Elamur Research⌕
☰
home / papers / 2022 / Using Item Response Theory to Measure Gender and Racial Bias of a BERT-based Automated English Speech Assessment System

Using Item Response Theory to Measure Gender and Racial Bias of a BERT-based Automated English Speech Assessment System

Paper · 2022 · Alexander Kwako, Yixin Wan, Jieyu Zhao et al.

Summary

The paper asks whether a transformer-based scorer for spoken English treats children differently by gender or by race. The authors built a full pipeline. Google speech-to-text transcribes the audio, and a fine-tuned BERT model assigns a score of 0–2. They tested it on one ELPA21 speaking item answered by students in grades 2–3. Home language served as a finer-grained proxy for race. Instead of the usual NLP bias probes, they borrowed a tool from psychometrics: differential item functioning (DIF). DIF checks whether groups of equal underlying proficiency get systematically different scores. It was estimated with a Graded 2PL item response theory model and the Improved Wald Test, with Benjamini-Hochberg control of the false discovery rate. The automated scores were about as reliable as human ratings, and slightly more consistent. No significant DIF appeared for gender or for any home-language group. The authors stress that this is a narrow, preliminary finding. Its main value is a statistically principled audit method that NLP bias research has rarely used.

Key points

  • Data:
    • 10,147 recorded responses to one short ELPA21 item, with a median length of 4.8 s.
    • 642 blank Google transcripts were dropped, leaving 9,505 responses.
    • Home-language groups with fewer than 200 responses were excluded.
  • Transcription quality: on 100 hand-transcribed responses, the word error rate was 22.3%. This is near the human-parity range reported for non-native speech.
  • Scoring model:
    • BERT slightly beat RoBERTa, reaching 88.85% test accuracy averaged over 3 seeds.
    • Training used class-weighted cross-entropy, because 76.6% of responses had the top score.
  • Consistency:
    • Two BERT models trained with different seeds and splits reached Spearman ρ = 0.88 with each other.
    • Human raters reached ρ = 0.81 with each other, measured on 1,929 double-scored responses.
  • Bias result:
    • Reference groups were Male and Spanish. Focal groups were Female, Persian, Ukrainian, Arabic, Vietnamese, Chinese and Russian.
    • Both the discrimination (a) and difficulty (b) parameter tests stayed above the Benjamini-Hochberg critical p-values.
    • Human scores showed no DIF either.

Interesting ideas

  • I1. Bias is defined as a score gap between groups after conditioning on latent ability, which separates genuine group differences from unfair treatment by the scorer.
  • I2. Splitting the item model into a discrimination parameter (reliability) and difficulty parameters (harshness) lets one test say how a system is unfair, not just whether it is.
  • I3. Many group comparisons are tested together under Benjamini-Hochberg false-discovery control, which avoids spurious "findings" from repeated testing.
  • I4. The model's own seed-to-seed agreement is set beside the inter-annotator agreement of humans, which gives a reliability baseline without any gold labels.
  • I5. An upstream black-box component (ASR) is audited by hand-transcribing a small random sample and measuring its error rate before trusting downstream results.

For Mova

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

  • I1 (learn): when evaluating parser or reader accuracy across text genres or dialects, control for sentence difficulty first, so a gap reflects the model's behavior rather than harder input.
  • I2 (answer): fitting a 2PL-style model to school-course test items could show whether the readers fail on certain question types because those types are harder or because the readers are inconsistent on them.
  • I3 (learn): apply the same correction when comparing many per-relation or per-language LAS deltas or negative-control results, so that chance fluctuations aren't reported as regressions.
  • I4 (learn): compare the agreement between witness parsers on silver data with human inter-annotator agreement to judge how much trust silver trees deserve.
  • I5 (tree): spot-check a hundred sentences of converted or silver treebanks manually to estimate the noise that feeds into frames and the world model.

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 Maryland, College ParkCollege Park · United States141 works · 29,174 citations🎓University of California, Los AngelesLos Angeles · United States99 works · 56,330 citations

 Fields, topics and concepts

📚 Shelves 8The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksSafety Research · 626 worksDecision Sciences · 249 worksManagement Science and Operations Research · 134 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,379 worksBias, fairness & safety · 587 works

Cite

Cite as Alexander Kwako, Yixin Wan, Jieyu Zhao, Kai-Wei Chang, Li Cai and Mark Hansen. 2022. Using Item Response Theory to Measure Gender and Racial Bias of a BERT-based Automated English Speech Assessment System. Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2022.bea-1.1

@inproceedings{kwako2022using,
  title = {{Using Item Response Theory to Measure Gender and Racial Bias of a BERT-based Automated English Speech Assessment System}},
  author = {Alexander Kwako, Yixin Wan, Jieyu Zhao, Kai-Wei Chang, Li Cai and Mark Hansen},
  booktitle = {Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022)},
  year = {2022},
  doi = {10.18653/v1/2022.bea-1.1},
  url = {https://doi.org/10.18653/v1/2022.bea-1.1}
}

