Elamur Research⌕
☰
home / papers / 2023 / Does BERT Exacerbate Gender or L1 Biases in Automated English Speaking Assessment?

Does BERT Exacerbate Gender or L1 Biases in Automated English Speaking Assessment?

Paper · 2023 · Alexander Kwako, Yixin Wan, Jieyu Zhao et al.

Summary

Pretrained language models like BERT can score spoken English test answers about as accurately as human raters. Whether they also carry over or amplify unfairness toward certain groups has been studied far less. The authors examine gender and first-language (L1) bias in a large U.S. K-12 assessment, ELPA21. They compare human scores with scores from an off-the-shelf BERT model that reads only automatic transcripts of the students' speech. Bias is measured as differential item functioning (DIF): whether students of similar ability, judged on the non-speaking parts of the test, perform differently on speaking items depending on their group. Human scores show moderate L1-related DIF among older students (grades 9–12), and more DIF on longer items. BERT scores track human scores closely. BERT increases overall L1 DIF slightly in grades 9–12, but the practical effect is very small. For anyone building automated scorers, the takeaway is that accuracy alone does not settle fairness. A group-conditional audit is also needed.

Key points

  • Data: ELPA21, a consortium of 7 U.S. state education agencies, 2020–2021 school year.
    • Two grade bands (2–3 and 9–12), three speaking items per band (short, medium, long).
    • 15,000 sampled students across ten L1 groups.
    • Spanish (82.7% of all examinees) serves as the reference group.
  • Pipeline: Responses were transcribed by Amazon Web Services ("en-US", default settings), then scored by an off-the-shelf BERT model. No acoustic information was used.
  • DIF method:
    • The matching criterion is proficiency estimated from non-speaking domains only, so the ability estimate is not contaminated by the bias under study.
    • Students are split into ten strata.
    • Bias is summarised with a uniform DIF statistic, z, using Benjamini-Hochberg adjusted p-values.
  • Grade-band differences in L1 DIF: These grow with item length (Δz_abs ≈ .013, .054, .145 for items 1–3). Second-order human-vs-BERT differences are not significant, so BERT does not amplify the age or length trends.
  • Group patterns:
    • In grades 2–3, Marshallese and Mayan speakers are disadvantaged on speaking items.
    • In grades 9–12, nearly all focal L1 groups do better than equally proficient Spanish speakers.
    • No individual L1 group or item shows a significant human-vs-BERT gap.

Interesting ideas

  • I1. Condition group comparisons on an ability estimate built from a different skill domain, so the yardstick cannot share the bias being measured.
  • I2. Test whether a model amplifies a gap using a difference of differences: compute group disparity under human labels, compute it again under model labels, then test the gap between the two.
  • I3. Use a scorer that lacks one input channel (here, audio) as a natural ablation: if bias survives without that channel, the channel cannot be its sole cause.
  • I4. Choose test items deliberately across response lengths, since bias here grew with the amount of language produced.
  • I5. Report effect sizes against established severity bands alongside significance, so a statistically real but tiny difference is not overstated.

For Mova

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

  • I1 (learn): When checking whether the parser is worse on some genre or dialect, stratify texts by a difficulty measure computed independently of parser output (e.g. sentence length or lexical rarity) before comparing error rates.
  • I2 (learn): Compare per-subgroup error disparities of gparse against gold treebank disparities, to tell model-induced unfairness apart from difficulty that already exists in the data.
  • I3 (learn): Run negative-control variants of a reader with certain cues removed (e.g. lemma-only, no morphology) to locate which input actually drives a systematic error.
  • I4 (answer): Stratify school-test and fable-QA evaluations by passage and answer length, to reveal whether reasoning failures concentrate in longer texts.
  • I5 (learn): Attach practical-magnitude thresholds to LAS or accuracy deltas between training rounds, so trivial significant changes do not trigger repair rules.

