With an accuracy of 59.9%, the model exceedingly outperforms existing approaches on tweets and partial network information. Our best result 61.0% accuracy is achieved by using GCN with one-hot encoding for nodes, which is significantly higher than existing methods. This shows that GCN is able to exploit the rich topological information of network to learn social representations for users. We postulate that the GCN with Bio did not do better than just a one-hot encoding for nodes because the main users do not have Bios: so all the labeled nodes in the GCN have no Bios, which makes learning difficult. We visualize the GCN final layer representations of training set (big ovals) and test set (dark colored dots) in Figure We analyze the predictions on test samples made by GCN with Bio feature input and GCN with the one-hot encoded input. We find that 11% of the test set's main IDs are correctly classified by only one of the two GCNs. This suggests that Bio features provide complementary information to the one-hot encoded input. In this work, the acquired network is dense. In cases when network is sparse, one-hot representation of an ID seems infeasible while BOW may generalize for the larger graph. While occupational class prediction could be used to improve service quality, we note that the use of network information might result in unintended consequences such as racial and ethnicity based segregation in online spaces. To alleviate such concerns, it would be useful in future to incorporate explainable predictions with work such as Previous works have used tweets or a fraction of the network information to extract features for occupation classification. To analyze the importance of network information, we extended an existing Twitter dataset for a user's social media connections (follow information). We showed that by using only follow information as an input to graph convolutional networks, one can achieve a significantly higher accuracy on the prediction task as compared to the existing approaches utilizing tweet-only information or partial network structure. Directions of future research include adaptation of our methods to a large scale, sparsely connected social network. One might also want to investigate the inductive settings of GCN
| 915 | 785 | 915 |
Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation
|
Disentangling the content and style in the latent space is prevalent in unpaired text style transfer. However, two major issues exist in most of the current neural models. 1) It is difficult to completely strip the style information from the semantics for a sentence. 2) The recurrent neural network (RNN) based encoder and decoder, mediated by the latent representation, cannot well deal with the issue of the long-term dependency, resulting in poor preservation of non-stylistic semantic content. In this paper, we propose the Style Transformer, which makes no assumption about the latent representation of source sentence and equips the power of attention mechanism in Transformer to achieve better style transfer and better content preservation. Source code will be available on Github 1 .
|
Text style transfer is the task of changing the stylistic properties (e.g., sentiment) of the text while retaining the style-independent content within the context. Since the definition of the text style is vague, it is difficult to construct paired sentences with the same content and differing styles. Therefore, the studies of text style transfer focus on the unpaired transfer. Recently, neural networks have become the dominant methods in text style transfer. Most of the previous methods These methods focus on how to disentangle the content and style in the latent space. The latent representation needs better preserve the meaning of the text while reducing its stylistic properties. Due to lacking paired sentence, an adversarial loss 1) It is difficult to judge the quality of disentanglement. As reported in 2) Disentanglement is also unnecessary. 3) Due to the limited capacity of vector representation, the latent representation is hard to capture the rich semantic information, especially for the long text. The recent progress of neural machine translation also proves that it is hard to recover the target sentence from the latent representation without referring to the original sentence. 4) To disentangle the content and style information in the latent space, all of the existing approaches have to assume the input sentence is encoded by a fix-sized latent vector. As a result, these approaches can not directly apply the attention mechanism to enhance the ability to preserve the information in the input sentence. 5) Most of these models adopt recurrent neural networks (RNNs) as encoder and decoder, which has a weak ability to capture the long-range dependencies between words in a sentence. Besides, without referring the original text, RNN-based decoder is also hard to preserve the content. The generation quality for long text is also uncontrollable. In this paper, we address the above concerns of disentangled models for style transfer. Different from them, we propose Style Transformer, which takes Transformer Our contributions are summarized as follows: • We introduce a novel training algorithm which makes no assumptions about the disentangled latent representations of the input sentences, and thus the model can employ attention mechanisms to improve its performance further. • To the best of our knowledge, this is the first work that applies the Transformer architecture to style transfer task. • Experimental results show that our proposed approach generally outperforms the other approaches on two style transfer datasets. Specifically, to the content preservation, Style Transformer achieves the best performance with a significant improvement.
|
Recently, many text style transfer approaches have been proposed. Among these approaches, there is a line of works aims to infer a latent representation for the input sentence, and manipulate the style of the generated sentence based on this learned latent representation. Although, learning a latent representation will make the model more interpretable and easy to manipulate, the model which is assumed a fixed size latent representation cannot utilize the information from the source sentence anymore. On the other hand, there are also some approaches without manipulating latent representation are proposed recently. In this paper, we follow the second line of work and propose a novel method which makes no assumption about the latent representation of source sentence and takes the proven self-attention network, Transformer, as a basic module to train a style transfer system. To make our discussion more clearly, in this section, we will first give a brief introduction to the style transfer task, and then start to discuss our proposed model based on our problem definition. In this paper, we define the style transfer problem as follows: Considering a bunch of datasets {D i } K i=1 , and each dataset D i is composed of many natural language sentences. For all of the sentences in a single dataset D i , they share some specific characteristic (e.g. they are all the positive reviews for a specific product), and we refer this shared characteristic as the style of these sentences. In other words, a style is defined by the distribution of a dataset. Suppose we have K different datasets D i , then we can define K different styles, and we denote each style by the symbol s (i) . The goal of style transfer is that: given a arbitrary natural language sentence x and a desired style s ∈ {s (i) } K i=1 , rewrite this sentence to a new one x which has the style s and preserve the information in original sentence x as much as possible. To tackle the style transfer problem we defined above, our goal is to learn a mapping function f θ (x, s) where x is a natural language sentence and s is a style control variable. The output of this function is the transferred sentence x for the input sentence x. A big challenge in the text style transfer is that we have no access to the parallel corpora. Thus we can't directly obtain supervision to train our transfer model. In section 3.4, we employ two discriminator-based approaches to create supervision from non-parallel corpora. Finally, we will combine the Style Transformer network and discriminator network via an overall learning algorithm in section 3.5 to train our style transfer system. Generally, Transformer follows the standard encoder-decoder architecture. Explicitly, for a input sentence x = (x 1 , x 2 , ..., x n ), the Transformer encoder Enc(x; θ E ) maps inputs to a sequence of continuous representations z = (z 1 , z 2 , ..., z n ). And the Transformer decoder Dec(z; θ D ) estimates the conditional probability for the output sentence y = (y 1 , y 2 , ..., y n ) by auto-regressively factorized its as: p θ (y t |z, y 1 , ..., y t-1 ). (1) At each time step t, the probability of the next token is computed by a softmax classifier: where o t is logit vector outputted by decoder network. To enable style control in the standard Transformer framework, we add a extra style embedding as input to the Transformer encoder Enc(x, s; θ E ). Therefore the network can compute the probability of the output condition both on the input sentence x and the style control variable s. Formally, this can be expressed as: and we denote the predicted output sentence of this network by f θ (x, s). Suppose we use x and s to denote the sentence and its style from the dataset D. Because of the absence of the parallel corpora, we can't directly obtain the supervision for the case f θ (x, s) where s = s. Therefore, we introduce a discriminator network to learn this supervision from the nonparallel copora. The intuition behind the training of discriminator is based on the assumption below: As we mentioned above, we only have the supervision for the case f θ (x, s). In this case, because of the input sentence x and chosen style s are both come from the same dataset D, one of the optimum solutions, in this case, is to reproduce the input sentence. Thus, we can train our network to reconstruct the input in this case. In the case of f θ (x, s) where s = s, we construct supervision from two ways. 1) For the content preservation, we train the network to reconstruct original input sentence x when we feed transferred sentence y = f θ (x, s) to the Style Transformer network with the original style label s. 2) For the style controlling, we train a discriminator network to assist the Style Transformer network to better control the style of the generated sentence. In short, the discriminator network is another Transformer encoder, which learns to distinguish the style of different sentences. And the Style Transformer network receives style supervision from this discriminator. To achieve this goal, we experiment with two different discriminator architectures. Conditional Discriminator In a setting similar to Conditional GANs Different from the previous one, in this case, only one sentence is feed into discriminator d φ (x), and the discriminator aims to answer the style of this sentence. More concretely, the discriminator is a classifier with K + 1 classes. The first K classes represent K different styles, and the last class is stand for the generated data from f θ (x, s) , which is also often referred as fake sample. In discriminator training stage, we label the real sentences x and reconstructed sentences y = f θ (x, s) to the label of the corresponding style. And for the transferred sentence y = f θ (x, s) where s = s, is labeled as the class 0. In Style Transformer network learning stage, we train the network f θ (x, s) to maximize the probability of the class which is stand for style s. In this section, we will discuss how to train these two networks. And the training algorithm of our model can be divided into two parts: the discriminator learning and Style Transformer network learning. The brief illustration is shown in Figure Loosely speaking, in the discriminator training stage, we train our discriminator to distinguish between the real sentence x and reconstructed sentence y = f θ (x, s) from the transferred sentence y = f θ (x, s). The loss function for the discriminator is simply the cross-entropy loss of the classification problem. For the conditional discriminator: And for the multi-class discriminator: According to the difference of discriminator architecture, there is a different protocol for how to label these sentences, and the details can be found in Algorithm 1. The training of Style Transformer is developed according to the different cases of f θ (x, s) where s = s or s = s. Self Reconstruction For the case s = s , or equivalently, the case f θ (x, s). As we discussed before, the input sentence x and the input style s comes from the same dataset , we can simply train our Style Transformer to reconstruct the input sentence by minimizing negative log-likelihood: For the case s = s, we can't obtain direct supervision from our training set. So, we introduce two different training loss to create supervision indirectly. Cycle Reconstruction To encourage generated sentence preserving the information in the input sentence x, we feed the generated sentence y = f θ (x, s) to the Style Transformer with the style of x and training our network to reconstruct original input sentence by minimizing negative loglikelihood: Style Controlling If we only train our Style Transformer to reconstruct the input sentence x from transferred sentence y = f θ (x, s), the network can only learn to copy the input to the output. To handle this degeneration problem, we further add a style controlling loss for the generated sentence. Namely, the network generated sentence y is feed into discriminator to maximize the probability of style s. For the conditional discriminator, the Style Transformer aims to minimize the negative loglikelihood of class 1 when feed to the discriminator with the style label s: And in the case of the multi-class discriminator, the Style Transformer is trained to minimize the the negative log-likelihood of the corresponding class of style s: Combining the loss function we discussed above, the training procedure of the Style Transformer is summarized in Algorithm 2. Compute L style (θ) for y by Eq. ( Finally, we can construct our final training algorithm based on discriminator learning and Style Transformer learning steps. Similar to the training process of GANs Before finishing this section, we finally discuss a problem which we will be faced with in the training process. Because of the discrete nature of the natural language, for the generated sentence y = f θ (x, s), we can't directly propagate gradients from the discriminator through the discrete samples. To handle this problem, one can use REIN-FORCE We evaluated and compared our approach with several state-of-the-art systems on two review datasets, Yelp Review Dataset (Yelp) and IMDb Movie Review Dataset (IMDb). The statistics of the two datasets are shown in Table A goal transferred sentence should be a fluent, content-complete one with target style. To evaluate the performance of the different model, following previous works, we compared three different dimensions of generated samples: 1) Style control, 2) Content preservation and 3) Fluency. We measure style control automatically by evaluating the target sentiment accuracy of transferred sentences. For an accurate evaluation of style control, we trained two sentiment classifiers on the training set of Yelp and IMDb using fastText Due to the lack of parallel data in style transfer area, automatic metrics are insufficient to evaluate the quality of the transferred sentence. Therefore we also conduct human evaluation experiments on two datasets. We randomly select 100 source sentences (50 for each sentiment) from each test set for human evaluation. For each review, one source input and three anonymous transferred samples are shown to a reviewer. And the reviewer is asked to choose the best sentence for style control, content preservation, and fluency respectively. • Which sentence has the most opposite sentiment toward the source sentence? • Which sentence retains most content from the source sentence? • Which sentence is the most fluent one? To avoid interference from similar or same generated sentences, "no preference." is also an option answer to these questions. In all of the experiment, for the encoder, decoder, and discriminator, we all use 4-layer Transformer with four attention heads in each layer. The hidden size, embedding size, and positional encoding size in Transformer are all 256 dimensions. Another embedding matrix with 256 hidden units is used to represent different style, which is feed into encoder as an extra token of the input sentence. And the positional encoding isn't used for the style token. For the discriminator, similar to In the experiment, we also found that preforming random word dropout for the input sentence when computing the self reconstruction loss (Eq. ( Results using automatic metrics are presented in Table For the human evaluation, we choose two of the most well-performed models according to the automatic evaluation results as competitors: DeleteAndRetrieve (DAR) And the generated outputs from multi-class discriminator model is used as our final model. We have performed over 400 human evaluation reviews. Results are presented in Table To study the impact of different components on overall performance, we further did an ablation study for our model on Yelp dataset, and results are reported in Table For better understanding the role of different loss functions, we disable each loss function by turns and retrain our model with the same setting for the rest of hyperparameters. After we disable self-reconstruction loss (Eq. ( 3) the discriminator provides style supervision to help the model control the style of generated sentences. Another group of study is focused on the different type of samples used in the discriminator training step. In Algorithm 1, we used a mixture of real sentence x and generated sentence y as the positive training samples for the discriminator. By contrast, in the ablation study, we trained our model with only one of them. As the result shows, the generated sentence is the key component in discriminator training. When we remove the real sentence from the training data of discriminator, our model can also achieve a competitive result as the full model with only a small performance drop. However, if we only use the real sentence the model will lose a significant part of the ability to control the style of the generated sentence, and thus yields a bad performance in style accuracy. However, the model can still perform a style control far better than the input copy model discussed in the previous part. For the reasons above, we used a mixture of real sample and generated sample in our final version. In this paper, we proposed the Style Transformer with a novel training algorithm for text style transfer task. Experimental results on two text style transfer datasets have shown that our model achieved a competitive or better performance compared to previous state-of-the-art approaches. Especially, because our proposed approach doesn't assume a disentangled latent representation for manipulating the sentence style, our model can get better content preservation on both of two datasets. In the future, we are planning to adapt our Style Transformer to the multiple-attribute setting like
| 793 | 2,686 | 793 |
RobustQA: A Framework for Adversarial Text Generation Analysis on Question Answering Systems
|
Question answering (QA) systems have reached human-level accuracy; however, these systems are not robust enough and are vulnerable to adversarial examples. Recently, adversarial attacks have been widely investigated in text classification. However, there have been few research efforts on this topic in QA. In this article, we have modified the attack algorithms widely used in text classification to fit those algorithms for QA systems. We have evaluated the impact of various attack methods on QA systems at character, word, and sentence levels. Furthermore, we have developed a new framework, named RobustQA, as the first opensource toolkit for investigating textual adversarial attacks in QA systems. RobustQA consists of seven modules: Tokenizer, Victim Model, Goals, Metrics, Attacker, Attack Selector, and Evaluator. It currently supports six different attack algorithms. Furthermore, the framework simplifies the development of new attack algorithms in QA.
|
With the release of large and high-quality datasets in the field of question answering (QA) In the context of QA systems, the model's accuracy drops drastically when some adversarial sentences are added to the input paragraphs Adversarial attacks have been widely investigated in the field of text classification The contributions of the paper can be summarized as follows: 1) We modify the attack algorithms that have been widely used in the field of text classification for QA systems. 2) We show that these modified attack algorithms can easily be evaluated on QA systems in three different characters, words, and sentence levels. 3) We build a new open-source framework named RobustQA, aiming at simplifying the research on textual adversarial attacks in QA systems. 4) We have incorporated both adversarial text generation and data augmentation in RobustQA for being used in adversarial training methods to improve the robustness and generalization of QA models. In this paper, we introduce the related works in Section 2. We compare the QA task against text classification and describe a sample textual adversarial attack algorithm implemented for the QA task in Section 3. Next, we introduce the Ro-bustQA framework modules in detail in Section 4 and demonstrate the framework's usage in Section 5. We present our setup and experimental results in Section 6. Finally, our conclusions and future works are presented in Section 7.
