Elamur Research⌕
☰
home / papers / 2026 / BRIDGE the Gap: Mitigating Bias Amplification in Automated Scoring of English Language Learners via Inter-group Data Augmentation

BRIDGE the Gap: Mitigating Bias Amplification in Automated Scoring of English Language Learners via Inter-group Data Augmentation

Paper · 2026 · Yun Wang, Xuansheng Wu, Jingyuan Huang et al.

Summary

Automated scorers for student answers, built on deep models and LLMs, can widen score gaps between student groups beyond the gaps in the human labels they learn from. The authors call this bias amplification and study it for English Language Learners (ELLs). Their explanation is representation bias combined with ERM. High-scoring ELL answers are extremely rare, so their share of the training gradient is tiny. The high-score boundary therefore follows non-ELL phrasing, and ELLs with equal knowledge get lower scores. BRIDGE is their fix. It generates synthetic high-scoring ELL answers by taking rubric-relevant content from plentiful high-scoring non-ELL answers and expressing it in the linguistic style of real ELL answers. A trained discriminator then removes poor candidates. On five California Science Test (CAST) short-answer items, BRIDGE lowers group disparity and high-score bias, and overall scoring quality stays essentially unchanged. Its fairness gains match those from adding real human-labelled ELL data. This matters because collecting rare high-scoring ELL samples is expensive.

Key points

  • High-scoring ELL answers are about 3.76% of ELL responses and under 1% of all responses. The authors decompose the ERM gradient to show this subgroup barely affects training.
  • The method has two stages. Stage 1 is inter-group stylistic reformulation (non-ELL content in ELL linguistic patterns). Stage 2 is discriminative filtering. Fairness is measured with BiasAmp and MSG Gap.
  • BRIDGE is compared against an oracle that adds real data. On Dataset 5, BRIDGE reaches a BiasAmp of 0.0050 while the oracle reaches 0.0131. BRIDGE is numerically better than the oracle on Datasets 1, 3 and 5.
  • The intra-group baselines often fail. On Dataset 4, Oversampling raises BiasAmp by 28.21% and Paraphrasing raises it by 42.95%.
  • BRIDGE barely changes scoring quality. Accuracy moves between −0.11% and +0.24%, and QWK between −0.27% and +0.15%. Both ranges are within the baseline's standard deviation.

Interesting ideas

  • I1. When a combination such as "rare group × high score" has almost no examples, build it from two abundant sources: substance from one group and surface form from another. This avoids rewriting the handful of real examples.
  • I2. A generator followed by a separate learned filter is a strong design. In the ablation, generation alone did little on Dataset 3 (BiasAmp 0.0073 → 0.0071), but adding the filter brought it to 0.0031, because the filter mattered most where generator output was noisy.
  • I3. Bias amplification is defined as the prediction gap between groups exceeding the gap already present in the gold labels. This separates model-induced unfairness from real differences in the data.
  • I4. Duplicating or lightly paraphrasing a tiny subgroup can make things worse. Naive augmentation should therefore be a standard baseline that a new method must beat.
  • I5. The paper treats content and style as separable but admits this breaks down when language itself is what is being assessed. That is a useful reminder to check whether the separation holds for a given task.

For Mova

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

  • I1 (learn): Rare syntactic constructions could be synthesized by recasting well-parsed, common content into the target construction. The resulting silver sentences would fill thin regions of the training data.
  • I2 (learn): LLM-teacher outputs could pass through a trained discriminator before entering silver data. The filter's effect should be reported separately, as in this ablation.
  • I3 (learn): Evaluation could compare Mova's error gap on non-native or nonstandard English against the gap in gold annotations. That would show whether training widens the difference.
  • I4 (tree): Before oversampling rare constructions in treebanks for the perceptron parser or gparse, test whether this actually helps LAS on those constructions or only causes overfitting.
  • I5 (answer): School-course readers could be tested on learner-style rewordings of the same questions. Answers should follow the meaning of the question, not its surface phrasing.

