poltextlab/xlm-roberta-large-pooled-cap-media1-v2

xlm-roberta-large-pooled-cap-media-v2

Model description

An xlm-roberta-large model finetuned on multilingual (english, german, hungarian, spanish, slovakian) training data labelled with major topic codes from the Comparative Agendas Project. Furthermore we used 7 additional media codes, following Boydstun (2013):

• State and Local Government Administration (24)
• Weather and Natural Disaster (26)
• Fires(27)
• Sports and Recreation (29)
• Death Notices (30)
• Churches and Religion (31)
• Other, Miscellaneous and Human Interest (99)

How to use the model

from transformers import AutoTokenizer, pipeline

tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-large")
pipe = pipeline(
    model="poltextlab/xlm-roberta-large-pooled-cap-media1-v2",
    task="text-classification",
    tokenizer=tokenizer,
    use_fast=False,
    token="<your_hf_read_only_token>"
)

text = "We will place an immediate 6-month halt on the finance driven closure of beds and wards, and set up an independent audit of needs and facilities."
pipe(text)

Gated access

Due to the gated access, you must pass the token parameter when loading the model. In earlier versions of the Transformers package, you may need to use the use_auth_token parameter instead.

Overall Performance:

• Accuracy: 74%
• Macro Avg: Precision: 0.76, Recall: 0.74, F1-score: 0.73
• Weighted Avg: Precision: 0.76, Recall: 0.74, F1-score: 0.73

Per-Class Metrics:

Unnamed: 0	precision	recall	f1-score	support
1: Macroeconomics	0.773585	0.82	0.796117	50
2: Civil Rights	0.714286	0.6	0.652174	50
3: Health	0.803922	0.82	0.811881	50
4: Agriculture	0.857143	0.84	0.848485	50
5: Labor	0.666667	0.68	0.673267	50
6: Education	0.86	0.86	0.86	50
7: Environment	0.829787	0.78	0.804124	50
8: Energy	0.851852	0.92	0.884615	50
9: Immigration	0.888889	0.8	0.842105	50
10: Transportation	0.661765	0.9	0.762712	50
12: Law and Crime	0.679245	0.72	0.699029	50
13: Social Welfare	0.842105	0.64	0.727273	50
14: Housing	0.666667	0.8	0.727273	50
15: Banking, Finance, and Domestic Commerce	0.714286	0.6	0.652174	50
16: Defense	0.596154	0.62	0.607843	50
17: Technology	0.709091	0.78	0.742857	50
18: Foreign Trade	0.88	0.88	0.88	50
19: International Affairs	0.534483	0.62	0.574074	50
20: Government Operations	0.790698	0.68	0.731183	50
21: Public Lands	0.808511	0.76	0.783505	50
23: Culture	0.678571	0.76	0.716981	50
24: State and Local Government Administration	0.587302	0.74	0.654867	50
26: Weather and Natural Disasters	0.913043	0.84	0.875	50
27: Fires	0.942857	0.66	0.776471	50
29: Sports and Recreation	0.843137	0.86	0.851485	50
30: Death Notices	0.956522	0.88	0.916667	50
31: Churches and Religion	0.782609	0.72	0.75	50
99: Other, Miscellaneous, and Human Interest	0.378947	0.72	0.496552	50
998: No Policy and No Media Content	0.75	0.06	0.111111	50
accuracy	0.736552	0.736552	0.736552	0.736552
macro avg	0.757315	0.736552	0.731373	1450
weighted avg	0.757315	0.736552	0.731373	1450

Inference platform

This model is used by the CAP Babel Machine, an open-source and free natural language processing tool, designed to simplify and speed up projects for comparative research.

Cooperation

Model performance can be significantly improved by extending our training sets. We appreciate every submission of CAP-coded corpora (of any domain and language) at poltextlab{at}poltextlab{dot}com or by using the CAP Babel Machine.

Debugging and issues

This architecture uses the sentencepiece tokenizer. In order to run the model before transformers==4.27 you need to install it manually.

If you encounter a RuntimeError when loading the model using the from_pretrained() method, adding ignore_mismatched_sizes=True should solve the issue.