|
Adversarial attacks have been extensively studied on continuous data Adversarial attacks can be categorized based on different aspects. Attacks are primarily divided into two types of white and black boxes. In white box attacks, the attacker has full access to the model and its parameters. In this type of attack, the gradient of the cost function relative to the input is used to generate an adversarial example Adversarial attacks are also divided into untargeted and targeted categories. In untargeted attacks, the goal is merely to cause the model to produce an incorrect output label Textual adversarial attacks are divided into three categories in terms of perturbation levels, i.e., character, word, and sentence. Character-level attacks usually manipulate characters based on insertion, deletion, swap, substitution, and repetition operations There have been only a few research initiatives focused on textual adversarial attacks in the field of QA. It has been demonstrated that paraphrasing the questions is an alternative method for generating adversarial sentences. In this regard, Ribeiro et al. (2018) used the back translation technique to obtain question paraphrase rules. Also, Gan and Ng (2019) used a transformer model to produce paraphrased questions and introduced two types of adversarial questions. The autoencoder structure was utilized in another recent research to generate adversarial sentences There are several open-source libraries for building adversarial examples on continuous data. The most notable ones are CleverHans 3 Question Answering vs. Text Classification In QA systems, the question and context are represented as a sequence of tokens, Q = {q 1 , q 2 , q 3 , ..., q n } and C = {c 1 , c 2 , c 3 , ..., c n }, respectively. In these systems, the main goal is to predict the answer, A, in the form of a span within the context, A = {c j , ..., c j+k }. The returned span includes a specific start and an end token indices of the context paragraphs. F1 score and exact match (EM) criteria are the two common metrics for evaluating QA systems. On the other hand, in the text classification task, the main goal is to recognize the correct class of an input text. Due to the substantial differences between QA and text classification tasks, the algorithms designed for dealing with the attacks in the text classification are not directly applicable to the QA problems. The main distinctions are related to their differences in the structure of the input data and the goal function of attack scenarios. In text classification of an input text X with the corresponding ground truth label Y and the victim model F, the goal of an attack scenario is to have an attack set up that transforms X to X with the minimum perturbation in such a way that the victim model predicts an incorrect label Ỹ , where Y ̸ = Ỹ . In QA tasks, every input X is composed of a question Q and a context C. The ground truth label Y, which is a part of the given context with specific start and end tokens, represents the correct answer to the given question. The predicted answer Ỹ is computed by considering the maximum probability for the start and end tokens. Akin to the text classification, the goal of an attack scenario here is to have an attack set up that transforms X to X with the minimum perturbation ∆x in a way that the victim model predicts an incorrect answer span Ỹ, where In both text classification and QA tasks, the goal function of an attack scenario determines the success of the attack on a given victim model. In text classification, an attack scenario for a given input example is regarded as successful if the model prediction for the example is not equal to its corresponding ground truth label. In this task, the goal function can be simply evaluated by a single criterion. However, in QA tasks, since a prediction label includes two items (i.e., the start and the end tokens of the predicted answer span), the goal function is usually evaluated by both F1 and EM criteria. To demonstrate the required modifications of an attack method to cope with the mentioned differences, we discuss the details of the changes applied to the TextFooler algorithm The TextFooler algorithm is a score-based textual adversarial attack that consists of two primary steps. The first step is the word importance ranking, in which words are sorted according to their importance. The second step is the word transformation, which produces suitable substitutes for the words with the highest importance level obtained from the first step to generate an adversary example. Algorithm 1 shows the pseudo-code of a revised version of the TextFooler algorithm, which is compatible with QA tasks. Word Importance Ranking (line 1-11) The input example X, which includes context C and question Q, accompanied by its corresponding ground truth label Y, is passed to the algorithm. The goal is to confuse the victim model by generating a new question Q, with the minimum perturbation to Q. One metric among the F1 score and EM measure is used for marking an adversarial example. In Algorithm 1, we use the F1 score and δ, a threshold value empirically set to 0.9, as the goal function criterion. Compute the importance score of the start and end answer span, Create a set W of all words w i ∈ Q sorted by the descending order of their importance score, either using start I s w i or average (I s w i + I e w i )/2 importance score. Filter out the stop words in W. for each word w j in W do Initiate the set of candidates CANDIDATES by extracting the top N synonyms using CosSim(Emb w j , Emb word ) for each word in Vocab. if there exists c k where α > δ then In FINCANDIDATES, only keep the candidates c k where α > δ First, a copy of X is taken as a potential adversarial example X. Then, question Q is systematically altered for n number of times by in turn deleting the token w i . Each altered question is then passed to the victim model to predict the answer span based on the highest probability values of the start and end tokens. Next, the model predictions for the initial question (i.e., Y) and that of each altered question (i.e., Ȳ) are compared. Accordingly, the importance score of the start and end tokens of the answer span I w i = (I s w i , I e w i ) for each altered question is computed by either line 5 (i.e., in the case of equality) or line 7 (i.e., otherwise). P Y (X) and P Ȳ(X) respectively represent the probability values of the start and end tokens of the answer span provided by the ground truth label Y and that of the label Ȳ predicted by the attacked model for the original question X. Similarly, P Y (X \w i ) and P Ȳ(X \w i ) respectively represent the probability values of the start and end tokens of the answers predicted by the original and attacked model for the perturbed question, in which w i has been omitted from the original question. In line 10, a set of W of all words w i ∈ Q is created and sorted by the descending order of their importance score (i.e., using I s w i or (I s w i + I e w i )/2). In our experiments, we have chosen I s w i to compute the importance score. Word Transformation (line 12-34) In lines 12-14, using the Cosine similarity metric, a set of candidates CANDIDATES is created by extracting the top N synonyms of word w j with the same part of speech as that of w j . In lines 15 to 23, each word w j is in turn substituted by a candidate (i.e., c k ) to create an altered example (i.e., X ′ ). Among all the candidates, those that cause the similarity between the potential adversarial question (i.e., Q) and the altered question (i.e., Q ′ ) to exceed a predefined threshold (that we empirically set it to 0.7), are considered as final candidates. Each final candidate along with its predicted label (i.e., Y k ) and the probability values of its start and end tokens (i.e., P k ) are stored. In lines 24-33, at first, the eligibility of each final candidate as an adversarial example is determined by computing an α value for the candidate and comparing the value against a predefined threshold δ. If a candidate modifies the initial question Q in a way that results in an altered question Q ′ having the maximum semantic similarity with Q, then Q ′ will be chosen as an adversarial question. However, if a candidate does not satisfy this condition, one of the final candidates with the least confidence score is instead selected. In RobustQA, we modified TextFooler We have developed a new attack framework named RobustQA for applying text adversarial attack algorithms to QA systems. This framework is an extension of OpenAttack Tokenizer. The tokenizer module of RobustQA supports multiple tokenization approaches, including word-, sub-word-, and character-level tokenization. It maintains the consistency between the tokenization of the original sample and that of the adversarial one, enabling the effective evaluation of the attack algorithms. Furthermore, it currently supports the Stanford question answering dataset (SQuAD) dataset Victim Model. The victim model module supports the QA-based models. An extended version of this module is implemented to integrate Hug-gingFace Transformer-based models * . This module contains multiple methods required for executing different adversarial attack scenarios in Ro-bustQA. These methods can be overridden or extended for any desired customized attack, as they have access to all the sub-layers of the model's output and perform their operation as middleware. Goals. The primary target of the goal module is to determine if an input sample is eligible as an adversarial candidate. The candidate sample is regarded as an eligible one if it can confuse the victim model and diminish its performance in terms of EM or F1 score metrics. Defining a custom goal for new QA attacks is achievable by extending the goal module. Metrics. The evaluation metrics of the attack scenarios can be selected or extended with this module. As discussed in Appendix B, the evaluation metrics specific to the QA task (i.e., EM and F1 score) are enabled by default. Other metrics such as edit distance, fluency, grammatical errors, modification rate, and semantic similarity are available for selection. Attacker. The attacker is an abstract module with a default implementation of all the required tools and logic to define an attack algorithm on a given QA victim model. Based on the F1 score metric and a predefined threshold value, an attack goal specific to the given QA task is defined and used as a criterion to determine the adversary potential of different input examples. The primary method of creating a custom QA attack algorithm is to extend the QA attacker module. Various types of adversarial attack algorithms are derived from this module in RobustQA, ready for experimentation. Attack Selector. The attack selector module facilitates the initiation of an attack scenario. This module enables effortless selection and instantiation of the victim model, tokenizer, dataset, attacker, and evaluation metrics. It also performs data sampling and preparation. An attack scenario is easily configured by passing the preferred settings to the attack selector module. Further comprehensive analysis of the attack algorithms is possible by providing additional customized metrics to this module. Evaluator. The execution and evaluation of the QA attack algorithms take place in the evaluator module. Attacks performance is evaluated from different aspects: (1) the attack success rate indicates the percentage of the attacks that fool the victim model and produce false predictions; (2) the modification rate is the percentage of the modified tokens in an adversarial example compared to the input example; (3) the fluency of adversarial examples are computed by perplexity by GPT-2 The RobustQA interface empowers users to execute attack scenarios either programmatically, utilizing the Python programming language, or via a command-line prompt. Appendix D demonstrates an example of the toolkit usage through commandline interface and code. Moreover, some adversarial examples generated by different attack algorithms are depicted in Appendix E. Utilizing RobustQA, we have evaluated the performance of six different adversarial attack algorithms on the large uncased Bidirectional Encoder Representations from Transformer (BERT) model In this article, we showed the effect of various textual adversarial attack algorithms in character, word, and sentence levels on QA systems. We also developed an open-source framework, named RobustQA, for the field of textual adversarial attack on QA systems, which consists of seven primary modules. This new framework offers different features that are easily customizable for applying existing or designing new algorithms, along with efficient analysis of attack scenarios. As our future work, this framework can be further extended to include other attack algorithms. We can also provide more functions and tools for further research in the context of attacks and defense within QA systems. The source code and documentation of RobustQA are available at Although RobustQA is reliable for implementing and evaluating textual adversarial attacks on QA models, a limitation may arise in certain attack algorithms due to their high resource requirements. Specifically, in some cases, the execution of the attack algorithms requires a high level of GPU resources and CPU iterations. Like many other deep learning algorithms, adversarial text generation and adversarial training heavily rely on GPU resources. As the augmented training set grows, the mentioned procedures demand a substantial share of GPU power. This requirement imposed some constraint on the extent of our experiments. Due to the intricacies of the QA domain and the diverse nature of attacks in this domain, it was not feasible for us to seamlessly integrate all of them. Some algorithms could perfectly align with specific QA architectures, while others might require some customizations. Although the required tools for implementing any adversarial attack algorithm can be embedded within the RobustQA framework, the challenge of adapting all the attack algorithms hindered the variety of our experiments conducted in this study. RobustQA is the first open-source framework for textual adversarial attack analysis in QA systems. As shown in Figure For the evaluation purpose, we have employed EM and F1 score criteria, which are regarded as the standard metrics for evaluating QA systems MODEL="bert-large-uncased-whole-word-masking\ -finetuned-squad" python qa.py \ --use_cuda \ --victim_model_or_path "$MODEL" \ --victim_tokenizer_or_path "$MODEL" \ --dataset "squad" \ --dataset_split "validation[0:1000]" \ --attack_recipe "textfooler" \ --batch_size 8 \ --language "english" \ --use_metric_f1_score \ --use_metric_exact_match \ --use_metric_edit_distance \ --use_metric_fluency \ --use_metric_grammatical_errors \ --use_metric_modification_rate \ --use_metric_semantic_similarity \ --use_metric_jaccard_char_similarity \ --use_metric_jaccard_word_similarity The SQuAD v1.1, introduced in 2016 by In this section, the generated adversarial examples of three attack algorithms are presented. The original and adversary questions are depicted in Table E.1. Other fields of the generated adversarial examples, such as "context" and "answers", are the same as the original instance from the SQuAD dataset. In this section, we present the evaluation results of six different adversarial attack algorithms implemented with the RobustQA framework. The computational experiments in this study were conducted on a system with an Intel Core i7-8700K CPU 3.70GHz 6-Core, a GeForce GTX 1080 8GB vRAM, and 64GB of RAM.
| 964 | 1,435 | 964 |
HuCurl: Human-induced Curriculum Discovery
|
We introduce the problem of curriculum discovery and describe a curriculum learning framework capable of discovering effective curricula in a curriculum space based on prior knowledge about sample difficulty. Using annotation entropy and loss as measures of difficulty, we show that (i): the top-performing discovered curricula for a given model and dataset are often non-monotonic as apposed to monotonic curricula in existing literature, (ii): the prevailing easy-to-hard or hard-to-easy transition curricula are often at the risk of underperforming, and (iii): the curricula discovered for smaller datasets and models perform well on larger datasets and models respectively. The proposed framework encompasses some of the existing curriculum learning approaches and can discover curricula that outperform them across several NLP tasks.
|
Annotation information has been extensively used by previous research in NLP to devise strategies for further data collection A curriculum is a planned sequence of learning materials and an effective one can improve training of NLP systems Existing CL approaches are designed to learn a single curriculum that works best for a given model and dataset. However, effective training could be achieved in multiple ways. In addition, existing approaches quantify sample difficulty through model behavior during training. Although efficient and effective, model behavior can be affected by initialization and training dynamics This paper advocates a re-imagining of CL paradigms by introducing and formalizing the task of curriculum discovery, which aims to find effective curricula for a given model and dataset over a curriculum space. The present work specifically focuses on determining when and in which difficulty order text data samples should be learned for effective training of NLP systems. We propose a framework that employs prior knowledge about sample difficulty, such as entropy in human annotations, to inform an effective and flexible sample weighting scheme for curriculum discovery. The framework is capable of discovering optimal curricula (within the space of its weight functions) for any given model and dataset by optimizing the weight functions and adjusting the difficulty group of data samples as training progresses. The discovered curricula provide useful insights about datasets and models, such as the relative importance of different groups of samples for models or knowledge dependency among samples. We illustrate that the proposed framework has the potential to encompass some of the existing CL approaches. Experimental results show that (a): the topperforming discovered curricula for the same model and dataset can be fundamentally dissimilar in their training strategies, indicating that effective training can be achieved in multiple ways; (b): the discovered curricula are often non-monotonic and greatly differ from the known strategies reported in existing literature, indicating that existing curricula, including easy-to-hard transition curricula, are at the risk of underperforming; and (c): the curricula discovered on small datasets and models perform exceptionally well on larger datasets and models respectively, illustrating the transferability of the discovered curricula. The paper presents a new curriculum learning approach that unlike existing approaches can discover multiple high-performing (and often diverse) curricula for each given NLP model and dataset, provide interpretable curricula in terms of sample difficulty, and encompass some of the existing curriculum learning approaches.
