Elamur Research⌕
☰
home / papers / 2022 / Roadblocks in Gender Bias Measurement for Diachronic Corpora

Roadblocks in Gender Bias Measurement for Diachronic Corpora

Summary

The paper examines whether a widely used way of measuring gender bias in word embeddings still works when the corpus covers many centuries. The method comes from Bolukbasi et al. (2016) and was extended to nine languages by Chen et al. (2021). It derives a "gender direction" from a fixed defining set of gendered word pairs, then scores a fixed list of profession words against that direction. The authors apply this setup to diachronic Arabic and Chinese material:

  • the Shamela Library of Arabic books;
  • a collection of Arabic poetry covering 11 distinct eras;
  • Google Books Ngram data;
  • a small collection of Chinese classics.

The method depends on stable word lists, and those lists turn out to drift. Many profession words are absent from older texts. Even basic gendered words change in frequency or disappear. Arabic female profession forms often carry unrelated meanings as well. The authors conclude that both word sets would need to change over time, so a lexicon-based bias score cannot be taken unchanged from modern corpora such as Google News or Wikipedia.

Key points

  • Baseline method. Bolukbasi et al. use 10 gendered pairs (e.g., she/he, mother/father) and 327 profession words. They take the first PCA component of the pair differences as the g direction and measure bias as the cosine with that direction (DirectBias).
  • Chen et al.'s changes. For translatability, they swapped hard-to-translate pairs (she/he, Mary/John) for pairs such as queen/king and madam/sir. For grammatically gendered nouns, they weight each male, female or neutral variant's score by its occurrence count.
  • Shamela Library size. It contains 6,527 books and about 703M tokens, with a vocabulary of about 2.5M. Of these books, 3 are pre-Islamic, 2,820 date before 1900, 773 date after 1900, and 2,931 are of unknown date.
  • Chinese evidence. Ngram curves for 1500–2019 show that 女子, 女人 and 妇女 compete as renderings of "woman", with 女子 fading in modern writing. In classics dating from 475 BC to 1992, roughly half of Chen et al.'s profession words do not occur, and neither do the defining words "boy" and "madam".
  • Arabic polysemy. In the Arabic poems, the female form for "bartender" can also mean a water creek or aqueduct, which confounds frequency-based scoring.

Interesting ideas

  • I1. Before trusting a lexicon-driven metric on a new corpus, first check that every seed word actually occurs there, because missing anchors silently distort the score.
  • I2. A concept like "woman" maps to several competing word forms whose popularity shifts over time, so the anchor should be a set of variants rather than a single word.
  • I3. Gendered or derived word forms can collide with unrelated senses (bartender vs. aqueduct), so counting surface forms without sense resolution gives misleading statistics.
  • I4. Chen et al. combine morphological variants by weighting each variant's score by its frequency, instead of choosing one canonical form.
  • I5. Fixed word lists encode the assumptions of one era, so any compiled list should carry an explicit note on the period or register it is valid for.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Evaluation suites and negative controls could include a coverage report showing which probe words and frames appear in the target text type, such as old-fashioned fables, before scores are compared.
  • I2 (tokens): Lemma normalization could map archaic or rare variants found in tales onto one concept key, so that downstream steps do not lose them.
  • I3 (world): Knowledge entries for professions and roles should be sense-tagged via WordNet, and an entity should be typed as a person only when the frame context supports that reading.
  • I4 (world): Statistics on roles or holdings could be pooled across gendered or inflected variants in the same frequency-weighted way, while each per-variant count stays inspectable for explanation.
  • I5 (world): Verb-class and ontology lists could record their intended register, letting the system flag a text whose vocabulary falls outside that range.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Topics Bias, fairness & safety Historical linguistics & language change

 Institutions

Also on the byline: Clarkson University.