References

26 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sunipa Dev, Tao Li, Jeff M. Phillips, and Vivek Srikumar. 2020. On measuring and mitigating biased inferences of word embeddings. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):7659– 7666. ↗
  2. Lei Chen, Klaus Zechner, Su-Youn Yoon, Keelan Evanini, Xinhao Wang, Anastassia Loukina, Jidong Tao, Lawrence Davis, Chong Min Lee, Min Ma, Robert Mundkowsky, Chi Lu, Chee Wee Leong, and Binod Gyawali. 2018. Automated scoring of nonnative speech using the¬†speechratersm v. 5.0 engine. ETS Research Report Series, 2018(1):1–31. find
  3. Jo-Kate Collier and Becky Huang. 2020. Test Review: Texas English Language Proficiency Assessment System (TELPAS). Language Assessment Quarterly, 17(2):221–230. Publisher: Taylor & Francis. ↗
  4. William H. Angoff. 1993. Perspectives on Differential Item Functioning Methodology. In Paul W. Holland and Howard Wainer, editors, Differential Item Functioning, pages 3–23. Lawrence Erlbaum Associates, Inc., Hillsdale, NJ. find
  5. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.
  6. Yoav Benjamini and Yosef Hochberg. 1995. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal statistical society: series B (Methodological), 57(1):289–300. Publisher: Wiley Online Library. ↗
  7. Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of" bias" in nlp. arXiv preprint arXiv:2005.14050. ↗
  8. Educational Testing Service. 2005. Test and Score Data Summary: 2004-05 Test Year Data Test of English as a Foreign Language. Technical report. find
  9. Michael C. Edwards and Maria Orlando Edelen. 2009. Special topics in item response theory. The SAGE handbook of quantitative methods in psychology, pages 178–198. Publisher: Sage Publications New York, NY. ↗
  10. L. Cai. 2012. flexMIRT: Flexible multilevel item factor analysis and test scoring [Computer software]. Seattle, WA: Vector Psychometric Group, LLC. find
  11. Li Cai, Kilchan Choi, Mark Hansen, and Lauren Harrell. 2016. Item response theory. Annual Review of Statistics and Its Application, 3:297–321. Publisher: Annual Reviews. ↗
  12. Keelan Evanini, Maurice Cogan Hauck, and Kenji Hakuta. 2017. Approaches to automated scoring of speaking for k–12 english language proficiency assessments. ETS Research Report Series, 2017(1):1Р11. find
  13. David M. Williamson, Xiaoming Xi, and F. Jay Breyer. 2012. A framework for evaluation and use of automated scoring. Educational measurement: issues and practice, 31(1):2–13. Publisher: Wiley Online Library. ↗
  14. Paul W. Holland, Howard Wainer, and William H. Angoff. 1993. Perspectives on Differential Item Functioning Methodology, page 3–23. Lawrence Erlbaum Associates. find
  15. Becky H. Huang and Belinda Bustos Flores. 2018. The English language proficiency assessment for the 21st century (ELPA21). Language Assessment Quarterly, 15(4):433–442. Publisher: Taylor & Francis. ↗
  16. Carol M. Woods, Li Cai, and Mian Wang. 2013. The Langer-improved Wald test for DIF testing with multiple groups: Evaluation and comparison to two-group IRT. Educational and Psychological Measurement, 73(3):532–547. Publisher: Sage Publications Sage CA: Los Angeles, CA. ↗
  17. Svetlana Kiritchenko and Saif M. Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. arXiv preprint arXiv:1805.04508. ↗
  18. Klaus Zechner. 2009. What did they actually say? Agreement and disagreement among transcribers of non-native spontaneous speech responses in an English proficiency test. find
  19. Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minnie Quartey, Zion Mengesha, Connor Toups, John R. Rickford, Dan Jurafsky, and Sharad Goel. 2020. Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14):7684–7689. Publisher: National Acad Sciences. ↗
  20. Klaus Zechner. 2020. Summary and Outlook on Automated Speech Scoring. In Klaus Zechner and Keelan Evanini, editors, Automated Speaking Assessment: Using Language Technologies to Score Spontaneous Speech, volume 3 of Innovations in Language Learning and Assessment at ETS, pages 192–204. Routledge, New York, NY. ↗
  21. Christopher M. Ormerod, Akanksha Malhotra, and Amir Jafari. 2021. Automated essay scoring using efficient transformer-based language models. arXiv preprint arXiv:2102.13136. ↗
  22. Haoran Zhang, Amy X. Lu, Mohamed Abdalla, Matthew McDermott, and Marzyeh Ghassemi. 2020. Hurtful words: quantifying biases in clinical contextual word embeddings. In proceedings of the ACM Conference on Health, Inference, and Learning, pages 110–120. ↗
  23. Steven J. Osterlind and Howard T. Everson. 2009. Differential item functioning. SAGE. ↗
  24. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 9-11, 2017, pages 2979–2989. Association for Computational Linguistics.
  25. Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukas Burget, Ondrej Glembek, Nagendra Goel, Mirko Hannemann, Petr Motlicek, Yanmin Qian, Petr Schwarz, Jan Silovsky, Georg Stemmer, and Karel Vesely. 2011. The kaldi speech recognition toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Processing Society. IEEE Catalog No.: CFP11SRW-USB. find
  26. A Steven J. Spencer, Christine Logel, and Paul G. Davies. 2016. Stereotype threat. Annual review of psychology, 67:415–437. Publisher: Annual Reviews. ASR Systems Comparison find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kwako-2022-item-response-theory-measure-gender/