|
Existing CL approaches are designed to learn a single curriculum that works best for a given model and dataset. They estimate sample difficulty through model behavior during training, quantified by the instantaneous loss for training We consider the training dataset D = {(x 1 , y 1 ), . . . , (x n , y n )} of size n, where x i denotes the ith training sample with the groundtruth label y i and ψ ∈ [0, 1] n indicates the initial difficulty estimates of training samples, see §3.4. The data is initially clustered into k groups of increasing difficulty, e.g. {easy, medium, hard} groups for k = 3, which can be achieved using difficulty score percentiles or 1-dimensional K-means applied to ψ. As Figure where l i is the instantaneous loss of sample i, and w(t; r c , s c ) is the weight of sample i in its difficulty group c at training iteration t, with class-specific weight function parameters r c and s c (see below). We define a curriculum using the generalized logistic function where r ∈ R is the rate-of-change parameter, which specifies how fast the weight can increase (r > 0) or decrease (r < 0); t ∈ [0, 1] is the training progress (typically iteration number divided by max iterations); and s ∈ R shifts the pivot weight of the logistic function (w(.) = .5) to the left or right such that at t = s the weight is 0.5. Figure Although the generalized logistic function in (2) can lead to effective curricula, monotonic functions are limited in their coverage capacity. For example, they do not allow easy samples with low weights to become important again (receive high weights) at later stages of training to mitigate forgetting, which is a major challenge for effective curriculum learning We find the optimal curriculum parameters (r, s) for each difficulty group using the Tree-structured Parzen Estimator (TPE) algorithm The discovered curricula are optimal within our search space, as defined by the weight functions and searchable parameters. However, in practice, we observed that the change in performance across the missing regions in the search space is minor. Given that our weight functions can approximate other curricula learned by existing CL models, see §4.7, we expect the optimum curriculum within our search space closely approximates the optimal curriculum for each dataset and model pair. Annotation entropy is a natural measure of difficulty (for humans) and may serve as a reliable difficulty metric for models. Entropy of each sample x i is calculated asl p c log p c Furthermore, moving average of a sample's instantaneous loss is a good metric for difficulty Figure We acknowledge that while multiple annotations per sample may not be readily available for many NLP datasets, such annotations were collected for most NLP datasets at their dataset development time. Our work shows that such information can be used to find effective curricula for NLP models and encourages dataset creators to publish their full annotation information. In addition, our curriculum discovery framework is independent of annotation information. In fact, we evaluated our approach with both annotation entropy and loss as two choices for sample-level difficulty estimation. For the purpose of our experiments, we chose datasets for which several annotations per sample are available. Such annotator-level information is often available at the creation time of most NLP datasets and provide rich information for effective learning. Before training, we partition each dataset into k difficulty groups using { i k } i=k i=0 quantiles. SNLI ChaosNLI Twitter Reddit. We developed this dataset to obtain population-level statistics of cancer patients. It contains 3.8k Reddit posts annotated by at least three annotators for relevance to specific cancer types. We define a multi-class classification task based on post relevance and cancer type, see Appendix A. We randomly split the data into 2.2k, 765, and 765 training, development and test sets respectively. ChaosNLI is balanced in its difficulty groups. We create difficulty-balanced versions of SNLI, Twitter and Reddit by collecting an equal number of samples from each difficulty group. The resulting datasets contain 1.7K to 2.3K samples. No-CL The conventional training approach, which involves utilizing all samples for training in each iteration. Self-paced Learning (SPL) where l i is the loss of instance i parameterized by θ, v i is a trainable weight parameter assigned to each instance, and f is a regularization function for the weights. The model finds v that minimizes its loss under the constraint of f . The binary scheme SPL is defined by the regularization function f (v; λ) = -λ∥v∥ 1 ; if l i < λ, v i = 1, otherwise v i = 0, i.e., only easy samples are selected at each step. Mentornet Difficulty Prediction (DP) where ŷi is the ground truth label and f measures the Spearman's rank correlation coefficient between labels produced by experts and non-experts. The model re-weights samples for performance improvement using a pre-defined threshold τ ,: SuperLoss (SL) where τ is the moving average of loss (as the measure of difficulty) and σ is sample confidence. The model emphasizes easy samples (those with small losses) throughout the training. Our approach employs two difficulty scoring functions and two curriculum types for each dataset. The difficulty scoring functions are Loss and Ent (entropy) described in §3.4. The first curriculum type (inc) is the off-the-shelf gradually increasing approach in Figure We use bayesian optimization to tune the parameters λ of SL and α and τ of DP on development data. The optimal values found are λ = 1.2, α = 0.9 and τ is set dynamically upon loading the dataset to the 50 percentile difficulty value of the training data. We use twitter-roberta-base for Twitter and roberta-base for other datasets, both from In addition, we set the search space for the rate (r) and shift (s) parameters to [-10, 10] with a step of 2 and [-0.5, 1.5] with a step of 0.25 respectively. The search is run for at least 100 trials using the method described in ( §3.3). Each trial is run with three seeds and the result is averaged. The search objective is to maximize accuracy over development data. The trial number in which the best parameters are found is reported in Appendix C. We only search for curricula with three difficulty groups to ease interpretability and improve readability, and to minimize the number of search parameters. However, in case of inc curriculum, the optimal number of difficulty groups for ChaosNLI, SNLI, Twitter, Reddit are 12, 3, 28, and 12 respectively; in all cases, we tune the number of groups on the development set and evaluate on the best performing one. Appendix B includes the results of tuning the number of groups. Table Figure Figure No-CL the computational expenses of using full datasets. Fourth, as noted previously, instances of the Reddit dataset consist of long paragraphs, causing high variance in models trained using the dataset. Consequently, the curricula obtained using the Reddit and loss as measure of difficulty are of lower quality and perform poorly. Appendix D reports the results of all configurations. Table Figure The framework presented in this paper is capable of representing curriculum learning approaches that prune noisy data, e.g. where l i and σ i are the instantaneous loss and confidence of sample i respectively. Therefore, the proposed framework can also represent CL approaches with a confidence scale larger than one. We introduce an effective curriculum learning framework that employs prior knowledge about sample difficulty in its training paradigm for curriculum discovery. The proposed framework initially partitions its input data into several groups of increasing difficulty, defines parameterized func- tions to weight sample losses in each difficulty group, moves samples across difficulty groups based on their learning progress, and enables tuning the parameters of the weight function to discover novel curricula. We demonstrate that this framework is capable of representing several categories of curriculum learning approaches. The task of curriculum discovery alleviates the limitations imposed by selecting a single curriculum strategy, and instead, focuses on finding and analyzing different curricula that work equally-well for a given model and dataset. In addition, the discovered curricula provide insight into how different portions of the dataset contribute toward learning at different stages of training a model, which, in turn, provide knowledge about the learning dynamics of different models. The task of curriculum discovery could be costly on large datasets, in particular, when the goal is to find optimal curricula for different models and datasets. To mitigate the computational (c) HNM The shaded area is the 95% CI over the datasets with five random seeds each. The curves are monotonic for most parts, and can be approximated by our framework. cost, we show that it is possible to rapidly discover a curriculum on a small subset of the dataset (or a smaller version of the model with significantly less number of parameters) and apply the resulting curriculum to the full dataset. There are several promising areas for future work. These include approaches for learning new difficulty indicators from data (e.g., linguistic difficulty including lexical, syntactic and semantic difficulty), prioritizing medium level instances and those with greatest progress during training, and developing challenge datasets that contain diverse data samples with different levels of difficulty. Finally, investigating diverse curricula that are suitable for general use and across datasets through curriculum discovery and generalization is a promising area for research. Figure
| 838 | 2,740 | 838 |
Length is a Curse and a Blessing for Document-level Semantics
|
In recent years, contrastive learning (CL) has been extensively utilized to recover sentence and document-level encoding capability from pre-trained language models. In this work, we question the length generalizability of CLbased models, i.e., their vulnerability towards length-induced semantic shift. We verify not only that length vulnerability is a significant yet overlooked research gap, but we can devise unsupervised CL methods solely depending on the semantic signal provided by document length. We first derive the theoretical foundations underlying length attacks, showing that elongating a document would intensify the high intra-document similarity that is already brought by CL. Moreover, we found that isotropy promised by CL is highly dependent on the length range of text exposed in training. Inspired by these findings, we introduce a simple yet universal document representation learning framework, LA(SER) 3 : length-agnostic self-reference for semantically robust sentence representation learning, achieving state-of-theart unsupervised performance on the standard information retrieval benchmark. Our code is publicly available.
|
In recent years, contrastive learning (CL) has become the go-to method to train representation encoder models In this work, we provide an extensive analysis of length generalizability of standard contrastive learning methods. Our findings show that, with default contrastive learning, models' capability to encode document-level semantics largely comes from their coverage of length range in the training. We first depict through derivation the theoretical underpinnings of the models' vulnerability towards length attacks. Through attacking the documents by the simple copy-and-concatenating elongation operation, we show that the vulnerability comes from the further intensified high intra-document similarity that is already pronounced after contrastive learning. This hinders a stable attention towards the semantic tokens in inference time. Further, we show that, the uniformity/isotropy promised by contrastive learning is heavily lengthdependent. That is, models' encoded embeddings are only isotropic on the length range seen in the training, but remain anisotropic otherwise, hindering the same strong expressiveness of the embeddings in the unseen length range. In the quest to bridge these unideal properties, we propose a simple yet universal framework, LA(SER) 3 : Length-Agnostic SElf-Reference for SEmantically Robust SEntence Representation learning. By providing the simple signal that "the elongated version of myself 1) should still mean myself, and thus 2) should not become more or less similar to my pairs", this framework could not only act as an unsupervised contrastive learning method itself by conducting self-referencing, but could also be combined with any contrastive learning-based text encoding training methods in a plug-and-play fashion, providing strong robustness to length attacks and enhanced encoding ability. We show that, our method not only improves contrastive text encoders' robustness to length attack without sacrificing their representational power, but also provides them with external semantic signals, leading to state-of-the-art unsupervised performance on the standard information retrieval benchmark.
|
Length preference of text encoders has been observed in the context of information retrieval In this section, we first formalize the problem of length attack, and then analyze the most important pattern (misaligned intra-document similarity) that gives rise to this vulnerability, and take an attention mechanism perspective to derive for the first time the reason why contrastive learning-based text encoders can be attacked. Problem Formulation: Simple Length Attack Given a sentence S with n tokens {x 1 , x 2 , ..., x n }, we artificially construct its elongated version by copying it m times, and concatenating it to itself. For instance, if m = 2, this would give us S = {x 1 , ..., x n , x 1 , ..., x n }. Loosely speaking, we expect the elongation to be a "semanticspreserved" operation, as repeating a sentence m times does not change the semantics of a sentence in most cases. For instance, in the context of information retrieval, repeating a document d by m time should not make it more similar to a query q. In fact, using pure statistical representation such as tf-idf where f (•) denotes the elongation operator, and m is a random integer. Therefore, no matter according to the semanticspreserved assumption discussed previously, or reference from statistics-based methods (3) We verify this phenomenon with Proof of Concept Experiment 1 (Figures To validate Observation 1, we fine-tune a vanilla MiniLM With the fine-tuned model, we first construct two extreme cases: one with a false positive pair ("what is NLP?" v.s., "what is computer vision?"), one with a positive pair ("what is natural language processing?" v.s., "what is computational linguistics?"). We compute cosine similarity between mean-pooled embeddings of the original pairs, and between the embeddings attained after conducting an elongation attack with m = 100 (Eq. 1). We found surprisingly that, while "what is NLP?" and "what is computer vision?" have 0.06 cosine similarity, their attacked versions achieve 0.42 cosine similarity -successfully attacked (cf. Figure On a larger scale, we then construct an inference set with all the document pairs from Semantic Textual Similarity benchmark (STS-b) Observation 2: Intra-document token interactions experience a pattern shift after elongation attacks. Taking an intra-document similarity perspective A significant increase on intra-document similarity of contrastive learning-based models is observed by The attention that any token x i in the sentence S gives to the dominant tokens can be expressed as: where q i is the query vector produced by x i , k T dominant is the transpose of the key vector produced by x dominant , and k T n is the transpose of the key vector produced by every token x n . We omit the V matrix in the attention formula for simplicity. After elongating the sentence m times with the copy-and-concat operation, the attention distribution across tokens shifts, taking into consideration that the default prefix [cls] token is not elongated. Therefore, in inference time, [cls] tokens share less attention than in the original sentence. To simplify the following derivations, we further impose the assumption that positional embeddings contribute little to representations, which loosely hold empirically in the context of contrastive learning With this in mind, after elongation, the same token in different positions would get the same attention, because they have the same token embedding without positional embeddings added. Therefore: Based on Eq. 6, we can see that attentions towards dominant tokens would increase after document elongation attack. However, we can also derive that the same applies to non-dominant tokens: In fact, every unique token except [cls] would experience an attention gain. Therefore, we have to prove that, the attention gain G d of dominant tokens (denoted as x d ) outweighs the attention gain G r of non-dominant (regular, denoted as x r ) tokens. To this end, we define: d k be l n , and e d k be a l c , we get: Similarly, we get: Also note that l d > l r : that's why they are called "dominating tokens" in the first place As a result, with elongation operation, every token is going to assign even more attention to the embeddings of the dominating tokens. And this effect propagates throughout layers, intensifying the high intra-document similarity ("entourage effect") found in With the derivations, we conduct PoC Experiment 2, aiming to demonstrate that intra-document similarity experiences a pattern shift after elongation attack, intensifying the "entourage effect", for contrastive fine-tuned models. Taking the same fine-tuned MiniLM checkpoint from PoC Experiment 1, we compute the intradocument similarity of all the model outputs on STS-b. For each document, we first compute its document embedding by mean-pooling, then compute the average cosine similarity between each token embedding and the document embedding. When using the vanilla MiniLM checkpoint, the intra-document similarity pattern is again reversed. This opposite pattern is well-aligned with the findings of Complementing the intensified intra-document similarity, we also display an isotropy misalignment before and after elongation attacks in embeddings produced by the fine-tuned checkpoint start becoming anisotropic (Figure On the other hand, elongation attacks seem to be able to help vanilla pre-trained models to escape from anisotropy, interestingly (Figure 3 Method: LA(SER) 3 After examining the two fundamental reasons underlying the built-in vulnerability brought by standard contrastive learning, the formulation of our method emerges as an intuitive outcome. Naturally, we explore the possibility of using only length as the semantic signal to conduct contrastive sentence representation learning, and propose LA(SER) 3 : Length-Agnostic Self-Reference for Semantically Robust Sentence Representation Learning. LA(SER) 3 builds upon the semanticspreserved assumption that "the elongated version of myself 1) should still mean myself, and thus 2) should not become more or less similar to my pairs". LA(SER) 3 leverages elongation augmentation during the unsupervised constrastive learning to improve 1) the robustness of in-document interaction pattern in inference time; 2) the isotropy of larger length range. We propose two versions of reference methods, for different format availability of sentences in target training sets. Self-reference In LA(SER) 3 self-ref setting, we take a sentence from the input as an anchor for each training input, and construct its positive pair by elongating the sentence to be m times longer. Intra-reference LA(SER) 3 intra-ref conducts intrareference within the document. The two components of a positive pair are constructed from different spans of the same document. Since we are only to validate effectiveness of LA(SER) 3 intra-ref , we implement this in the simple mutually-excluded span setting. In other words, the LA(SER) 3 intra-ref variant takes a sentence (either the first or a random sentence) from the text as an anchor, uses the rest of the text in the input as its positive pair, and elongates the anchor sentence m times as the augmented anchor. For both versions, we use the standard infoNCE loss Training datasets We conduct our experiments on two training dataset settings: 1) training wiki uses 1M sentences sampled from Wikipedia, in line with previous works on contrastive sentence representation learning The trained models are mainly evaluated on the BEIR benchmark The reasons why we do not follow the de facto practice, which mainly focuses on cherry-picking the best training setting that provides optimal performance on STS-b are as follows: Firstly, performances on STS-b do not display strong correlations with downstream tasks beyond semantic textual similarity tasks. However, for completeness, we also provide the results of STS-b in Appendix A. Baselines We compare our methods in two settings, corresponding to the two versions of LA(SER) 3 : 1) Self-Reference. Since we assume using the input itself as its positive pair in this setting, it is natural to compare LA(SER) 3 self-ref to the strong baseline SimCSE 2) Intra-Reference. The baseline method in this case is: taking a sentence (random or first) from a document as anchor, then use the remaining content of the document as its positive pair. Notably, this baseline is similar to the unsupervised pretraining part of COCO-DR Implementation Details We evaluate the effectiveness of our method with BERT Notably, previous works on contrastive sentence representation learning For the selection of the anchor sentence, we take the first sentence of each document in the main experiment (we will discuss taking a random sen-tence instead of the first sentence in the ablation analysis in §5.1). For LA(SER) 3 self-ref , we elongate the anchor sentence to serve as its positive pair; for LA(SER) 3 intra-ref , we take the rest of the document as its positive pair, but then elongate the anchor sentence as the augmented anchor. For the selection of the elongation hyperparameter m, we sample a random number for every input depending on its length and the max length of 256. For instance, if a sentence has 10 tokens excluding [cls], we sample a random integer from The main results are in Tables Our method has the exact same setting (training a vanilla BERT on the same training wiki ) with the rest of the baselines except InfoCSE, which further benefits from the training of an auxiliary network. Note that with a batch size of 64, our method already outperforms all the baselines to a large margin except InfoCSE. Since we train with a max sequence length of 256 (all baselines are either 32 or 64), we find that training with a larger batch size (128) further stabilizes our training, achieving state-of-the-art results. Moreover, we achieve stateof-the-art with only a BERT base . In general, we find that our performance gain is more pronounced when the length range of the dataset is large. On BERT-base experiments, large nDCG@10 performance gain is seen on NFCorpus (doc. avg. length 232.26, SimCSE: 0.1048 -> LA(SER) 3 : 0.1919), Scifact (doc. avg. length 213.63, SimCSE: 0.2492 -> LA(SER) 3 : 0.4317), Arguana (doc. avg. length 166.80, SimCSE: 0.2796 -> LA(SER) 3 : 0.4227). On the other hand, our performance gain is limited when documents are shorter, such DBPedia (avg. length 49.68) and Table We also found that, even though MiniLM-L6 shows great representational power if after supervised contrastive learning with high-quality doc-ument pairs (see popular Sentence Transformers checkpoint all-MiniLM-L6-v2), its performance largely falls short under unsupervised training settings, which we speculate to be due to that the linguistic knowledge has been more unstable after every second layer of the model is taken (from 12 layers in MiniLM-L12 to 6 layers). Under such setting, LA(SER) 3 intra-ref largely outperforms LA(SER) 3 self-ref , by providing signals of more lexical differences in document pairs. In this section, we ablate two important configurations of LA(SER) 3 . Firstly, the usage of LA(SER) 3 involves deciding which sentence in the document to use as the anchor ( § 5.1). Secondly, how do we maximize the utility of self-referential elongation? Is it more important for the model to know "me * m = me", or is it more important to cover a wider length range ( § 5.2)? If a document consists of more than one sentence, LA(SER) 3 requires deciding which sentence in the document to use as the anchor. We ablate this with both LA(SER) The results are in Table We hypothesize that as LA(SER) 3 provides augmented semantic signals to contrastive learning, it would be hurt by overly noisy in-batch inputs. By contrast, LA(SER) 3 intra-sim behaves robustly to this noise because the rest of the document apart from the anchor could serve as a stabilizer to the noise. With the validated performance gain produced by the framework, we decompose the inner-workings by looking at the most important component, elongation. A natural question is: is the performance gain only brought by coverage of larger trained length range? Or does it mostly rely on the semantic signal that, "my-longer-self" still means myself? Table 6 Auxiliary Property Analysis Recalling in Observation 2 and PoC experiment 2, we focused on analyzing the effect of elongation attack on intra-sentence similarity, which is already high after CL We analyze the positional (in)sensitivity of 4 models (MiniLM We take the sentence pairs from STS-b test set as the inference set, and compute each model's perceived cosine similarity on the sentence pairs (distribution 1). We then randomly shuffle the word orders of all sentence 1s in the sentence pairs, and compute each model's perceived cosine similarity with sentence 2s again (distribution 2). The divergence of the two distributions for each model can serve as a proxy indicator of the model's sensitivity towards word order, and thus towards positional shift. The lower the divergence, the more insensitive that a model is about positions. We find that the Jenson Shannon divergence yielded by MiniLM has gone from 0.766 (vanilla) to 0.258 (after contrastive learning). And the same for mpnet goes from 0.819 (vanilla) to 0.302 (after contrastive learning). This finding shows that contrastive learning has largely removed the contribution of positions towards document embeddings, even in the most extreme case (with random shuffled word orders). This has made contrastivelylearned models acting more like bag-of-words models, aligning with what was previously found in vision-language models Moreover, MiniLM uses absolute positional embeddings while mpnet further applies relative positional embeddings. We believe that the positional insensitivity pattern holds for both models can partly make the pattern and LA(SER) 3 's utility more universal, especially when document encoders are trained with backbone models that have different positional encoding methods. In this work, we questioned the length generalizability of contrastive learning-based text encoders. We observed that, despite their seemingly strong representational power, this ability is strongly vulnerable to length-induced semantic shifts. we formalized length attack, demystified it, and defended against it with LA(SER) 3 . We found that, teaching the models "my longer-self = myself" provides a standalone semantic signal for more robust and powerful unsupervised representation learning.