Cite

Cite as Saied Alshahrani, Esma Wali, Abdullah R Alshamsan, Yan Chen and Jeanna Matthews. 2022. Roadblocks in Gender Bias Measurement for Diachronic Corpora. Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2022.lchange-1.15

@inproceedings{alshahrani2022roadblocks,
  title = {{Roadblocks in Gender Bias Measurement for Diachronic Corpora}},
  author = {Saied Alshahrani, Esma Wali, Abdullah R Alshamsan, Yan Chen and Jeanna Matthews},
  booktitle = {Proceedings of the 3rd Workshop on Computational Approaches to Historical Language Change},
  year = {2022},
  doi = {10.18653/v1/2022.lchange-1.15},
  url = {https://doi.org/10.18653/v1/2022.lchange-1.15}
}

References

24 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pages 298–306. our card
  2. Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? New York, NY, USA. Association for Computing Machinery. find
  3. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29. our card
  4. Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183–186. find
  5. Yan Chen, Christopher Mahoney, Isabella Grasso, Esma Wali, Abigail Matthews, Thomas Middleton, Mariama Njie, and Jeanna Matthews. 2021. Gender bias and under-representation in natural language processing across human languages. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pages 24–34. find
  6. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781. ↗ our card
  7. Hamdy Mubarak. 2017. Build fast and accurate lemmatization for arabic. arXiv preprint arXiv:1710.06700. ↗ our card
  8. Marwa Naili, Anja Habacha Chaibi, and Henda Hajjami Ben Ghezala. 2017. Comparative study of word embedding methods in topic segmentation. Procedia Computer Science, 112:340–349. Knowledge-Based and Intelligent Information Engineering Systems: Proceedings of the 21st International Conference, KES-20176-8 September 2017, Marseille, France. find
  9. Ossama Obeid, Nasser Zalmout, Salam Khalifa, Dima Taji, Mai Oudah, Bashar Alhafni, Go Inoue, Fadhl Eryani, Alexander Erdmann, and Nizar Habash. 2020. Camel tools: An open source python toolkit for arabic natural language processing. In Proceedings of the 12th language resources and evaluation conference, pages 7022–7032. our card
  10. Abu Dhabi Department of Culture and Tourism. 2016. Poetry encyclopedia. https://poetry.dctabudhabi.ae. ↗
  11. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. ↗ our card
  12. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. ↗ our card
  13. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations." arxiv preprint. arXiv preprint arXiv:1802.05365. ↗ our card
  14. Diwan. 2013. Poetry dataset. https://www.aldiwan.net/. ↗
  15. Maja Rudolph and David Blei. 2018. Dynamic embeddings for language evolution. In Proceedings of the 2018 World Wide Web Conference, pages 1003–1011. find
  16. Shamela Library Foundation. 2012. Shamila library dataset. https://shamela.ws/page/download. ↗
  17. Michael Grosvald, Sarah Al-Alami, and Ali Idrissi. 2019. Word reading in arabic: Influences of diacritics and ambiguity. In 36th West Coast Conference on Formal Linguistics, pages 176–181. Cascadilla Proceedings Project. find
  18. Nizar Habash, Fadhl Eryani, Salam Khalifa, Owen Rambow, Dana Abdulrahim, Alexander Erdmann, Reem Faraj, Wajdi Zaghouani, Houda Bouamor, Nasser Zalmout, et al. 2018. Unified guidelines and resources for arabic dialect orthography. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). our card
  19. Youssef Kadri and Jian-Yun Nie. 2006. Effective stemming for arabic information retrieval. In Proceedings of the International Conference on the Challenge of Arabic for NLP/MT. find
  20. Marzieh Karch. 2021. How to use the ngram viewer tool in google books. In https://www.lifewire.com/google-books-ngram-viewer-1616701. ↗
  21. Esma Wali, Yan Chen, Christopher Mahoney, Thomas Middleton, Marzieh Babaeianjelodar, Mariama Njie, and Jeanna Neefe Matthews. 2020. Is machine learning speaking my language? a critical look at the nlp-pipeline across 8 human languages. arXiv preprint arXiv:2007.05872. ↗
  22. Melvin Wevers. 2019. Using word embeddings to examine gender bias in dutch newspapers, 1950-1990. arXiv preprint arXiv:1907.08922. ↗
  23. Yang Xu, Jiasheng Zhang, and David Reitter. 2019. Treat the word as a whole or look inside? subword embeddings model language change and typology. In Proceedings of the 1st International Workshop on Computational Approaches to Historical Language Change, pages 136–145. find
  24. Waleed A. Yousef, Omar M. Ibrahime, Taha M. Mad-bouly, Moustafa A. Mahmoud, Ali H. El-Kassas, Ali O. Hassan, and Abdallah R. Albohy. 2018. Arabic poem comprehensive dataset. https://hci-lab.github.io/ArabicPoetry-1-Private/PCD. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alshahrani-2022-roadblocks-gender-bias-measurement-diachronic/