Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Cleaning stereotypes out of training data for one group often worsens them for others, even unrelated ones.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 15:05 UTC pith:YWZXEHMJ

load-bearing objection Solid empirical grid showing that common data-level debiasing often redistributes stereotype scores across groups; single-seed and proxy-score limits are already flagged by the authors. the 3 major comments →

arxiv 2607.07937 v1 pith:YWZXEHMJ submitted 2026-07-08 cs.CL

When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigation

Yahan Zheng , John Guerrerio , Soroush Vosoughi , Weicheng Ma This is my paper
classification cs.CL
keywords stereotype mitigationpreprocessing debiasingside effectsStereoSetCrowS-Pairsattention rolloutlanguage modelsWikipedia
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Preprocessing-based debiasing—removing stereotypical sentences, stripping group mentions, or swapping identity references in Wikipedia before pre- or post-training—reliably lowers stereotype scores for the targeted demographic. The paper shows this frequently comes with side effects: stereotyping or counter-stereotyping rises for non-target groups, including across gender, race, and religion categories that share no obvious content overlap. The pattern holds for both encoder-only and decoder-only models, at full and 5% data scales, and even appears in a larger model after limited post-training. Standard aggregate benchmarks often miss the shifts, and attention-rollout maps change little, so the redistribution is not explained by large attention rewiring. The authors therefore treat data-level mitigation as an intervention whose collateral effects must be measured, not as a clean fix.

Core claim

Interventions that reduce measurable stereotypes for a chosen demographic on Wikipedia-trained models systematically induce side effects: undesired movement of stereotype scores away from neutrality for other groups, including across unrelated categories. These side effects are robust across three common preprocessing recipes, pre- and post-training, two model families, and reduced data scales, and they are poorly captured by overall benchmark scores or by attention-flow diagnostics.

What carries the argument

Side-effect definition and measurement: a mitigation that improves the target group's stereotype score while moving one or more non-target groups farther from the neutral baseline of 50 on StereoSet (or the corresponding CrowS-Pairs category scores), tracked by directional change relative to the matched base model under fixed seed and scale.

Load-bearing premise

Directional shifts of StereoSet and CrowS-Pairs scores away from 50, measured once under a fixed seed on Wikipedia-derived data, are treated as a trustworthy signal of real stereotype redistribution rather than an artifact of benchmark makeup or detector noise.

What would settle it

Re-run the same three preprocessing recipes on the same Wikipedia snapshot with multiple random seeds and a second independent stereotype detector; if non-target score movements reverse sign or vanish while target improvements remain, the claimed side-effect phenomenon collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper argues that preprocessing-based stereotype mitigation (DG: remove stereotypical sentences; RG: remove group mentions; SR: swap group references) on Wikipedia, used for pre- or post-training of TinyBERT and GPT-2 (plus one LLaMA-2-7B RG-female slice), reliably lowers StereoSet SS / CrowS-Pairs scores for the target demographic while frequently moving non-target scores away from the neutral baseline of 50, including across gender/race/religion categories. These side effects appear under both full and 5% data scales, are not explained by simple stereotype/anti-stereotype content redistribution, and are not accompanied by large attention-rollout shifts (PD/SD/JSD/L2). The authors supply the full experimental grid, extensive appendix tables (A1–A42), public code, and an explicit Limitations section, and recommend side-effect-aware diagnostics and evaluation.

Significance. If the directional pattern holds, the result is practically important: data-level debiasing is widely used precisely because it is simple and inference-cost-free, yet the paper shows that target-group gains can be purchased by redistributing harm to other groups, including groups outside the intervention category and outside the guiding benchmarks. Strengths that raise the contribution include the systematic grid (two architectures, three interventions, pre-/post-training, two scales), the operational definition of side effects, the public code release, the attention-rollout negative result that avoids over-claiming mechanism, and the open Limitations (single seed, compact models + one large slice, English Wikipedia only, possible DG false positives and SR artifacts). The work therefore supplies both a cautionary empirical finding and concrete evaluation recommendations for the fairness community.