| 1,151 | 2,153 | 1,151 |
Reasoning with Multimodal Sarcastic Tweets via Modeling Cross-Modality Contrast and Semantic Association
|
Sarcasm is a sophisticated linguistic phenomenon to express the opposite of what one really means. With the rapid growth of social media, multimodal sarcastic tweets are widely posted on various social platforms. In multimodal context, sarcasm is no longer a pure linguistic phenomenon, and due to the nature of social media short text, the opposite is more often manifested via cross-modality expressions. Thus traditional text-based methods are insufficient to detect multimodal sarcasm. To reason with multimodal sarcastic tweets, in this paper, we propose a novel method for modeling cross-modality contrast in the associated context. Our method models both cross-modality contrast and semantic association by constructing the Decomposition and Relation Network (namely D&R Net). The decomposition network represents the commonality and discrepancy between image and text, and the relation network models the semantic association in cross-modality context. Experimental results on a public dataset demonstrate the effectiveness of our model in multimodal sarcasm detection.
|
Sarcasm is a sophisticated linguistic phenomenon, defined by Merriam-Webster Dictionary as 'The use of words that mean the opposite of what you really want to say, especially in order to insult someone, to show irritation, or to be funny'. It can not only disguise the hostility of the speaker, but also enhance the effect of mockery or humor on the listener With the fast growing and diverse trend of social media, multimodal sarcastic tweets which convey abundant user sentiment are widely posted on various social platforms. There is a great demand for multimodal sarcasm detection to facilitate various applications. However, traditional text-based methods are not applicable to detect multimodal sarcastic tweets (Fig. Several related work on multimodal sarcasm detection has been proposed To tackle the above challenges, in this paper, we propose a novel method to model both crossmodality contrast and semantic association by constructing the Decomposition and Relation Network (i.e. D&R Net) for multimodal sarcasm detection task. The decomposition network implicitly models cross-modality contrast information via representing the commonality and discrepancy between image and text in tweets. The relation network explicitly captures the semantic association between image and text via a cross-modality attention mechanism. The main contributions of our work are as follows: • We identify the essential research issue in multimodal sarcasm detection, and propose a method to model cross-modality contrast in the associated context of multimodal sarcastic tweets. • We construct the Decomposition and Relation Network (D&R Net) to implicitly represent the contextual contrast and explicitly capture the semantic association between image and text, which provides the reasoning ability and word-level interpretability for multimodal sarcasm detection. • We compare our model with the existing stateof-the-art methods, and experimental results on a publicly available dataset demonstrate the effectiveness of our model in multimodal sarcasm detection. 2 Related Work
|
Traditional sarcasm detection takes text-based approaches, including feature engineering, context based and neural network models. Earlier feature engineering approaches are based on the insight that sarcasm usually occurs with specific signals, such as syntactic patterns (e.g. using highfrequency words and content words) Recently, researchers utilize the powerful techniques of neural networks to get more precise semantic representations of sarcastic text and model the sequential information of sarcastic context. Some approaches consider the contextual tweets of target tweet, using RNN model for contextual tweets representation and modeling the relationship between target and contextual tweets for sarcastic text classification With the prevalence of multimodal tweets, multimodal sarcasm detection has gained increasing research attention recently. In this paper, we propose a novel method to model the cross-modality contrast and semantic association in multimodal context by constructing the Decomposition and Relation Network (D&R Net), which enables our model to reason with multimodal sarcastic tweets and provides pertinent evidence for interpretation. Fig. Standard image, text and visual attributes (e.g. sunnet, scene, snow) are utilized in the previous multimodal sarcasm detection where, T ext = [W j ] T j , T is the length of text sequence; AN P s = [P i ] N i , N is the number of adjective-noun pair, in which each pair P i contains an adjective word A i , a noun word N i and the probability value p i of this kind of ANP existing in the attached Image, In encoding module, we map these triple inputs into hidden representations. All textual words W j , A i , N i are firstly mapped into embedding vectors w j , a i , n i ∈ R d . For each text, we utilize the bi-directional long short term memory (BiLSTM) network to represent textual sequence into a hidden representation vector and incorporate the contextual information. It maps word embedding w j into hidden state h w j ∈ R d . For each ANP, we directly compute the maxpooling result of its adjective and noun word embeddings as the hidden representation. For each image, we adopt a pre-trained convolutional neural network to extract image feature and also encode the result into d-dimensional space. We focus on contextual contrast of multimodal sarcastic tweets and design the decomposition network (D-Net) to represent the commonality and discrepancy of image and text in high-level spaces. The D-Net breaks down the raw visual or textual representation into a shared subspace and unique visual or textual subspace through three layers. The shared layer tends to extract invariant shared features f * shared of image and text, and image or text layer is forced to decompose image or text into unique variant contrast features f * unique , which can be defined as where f * is the feature of input modality * ∈ {image, text}, f image is the raw image encoding representation H m , f text is the last hidden state h w T of BiLSTM which is used as the overall representation of text, and W shared ∈ R ds×d , P * ∈ R du×d are projection matrices of shared space, unique visual space and textual space. In multimodal sarcastic tweets, we expect our model to focus more on the opposite between different modality information. Thus, we reinforce discrepancy between image and text, and on the contrary, weaken their commonality. Specifically, we combine the above unique variant contrast features as the cross-modality contrast representation. where ⊕ denotes the concatenation operation. We propose the relation network (R-Net) to fully capture the contextual association between image and text from multiple views. The relationship between image and text is usually multi-coupled, that is text may involve multiple entities in images, whereas different regions of the image may also involve different text words. We have already extracted multiple ANPs as the visual semantic information, which is beneficial to model multi-view associations between image and text according to different views of ANPs. Thus, we propose the ANP-aware cross-modality attention layer to align textual words and ANPs via utilizing each ANP to query each textual word and computing their pertinence. We first calculate the cross interactive attention matrix S ∈ R N ×T to measure how text words and image ANPs relate. where W ∈ R d×d is the parameter of bi-linear function, and each score s ij ∈ S indicates the semantic similarity between i-th ANP encoding h p i ∈ H p and j-th text word encoding h w j ∈ H w . We then compute the cross-modality attention weight α i j of i-th ANP for j-th textual word by normalizing the i-th row of attention matrix S, and calculate the weighted average of textual hidden states as the i-th ANP-aware textual representation r i ∈ R d : Thus, we query the text N times with different ANPs to get multi-view textual representations [r 1 , r 2 , . . . , r N ]. Our proposed ANP-aware cross-modality attention mechanism is a variant of multi-head attention We extract ANPs from each image and only select the Top N ANPs according to their extracted probability values [p 1 , p 2 , . . . , p N ]. Hence, different textual representations should be influenced by different ANP probability values. Thus, we get the final cross-modality association representation r rel ∈ R d by calculating weighted average of these ANP-aware textual representations [r 1 , r 2 , . . . , r N ] according to the related normalized ANP probability distributions. Finally, we feed the above acquired cross-modality contrast and semantic association representations, denoted as r dec and r rel respectively, into the top fully-connected layer and use the sigmod function for binary sarcasm classification. where w s ∈ R 1×(2du+d) , b s ∈ R 1 are the parameters of fully-connected layer. Our model optimizes two losses, including classification loss and orthogonal loss. We use cross entropy loss function as the sarcasm classification loss: where y i is the ground truth of i-th sample (i.e., 1 for sarcasm and 0 for non-sarcasm ), and ŷi is the predicted label of our model. In D-Net (Subsection 3.3), we share the same matrix for both image and text to ensure projecting them into the same subspace. Besides, in initialization and training process, to ensure that the decomposed unique subspaces are unrelated or in conflict with each other, we impose their projection matrices P * with the additional orthogonal constraint for the shared projection matrix W shared . We convert these orthogonal constraints into the following orthogonal loss: where • 2 F denotes the Frobenius norm. We finally minimize the combined loss function: where λ is the weight of orthogonal loss. We use a publicly available dataset constructed by For fair comparison, we adopt the same data preprocessing used in Our work focus on the multimodal sarcasm detection using image and text modalities. Thus, we compare our model with the only two existing related models using the same modalities. • MLP+CNN 1 ee.columbia.edu/ln/dvmm/vso/download/sentibank.html • Hierarchical FM We compare our model with multimodal baseline models with the F1-score and Accuracy metrics. Table To further explore the effects of multimodal inputs for sarcasm detection, we compare our model with the representative text-based sarcasm detection models and an image-based baseline model. • ResNet • CNN • BiLSTM We use F1-score and Accuracy as the evaluation metrics. Table To evaluate the performance of each component used in our D&R Net, we conduct the detailed ablation studies on various variants of our model. The ablation results are shown in Table In general, we find those variants underperform our model. The most obvious declines come from the direct removal of our two core modules, D-Net and R-Net (see row 1, 3). Comparing these two variants, we find that removing D-Net has greater performance drop than removing R-Net. This suggests that modeling the cross-modality contrast in D-Net is more useful than cross-modality association in R-Net. After removing the D-Net, the model only accepts the text and ANPs inputs. Thus we further incorporate image information via directly concatenating image encoding in the final fusion layer (see row 2). The improvement compared with -D-Net shows the effectiveness of using image modality for multimodal sarcasm detection. Similarly, we also add the representation of ANPs to the fusion layer after removing the R-Net module (see row 4). However, the performance unexpectedly continues to decrease. One possible reason for this is that the fusion of ANPs affects the original decomposition results in spite of using triple inputs. It is worth mentioning that replacing our ANPs with noun attributes used in In this section, we provide case studies through several practical examples to illustrate that our D&R Net really learns to reason multimodal sarcastic tweets with interpretability. Fig. For those text-only or image-only models, it's almost impossible to detect the sarcasm intention of Fig. Our proposed ANP-aware cross-modality attention mechanism explicitly calculates the cross interactive attention between text words and image ANPs, providing the explainable reasoning evidence for sarcasm detection. We further illustrate this attention mechanism by visualizing its outputs on two multimodal sarcastic tweets in Fig. In this paper, we identify the essential research issue in multimodal sarcasm detection. To model the cross-modality contrast in the associated context of multimodal sarcastic tweets, we propose the D&R Net to represent the commonality and discrepancy between image and text and multi-view semantic associations in cross-modality context. Our model is capable of reasoning multimodal sarcastic tweets with word-level interpretation. Experimental results on a public dataset show that our model achieves the state-of-the-art performance compared with the existing models.
| 1,077 | 2,072 | 1,077 |
Rethinking Style Transformer with Energy-based Interpretation: Adversarial Unsupervised Style Transfer using a Pretrained Model
|
Style control, content preservation, and fluency determine the quality of text style transfer models. To train on a nonparallel corpus, several existing approaches aim to deceive the style discriminator with an adversarial loss. However, adversarial training significantly degrades fluency compared to the other two metrics. In this work, we explain this phenomenon using energy-based interpretation, and leverage a pretrained language model to improve fluency. Specifically, we propose a novel approach which applies the pretrained language model to the text style transfer framework by restructuring the discriminator and the model itself, allowing the generator and the discriminator to also take advantage of the power of the pretrained model. We evaluated our model on three public benchmarks GYAFC, Amazon, and Yelp and achieved state-of-the-art performance on the overall metrics.
