Elamur Research⌕
☰
home / papers / 2017 / Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English

Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English

Paper · 2017 · Su Lin Blodgett and Brendan O'Connor

Summary

The paper argues that algorithmic fairness should cover more than decision-making systems. It should also cover how accurately language technology handles text from different social groups. The authors study language identification, the gate that decides which messages reach the rest of an NLP pipeline. They test it on Twitter text tied to African-American English (AAE).

The data comes from their earlier geolocated Twitter corpus. A mixed-membership topic model links words to U.S. Census race/ethnicity statistics for the blockgroup where each tweet was sent. Messages whose tokens mostly come from the African-American language model are labelled AA-aligned, and white-aligned messages are chosen the same way. The authors then sample equal numbers of tweets per length bin so that message length cannot explain the results. They run four off-the-shelf identifiers on this sample.

In almost every condition, AA-aligned tweets are recognized as English less often than white-aligned ones. The gap is widest for very short messages. This matters because dropped messages silently remove a community's voice from any downstream analysis, such as opinion measurement.

Key points

  • The source corpus has 59.2 million public U.S. tweets from 2013. Messages count as "aligned" when more than 0.8 of their tokens are attributed to that group's language model.
  • The test set has four length bins with 2,500 AA-aligned and 2,500 white-aligned tweets each, giving 20,000 messages in total. The set is publicly released.
  • The classifiers tested were langid.py, IBM Watson (62 languages), Microsoft Azure (120 languages), and Twitter's own metadata label. Google's service was dropped because every request failed.
  • For tweets of at most 5 tokens, the English-recognition gap reached 19.7 points for Twitter's label and 15.1 for IBM Watson. About 41.7% of AA-aligned messages in the corpus fall in that short range.
  • langid.py tagged two sample AAE tweets as Danish with over 99.9% confidence. These tweets contained a null copula, "af", "den" for "then", and the future auxiliary "finna".

Interesting ideas

  • I1. Bin the test items by length and sample the same number per group in each bin, so that a group difference cannot be explained by one group simply writing shorter texts. This is a cheap way to remove a known confound before claiming a disparity.
  • I2. Weak dialect labels can be obtained without annotators. Pair a topic-style model with an external prior, here census demographics, and let each message carry its own mixture so that one author can switch varieties.
  • I3. Audit the very first gate in a pipeline, not just the final task. An early filter that discards input fails silently, and every later component inherits the loss.
  • I4. Probe systems with a few hand-picked examples that carry known dialect features, such as a missing copula or a dialect-specific tense auxiliary. These examples show why a system fails, not just how often.
  • I5. Leave the noisy surface features of social media text (mentions, emoji, hashtags) in place during evaluation, so the results reflect real input rather than a cleaned-up version.

For Mova

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

  • I1 (learn): Evaluation reports could stratify sentences by length and compare per-bin scores across genres or dialects, so that a gap is not just a length effect.
  • I2 (learn): A similar soft-label scheme could flag dialectal or informal sentences in raw text, which could then go to the witness parsers or LLM teacher for targeted silver data.
  • I3 (tokens): Any English/non-English or "is this parseable" check before the tokenizer should be tested on dialectal and very short inputs, logging what gets rejected rather than dropping it quietly.
  • I4 (tree): A small probe set with null copula and auxiliaries like "finna" could check whether the UD parser and dialect conversion still produce a sensible predicate and aux structure.
  • I5 (tokens): Tokenizer and tagger tests could include uncleaned social-media strings, so that robustness to informal orthography such as "den" or "af" is measured rather than assumed.

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

 Institutions

🎓University of Massachusetts AmherstAmherst Center · United States67 works · 64,591 citations

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 worksSocial Sciences · 744 worksSafety Research · 221 works

🧭 Topics 1Research subjects and languages the work is about.English · 1,373 works

Cite

Cite as Su Lin Blodgett and Brendan O'Connor. 2017. Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English. arXiv:1707.00061.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1707.00061 · arXiv 1707.00061

@misc{blodgett2017racial,
  title = {{Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English}},
  author = {Su Lin Blodgett and Brendan O'Connor},
  year = {2017},
  doi = {10.48550/arxiv.1707.00061},
  eprint = {1707.00061},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1707.00061}
}