major comments (3)
  1. §3 Evaluation protocol and Limitations: all SS/CrowS-Pairs directional claims rest on a single fixed seed (42) with no multi-seed uncertainty or significance tests. Because the central claim is the frequency and robustness of non-target shifts (not merely that a target SS can move toward 50), the manuscript should either report multi-seed runs for a representative subset of the grid or supply bootstrap/permutation intervals on the directional changes so that readers can judge how often the side-effect pattern survives sampling variation.
  2. §5.1–5.2 and Tables A1–A21: the paper correctly notes that side effects are not fully explained by altered stereotype/anti-stereotype content ratios, yet it never quantifies those ratios (or co-occurrence statistics) before vs. after each DG/RG/SR intervention. Without that measurement, the claim that distributional explanations are insufficient remains qualitative; a short appendix table of pre-/post-intervention stereotype-sentence counts or PMI shifts for the six groups would make the argument load-bearing rather than suggestive.
  3. §6 and Figure 5: attention-rollout distances are reported as small (max PD 0.0061, JSD 0.0925, etc.) even when SS moves substantially, which is a useful negative result. However, the paper treats rollout as the primary mechanistic probe without a positive control (e.g., a known semantic intervention that does produce large rollout shifts on the same StereoSet-derived bench). Adding one such control, or explicitly bounding what rollout can and cannot detect, would strengthen the conclusion that “semantic routing alone does not explain the effect.”
minor comments (5)
  1. Figures 1–2 and the corresponding appendix tables mix “closer to 50” (underlined) and “farther from 50” (bold) conventions; a single consistent legend and color scale across all SS plots would improve readability.
  2. §3: the DG detector is cited with F1=0.98 on CrowS-Pairs; a brief note on its false-positive rate on Wikipedia-style prose (or a small manual audit) would help readers assess contamination risk for the DG condition.
  3. §7.2 Table 1: LLaMA-2 results are reported for only the RG-female setting; even a one-sentence statement of why other interventions were computationally infeasible would clarify the scope of the “massive models” claim.
  4. Typos and notation: “canincrease” (abstract), occasional inconsistent hyphenation of “pre-/post-training,” and the use of both “SG-” and “SR-” labels in Figure 5 should be cleaned for camera-ready.
  5. Ethics / Limitations: the recommendation to document non-targeted groups in model cards is valuable; a short checklist or example diagnostic (e.g., “report ΔSS for all six groups after any single-group intervention”) would make the actionable advice more concrete.

Circularity Check

0 steps flagged

No significant circularity: purely empirical measurement of side effects against external benchmarks with no derivation that reduces to inputs by construction.

full rationale

The paper's central claim is an empirical observation from a full experimental grid (TinyBERT/GPT-2, pre-/post-training, full/5% Wikipedia, DG/RG/SR interventions, six demographic groups, plus one LLaMA-2 slice): preprocessing that moves target-group StereoSet SS or CrowS-Pairs scores toward 50 frequently moves non-target scores away from 50, including across categories. Side effects are defined operationally as undesired directional shifts relative to matched base models; evaluation uses public external benchmarks (StereoSet, CrowS-Pairs) whose instance composition is independent of the Wikipedia preprocessing recipes. Attention-rollout distances are reported as a diagnostic probe only and are not used to derive the side-effect claim. There are no equations, fitted parameters renamed as predictions, uniqueness theorems, ansatzes, or load-bearing self-citations that force the result by construction. Self-citations appear only as background on prior bias literature. The work is self-contained against external data; Limitations already flag single-seed and detector caveats without circular dependence. Score 0 is therefore the correct, proportionate finding.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 1 invented entities

The paper is empirical; load-bearing premises are domain conventions of bias measurement and the operational definitions of the three cleaning operators. No free parameters are fitted to produce the central claim; the invented entity is the operational 'side effect' itself.

axioms (3)
  • domain assumption StereoSet SS = 50 is the neutral baseline; movement away from 50 is undesirable stereotyping or anti-stereotyping.
    Used throughout Sections 3–5 and all appendix tables to label side effects.
  • domain assumption A pretrained sentence-level stereotype detector (F1=0.98 on CrowS-Pairs) correctly flags sentences for the DG operator.
    Section 3; false positives would alter the cleaned corpora and therefore the measured side effects.
  • ad hoc to paper Attention-rollout distances (PD, SD, JSD, L2) are a valid probe for whether semantic routing changes accompany stereotype shifts.
    Section 6; the paper itself notes the probe is surface-level and does not locate circuits.
invented entities (1)
  • side effect (of preprocessing-based mitigation) no independent evidence
    purpose: Operational label for any undesired SS movement of a non-target group after a target-group intervention.
    Defined in the abstract and Section 1; the entire empirical contribution is the demonstration that such side effects occur frequently and unpredictably.