|
Text style transfer is the task of converting a sentence from one style to another while preserving style-agnostic semantics. In solving the text style transfer task, three criteria must be considered: 1) style control, how well a style is transferred from the original sentence to the generated one, 2) content preservation, how well the generated sentence has retained the semantics of the original, and 3) fluency, how natural the generated sentence is. Text style transfer is challenging, since fluently converting the style of a sentence often conflicts with content preservation transfer approaches have become popular, including those using an autoencoder In this paper, we review Style Transformer to investigate the reason behind the fluency degradation in adversarial models. To more precisely interpret what fluency is, we introduce the notion of energy
|
Mismatch Figure have nothing to do with formality (e.g., political expressions) would have high energy in the corpus. Accordingly, we define fluency as having low energy in a particular corpus, in which the fluent sentences express one of the styles in the corpus. As illustrated in Figure To apply the LM-based regularizers, we can leverage pretrained models such as GPT-2 to generate fluent sentences. Moreover, fluency is expected to further improve when the generator and the discriminator are also replaced with a pretrained model. However, as shown in Figure Our contributions can be summarized as follows: • We analyze the fluency degradation in adversarial training with an energy-based interpretation, and propose a regularizer leveraging a language model to prevent fluency degradation. • We reconstruct the discriminator and language model such that the single pretrained language model can be employed in the text style transfer framework. • We achieve new state-of-the-art results on GYAFC, Amazon, and Yelp datasets and carefully analyze the contribution of each component of our model. Many of the previous studies have attempted to learn disentangled representations of text by separating representations of content and style in a latent space. For instance, In contrast, there are methods without disentangled representations that do not explicitly disentan-gle the content and style of text using reinforcement learning Recently, pretrained models have achieved great success on various NLP tasks such as machine translation In addition, several works have focused on style transfer in a specific domain, or for leveraging an additional corpus. To transfer writing styles between authors, The conventional probabilistic model outputs the normalized probability p(x) for input variable x. In contrast, the energy-based model outputs the non-normalized scalar value E(x) denoted as energy Several works have leveraged the energy-based model for image generation In an unsupervised setting, we assume the nonparallel corpus X = {x (0) , x (1) , • • • , x (m) } and X ′ = {x ′(0) , x ′(1) , • • • , x ′(n) }, and denote each style of the corpus as s and s ′ . The objective is to train a style transfer model G in an unsupervised way such that a sentence x is turned into a sentence x ′ having similar content but the other style. Style Transformer (1) which is the cross entropy between the reconstructed sentence x and source sentence x. While transferring the sentence to the target style in x ′ ∼ G (x, s ′ ), the content of the sentence should be preserved. As in previous studies (2) which regularizes the generated sentence so that it is identical to the source sentence when retransferred to the original style. For style control, Style Transformer leverages an external model that discriminates the style. The discriminator D judges the consistency between the given sentence x and style s. The discriminator is trained separately from the generator and takes the generated sentences along with the original sentences. The training process for the discriminator optimizes where labeling {(x, s), ( x, s)} in positive as c = 1, {(x, s ′ ), ( x ′ , s ′ )} in negative as c = 0. Style Transformer attempts to deceive this discriminator into classifying the generated sample ( x ′ , s ′ ) as c = 1: (4) The upper part of Figure In the cycle and style loss, the gradients should be propagated into the generated sentences, but the nature of language discreteness prevents a trivial solution. To propagate the gradients directly, Style Transformer feeds the generated sentences to the discriminator in the form of a softmax distribution for each token. This soft representation of the sentences empirically reports better performance than REINFORCE BART Style Transformer follows the transformer encoder-decoder structure and initializes weights by training the dataset in an autoencoding manner. In contrast, we leverage BART In our preliminary study, there is a significant gap between the perplexity of the corpus in the target style and the generated sentences. Based on the energy-based interpretation The overall structure of our model. We concatenate the style label in front of the input of the BART encoder and decoder except the LM depicted in which is the exponential ratio of the negative energy E between the transferred sentence x ′ and style s or s ′ . This expression matches the real implementation as the discriminator takes the sentence x and style s as input and outputs of two logits. Each logit value means the negative energy of style s and s ′ , and the discriminator calculates the softmax output between them. To deceive the style discriminator, the generator needs to minimize E( x ′ , s ′ ) while maximizing E( x ′ , s). Meanwhile, the energy between the sentence and style can be interpreted as the perplexity or entropy of the sentence with the original style in E( x ′ , s) ≈ PPL s ( x ′ ). Maximizing the perplexity with the original style degrades the fluency of the generated sentences because both styles are from the corpus, sharing syntactic and semantic attributes. Figure Inspired by the work of in advance, and optimize the cross entropy of the generated sentence during training along with other losses as ). We report and analyze the fluency enhancement with this loss in Section 4.7. Finally, the total loss of our model is where each λ implies the coefficient for each loss. For fluent generation, it is desirable to apply a pretrained model There is, however, a rather simple solution to this problem of mismatching tokenizers: We used the same pretrained model for the generator, discriminator, and LM. In this way, we leverage the BART classifier proposed in the original BART paper. The BART classifier takes the same sequence x in the encoder and decoder, and predicts the class label at the <eos> token position at the decoder. For the LM, we adopt BART again to share the same vocab and tokenizer, and also take advantage of the BART decoder, which works as the language model in the text infilling task For the experiments, we used three widely-used English datasets The GYAFC dataset (Rao and Tetreault, 2018) was originally a question-and-answer dataset on an online forum, consisting of informal and formal sentences from the two categories: Entertainment & Music (EM) and Family & Relationships (FR). The Amazon dataset is a product review dataset, labeled as either a positive or negative sentiment. The Yelp dataset 1 is a restaurant and business review dataset with positive and negative sentiments. We chose the four unsupervised baselines, CrossAlign An ideal output is a sentence whose style is transferred to the target style while preserving the original content without losing fluency. Therefore, performance is measured using three criteria: 1) style transfer accuracy, 2) content preservation, and 3) fluency. Style transfer accuracy This metric indicates how many generated sentences are correctly transferred into the target style. Following We measure the similarity between sentences with a subword embedding model Fluency Following Overall metric Following the work of our model and baseline, and we select the model with the highest score in the dev dataset and plot the selected model along with the fluency measure. Table The perplexity of the source copy is not extremely high when compared to human references, as shown in Table Style Transformer typically shows convincing accuracy, but it reported high perplexity because the generated sentences deviated from the corpus distribution. On the other hand, especially for GYAFC EM & FR, our method reported state-ofthe-art performance on the overall metric of style transfer and content preservation, while reporting higher or similar fluency scores than the others. On the other hand, on the Amazon dataset, our model showed lower content preservation than the baseline. However, the human reference in the Amazon dataset only reports 54.6% style accuracy, and this implies the reference may not represent the target distribution, so the similarity score to the Amazon human reference is questionable. For the Yelp dataset, the similarity score was significantly lower than that of Masker, but it should be noted that this self-similarity is calculated by comparing the original sentence, so the high similarity implies the generator did not make any changes to the sentences. Nevertheless, our model exhibited the best J(A, S) score for the Amazon and Yelp datasets, which indicates the overall performance of style transfer and content preservation. Therefore, our model performed better on style transfer than other baselines, while maintaining fluency. To analyze the experimental results qualitatively, we also conduct human evaluation on GYAFC EM and Amazon datasets. As with automatic evaluation, we evaluate three criteria: fluency, style control, and content preservation. We evaluate the GYAFC EM dataset from informal to formal style, and the Amazon dataset from negative to positive style. For each dataset, a total of 20 source sentences are randomly selected. For each source sentence, four sentences are presented as answer options, one from our model and three from each baseline except CrossAlign. The results of the GYAFC EM and Amazon datasets are shown in Table Table The quality of materials and workmanship is noticeably less. The quality of materials and workmanship is noticeably better. The quality of materials and workmanship is noticeably superior. We conducted an ablation study to understand the contribution of each component in our proposed method. The results of the ablation study on the GYAFC FR dataset are shown in Table As our model uses adversarial learning, training is somewhat unstable, like GAN Although we have paid attention to model selection to compensate for the unstable training, the instability of adversarial learning remains an issue. In addition, we have only conducted experiments on widely used datasets, to compare our work with previous studies. These datasets are composed of binary style classes, such as positive and negative sentiments. Therefore, conducting experiments using multi-class datasets Our model may generate negative and rude expressions about a specific person or a commercial site because of the data distribution of the Yelp, Amazon, and GYAFC datasets. However, we propose our work in anticipation of positive applicability as shown in previous studies. Using energy-based interpretation, we found that fluency is inevitably degraded when deceiving the discriminator in Style Transformer Architecture details Our implementation is based on bart-base Training details For model selection, we record the model checkpoint per 500 steps, and the model with the highest J(A, S) in a single run is selected as our final model. Our model takes about 14 hours on a single NVIDIA RTX A6000 machine to train the GYAFC dataset. We use the AdamW optimizer (Loshchilov and Hutter, 2019) with a batch size of 64. The initial learning rate of the un-pretrained layers, such as the last linear layer of the discriminator, are set to 2e -4 , and all the others are set to 3e -5 . In addition, a cross entropy of the last linear layer in the discriminator is labelsmoothed Previous studies based on the Amazon and Yelp datasets Dataset For the Amazon dataset, we borrowed the preprocessed dataset provided by Baseline For a fair comparison between baselines in raw text, the tokenizer of each model was changed to a subword tokenizer The quality mixer cord and workmanship is noticeably less.
| 887 | 864 | 887 |
ReTAG: Reasoning Aware Table to Analytic Text Generation
|
The task of table summarization involves generating text that both succinctly and accurately represents the table or a specific set of highlighted cells within a table. While significant progress has been made in table to text generation techniques, models still mostly generate descriptive summaries, which reiterates the information contained within the table in sentences. Through analysis of popular table to text benchmarks (ToTTo (Parikh et al., 2020) and InfoTabs (Gupta et al., 2020)) we observe that in order to generate the ideal summary, multiple types of reasoning is needed coupled with access to knowledge beyond the scope of the table. To address this gap, we propose RETAG, a table and reasoning aware model that uses vector-quantization to infuse different types of analytical reasoning into the output. RETAG achieves 2.2%, 2.9% improvement on the PARENT metric in the relevant slice of ToTTo and InfoTabs for the table to text generation task over state of the art baselines. Through human evaluation, we observe that output from RETAG is upto 12% more faithful and analytical compared to a strong table-aware model. To the best of our knowledge, RETAG is the first model that can controllably use multiple reasoning methods within a structure-aware sequence to sequence model to surpass state of the art performance in multiple table to text tasks. We extend (and open source 35.6K analytical, 55.9k descriptive instances) the ToTTo, InfoTabs datasets with the reasoning categories used in each reference sentences.
|
In the task of Table to Text Generation Most table to text models, use structure-aware architectures or pretraining strategies to internally equip the model with one or more types of analyt-ical reasoning capabilities. We draw attention to two practical scenarios where reasoning categories should not be statically baked into a table to text model and more inference time control is needed. 1. Dataset Diversity: Tables like financial charts that consists of pre-dominantly numbers usually need numerical and temporal reasoning. On the other hand, tables like biographic information would need entity and common-sense knowledge more often. Since the same model would be used to summarize diverse tables, it becomes essential to be able to pick the appropriate reasoning categories based on the input table. 2. Usage and Context Diversity: Depending on the context of usage, a basketball match scorecard table can be summarized in two distinct ways. For avid sports experts, temporal and tabular reasoning can be used to summarize interesting patterns in the table. On the other hand, for a newspaper article the focus would shift to entity knowledge and crisp tabular summary of match results. Following from the examples above, we argue that explicit control is needed to dynamically pick the reasoning categories. during inference time. To the best of our knowledge, there is no prior literature on table to text summarization with explicit inference-time reasoning control. Next, we study some key patterns in the human generated references in ToTTo In this paper, we make the following contributions to address the problems described above. • We release an extended version of the ToTTo and the InfoTabs dataset that are annotated with the following popular reasoning categories: numerical, temporal, commonsense, table reasoning and entity knowledge as defined in
|
To the best of our knowledge, none of existing techniques provide controllability with multicategory reasoning. This is the focus of our work. We discuss other prior literature on table to text datasets and controllability in Appendix G. In this section, through a systematic human evaluation we first justify the choice of six reasoning categories (five analytical and the descriptive) that we introduce in Section 1. Next, we proceed to formally state the problem of Reasoning-Aware Table InfoTabs We sample roughly 20% of instances from the validation set of each dataset (1750 instances in ToTTo and 180 instances in InfoTabs respectively). The annotation below was done by a team of trained human experts and the details of the process is discussed in further detail in Section 6. All instances were now annotated with the 14 categories as described in We observe that three categories (Ellipsis, Negation, Subjective) are very rare in both datasets. Due to the sparsity of training data in these categories, we do not consider them for the modeling task. The categories (Coreference, Lexical reasoning, Syntactic Alternations) are well covered by linguistic pre-training datasets and hence we do not explicitly try to model them either. We focus on the remaining 8 categories. Due to similarity of the categories and to coalesce training data, we combine (Numerical Reasoning, Quantification) into one category. Similarly we combine Entity Type and Named Entities into one category leaving us with a total of six categories. Please note that the above methodology of constructing the six categories is an extension to Informed by the above analysis, we annotated the entire dataset of InfoTabs with the six categories. The training dataset for ToTTo is large (120K samples), therefore we using a filtering heuristic (refer Appendix-A) to annotate a subset of the training set. The test set of ToTTo is not directly available as it is an online benchmark. Hence we annotated only the validation dataset for ToTTo. Please note that no hyper-parameter tuning was done using the ToTTo validation set and was used for performance measurement alone. Let R be the set of six categories defined in Section 1: Descriptive, Tabular, Numerical, Temporal, Common-Sense and Entity reasoning. Given a table T , set of cells {C ij } contained within T and a reasoning category set r ⊆ R, the task of reasoning aware table to text is to generate a summary that uses all the information contained in {C ij }, applies analytical reasoning r and then produces a truthful summary S confirming to r. We use several automatic metrics to measure the relative quality of the generated sentences compared to the reference: BLEU In addition we use three human eval metrics defined by the following questions: 1.Reasoning: Does the generated summary use all reasoning categories required in r? 2. Faithful: Is the generated summary hallucination free (i.e., faithful to the table T ?) 3. Coverage: Are all cells in {C ij } used in the generated summary? We propose our model RETAG for controllable multi-skill reasoning table to text generation. We embed the two key aspects required to generate analytical sentences: control on reasoning categories, and being reasoning aware. First, to better model each reasoning category with control, we use a vector quantization (VQ) strategy in RETAG ( §4.2 and §4.4). Second, to precisely model each reasoning category, we use a pretraining strategy to better learn the reasoning categories from structured tables and free-form text data ( §4.3). For any table-to-text model, a basic property required is the efficient understanding of table structure. To infuse this aspect, we use the pretrained TAPEX model It is important to note that our model architecture contributions are not limited to TAPEX , and thus we integrate the same modules in T5 We denote the encoder and decoder as E and D. We use the question q to pass the reasoning tags in the input: Generate a sentence with TAG reasoning based on the following table? for analytical generation; Generate a descriptive sentence based on the following table? for the descriptive generation task. We concatenate q with the linearized table t to form the input x. The encoded vector is E(x) ∈ R N ×H , where N is the number of tokens in x, and H is the latent dimension. For our base model, the decoder D generates the corresponding output sentence y = D(E(x)). Our primary objective is to incorporate reasoning level control in our model RETAG . However, one of the main challenges of generating analytical sentences is to learn category specific aspects, which can be used to perform interaction between these categories for complex reasoning. To achieve this, we sandwich a vector quantization module between the encoder and the decoder. Each reasoning category has its own codebook on which the vector quantization operation is performed. We use the encoded representation E(x) to intervene reasoning specific knowledge from codebooks to create a new reasoning aware representation which is then passed to the decoder D. A codebook is denoted as c ∈ R K×H , which is a latent embedding space of size K, with H being the dimension of each of the codes c i in c Now, to model multi-category reasoning for generating analytical sentences, we propose the following weighted summation technique: Here, w i represents scalar weights for each reasoning category, predicted from the last layer of the encoder E, through an additional head layer. The binary labels 1 R simplify the equation so that the codebooks used are restricted only to the specified reasoning labels. Furthermore, we add a residual connection between the reasoning based representation and the original encoder representation. We then pass the resultant vector to the decoder D to generate the analytical sentence y a . We also generate the descriptive sentence y d in a similar manner using a residual connection between the encoded vector and the quantized representation from the descriptive codebook. For ease of understanding, we refer to Q r (x) as Q(x). In order to generate analytical sentences with our proposed architecture, it is crucial that the codebooks are rich in representing each of the reasoning categories efficiently. The five reasoning cat-egories we use extends beyond performing inferences on specific tables. Therefore, we explore pretraining strategies with various free-form and structured-data based datasets having the specific reasoning components. We collect the