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

🎓Hong Kong Polytechnic UniversityHong Kong · China84 works · 4,956 citations🤝Educational Testing ServicePrinceton · United States27 works · 2,212 citations🎓University of GeorgiaAthens · United States17 works · 26,331 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksInformation Systems · 1,437 worksSafety Research · 543 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,263 worksGrammatical error correction & learner language · 156 works

Cite

Cite as Yun Wang, Xuansheng Wu, Jingyuan Huang, Lei Liu, Xiaoming Zhai and Ninghao Liu. 2026. BRIDGE the Gap: Mitigating Bias Amplification in Automated Scoring of English Language Learners via Inter-group Data Augmentation. arXiv:2602.23580.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2602.23580 · arXiv 2602.23580

@misc{wang2026bridge,
  title = {{BRIDGE the Gap: Mitigating Bias Amplification in Automated Scoring of English Language Learners via Inter-group Data Augmentation}},
  author = {Yun Wang, Xuansheng Wu, Jingyuan Huang, Lei Liu, Xiaoming Zhai and Ninghao Liu},
  year = {2026},
  doi = {10.48550/arxiv.2602.23580},
  eprint = {2602.23580},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2602.23580}
}

References

36 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abedi, J.: Psychometric issues in the ell assessment and special education eligibility. Teachers College Record 108(11), 2282–2303 (2006) find
  2. Alkhawaldeh, I.M., Albalkhi, I., Naswhan, A.J.: Challenges and limitations of synthetic minority oversampling techniques in machine learning. World journal of methodology 13(5), 373 (2023) find
  3. American Educational Research Association, American Psychological Association, National Council on Measurement in Education: Standards for Educational and Psychological Testing. American Educational Research Association, Washington, DC (2014) find
  4. Andersen, N., Mang, J., Goldhammer, F., Zehner, F.: Algorithmic fairness in automatic short answer scoring. International Journal of Artificial Intelligence in Education pp. 1–38 (2025) find
  5. Chai, Y., Xie, H., Qin, J.S.: Text data augmentation for large language models: A comprehensive survey of methods, challenges, and opportunities. arXiv preprint arXiv:2501.18845 (2025) Y. Wang et al. ↗
  6. Chaudhari, R., Patel, M.: Deep learning in automated short answer grading: A comprehensive review. In: ITM Web of Conferences. vol. 65, p. 03003. EDP Sciences (2024) find
  7. Chen, D., Yin, C.: Data augmentation for intent classification (2022), https://arxiv.org/abs/2206.05790 ↗
  8. Crossley, S.A., McNamara, D.S.: Predicting second language writing proficiency: The roles of cohesion and linguistic sophistication. Journal of Research in Reading 35(2), 115–135 (2012) find
  9. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. In: Proc. of NAACL-HLT (2019)
  10. Dikli, S., Bleyle, S.: Automated essay scoring feedback for second language writers: How does it compare to instructor feedback? Assessing writing 22, 1–17 (2014) find
  11. Emirtekin, E.: Large language model-powered automated assessment: A systematic review. Applied Sciences 15(10), 5683 (2025) find
  12. Guo, S., Wang, Y., Yu, J., Wu, X., Ayik, B., Watts, F.M., Latif, E., Liu, N., Liu, L., Zhai, X.: Artificial intelligence bias on english language learners in automatic scoring. In: International Conference on Artificial Intelligence in Education. pp. 268–275. Springer (2025)
  13. Hashimoto, T., Srivastava, M., Namkoong, H., Liang, P.: Fairness without demographics in repeated loss minimization. In: International Conference on Machine Learning. pp. 1929–1938. PMLR (2018) find
  14. Hurst, A., Lerer, A., Goucher, A.P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al.: Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024) ↗