pith-pipeline@v1.1.0-grok45 · 29192 in / 2311 out tokens · 26479 ms · 2026-07-10T15:05:13.928109+00:00 · methodology

0 comments
read the original abstract

Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often induce unintended shifts-side effects, where stereotyping or counter-stereotyping can increase relative to neutral baselines for other demographics, including across unrelated demographic categories. We demonstrate these side effects across two model families (encoder-only and decoder-only), multiple preprocessing strategies (removing stereotypical sentences, removing group mentions, and swapping group references), and both pre- and post-training at different data scales on Wikipedia. Standard benchmarks frequently miss these shifts. Using attention-rollout analysis, we observe that such side effects are not accompanied by large changes in attention flow, complicating mechanistic explanations. We discuss implications for evaluation, provide actionable diagnostics, and argue for side-effect-aware, transparent mitigation practices.

Figures

Figures reproduced from arXiv: 2607.07937 by John Guerrerio, Soroush Vosoughi, Weicheng Ma, Yahan Zheng.

Figure 1
Figure 1. Figure 1: StereoSet stereotype scores (SS; 50 is neutral, closer to 50 indicates [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: StereoSet stereotype (SS) for models post-trained on debiased Wikipedia. Full numeric results appear in Appendix Tables A10-A12 and A16-A18. ward females (Figures 1a and 2a), whereas mod￾els trained on RG-male exhibit inconsistent pat￾terns across pre- versus post-training (Figures 1b and 2b). Unintended effects also extend beyond directly opposing groups. For instance, pre-training on RG￾Muslim increases … view at source ↗
Figure 4
Figure 4. Figure 4: Example of cross-group stereotype shifts [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distances/divergences between attention-rollout distributions of stereotype-mitigated models and their [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 1 internal anchor

  1. [1]

    Science , volume=

    Semantics derived automatically from language corpora contain human-like biases , author=. Science , volume=. 2017 , publisher=

    work page 2017

  2. [2]

    and Saligrama, Venkatesh and Kalai, Adam T

    Bolukbasi, Tolga and Chang, Kai-Wei and Zou, James Y. and Saligrama, Venkatesh and Kalai, Adam T. , title =. Advances in Neural Information Processing Systems , volume =

    work page

  3. [3]

    Gender Bias in Contextualized Word Embeddings

    Zhao, Jieyu and Wang, Tianlu and Yatskar, Mark and Cotterell, Ryan and Ordonez, Vicente and Chang, Kai-Wei. Gender Bias in Contextualized Word Embeddings. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019. doi:10.18653/v1/N19-1064

    work page doi:10.18653/v1/n19-1064 2019

  4. [4]

    , year =

    Nangia, Nikita and Vania, Clara and Bhalerao, Rasika and Bowman, Samuel R. C row S -Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.154

    work page doi:10.18653/v1/2020.emnlp-main.154 2020

  5. [5]

    Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods

    Zhao, Jieyu and Wang, Tianlu and Yatskar, Mark and Ordonez, Vicente and Chang, Kai-Wei. Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods. Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers). 2018. doi:10.18653/v1/N18-2003

    work page doi:10.18653/v1/n18-2003 2018

  6. [6]

    Proceedings of the 59th

    Nadeem, Moin and Bethke, Anna and Reddy, Siva. S tereo S et: Measuring stereotypical bias in pretrained language models. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021. doi:10.18653/v1/2021.acl-long.416

    work page doi:10.18653/v1/2021.acl-long.416 2021

  7. [7]

    ROBBIE : Robust Bias Evaluation of Large Generative Language Models

    Esiobu, David and Tan, Xiaoqing and Hosseini, Saghar and Ung, Megan and Zhang, Yuchen and Fernandes, Jude and Dwivedi-Yu, Jane and Presani, Eleonora and Williams, Adina and Smith, Eric. ROBBIE : Robust Bias Evaluation of Large Generative Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.1...