References

20 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Timothy Baldwin, Paul Cook, Marco Lui, Andrew MacKinlay, and Li Wang. 2013. How Noisy Social Media Text, How Diffrnt Social Media Sources?. In International Joint Conference on Natural Language Processing. 356–364.
  2. Timothy Baldwin and Marco Lui. 2010. Language identification: The long and the short of the matter. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics. Association for Computational Linguistics, 229–237. find
  3. Su Lin Blodgett, Lisa Green, and Brendan O’Connor. 2016. Demographic Dialectal Variation in Social Media: A Case Study of African-American English. Proceedings of EMNLP (2016).
  4. Michael Feldman, Sorelle A Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkatasubramanian. 2015. Certifying and removing disparate impact. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 259–268. ↗
  5. Sharad Goel, Maya Perelman, Ravi Shroff, and David Alan Sklansky. 2017. Combatting police discrimination in the age of big data. New Criminal Law Review: In International and Interdisciplinary Journal 20, 2 (2017), 181–232. ↗
  6. Lisa J. Green. 2002. African American English: A Linguistic Introduction. Cambridge University Press. ↗
  7. Dirk Hovy and L. Shannon Spruit. 2016. The Social Impact of Natural Language Processing. In Proceedings of ACL. ↗
  8. Baden Hughes, Timothy Baldwin, Steven Bird, Jeremy Nicholson, and Andrew MacKinlay. 2006. Reconsidering Language Identification for Written Language Resources. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06). European Language Resources Association (ELRA). http://aclweb.org/anthology/L06-1274 ↗
  9. Taylor Jones. 2015. Toward a Description of African American Vernacular English Dialect Regions Using “Black Twitter”. American Speech 90, 4 (2015), 403–440. 12 https://newsroom.fb.com/news/2016/07/facebook-diversity-update-positive-hiring-trends-show-progress/https://www.google.com/diversity/13 https://www.census.gov/quickfacts/table/RHI225215/00 ↗
  10. Anna Jørgensen, Dirk Hovy, and Anders Søgaard. 2016. Learning a POS tagger for AAVE-like language. In Proceedings of NAACL. Association for Computational Linguistics.
  11. Anna Katrine Jørgensen, Dirk Hovy, and Anders Søgaard. 2015. Challenges of studying and processing dialects in social media. In Proceedings of the Workshop on Noisy User-generated Text. 9–18. find
  12. Matthew Joseph, Michael Kearns, Jamie Morgenstern, Seth Neel, and Aaron Roth. 2016. Rawlsian fairness for machine learning. arXiv preprint arXiv:1610.09559 (2016). ↗
  13. Matthew Joseph, Michael Kearns, Jamie H Morgenstern, and Aaron Roth. 2016. Fairness in Learning: Classic and Contextual Bandits. In Advances in Neural Information Processing Systems 29, D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (Eds.). Curran Associates, Inc., 325–333. http://papers.nips.cc/paper/6355-fairness-in-learning-classic-and-contextual-bandits.pdf ↗
  14. William Labov. 1972. Language in the inner city: Studies in the Black English vernacular. Vol. 3. University of Pennsylvania Press. find
  15. Maider Lehr, Kyle Gorman, and Izhak Shafran. 2014. Discriminative pronunciation modeling for dialectal speech recognition. In Proc. Interspeech. find
  16. M. Lui and T. Baldwin. 2012. langid. py: An Off-the-shelf Language Identification Tool. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (ACL 2012), Demo Session, Jeju, Republic of Korea. http://www.aclweb.org/anthology-new/P/P12/P12-3005.pdf ↗
  17. Fred Morstatter, Jrgen Pfeffer, Huan Liu, and Kathleen Carley. 2013. Is the Sample Good Enough? Comparing Data from Twitter’s Streaming API with Twitter’s Firehose. In International AAAI Conference on Weblogs and Social Media. http://www.aaai.org/ocs/index.php/ICWSM/ICWSM13/paper/view/6071 ↗
  18. John Russell Rickford. 1999. African American vernacular English: Features, evolution, educational implications. Wiley-Blackwell. find
  19. Ian Stewart. 2014. Now We Stronger than Ever: African-American English Syntax in Twitter. In Proceedings of the Student Research Workshop at the 14th Conference of the European Chapter of the Association for Computational Linguistics. Association for Computational Linguistics, Gothenburg, Sweden, 31–37. http://www.aclweb.org/anthology/E14-3004 ↗
  20. Rachael Tatman. 2017. Gender and Dialect Bias in YouTube’s Automatic Captions. In Proceedings of the First ACL Workshop on Ethics in Natural Language Processing. Association for Computational Linguistics, Valencia, Spain, 53–59. http://www.aclweb.org/anthology/W/W17/W17-1606 ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/blodgett-2017-racial-disparity-natural-language-processing/