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 States132 works · 25,380 citations🎓University of California, Los AngelesLos Angeles · United States87 works · 54,961 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Arts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 worksDecision Sciences · 209 worksManagement Science and Operations Research · 108 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Alexander Kwako, Yixin Wan, Jieyu Zhao, Mark Hansen, Kai-Wei Chang and Li Cai. 2023. Does BERT Exacerbate Gender or L1 Biases in Automated English Speaking Assessment?. Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.bea-1.54

@inproceedings{kwako2023does,
  title = {{Does BERT Exacerbate Gender or L1 Biases in Automated English Speaking Assessment?}},
  author = {Alexander Kwako, Yixin Wan, Jieyu Zhao, Mark Hansen, Kai-Wei Chang and Li Cai},
  booktitle = {Proceedings of the 18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023)},
  year = {2023},
  doi = {10.18653/v1/2023.bea-1.54},
  url = {https://doi.org/10.18653/v1/2023.bea-1.54}
}

References

46 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Nancy L Allen, John R Donoghue, and Terry L Schoeps. 2001. The naep 1998 technical report. Education Statistics Quarterly, 3(4):95–98. find
  2. American Educational Research Association, American Psychological Association, and National Council on Measurement in Education. 2014. Standards for educational and psychological testing. American Educational Research Association. find
  3. H Douglas Brown et al. 2000. Principles of language learning and teaching, volume 4. Longman New York. find
  4. Keith Brown. 2005. Encyclopedia of language and linguistics, volume 1. Elsevier. find
  5. 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. ↗
  6. Joseph R Cimpian, Karen D Thompson, and Martha B Makowski. 2017. Evaluating english learner reclassification policy effects across districts. American Educational Research Journal, 54(1_suppl):255S– 278S. ↗
  7. Jo-Kate Collier and Becky Huang. 2020. Test review: Texas english language proficiency assessment system (telpas). Language Assessment Quarterly, 17(2):221–230. ↗
  8. William H Angoff. 1993. Perspectives on differential item functioning methodology. find
  9. Katrijn Denies, Liesbet Heyvaert, Jonas Dockx, and Rianne Janssen. 2022. Mapping and explaining the gender gap in students’ second language proficiency across skills, countries and languages. Learning and Instruction, 80:101618. ↗
  10. 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. ↗
  11. Tracey M Derwing and Murray J Munro. 2013. The development of l2 oral language skills in two l1 groups: A 7-year study. Language learning, 63(2):163–185. ↗
  12. Su Blodgett, Solon Barocas, Hal Daumé, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of "bias" in nlp.
  13. 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. ↗
  14. Neil J Dorans and Edward Kulick. 1986. Demonstrating the utility of the standardization approach to assessing unexpected differential item performance on the scholastic aptitude test. Journal of educational measurement, 23(4):355–368. ↗
  15. Neil J Dorans and Karin Zeller. 2004. Examining freedle’s claims about bias and his proposed solution: Dated data, inappropriate measurement, and incorrect and unfair scoring. ETS Research Report Series, 2004(2):1–33. ↗
  16. Educational Testing Service. 2005. Test and score data summary: 2004-05 test year data test of english as a foreign language. find
  17. Bradley Efron and Robert J Tibshirani. 1994. An introduction to the bootstrap. CRC press. ↗
  18. G Engelhard. 2002. Monitoring raters in performance assessments. Large-scale assessment programs for all students: Validity, technical adequacy, and implementation, pages 261–287. find
  19. Roy Freedle. 2003. Correcting the sat’s ethnic and social-class bias: A method for reestimating sat scores. Harvard Educational Review, 73(1):1–43. ↗
  20. Anthony G Greenwald and Mahzarin R Banaji. 1995. Implicit social cognition: attitudes, self-esteem, and stereotypes. Psychological review, 102(1):4. ↗
  21. 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. ↗
  22. Xiaoting Huang, Mark Wilson, and Lei Wang. 2016. Exploring plausible causes of differential item functioning in the pisa science assessment: language, curriculum or culture. Educational Psychology, 36(2):378– 390. find