  15. Leino, K., Black, E., Fredrikson, M., Sen, S., Datta, A.: Feature-wise bias amplification. arXiv preprint arXiv:1812.08999 (2018) ↗
  16. Li, Z., Wu, X., Li, Y., Hu, L., Liu, N.: Less is enough: Synthesizing diverse data in feature space of llms. arXiv preprint arXiv:2602.10388 (2026) ↗
  17. Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., Galstyan, A.: A survey on bias and fairness in machine learning. ACM computing surveys (CSUR) 54(6), 1–35 (2021)
  18. Messick, S.: Validity. ETS research report series 1987(2), i–208 (1987) find
  19. Palermo, C.: Rater characteristics, response content, and scoring contexts: Decomposing the determinates of scoring accuracy. Frontiers in Psychology 13, 937097 (2022) find
  20. Park, S., Hong, Y., Heo, B., Yun, S., Choi, J.Y.: The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6887–6896 (2022) find
  21. Plecko, D., Bareinboim, E.: Mind the gap: A causal perspective on bias amplification in prediction & decision-making. Advances in Neural Information Processing Systems 37, 84384–84408 (2024) find
  22. Sagawa, S., Koh, P.W., Hashimoto, T.B., Liang, P.: Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731 (2019) ↗
  23. Schwartz, R., Schwartz, R., Vassilev, A., Greene, K., Perine, L., Burt, A., Hall, P.: Towards a standard for identifying and managing bias in artificial intelligence, vol. 3. US Department of Commerce, National Institute of Standards and Technology . . . (2022) Mitigating Bias Amplification via Inter-group Data Augmentation find
  24. Shahbazi, N., Lin, Y., Asudeh, A., Jagadish, H.: Representation bias in data: A survey on identification and resolution techniques. ACM Computing Surveys 55(13s), 1–39 (2023) find
  25. Sourati, Z., Ziabari, A.S., Dehghani, M.: The homogenizing effect of large language models on human expression and thought. arXiv preprint arXiv:2508.01491 (2025) ↗
  26. Vapnik, V.: Principles of risk minimization for learning theory. Advances in neural information processing systems 4 (1991) find
  27. Wang, A., Russakovsky, O.: Directional bias amplification. In: International Conference on Machine Learning. pp. 10882–10893. PMLR (2021) find
  28. Wang, Y., Ding, Z., Wu, X., Sun, S., Liu, N., Zhai, X.: Autoscore: Enhancing automated scoring with multi-agent large language models via structured component recognition. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 40898–40906 (2026) find
  29. Weigle, S.C.: English language learners and automated scoring of essays: Critical considerations. Assessing Writing 18(1), 85–99 (2013) find
  30. Welch, C., Dunbar, S., Ji, J., Vernon, A., Park, J.: Response time for english learners on large-scale writing assessments. Assessing writing 66, 100979 (2025) find
  31. Wilson, J., Huang, Y.: Validity of automated essay scores for elementary-age english language learners: Evidence of bias? Assessing Writing 60, 100815 (2024) find
  32. Xu, H., Gan, W., Qi, Z., Wu, J., Yu, P.S.: Large language models for education: A survey. arXiv preprint arXiv:2405.13001 (2024) ↗
  33. Zhai, X., Haudek, K.C., Stuhlsatz, M.A., Wilson, C.: Evaluation of construct-irrelevant variance yielded by machine and human scoring of a science teacher pck constructed response assessment. Studies in Educational Evaluation 67, 100916 (2020) find
  34. Zhai, X., Pellegrino, J.W.: Large-scale assessment in science education. In: Handbook of research on science education, pp. 1045–1097. Routledge (2023) find
  35. Zhao, J., Wang, T., Yatskar, M., Ordonez, V., Chang, K.W.: Men also like shopping: Reducing gender bias amplification using corpus-level constraints. arXiv preprint arXiv:1707.09457 (2017) ↗
  36. Zhu, X., Wu, H., Zhang, L.: Automatic short-answer grading via bert-based deep neural networks. IEEE Transactions on Learning Technologies 15(3), 364–375 (2022). https://doi.org/10.1109/TLT.2022.3175537 ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-bridge-gap-mitigating-bias-amplification/