    work page doi:10.18653/v1/2023.emnlp-main.230 2023

  8. [8]

    Proceedings of the 2021

    Dhamala, Jwala and Sun, Tony and Kumar, Varun and Krishna, Satyapriya and Pruksachatkun, Yada and Chang, Kai-Wei and Gupta, Rahul , title =. 2021 , isbn =. doi:10.1145/3442188.3445924 , booktitle =

    work page doi:10.1145/3442188.3445924 2021

  9. [9]

    On Measuring Social Biases in Prompt-Based Multi-Task Learning

    Aky. On Measuring Social Biases in Prompt-Based Multi-Task Learning. Findings of the Association for Computational Linguistics: NAACL 2022. 2022. doi:10.18653/v1/2022.findings-naacl.42

    work page doi:10.18653/v1/2022.findings-naacl.42 2022

  10. [10]

    Computational Linguistics , author =

    Gallegos, Isabel O. and Rossi, Ryan A. and Barrow, Joe and Tanjim, Md Mehrab and Kim, Sungchul and Dernoncourt, Franck and Yu, Tong and Zhang, Ruiyi and Ahmed, Nesreen K. Bias and Fairness in Large Language Models: A Survey. Computational Linguistics. 2024. doi:10.1162/coli_a_00524

    work page doi:10.1162/coli_a_00524 2024

  11. [11]

    Queer People are People First: Deconstructing Sexual Identity Stereotypes in Large Language Models

    Dhingra, Harnoor and Jayashanker, Preetiha and Moghe, Sayali and Strubell, Emma , title =. arXiv preprint arXiv:2307.00101 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Text Style Transfer for Bias Mitigation using Masked Language Modeling

    Tokpo, Ewoenam Kwaku and Calders, Toon. Text Style Transfer for Bias Mitigation using Masked Language Modeling. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop. 2022. doi:10.18653/v1/2022.naacl-srw.21

    work page doi:10.18653/v1/2022.naacl-srw.21 2022

  13. [13]

    UniBias: Unveiling and Mitigating

    Hanzhang Zhou and Zijian Feng and Zixiao Zhu and Junlang Qian and Kezhi Mao , booktitle=. UniBias: Unveiling and Mitigating. 2024 , url=

    work page 2024

  14. [14]

    Deciphering Stereotypes in Pre-Trained Language Models

    Ma, Weicheng and Scheible, Henry and Wang, Brian and Veeramachaneni, Goutham and Chowdhary, Pratim and Sun, Alan and Koulogeorge, Andrew and Wang, Lili and Yang, Diyi and Vosoughi, Soroush. Deciphering Stereotypes in Pre-Trained Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2...

    work page doi:10.18653/v1/2023.emnlp-main.697 2023

  15. [15]

    NeurIPS Safe Generative AI Workshop , year =

    Tong, Schrasing and Zemour, Eliott and Lohanimit, Rawisara and Kagal, Lalana , title =. NeurIPS Safe Generative AI Workshop , year =

    work page

  16. [16]

    BOLT : Fast Energy-based Controlled Text Generation with Tunable Biases

    Liu, Xin and Khalifa, Muhammad and Wang, Lu. BOLT : Fast Energy-based Controlled Text Generation with Tunable Biases. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2023. doi:10.18653/v1/2023.acl-short.18

    work page doi:10.18653/v1/2023.acl-short.18 2023

  17. [17]

    First the Worst: Finding Better Gender Translations During Beam Search

    Saunders, Danielle and Sallis, Rosie and Byrne, Bill. First the Worst: Finding Better Gender Translations During Beam Search. Findings of the Association for Computational Linguistics: ACL 2022. 2022. doi:10.18653/v1/2022.findings-acl.301

    work page doi:10.18653/v1/2022.findings-acl.301 2022

  18. [18]