following datasets: (i) numerical and textual data (ND, TD) from We have a total of 276k instances from the above datasets spanning over the five reasoning categories. We formulate a seq-to-seq text generation task and pretrain our model (encoder, decoder, codebook) on the reasoning-aware dataset. We detail the model training strategy in Section 4.4. To further improve upon reasoning based representations, we add a classifier network M on top of the residual features u a and u d , which classifies it into analytical and descriptive classes. This classification constraint helps the model to broadly learn the difference between analytical and descriptive sentence. We term this strategy as CI (Classification of Intermediate activations). We show later that this classification strategy helps in improved generation for both descriptive and analytical sentences. The overall loss function for a batch consisting of both analytical and descriptive references, is as follows: The loss function consists of four components: Generative Loss: The first term is the crossentropy loss over the vocabulary for generating the gold descriptive and analytical sentences ŷd and ŷi . Classifier Loss: The second term is the cross-entropy loss from the classification constraint. We denote -and + as the descriptive and analytical class. Codebook Loss: stop-gradient (sg) in the third term is required for training the codebooks as the quantization process with argmin operation in Equation ( Den Oord et al. ( We evaluate the performance of RETAG as follows: (i) Performance comparison against strong baselines (ii) Ablation study of design choices used in RETAG (iii) Effectiveness of Vector Quantization (iv) Reasoning category wise quantitative analysis and (v) Human evaluations for faithfulness and reasoning control. Since our primary contribution is on multi-category reasoning, we benchmark RETAG 's performance for ToTTo valid and InfoTabs test datasets which have heterogeneous reasoning categories. Hence, we do not evaluate against datasets that are specific to one type of reasoning, such as LogicNLG, Numerical-NLG, etc. We use the following notations for our experiments: given a question q with a linearized table t, we concatenate them to form the input x (as mentioned in Section 4.1). The table may contain some highlighted cells, which can be enclosed within special indicators such as <hl> or can be addition-ally mentioned at the beginning or end of the table string. We assume that the highlighted cells would be indicated in either of these ways within the linearized table t. We use the strategy devised in We use the following models in our experiments: T5-Large (768M parameters) 1. We use only x to directly generate y. No information about r is consumed by the model. This is the usual seq-to-seq baseline. We denote this strategy as NO TAGS in Table 2. We use information about r as part of the question q. Then we train the models to generate y from x. The category information is thus used as part of the input string x. We deonte this as TAGS in Table 3. We use information about r with the codebook selection strategy. This is our proposed RE-TAG method with pretraining as mentioned earlier in Section 4 and Section 4.3. ToTTo: We observe that the models with the TAGS or RETAG approach generally outperform the models with NO TAGS. We observe this result across most of the evaluation metrics. In particular the TAPEX RETAG model achieves around 1% improvement in BLEU-1, ROUGE-L and around 2% improvement over the NO TAGS models for the overall performance in the validation set. The performance improvement in the analytical set is slightly more prominent compared to the descriptive set in the BLEU-1 and PARENT metrics. We postulate that the tag-based distinction between analytical and descriptive control improves the performance for descriptive sentences, as the model gets a clear signal of when to describe the content versus when to reason. We further study the importance of augmenting the input table with reasoning categories for finetuning the models in the Tags group of results. We observe that it leads to increment in performance across the BLEU-1 and PARENT metrics for the overall set. the improvement in performance is around 1% across for BLEU-1 and PARENT for all the models. InfoTabs We achieve considerable improvement with the TAGS and RETAG approach for overall performance and analytical set performance in Infotabs. TAPEX model has superior performance over the T5 family model in NO TAGS as the TAPEX is trained on table corpora on table understanding tasks. However, the performance of the comparatively poorer T5 and FLAN-T5 model is significantly improved with the use of categorical information. It re-iterates the importance of adding reasoning based control in various models. Our proposed TAPEX RETAG model still outperforms the TAPEX TAG model by more than 1% for BLEU-1 and PARENT for overall set, and around 3% and 1% for BLEU and PARENT for the analytical set. RETAG consists of three main components: the codebooks, the classification objective to differentiate analytical and descriptive and the pretraining technique of the codebooks. In this section, we study the effect of these three components on the We observe that six codebook setup consistently outperforms two codebook setup across the various metrics for both ToTTo and InfoTabs datasets (first and third rows in Table In Table 3 we also study the effect of classifying the residual features u a and u d into analytical and descriptive classes with the CI classification loss (Section 4.4). We observe that the RETAG performance improves consistently with CI constraint for both the two and six codebook setups on both the ToTTo and InfoTabs dataset. In Table In this section, we analyze the effectiveness of the codebooks for analytical generations as follows. InfoTabs TAPEX Label Type ToTTo Infotabs 1. Category-Wise Performance We evaluate results across reasoning categories in Table 2. Multi-Category Reasoning: We also study TAPEX RETAG for complex analytical sentences that involves two or more reasoning categories. We report average results for instances having two or three categories in the R.H.S of Table In Table We conclude that RETAG models capture reasoning-specific information in each codebook through pretraining, which it uses effectively for both single and multi-category analytical sentence generation. We sample 500 instances from the ToTTo validation set and generate their corresponding analytical sentences from four different models specified in Table We ask the annotators to provide a label between -yes (score 1), partially (score 0.5), or no (score 0). We compare TAPEX RETAG against the TAPEX model without tags for human evaluation to quantify faithfulness and reasoning control. We collect 3 ratings for each sample and compile the majority label results in percentage scale in Table As explained in Section 3.1, we will release 31.4K analytical instances and 50.6K for the ToTTo train and validation set. We will also release 4.2K analytical instances and 5.3K descriptive sentences over the entire InfoTabs dataset. This section explains the human labeling methodology and the corresponding performance metrics. We prepared detailed annotation instructions, qualifying questions and trained a pool of 14 crowdsource annotators. The annotators are based in India and are fluent in English. The annotators were paid at rates above required by local wage laws. We instructed the annotators to choose one or more of the five reasoning categories for analytical sentences. We instructed them to keep the five reasoning categories and the Descriptive category exclusive i.e. a sentence is descriptive only when it does not use any of the other five reasoning categories. Three annotators labeled every instance and we keep only those label voted by atleast two raters. The annotators reached a high consensus agreement on the task. 86.81% of ratings had all three raters agree on the binary class for categorizing between descriptive and analytical. 75.12% all three raters agreed on the exact same set of categories for choosing the analytical categories. In this paper, we presented the case for Reasoningaware table-to-text models. We introduced RETAG, a vector quantized approach for encoder-decoder table-to-text models with explicit control across reasoning categories. RETAG beats SOTA models for ToTTo and InfoTabs datasets for analytical and descriptive sentence generation. We will also release close to 35.6K instances of reasoning category tagged analytical abd 55.9k instances of descriptive table to text data. Some of the limitations of our work are as follows. First, the dataset curation and performance evaluation was restricted to datasets in the English language, and does not extend to non-dominant languages. Second, several advanced methods have been introduced for numerical reasoning. Our current strategy to incorporate reasoning is datacentric. However, we would like to emphasize that the explicit reasoning control is complementary to the existing methods and in future works, advance methods to infuse reasoning can be used alongside our method. Third, to gain explicit reasoning control for newer domain/reasoning category, involves few examples to be annotated to bootstrap the model using our method. Fourth, although RETAG is designed for multiple skill reasoning, in future work we will also benchmark RETAG against reasoning specific datasets such as Logic-NLG. ToTTo InfoTabs We show some examples of table, sentence pairs and the corresponding reasoning categories Figure We use beam search to generate outputs from the our generative models. We used a beam length of 10 is used. All models were trained with the AdamW optimizer with a learning rates of 1e-6, 3e-6, 5e-6, 1e-5. We used Quadro RTX 8000 GPU for our experiments. We train all our models for 10 epochs, which takes 3 hours for ToTTo and 1.5
| 1,537 | 1,869 | 1,537 |
Mind the Gap between the Application Track and the Real World
|
Recent advances in NLP have led to a rise in inter-disciplinary and application-oriented research. While this demonstrates the growing real-world impact of the field, research papers frequently feature experiments that do not account for the complexities of realistic data and environments. To explore the extent of this gap, we investigate the relationship between the realworld motivations described in NLP papers and the models and evaluation which comprise the proposed solution. We first survey papers from the NLP Applications track from ACL 2020 and EMNLP 2020, asking which papers have differences between their stated motivation and their experimental setting, and if so, mention them. We find that many papers fall short of considering real-world input and output conditions due to adopting simplified modeling or evaluation settings. As a case study, we then empirically show that the performance of an educational dialog understanding system deteriorates when used in a realistic classroom environment.
|
Modern NLP systems, powered by large language models (LLMs), now have the ability to perform well at foundational natural language understanding and generation tasks Recently published research from these tracks includes work on complex and important tasks such as synthesizing code for visualization Motivated by this observation and in line with the ACL 2023 theme track, we set out to investigate the relationship between the motivation described in the introductions and the actual experiments in application-focused NLP papers. We survey papers from the NLP applications tracks of ACL 2020 and EMNLP 2020. Specifically, we ask if there are gaps between motivation and experimentation, in the form of i) sub-tasks that are required for the application, but haven't been mentioned in the paper ii) data distributions that are expected in real-world conditions, but haven't been included in the paper's modeling or evaluation. We find that authors do not always explicitly mention assumptions they make, and often operate in con- strained scenarios highly different from their intended motivation. To empirically demonstrate the severity of this problem, we then present a case study investigating the performance of an educational dialog system, when the inputs are changed from manually transcribed data to transcripts from a state-of-the-art ASR system. The purpose of the system is to classify utterances made by a student in a classroom into talkmoves
|
For the survey of application-oriented research papers, we look at all papers from the NLP Applications track of two recent NLP conferences, ACL 2020 and EMNLP 2020, which have a total of 115 papers. These conferences, which were conducted virtually, provide publicly available interfaces, We then manually filter papers to identify those that propose and work on new tasks. We choose these since papers that tackle existing tasks, such as fact checking, might be restricted to existing benchmarks and datasets that are established in a topic We then answer four questions about each paper: 1. Does the paper comprehensively describe the use case for a reader to understand? This question helps us establish that the motivations of the authors are clear to us before proceeding with the survey. We discard papers if the answer is no here. 2. Is the paper dealing with an entire task or a sub-task only? An example of the sub-task only would be if the desired application was assisting students with writing by providing feedback, but the actual task worked on was detecting errors in writing, with the task of formulating feedback being a sub-task for future work. 3. Does the paper mention the other missing subtasks explicitly? We investigate if the authors either mention existing systems that work on the other sub-tasks, or explicitly describe the remaining steps as future work. This is only collected when the answer to Q2 is "sub-task only". 4. Is the downstream evaluation realistic? An example of the answer being No, is if the expected use-case requires classifying spoken dialog in real-time, but the paper only evaluates on manually transcribed data. The survey is conducted by three authors of this paper, who have all been working on NLP for 3+ years. In cases where agreement is not perfect, we report the majority answer. While all four questions take either yes or no for an answer, we optionally collect reasons for answering no on Questions 1 and 4. We only accept unsure as an answer when no decision can be made. The results of the survey are presented in Table We compute inter-rater agreement using Krippendorff's α, used when there are more than two annotators In this section, we present a case study of an application from the domain of education. The task involves classifying student utterances into talk moves Dataset The data consists of conversations among middle school students performing collaborative work in science classrooms, documented in more detail in Model Following the state-of-the-art model for classifying teacher talk moves However, when deploying our models in the classroom, we do not have access to clean human transcripts, and instead need to work with the outputs of ASR systems. To compare the differences between both, we look at two state-of-the-art ASR systems: Google To show the performance gap caused by the above distribution shift, we evaluate our model on both human transcriptions and transcriptions from the two ASR systems. For each ASR transcript, we report both performances on their filtered version (Google filter , Whisper filter ) and the all ver- sion (Google all , Whisper all ). We report macro F1 as well as class-wise F1 for all models, as shown in Table When considering real-world deployment, the potential for such a dramatic drop in performance should be taken into account by both the designer (including researchers) and the user (such as teachers). However, for similar applications based on classroom discourse analysis, such as classifying teacher talk moves Through the above survey and case study, we qualitatively and quantitatively examine the gap between task-focused solutions in NLP research, and realistic use cases. We first acknowledge that there has existed a long-standing tradition in NLP to contextualize current research efforts through potential future applications. Looking at task-oriented dialog systems for example, early work such as To bridge this gap, we propose a few initial recommendations: i) we suggest including a question on the Responsible NLP Checklist Our paper adds to a body of work on meta-analysis of NLP papers and the state of NLP research, particularly from the recently introduced theme tracks at *ACL conferences In addition, We investigate the "gap" between the motivations of application-focused NLP papers and their actual experimental setting. Through a survey of NLP Applications papers from two NLP conferences, we find that i) necessary components for the application get overlooked when papers focus on subtasks and ii) realistic input sources such as ASR are not being considered in downstream evaluations. We further highlight the severity of the latter issue through a case study on a dialog understanding system intended for classrooms, showing the drop in performance when ASR input, expected in the real-world, is used. While we outline potential strategies to address this issue, we hope our work will spur further discussion about future steps. One of the limitations of our survey is that it covers a limited sample space of 15 papers from EMNLP 2020 and ACL 2020. While a larger sample would be helpful in gathering more evidence, access to specific tracks is limited at NLP conferences, unless hosted online via a virtual or hybrid system. With respect to our case study, we evaluate on the ASR utterances, but with labels corresponding to the original manual transcriptions. For a perfect comparison, the ASR utterances would need to be re-annotated as the talk move could change based on the severity of transcription errors. Fan Zhou, Shengming Zhang, and Yi Yang. 2020. Interpretable operational risk classification with semisupervised variational autoencoder. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 846-852, Online. Association for Computational Linguistics. A Talk Move and Label Clustering D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank. D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank. D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank.
| 1,014 | 1,458 | 1,014 |
SEM-F 1 : an Automatic Way for Semantic Evaluation of Multi-Narrative Overlap Summaries at Scale
|
Recent work has introduced an important yet relatively under-explored NLP task called Semantic Overlap Summarization (SOS) that entails generating a summary from multiple alternative narratives which conveys the common information provided by those narratives. Previous work also published a benchmark dataset for this task by collecting 2, 925 alternative narrative pairs from the web and manually annotating 411 different reference summaries by engaging human annotators. In this paper, we exclusively focus on the automated evaluation of the SOS task using the benchmark dataset. More specifically, we first use the popular ROUGE metric from text-summarization literature and conduct a systematic study to evaluate the SOS task. Our experiments discover that ROUGE is not suitable for this novel task and therefore, we propose a new sentencelevel precision-recall style automated evaluation metric, called SEM-F 1 (Semantic F 1 ). It is inspired by the benefits of the sentence-wise annotation technique using overlap labels reported by the previous work. Our experiments show that the proposed SEM-F 1 metric yields a higher correlation with human judgment and higher inter-rater agreement compared to the ROUGE metric.
|
Human beings can be viewed as subjective sensors who observe real word events and report relevant information through their narratives Given two distinct narratives N 1 and N 2 of an event e, how can we automatically generate a single summary about e which conveys the common information provided by both N 1 and N 2 ? Multiple-perspective alternative narratives are frequent in a variety of domains, including education, the health sector, military intelligence, content analysis and privacy. Therefore, automatic summarization of multiple-perspective narratives has become a pressing need in this information explosion era and can be highly useful for digesting such multi-narratives at scale and speed. Figure The goal of SOS task is to generate a summary that conveys the common/overlapping information provided by the green text, i.e., the terrorist leader has been located. At first glance, the SOS task may appear similar to a traditional multi-document summarization task where the goal is to provide an overall summary of the (multiple) input documents; however, the difference is that, for SOS, the goal is to provide summarized content with an additional constraint, i.e., the commonality criteria. There is no current baseline method that exactly matches our task; more importantly, it is unclear how to properly evaluate this task in an automated fashion. Therefore, as a starting point, we frame the SOS task as a constrained seq-to-seq problem where the goal is to generate a summary from two input documents that convey the overlapping information present in both input text documents. However, the bigger challenge we need to first address is the evaluation of the task. To address these challenges, we make the following contributions in this paper. 1. We frame Semantic Overlap Summarization (SOS) (from multiple alternative narratives) as a constrained multi-seq-to-seq problem and exclusively study how automatic evaluation of this task can be performed at a large scale. 2. As a starting point, we experiment with ROUGE, a widely popular metric for evaluating text summarization tasks, and demonstrate that ROUGE is NOT suitable for the automatic evaluation of SOS task. 3. Based on the findings of our previous work, we propose a new precision-recall style evaluation metric, SEM-F 1 (Semantic F 1 ), for evaluating the SOS task. Extensive experiments show that new SEM-F 1 improves the interrater agreement compared to the traditional ROUGE metric, and also, shows a higher correlation with human judgments.