  23. Okim Kang and Katherine Yaw. 2021. Social judgement of l2 accented speech stereotyping and its influential factors. Journal of Multilingual and Multicultural Development, pages 1–16. ↗
  24. Svetlana Kiritchenko and Saif M Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. arXiv preprint arXiv:1805.04508. ↗
  25. Antony John Kunnan. 2017. Evaluating language assessments. Taylor & Francis. ↗
  26. Alexander Kwako. 2023. Mitigating Gender and Racial Bias in Automated English Speaking Assessment. University of California, Los Angeles. find
  27. 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. In Proceedings of the 17th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2022), pages 1–7.
  28. Stephanie Lindemann and Nicholas Subtirelu. 2013. Reliably biased: The role of listener expectation in the perception of second language speech. Language Learning, 63(3):567–594. ↗
  29. Nathan Mantel. 1963. Chi-square tests with one degree of freedom; extensions of the mantel-haenszel procedure. Journal of the American Statistical Association, 58(303):690–700. ↗
  30. Michalis P Michaelides. 2008. An illustration of a mantel-haenszel procedure to flag misbehaving common items in test equating. Practical Assessment, Research, and Evaluation, 13(1):7. find
  31. Christopher Ormerod. 2022. Short-answer scoring with ensembles of pretrained language models. arXiv preprint arXiv:2202.11558. ↗
  32. Christopher Ormerod, Susan Lottridge, Amy E Harris, Milan Patel, Paul van Wamelen, Balaji Kodeswaran, Sharon Woolf, and Mackenzie Young. 2022. Automated short answer scoring using an ensemble of neural networks and latent semantic analysis classifiers. International Journal of Artificial Intelligence in Education, pages 1–30. ↗
  33. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, and Luca Antiga. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026–8037. ↗
  34. Python Software Foundation. 2022. The python language reference. find
  35. Maria Veronica Santelices and Mark Wilson. 2010. Unfair treatment? the case of freedle, the sat, and the standardization approach to differential item functioning. Harvard Educational Review, 80(1):106–134. ↗
  36. Gabriel Stanovsky, Noah A Smith, and Luke Zettlemoyer. 2019. Evaluating gender bias in machine translation. arXiv preprint arXiv:1906.00591. ↗
  37. Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. arXiv preprint arXiv:1906.08976. ↗
  38. Xinhao Wang, Keelan Evanini, Yao Qian, and Matthew Mulholland. 2021. Automated scoring of spontaneous speech from young learners of english using transformers. In 2021 IEEE Spoken Language Technology Workshop (SLT), pages 705–712. IEEE. find
  39. Zhen Wang, Klaus Zechner, and Yu Sun. 2018. Monitoring the performance of human and automated scores for spoken responses. Language Testing, 35(1):101– 120. find
  40. Valerie SL Williams, Lyle V Jones, and John W Tukey. 1999. Controlling error in multiple comparisons, with examples from state-to-state differences in educational achievement. Journal of educational and behavioral statistics, 24(1):42–69. ↗
  41. Paula Winke, Susan Gass, and Carol Myford. 2013. Raters’ l2 background as a potential source of bias in rating oral performance. Language Testing, 30(2):231–252. ↗
  42. Thomas Wolf, Julien Chaumond, Lysandre Debut, Victor Sanh, Clement Delangue, Anthony Moi, Pierric Cistac, Morgan Funtowicz, Joe Davison, and Sam Shleifer. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45.
  43. 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. ↗
  44. Klaus Zechner. 2009. What did they actually say? agreement and disagreement among transcribers of non-native spontaneous speech responses in an english proficiency test. In International Workshop on Speech and Language Technology in Education. find
  45. Bruno D Zumbo. 2007. Three generations of dif analyses: Considering where it has been, where it is now, and where it is going. Language assessment quarterly, 4(2):223–233. ↗
  46. Rebecca Zwick, John R Donoghue, and Angela Grima. 1993. Assessment of differential item functioning for performance tasks. Journal of Educational Measurement, 30(3):233–251. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kwako-2023-does-bert-exacerbate-gender-or/