Skip to main content

...

Project description

HONEST: Measuring Hurtful Sentence Completion in Language Models

https://img.shields.io/pypi/v/honest.svg https://img.shields.io/travis/MilaNLProc/honest.svg Documentation Status Medium Blog Post

Large language models (LLMs) have revolutionized the field of NLP. However, LLMs capture and proliferate hurtful stereotypes, especially in text generation. We propose HONEST, a score to measure hurtful sentence completions in language models. It uses a systematic template- and lexicon-based bias evaluation methodology in six languages (English, Italian, French, Portuguese, Romanian, and Spanish) for binary gender and in English for LGBTQAI+ individuals.

See the papers for additional details:

Nozza D., Bianchi F., and Hovy D. “HONEST: Measuring hurtful sentence completion in language models.” The 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, 2021. https://aclanthology.org/2021.naacl-main.191

Nozza D., Bianchi F., Lauscher L., and Hovy D. “Measuring Harmful Sentence Completion in Language Models for LGBTQIA+ Individuals.” The Second Workshop on Language Technology for Equality, Diversity and Inclusion at the Annual Meeting of the Association for Computational Linguistics 2022. https://milanlproc.github.io/publication/2022-honest-hurtful-language-model-lgbtqia+/

Tutorials

Name

Link

Compute HONEST score (+Viz) (stable v0.2.0)

Open In Colab

Installing

pip install -U honest

Using

# Load HONEST templates
evaluator = honest.HonestEvaluator(lang)
masked_templates = evaluator.templates(data_set="binary") # or "queer_nonqueer" or "all"

# Load BERT model
tokenizer = AutoTokenizer.from_pretrained(name_model)
model = AutoModelForMaskedLM.from_pretrained(name_model)

# Define nlp_fill pipeline
nlp_fill = pipeline('fill-mask', model=model, tokenizer=tokenizer, top_k=k)

print("FILL EXAMPLE:",nlp_fill('all women likes to [M].'.replace('[M]',tokenizer.mask_token)))

# Fill templates (please check if the filled words contain any special character)
filled_templates = [[fill['token_str'].strip() for fill in nlp_fill(masked_sentence.replace('[M]',tokenizer.mask_token))] for masked_sentence in masked_templates.keys()]

honest_score = evaluator.honest(filled_templates)
print(name_model, k, honest_score)

Citation

Please use the following bibtex entries if you use this score in your project:

@inproceedings{nozza-etal-2021-honest,
title = {"{HONEST}: Measuring Hurtful Sentence Completion in Language Models"},
author = "Nozza, Debora and Bianchi, Federico  and Hovy, Dirk",
booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
month = jun,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.naacl-main.191",
doi = "10.18653/v1/2021.naacl-main.191",
pages = "2398--2406",
}

@inproceedings{nozza-etal-2022-measuring,
    title = {Measuring Harmful Sentence Completion in Language Models for LGBTQIA+ Individuals},
    author = "Nozza, Debora and Bianchi, Federico and Lauscher, Anne and Hovy, Dirk",
    booktitle = "Proceedings of the Second Workshop on Language Technology for Equality, Diversity and Inclusion",
    publisher = "Association for Computational Linguistics",
    year={2022}
}

Development Team

Software Details

Credits

This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template.

Note

Remember that this is a research tool :)

History

0.1.0 (2022-01-25)

  • First release on PyPI.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

honest-0.2.1.tar.gz (8.0 kB view hashes)

Uploaded Source

Built Distribution

honest-0.2.1-py2.py3-none-any.whl (6.2 kB view hashes)

Uploaded Python 2 Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page