|
As SOS can be viewed as a multi-document summarization task with additional commonality constraint, text summarization literature is the most relevant to our work. Over the years, many paradigms for document summarization have been explored Recently, encoder-decoder-based neural models have become really popular for abstractive summarization Alternatively, one could aim to recover verb predicate-alignment structure Along the evaluation dimension, ROUGE Here we first provide a brief description of the SOS task and the benchmark dataset that was introduced by To simplify notations, let us stick to having only two documents D A and D B as our input since it can easily be generalized in case of more documents using SOS repeatedly. Also, let us define the output as A human would mostly express the output in the form of natural language and thus, the SOS task is framed as a constrained multi-seq-to-seq (text generation) task where the output text only contains information that is present in both the input documents. Also, overlap summary should also have minimal repetition i.e. brevity is a desired property of Semantic Overlap Summarization. For example, if a particular piece of information or quote is repeated twice in both documents, we don't necessarily want it to be present in the output overlap summary two times. The output can either be an extractive summary or abstractive summary or a mixture of both, as per the use case. Additionally, SOS should follow the commutative property, i.e Here I i refers to the i th annotator where i ∈ {1, 2, 3, 4} and "Average" row represents the average correlation of the max values across annotators. Boldface values are statistically significant at p-value < 0.05. For 5 out of 6 annotator pairs, the correlation values are quite small (≤ 0.50), thus, implying the poor inter-rated agreement with regards to the ROUGE metric. One of the key challenges with SOS task As ROUGE we first conducted a case study with ROUGE as the evaluation metric for the SOS task. For methods, we experimented with multiple SoTA pretrained abstractive summarization models as naive baselines for Semantic-Overlap Summarizer (SOS). These models are 1) BART Generating the summary: In order to handle two input documents, we concatenate them and feed the concatenated input directly to the model. The maximum summary length model hyper-parameter was set to 300 based on the max words across samples in the training data. The default values were used for all other hyper-parameters for each respective model. Post-Processing: After the generation of model summaries, we did very basic post-processing. For example, for the Pegasus model, the new line character '<n>' was simply replaced by a blank space following the code from Huggingface. For evaluation, we first evaluated the machinegenerated overlap summaries for the 137 manually annotated testing samples using the ROUGE metric and followed the procedure mentioned in the paper 5 Sentence-wise Manual Scoring Overlap Labels: Label-annotators (L 1 , L 2 and L 3 ) were asked to look at each machine-generated sentence separately and determine if the core information conveyed by it is either absent, partially present or present in any of the four reference summaries (provided by I 1 , I 2 , I 3 and I 4 ) and respectively, assign the label A, PP or P. More precisely, annotators were provided with the following instructions: if the human feels there is more than 75% overlap (between each system-generated sentence and any reference-summary sentence), assign label P, else if the human feels there is less than 25% overlap, assign label A, and else, assign PP otherwise. This sentence-wise labelling was done for 50 different samples (with 506 sentences in total for system and reference summary), which resulted in a total of 3 × 506 = 1, 518 sentence-level ground-truth labels. To create the overlap labels (A, PP or P) for precision, we concatenated all 4 reference summaries to make one big reference summary and asked label-annotators (L 1 , L 2 and L 3 ) to use it as a single reference for assigning the overlap labels to each sentence within machine generated summary. We argue that if the system could generate a sentence conveying information which is present in any of the references, it should be considered a hit. For recall, label-annotators were asked to assign labels to each sentence in each of the 4 reference summaries separately (provided by (I 1 , I 2 , I 3 and I 4 )), with respect to the machine summary. Inter-Rater-Agreement: After annotating each system-generated sentence (for precision) and reference sentence (for recall) with the labels (A, PP or P), we used the Kendall rank correlation coefficient to compute the pairwise annotator agreements among these ordinal labels. Table y ← Column-wise-max(cosSs) 14: return (x, y) 15: end procedure larity between two sentence embeddings) to infer the semantic overlap between a system-generated sentence and a reference sentence from both precision and recall perspectives and then, combine them into the F 1 score. The SEM-F 1 metric computes cosine similarity scores between sentence pairs from both precision and recall perspectives. To verify whether the SEM-F 1 metric correlates with human judgement, we further converted the sentence-wise cosine similarity scores into Presence (P), Partial Presence (PP) and Absence (A) labels using user-defined thresholds as described in algorithm 2. This helped us to directly Table Algorithm 2 Threshold Function 1: procedure THRESHOLD(rawSs, T ) 2: initialize Labels ← [] 3: for each element e in rawSs do 4: if e ≥ tu% then 5: Labels.append(P ) 6: else if t l % ≤ e ≤ tu% then 7: Labels.append(P P ) 8: else 9: Labels.append(A) 10: end if 11: end for 12: return Labels 13: end procedure compare the SEM-F 1 inferred labels against the human annotated labels. We leveraged state-of-the-art sentence embedding models to encode sentences from both the model-generated summaries and the human-written reference summaries. To be more specific, we experimented with 3 sentence encoder models: Paraphrase-distilroberta-base-v1 (P-v1) Following the procedure in Table Here, we present the actual SEM-F 1 scores for the three models (BART, T5 and Pegasus) described in section 4 along with scores for two random baselines: 1) Random Reference, 2) Random Output. Random Reference: Here, the model-generated summary is compared against a random reference to compute SEM-F 1 scores. The random selection is done by sampling a reference summary from the pool of remaining 136 × 4 = 544 references. Random Output: In this case, a randomly generated output is compared against actual humanwritten reference summaries to compute SEM-F 1 scores. The random selection is done by sampling a machine-generated output from the pool of remaining 136 machine-generated outputs. As reported in table 8, abstractive summarization models achieve approximately 40-45 percent improvement over the random baseline scores suggesting SEM-F 1 can indeed distinguish the "good" from the "bad". Following the case-study based on ROUGE in section 4, we computed the Pearson's correlation coefficients between each pair of raw SEM-F 1 scores obtained using each of the 4 reference summaries. The corresponding correlations are shown in Table 9. For each annotator pair, we report the maximum (across 3 models) correlation value. The average correlation value across annotators is 0.49, 0.49 and 0.54 for P-V1, STSB, USE embeddings, respectively, suggesting a clear improvement over ROUGE. In this work, we proposed a more accurate metric, called SEM-F 1 , for evaluating the SOS task. This metric compares the model-generated overlap summaries with the reference summary on a per-sentence basis using overlap labels and com-bines them to generate F 1 scores. Our experiments show that SEM-F 1 is more robust and yields higher agreement with human judgement and most importantly, can be computed automatically making it suitable for large-scale evaluation. One particular limitation of this work is that we have used pre-trained abstractive summarization models as naive baselines / proxy for semantic overlap summarizer and did not attempt to develop a custom method which optimizes for the overlap constraint. However, the primary focus of this paper is the evaluation of the SOS task. Therefore, the design and optimization of methods is an orthogonal goal to this paper, which we will pursue as our immediate future work. We use the benchmark dataset proposed by
| 1,223 | 2,531 | 1,223 |
Bridging The Gap: Entailment Fused-T5 for Open-retrieval Conversational Machine Reading Comprehension
|
Open-retrieval conversational machine reading comprehension (OCMRC) simulates reallife conversational interaction scenes. Machines are required to make a decision of Yes/No/Inquire or generate a follow-up question when the decision is Inquire based on retrieved rule texts, user scenario, user question and dialogue history. Recent studies try to reduce the information gap between decision-making and question generation, in order to improve the performance of generation. However, the information gap still persists because these methods are still limited in pipeline framework, where decision-making and question generation are performed separately, making it hard to share the entailment reasoning used in decision-making across all stages. To tackle the above problem, we propose a novel one-stage end-to-end framework, called Entailment Fused-T5 (EFT), to bridge the information gap between decisionmaking and question generation in a global understanding manner. The extensive experimental results demonstrate that our proposed framework achieves new state-of-the-art performance on the OR-ShARC benchmark. Our model and code are publicly available 1 .
|
Open-retrieval conversational machine reading comprehension (OCMRC) Inquire based on retrieved rule texts, user scenario, user question and dialogue history. Previous studies ing decision-making, span extraction and question rephrasing. Different entailment reasoning strategies are utilized to improve the performance of decision-making. Span extraction and question rephrasing are conducted for question generation. These pipeline frameworks are either completely independent of the three stages However, the information gap still persists because these methods are still limited in pipeline framework, where decision-making and question generation are performed separately, making it hard to share the entailment reasoning used in decision-making across all stages. To tackle the above problem, we propose a novel one-stage end-to-end framework, called entailment fused-T5 (EFT) to bridge the information gap between decision-making and question generation in a global understanding manner. Specifically, our model consists of a universal encoder and a duplex decoder. The decoder consists of an activated entailment reasoning decoder and a fused answer generation decoder. The implicit reasoning chains of both decision-making and question generation in the multi-fused answer generation decoder are explicitly supervised by ac- tivated entailment reasoning through the shared entailment representation of our encoder. Moreover, a relevance-diversity fusion strategy is utilized to further improve the implicit reasoning abilities among the multiple retrieved rules for the fused answer generation decoder through the implicit ranking method. Thus, our model can reason in a global understanding manner. The extensive results, as illustrated in Figure Our contributions are summarized as follows: • We propose a novel one-stage end-to-end framework, called entailment fused-T5 (EFT) to bridge the information gap between decision-making and question generation through a global understanding manner. • We further investigate a relevance-diversity fusion strategy (RD strategy) to improve the implicit reasoning abilities of our model. • Extensive experiments demonstrate the effectiveness of our proposed framework on the OR-ShARC benchmark.
|
Conversation-based Reading Comprehension Conversation-based reading comprehension Conversational Machine Reading Comprehension CMRC In this paper, we focus on the open-retrieval conversational machine reading (OCMRC) task Due to the hybrid-form task, the previous methods In open-retrieval CMRC, the machines are first required to retrieve related rule texts in a knowledge base, given user question and user scenario. Then machines are required to make decisions or generate follow-up questions based on retrieved rule texts, user scenario, user question and dialogue history. Thus, we conduct a retriever to first retrieve related rule texts from the knowledge database, and then generate the final answer through our end-to-end reader EFT. The training procedure is shown in Algorithm 1. We first concatenate the user question and user scenario as the query to retrieve related rule texts in the knowledge base. The knowledge base is divided into the seen subset and the unseen subset. This is to simulate the real usage scenario: users will ask questions about rules they have already seen, or rules that are completely new. We only use seen rules in the training process. In this work, we utilize DPR In this stage, each item is formed as the tuple {R, S, Q, D}. R donates the rule text can- didates. R = {r 1 , r 2 , ..., r k }, where r donates the rule text item of R. S and Q represent user scenario and user questions, respectively. D donates the dialogue history. Given {R, S, Q, D}, EFT will directly generate a decision of Yes/No/Inquire or follow-up question when the decision is Inquire. EFT consists of a universal encoder and a duplex decoder. The Duplex decoder consists of an activated entailment reasoning decoder and a fused answer generation decoder. In this way, the whole implicit reasoning chains of the fused answer generation decoder will be fine-grained supervised by activated entailment reasoning with the shared entailment representation. Thus, the fused answer generation decoder will reason in a global understanding manner. The details are shown in Figure Given {R, S, Q, D}, we random sample k items in R, and concatenate each of them with S, Q, D as c, thus the item collection is formed as C = {c 1 , c 2 , ..., c k }. Specifically, each r in R is first parsed to elementary discourse unit (EDU) by a pre-trained model We utilize duplex decoding to explicitly supervise our answer generation stage, which introduced the explicit entailment reasoning information in implicit answer generation reasoning. The answer generation will either generate a decision of Yes/No/Inquire or a follow-up question when the decision is Inquire. The activated entailment reasoning decoder will reason the entailment states of the EDUs. The duplex decoder is trained in a multi-task form. And the activated entailment reasoning only activates in training stage. Reasoning Each EDU will be classified into one of three entailment states, including ENTAILMENT, CONTRADICTION and NEUTRAL. To get the noisy supervision signals of entailment states, we adopt a heuristic approach Previous studies typically introduce entailment reasoning in all rule text segmented EDUs. This will greatly increase the proportion of NEUTRAL labels and affect the model effect, because nearly all of the entailment states of EDUs in retrieved irrelevant rules are NEUTRAL, and introducing more noise in the training stage. In our method, entailment reasoning will only activate for the golden rule text. Utilizing this setting, one benefit is to balance the categories of entailment reasoning, and the other is to supervise the implicit reasoning of the fused decoder, which can help the fused decoder infer correct rule text from multiple retrieved rule texts. Given the sentence-level representation H s , we utilize inter-attention reasoning to fully interact with various information in r, including EDUs, user question, user scenario and dialogue history. We utilize an inter-sentence Transformer where the W c is trainable parameters, e i is the predicted score for the three labels of the i-th states. end for 8: θ e ← θ eτ g 10: θ a ← θ aτ g 12: end while the relevance-diversity fusion strategy (RD fusion strategy), which consists of relevance-diversity candidate rule texts, order information protection and fused answer generation. The rule text candidates are consists of top k relevant rule texts and randomly sampled rule texts, which are called RD candidate rule texts. Thus, the candidates are full-filled with relevant and diverse rule texts. On the premise of ensuring relevance among the rule texts, the diversity of learnable information sampling combinations is further improved. Moreover, the order of items fused in f w may lead to information leakage and affect the reasoning ability of the decoder in the training stage, so as we mentioned in the last section, we will randomly shuffle the order of items when inputting to the encoder to protect the order information. In the evaluation stage, only the top 5 unshuffled relevant rule texts will be utilized for answer generation. The fused answer generation is utilized to generate either the decision or the follow-up question. We employ T5 decoder as our answer generation decoder. Given encoder fused representation f w , and the final answer a, including decision or follow-up question, the answer is composed of the variable-length tokens x i , the probabilities over the tokens are shown in the blow: where θ donates the trainable parameters of our decoder. Activated Entailment Reasoning The activated entailment reasoning is supervised by crossentropy loss, by given the entailment stages c i : where r is the ground truth of entailment state. Fused Answer Generation The fused answer generation training objective is computed as illustrated in below: The overall loss function is: 4 Experiment and Analysis Dataset The experiment dataset is OR-ShARC Evaluation The evaluation consists of two parts: decision-making and question generation. We utilize Micro-Acc and Macro-Acc for the results of decision-making, and use F1 BLEU Where M is the total number of Inquire decisions made by our model. y i is the predicted question, ŷi is the corresponding ground truth prediction. The recall of BLEU is computed in a similar way. where N is the total number of Inquire decision from the ground truth annotation, The calculation of F1 BLEU is shown below: Implementation Details We utilize the T5-base The fused number k is set as 5 for fused answer generation for both training and inference. We use AdamW All results on the OR-ShARC benchmark are illustrated in improvement in BLEU scores. EFT outperforms OSCAR by 27.7% in F1 BLEU1 , 36.9% in F1 BLEU4 for the question generation on the dev set, and outperforms OSCAR by 20.8% in F1 BLEU1 , 26.5% in F1 BLEU4 for the question generation on the test set. We further to investigate the classwise accuracy performance of EFT, as shown in Table To further investigate the performance for our proposed EFT in seen and unseen settings, the performance of the split subset 3 is illustrated in Table EFT is 35.0% higher in micro-acc than seen subset, 35.3% higher in macro-acc than unseen subset. Our retrieval results are illustrated in Table The ablation studies of EFT on the dev set of OR-ShARC benchmark are shown in Table • EFT-wo/s trains the model without relevance-diversity (RD) candidate rule texts. Only top-5 randomly shuffled relevant rule texts are considered in the training stage. • EFT-wo/s+a trains this model additionally remove activated entailment reasoning. • EFT-wo/s+a+i trains this model further cancels random shuffle in the training stage. • EFT-wo/s+a+i+f trains this model without multi rule fused answer generation, only top-1 retrieved rule text is considered. We investigate the necessity of the RD candidate rule texts. This strategy is utilized to improve the implicit reasoning abilities of our decoder by improving the learning space of fused candidates. On the premise of ensuring the relevance among the rule texts, the diversity of learnable information sampling combinations is further improved. As shown in texts in enhancing the information-seeking abilities of our fused answer generation decoder. By removing RD candidate rule texts, the micro-acc is decreased by 0.5, the macro-acc is decrease by 0.4, the F1 BLEU1 is decreased by 1.7, and the F1 BLEU4 is decreased by 2.0. The above results emphasize the indispensability of RD candidate rule texts. EFT-wo/s+a trains this model additionally remove activated entailment reasoning. As illustrated in Table The order of fused representation used in fused answer generation decoder may lead to information leakage and affect the reasoning ability of the decoder. In order to avoid the problem of poor information seeking ability caused by excessive learning of position information of the model, we randomly shuffle the order of fused representation to protect the order information in the training stage. As illustrated in Table Analysis of Fused Generation Fused Generation is utilized to introduce the ability to pro- cess multiple rule contextualized information. The multiple rule contextualized information are fused as a single fused information. EFT-wo/s+a+i+f trains this model without multi rule fused answer generation, only top1 retrieved rule text is considered. In this manner, the performance is limited with the retrieval performance. Compared with EFT-wo/s+a+i+f, the performance of both decision-making and question generation of EFTwo/s+a+i are significantly improved by introducing fused answer generation strategy, the microacc is increased by 9.2 , the macro-acc is increased by 8.9, the F1 BLEU1 is increased by 12.0, and the F1 BLEU4 is increased by 11.2. The above results suggests the necessity of fused answer generation strategy. In this paper, we propose a novel end-to-end framework, called EFT, to bridge the information gap between decision-making and question generation through the shared entailment representation in a global understanding manner. Extensive experimental results on the OR-ShARC benchmark demonstrate the effectiveness of our proposed framework EFT. In our analysis, the implicit reasoning ability of both decision-making and question generation is significantly improved by sharing external explicit entailment knowledge through our novel framework EFT. As shown in Table Dev Set Test Set Top1 Top5 Top10 Top20 Top1 Top5 Top10 Top20 Limitations section after section 5 A2. Did you discuss any potential risks of your work? Limitations section after section 5 A3. Do the abstract and introduction summarize the paper's main claims? Abstract and section1 A4. Have you used AI writing assistants when working on this paper? We use chatGPT to verify the grammar of our abstract. B Did you use or create scientific artifacts? Appendix.C B1. Did you cite the creators of artifacts you used? Appendix.C B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix.C B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Not applicable. Left blank. B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 4.1 B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix.C and Section 4.2
| 1,159 | 2,245 | 1,159 |
GAINER: Graph Machine Learning with Node-specific Radius for Classification of Short Texts and Documents
|
Graphs provide a natural, intuitive, and holistic means to capture relationships between different text elements in Natural Language Processing (NLP) such as words, sentences, and documents. Recent advancements in the field of Graph Machine Learning (GML) have led to the development of numerous models to process text for various natural language applications, including but not limited to short-text classification, document classification, and others. At the heart of GML models, specifically those based on Graph Neural Networks (GNNs), lies the message passing operation which has shown to be an essential component for strong empirical performance in NLP. However, the number of message passing steps (often known as the radius) is fixed for all the nodes in existing GML models for NLP. Fixing the radius poses a fundamental restriction as nodes exhibit diverse properties and varying amounts of informative local structures in the input graph. This paper presents GAINER, a novel framework called Graph mAchine learnIng with Node-spEcific Radius, aimed at graph-based NLP. We propose non-neural and novel neural approaches built on the core ideas of GAINER. Through rigorous experimentation, we demonstrate the efficacy of GAINER in popular NLP tasks.
