elishowk's picture
Automatic correction of README.md metadata for keys. Contact [email protected] for any question
7377e74
metadata
language:
  - multilingual
  - pt
  - en
tags:
  - xlm-roberta-large
  - semantic role labeling
  - finetuned
  - dependency parsing
license: apache-2.0
datasets:
  - PropBank.Br
  - CoNLL-2012
  - Universal Dependencies
metrics: f1

XLM-R large fine-tuned in Portuguese Universal Dependencies and English and Portuguese semantic role labeling

Model description

This model is the xlm-roberta-large fine-tuned first on the Universal Dependencies Portuguese dataset, then fine-tuned on the CoNLL formatted OntoNotes v5.0 and then fine-tuned on the PropBank.Br data. This is part of a project from which resulted the following models:

For more information, please see the accompanying article (See BibTeX entry and citation info below) and the project's github.

Intended uses & limitations

How to use

To use the transformers portion of this model:

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("liaad/ud_srl-enpt_xlmr-large")
model = AutoModel.from_pretrained("liaad/ud_srl-enpt_xlmr-large")

To use the full SRL model (transformers portion + a decoding layer), refer to the project's github.

Limitations and bias

  • This model does not include a Tensorflow version. This is because the "type_vocab_size" in this model was changed (from 1 to 2) and, therefore, it cannot be easily converted to Tensorflow.
  • The model was trained only for 10 epochs in the Universal Dependencies dataset.
  • The model was trained only for 5 epochs in the CoNLL formatted OntoNotes v5.0.
  • The English data was preprocessed to match the Portuguese data, so there are some differences in role attributions and some roles were removed from the data.

Training procedure

The model was trained on the Universal Dependencies Portuguese dataset; then on the CoNLL formatted OntoNotes v5.0; then on Portuguese semantic role labeling data (PropBank.Br) using 10-fold Cross-Validation. The 10 resulting models were tested on the folds as well as on a smaller opinion dataset "Buscapé". For more information, please see the accompanying article (See BibTeX entry and citation info below) and the project's github.

Eval results

Model Name F1 CV PropBank.Br (in domain) F1 Buscapé (out of domain)
srl-pt_bertimbau-base 76.30 73.33
srl-pt_bertimbau-large 77.42 74.85
srl-pt_xlmr-base 75.22 72.82
srl-pt_xlmr-large 77.59 73.84
srl-pt_mbert-base 72.76 66.89
srl-en_xlmr-base 66.59 65.24
srl-en_xlmr-large 67.60 64.94
srl-en_mbert-base 63.07 58.56
srl-enpt_xlmr-base 76.50 73.74
srl-enpt_xlmr-large 78.22 74.55
srl-enpt_mbert-base 74.88 69.19
ud_srl-pt_bertimbau-large 77.53 74.49
ud_srl-pt_xlmr-large 77.69 74.91
ud_srl-enpt_xlmr-large 77.97 75.05

BibTeX entry and citation info

@misc{oliveira2021transformers,
      title={Transformers and Transfer Learning for Improving Portuguese Semantic Role Labeling}, 
      author={Sofia Oliveira and Daniel Loureiro and Alípio Jorge},
      year={2021},
      eprint={2101.01213},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}