|
Graphs present a natural, intuitive, and holistic representation for understanding the interactions that exist among different text elements, such as words, sentences, and documents. The use of graphs provides a wide array of options for effectively representing and tackling different problems in Natural Language Processing (NLP). For instance, worldlevel, sentence-level, and document-level graphs capture various aspects of text datasets. Recent breakthroughs in Graph Machine Learning (GML), notably driven by the progress made in Graph Neural Networks (GNNs) The prevalent approach in GNNs for NLP research, including very recent publications Inspired by the aforementioned fundamental limitations of existing GML models in graph NLP, our work makes the following contributions: • We propose GAINER, a novel framework called Graph mAchine learnIng with Node-spEcific Radius, aimed at graph NLP (Please see Figure • We propose novel approaches aimed at graph NLP, comprising Simple-GAINER (a nonneural approach) and Neural-GAINER, built upon the core idea of GAINER (Please see Sections 4.3 and 4.6). • We demonstrate the adaptability of GAINER and its efficacy in a wide range of tasks including short-text classication, document classification on text attributed graphs, and document coherence assessment. Our methods achieve statistically significant results on 5 of the 6 datasets evaluated (Please see Section 5).
|
We divide the related work into three subsections. The prevailing trend in machine learning models for graph-structured inputs involves the learning of representations for graph nodes Decoupled GNNs, characterised by the separation of the message passing operation and the feature transformation operation, have emerged as effective models in GML tasks Adaptive GNNs, equipped with gate/attention mechanisms or reinforcement learning, have been suggested by numerous learning-based approaches to dynamically aggregate information for each individual node Our proposed method merges the strengths of decoupled and adaptive approaches, offering a blend of simplicity and adaptability tailored to taskspecific applications. The presence of graph structures in a wide range of NLP problems has sparked a surge of interest in utilising GNNs as a promising approach to tackle several NLP tasks effectively In most of the existing literature on GNNs in NLP, a 2-layer GCN is commonly employed, which may work well for nodes with strong connections but falls short in effectively handling nodes with weak connections in the graph (e.g., low degree nodes connected to other low degree nodes). Our proposed idea of employing a node-specific radius is specifically tailored to tackle nodes characterised by a weak or inadequate connectivity structure. In this study, we investigate text classification and document processing tasks as illustrative examples and leave other tasks for future work. We present notation to introduce the method and discuss problems studied in the paper. We first delve into the notations used in this work, to establish a common understanding of the symbols and terminology used throughout the paper. Input Graph: Let G = (V, E) be an input undirected graph where Let à ∈ {0, 1} n×n be the adjacency matrix of G with self-loops, i.e., Ãv,v = 1 for all v ∈ V . Note that Ãv,u = 1 if and only if there exists an edge betwneen v ∈ V and u ∈ V . Let ∆ be a diagonal matrix consisting of the node degrees, i.e., ∆ v,v = n u=1 Ãv,u and zero entries elsewhere. We assign the symbol A to represent the symmetrically normalised adjacency matrix T ∈ R n×d denotes the input feature matrix. The superscript 0 in X (0) signifies that the features utilised in the GML model are not treated as hidden features but are instead directly incorporated as input. Many problem instances in Graph NLP are approached through the popular GCN model (Kipf and Welling, 2017) as a go-to solution, capitalising on its ability to integrate the graph G and the input node features X (0) . Leveraging an aggregation process, the GCN model merges the features of a node with the features of its neighbours, enabling the creation of smoother node representations. The process of an L-layer GCN can be defined as (1) where η(•) is the activation function and W (l) is a layer-specific trainable weight matrix at layer l. Within the scope of this paper, we analyse noteworthy NLP problems, drawing attention to the nodes and edges of the input graph G = (V, E), and the node features X (0) exploited by GNNs in NLP. 1) Short Text Classification: Based on recent research 2) Document Classification in Text Attributed Graphs: In the domain of text attributed graphs, the customary practice involves using nodes to represent documents for node classification purposes 3) Document Coherence Assessment: An alternative way to model the structural similarity of documents is by analysing the sentences within them In the aformentioned examples, the existing literature employs an L-layer GCN, which considers L-hop information around each node to propagate and smooth information across edges. The number of layers L, is considered a hyperparameter, and empirical results suggest that setting L = 2 yields the best performance in most cases. Figure While the approach proves effective for wellconnected nodes, such as (i) high-degree nodes, or (ii) low-degree nodes with high-degree neighbours, it falls short when it comes to poorly connected nodes, such as low-degree nodes connected to other low-degree nodes. Furthermore, as the value of The graph is the same in all the four images. In the first and third images, the node of interest is indexed by 1, while in the second and fourth images, it is the node with index 2. A GCN with only 2 layers might not capture information from a sufficient number of hops, leading to an inadequate representation of poorly-connected nodes (first image). By examining the second image, we can see that a well-connected node possesses a 2-hop neighborhood that spans a significant portion of the graph. Adding more GCN layers can lead to excessive smoothing, resulting in highly similar representations for the majority of nodes. Third and fourth images illustrate that by incorporating a node-specific radius, GAINER can flexibly adjust the degree of smoothing, leading to larger radii for poorly connected nodes (e.g., green nodes around 1) and smaller radii for well-connected nodes (e.g., blue nodes around 2). L increases, the hidden representations of wellconnected nodes become excessively smoothed, resulting in oversmoothing Our proposed approach to address this tradeoff revolves around the introduction of a node-specific radius, represented as r(v, τ ), as a replacement for the conventional number of layers L in GCNs. This radius is assigned to each node v ∈ V , and is complemented by a threshold value τ > 0. This motivates our framework referred to as GAINER An essential finding in the GCN process, as described by Equation (2) SGC has emprically shown to be highly competitive in terms of accuracy and offers substantial training speed improvements over GCN across various datasets, including NLP datasets. It is important to note that in Equation 2, the notation A L represents the matrix A raised to the power of L. The essence of GAINER becomes evident when we examine Equation 2 on a per-node basis, replacing L with r(v, τ ) for each node v ∈ V in the graph: Equation 3 employs the notation [M] v to represent the specific row of matrix M indexed by v. The model that emerges from this approach is referred to as Simple-GAINER, abbreviated as SGR. In clear contexts, s v is used to represent the particular row indexed by the vertex v in the matrix A r(v,τ ) X (0) , indicated as We are driven by the intuition of assigning a small value of r(v, τ ) to well-connected nodes, while providing poorly-connected nodes a larger value, thereby extracting the maximum value from the graph structure G. Additionally, we aim for the final smoothed features, s v , of each node to remain close to the original input features of the node to prevent excessive smoothing. The threshold τ is selected with the precise intention of ensuring that ||s vx v || 2 does not exceed τ , where || • || 2 represents the l 2 norm. The value r(v, τ ) is chosen so that The threshold τ , acting as a task-specific hyperparameter, empowers us to meticulously tailor the level of smoothing to meet the task's requirements. A central query we set out to investigate was whether we could formulate a neural counterpart of SGR, taking into account that GCN acts as the neural counterpart of SGC. One significant obstacle in this formulation is determining how to incorporate layer-specific weight matrices in Equation 1 when nodes possess highly varying radii. Nevertheless, by sharing the same weight matrix, say W, across all nodes and their radii, we introduce a novel GNN architecture known as Neural-GAINER, abbreviated as NGR. The process of NGR on a per-node basis is as follows: (5) The unrolling of Equation 5 allows the GNN to handle information at different radii, similar to the flexibility of recurrent neural networks (RNNs) which enables them to handle variable-size inputs. Unlike an RNN, our NGR aggregates node features from l hops away at every layer l, a unique characteristic that distinguishes the two architectures. Let R be the maximum radius r(v, τ ) of all the nodes v ∈ V of the input graph G = (V, E) and m = |E| be the number of edges in G. The time complexity of the key step of GAINER, i.e., computing node-specific radii is O(Rmd) where d is the number of input features. The time complexity of training and inference of SGR is O(nd 2 ) time where n = |V | is the number of nodes in G and those of NGR is O(Rnd 2 ). In this section, we empirically validate GAINER's efficacy by conducting extensive experiments including baseline comparison, training timeaccuracy tradeoff, memory consumption, sensitivity analyses, etc. The accuracy comparisons are shown in the main text while the other experiments are in the appendix. The tasks considered are 1. Inductive short-text classification, 2. Document classification on attributed graphs, 3. Document coherence assessment. We have utilised an NVIDIA Titan RTX GPU for training all the models. The training specifics are described in the appropriate subsection dedicated to the given task. Additional details regarding graph construction procedures, datasets, baselines, hyperparameters, and more are given in the appendix following the references. Short text classification (STC) is a crucial task that has been extensively studied in various NLP applications, including news tagging, efficient information retrieval, sentiment analysis, and query intent classification. In recent times, GNNs have demonstrated remarkable performance in STC by effectively exploiting relevant relational side information through message passing along edges. Recent observations Our approach closely follows the setup of a recent study The training corpus is used to construct a graph, which is then employed to train SGR and NGR. During the evaluation phase, new and unseen documents are introduced into the graph, and the model weights are employed to predict the coherence levels of these documents (inductive setting). A crossentropy loss function is used to train the models Figure Table Significance: The p-value of a Welch's t-test comparing the accuracy of our proposed models with the accuracy of the most competitive baselines in Tables Based on the table, it is clear that our proposed SGR and NGR methods excel in utilising the word graph structure to its potential, surpassing GNN-based methods with hop size fixed across all nodes. A Text-attributed Graph (TAG) represents a graph structure where nodes correspond to documents, ci-tations between documents serve as edges, and textual attributes such as title and abstract are used to build node features The node features consist of three distinct components: (i) a fine-tuned language model representation of the text sequence (title and abstract), (ii) a fine-tuned language model representation of the explanation generated by a large language model (LLM), such as ChatGPT, and (iii) the highestranked predictions of the document class provided by the LLM The concept of textual coherence involves creating a sense of flow and logical progression between sentences, ensuring they are not disjointed or randomly ordered, but instead well-connected and organised We delve into the fundamental aspect of our proposed methods: the node-specific radius, which serves as the distinguishing feature, enabling them to outperform existing approaches across tasks. In Figure Figure We have introduced GAINER, a novel graph-based learning framework that assigns a dedicated radius to each node, controlling information propagation depth. We propose Simple-GAINER and Neural-GAINER for graph NLP to harness the power of graph structures to advance graph NLP research. Extensive experiments on short text classification, document classification, and coherence assessment demonstrate the significance of GAINER. Our work lays the foundation for various potential extensions and future enhancements. More Challenging Structures: Our GAINER approach leverages the principle of homophily, which suggests that with similar labels tend to be connected in the graph, a characteristic commonly observed in our target tasks and datasets. In the heterophilic setting Multiple Modalities: In the context of expanding the scope of our work, there are several promising directions to explore. Firstly, considering multimodal or multi-graph settings could provide a richer representation of the data by incorporating diverse sources of information such as text, images, or knowledge graphs. This would enable us to capture more comprehensive relationships and dependencies within the data. Additionally, incorporating external knowledge sources, such as ontologies or domain-specific knowledge bases, could enhance the model's understanding and improve its performance on specific tasks. Transferability: Investigating the transferability of our methods across different domains or tasks would be valuable, as it could reveal the generalisability of our approaches and potentially enable knowledge transfer from one domain to another. Transferring the ideas of GAINER to more advancing models such as graph attention (PMI), whereas the second graph with short texts as nodes is learned during the training process through the construction of edges based on cosine similarity of trained embeddings. Notably, our focus in this work is primarily on leveraging our proposed GAINER techniques specifically for the word-level graph. However, an intriguing avenue for future investigation involves extending GAINER to hierarchical graph learning and/or incorporate edge learning within the short text graph, which holds potential for further advancements in this domain. We compare our SGR and NGR methods with: • Traditional two-step feature extraction and classification methods including TF-IDF+SVM, LDA+SVM • Pretrained BERT • Inductive GNN based text classification methods including TLGNN • Inductive STC Methods including HGAT-Inductive The sliding window size for caclulating PMI is 5 and the word embedding size is 200. We use the Adam optimiser with a learning rate of 0.001 to train for a maximum of 1000 epochs. The dropout rate is 0.9. The threshold for GAINER is selected based on grid search in the range τ ∈ {0.05, 0.075, 0.1, 0.125, 0.15}. training dataset size is extremely limited. We attribute this to the enhanced information propagation capabilities of SGR, NGR, allowing them to leverage the rich graph structure more efficiently, especially in scenarios with low supervision. The effect of varying embedding sizes on NGR performance is depicted in Figure Table • APPNP • AP-GCN • DAGNN • PPRGo Linear methods, in the context of graph machine learning, pertain to a category of approaches where the feature propagation over the graph follows a linear function of specific graph structural elements, such as the graph Laplacian, the adjacency matrix. • SGC • SIGN (Frasca et al., 2020) SIGN is a highly efficient and scalable graph embedding method that offers an alternative to graph sampling in GCN. It utilises various local graph operators tailored to different tasks. • S 2 GC (Zhu and Koniusz, 2021) introduces a modified Markov Diffusion Kernel to create a variant of GCN that balances low-pass and high-pass filtering. This unique approach enables the capturing of both global and local contexts for each node. The node embedding size of NGR is selected based on grid search in the range {32, 64, 128, 256}. We use the Adam optimiser with a learning rate of 0.001 to train for a maximum of 1000 epochs. The dropout rate is 0.5. The threshold for GAINER is selected in the range τ ∈ {0.05, 0.075, 0.1, 0.125, 0.15}. In this section, we delve into the specifics of the document coherence assessment task. In particular, we provide a detailed account of the dataset utilised, the experimental setup employed in our study, com-prehensive descriptions of the baseline methods employed, and an overview of the hyperparameters chosen. Our study utilizes the Grammarly Corpus of Discourse Coherence (GCDC) dataset We closely follow the setup of a recent study Our approach to representing a document as a directed sentence graph builds upon a prior work To process and segment a document, we employ the Stanza toolkit In this section, we focus on representing sentences through a subgraph set, allowing us to compare graph structures efficiently and enables document comparison based on structure. A subgraph of a graph is such that the nodes in it can be mapped to the nodes in the graph with the same connection relations. We only consider subgraphs without backward edges, as our approach processes documents from left to right. We use weakly connected and disconnected subgraphs, as they reflect document properties related to coherence. Given a sentence graph, we mine contained k-node subgraphs, filter out distant subgraphs, count their frequency, and identify isomorphic subgraphs to represent the sentence graph as a subgraph set. The aformentioned approach is inspired by a prior study In this section, we introduce the concept of the doc-subgraph graph, which is an undirected graph constructed at the corpus level. It connects structurally similar documents through their shared subgraphs. The graph consists of document nodes and subgraph nodes, with the total number of nodes being the sum of the number of documents and the number of distinct k-node subgraphs mined from the documents. Two types of edges are defined in the graph: edges between documents and subgraphs, and edges between subgraphs. The first type of edge is determined based on the presence of a subgraph in a document's subgraph set, with the edge weight being a combination of the subgraph's frequency in the set and its inverse document frequency. The second type of edge is constructed between subgraphs that co-occur in the same document's subgraph set, and its weight is calculated using the Pointwise Mutual Information (PMI) measure. The resulting doc-subgraph graph captures the structural relationships between documents and subgraphs, providing a comprehensive representation of the corpus. We apply GAINER methods, viz., SGR and NGR on the aformentioned docsubgraph-graph.
| 1,259 | 1,423 | 1,259 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.