    Debiasing Pre-Trained Language Models via Efficient Fine-Tuning

    Gira, Michael and Zhang, Ruisu and Lee, Kangwook. Debiasing Pre-Trained Language Models via Efficient Fine-Tuning. Proceedings of the Second Workshop on Language Technology for Equality, Diversity and Inclusion. 2022. doi:10.18653/v1/2022.ltedi-1.8

    work page doi:10.18653/v1/2022.ltedi-1.8 2022

  19. [19]

    and Li, Manling and Ji, Heng , title =

    Yang, Ke and Yu, Charles and Fung, Yi R. and Li, Manling and Ji, Heng , title =. 2023 , isbn =. doi:10.1609/aaai.v37i9.26279 , booktitle =

    work page doi:10.1609/aaai.v37i9.26279 2023

  20. [20]

    Debiasing Pretrained Text Encoders by Paying Attention to Paying Attention

    Gaci, Yacine and Benatallah, Boualem and Casati, Fabio and Benabdeslem, Khalid. Debiasing Pretrained Text Encoders by Paying Attention to Paying Attention. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.651

    work page doi:10.18653/v1/2022.emnlp-main.651 2022

  21. [21]

    Auto-Debias: Debiasing Masked Language Models with Automated Biased Prompts

    Guo, Yue and Yang, Yi and Abbasi, Ahmed. Auto-Debias: Debiasing Masked Language Models with Automated Biased Prompts. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. doi:10.18653/v1/2022.acl-long.72

    work page doi:10.18653/v1/2022.acl-long.72 2022

  22. [22]

    Ouyang, Long and Wu, Jeff and Jiang, Xu and Almeida, Diogo and Wainwright, Carroll L. and Mishkin, Pamela and Zhang, Chong and Agarwal, Sandhini and Slama, Katarina and Ray, Alex and Schulman, John and Hilton, Jacob and Kelton, Fraser and Miller, Luke and Simens, Maddie and Askell, Amanda and Welinder, Peter and Christiano, Paul and Leike, Jan and Lowe, R...

    work page 2022

  23. [23]

    Logic, language, and security: essays dedicated to Andre Scedrov on the occasion of his 65th birthday , pages=

    Gender bias in neural natural language processing , author=. Logic, language, and security: essays dedicated to Andre Scedrov on the occasion of his 65th birthday , pages=. 2020 , publisher=

    work page 2020

  24. [24]

    Gender-tuning: Empowering Fine-tuning for Debiasing Pre-trained Language Models

    Ghanbarzadeh, Somayeh and Huang, Yan and Palangi, Hamid and Cruz Moreno, Radames and Khanpour, Hamed. Gender-tuning: Empowering Fine-tuning for Debiasing Pre-trained Language Models. Findings of the Association for Computational Linguistics: ACL 2023. 2023. doi:10.18653/v1/2023.findings-acl.336

    work page doi:10.18653/v1/2023.findings-acl.336 2023

  25. [25]

    Deep learning on a healthy data diet: finding important examples for fairness , year =

    Zayed, Abdelrahman and Parthasarathi, Prasanna and Mordido, Gon. Deep learning on a healthy data diet: finding important examples for fairness , year =. doi:10.1609/aaai.v37i12.26706 , booktitle =

    work page doi:10.1609/aaai.v37i12.26706

  26. [26]

    Demographic-Aware Language Model Fine-tuning as a Bias Mitigation Technique

    Garimella, Aparna and Mihalcea, Rada and Amarnath, Akhash. Demographic-Aware Language Model Fine-tuning as a Bias Mitigation Technique. Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). 2022. doi...

    work page doi:10.18653/v1/2022.aacl-short.38 2022

  27. [27]

    Looking for a Handsome Carpenter! Debiasing GPT -3 Job Advertisements

    Borchers, Conrad and Gala, Dalia and Gilburt, Benjamin and Oravkin, Eduard and Bounsi, Wilfried and Asano, Yuki M and Kirk, Hannah. Looking for a Handsome Carpenter! Debiasing GPT -3 Job Advertisements. Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP). 2022. doi:10.18653/v1/2022.gebnlp-1.22

    work page doi:10.18653/v1/2022.gebnlp-1.22 2022

  28. [28]

    , title =

    Raffel, Colin and Shazeer, Noam and Roberts, Adam and Lee, Katherine and Narang, Sharan and Matena, Michael and Zhou, Yanqi and Li, Wei and Liu, Peter J. , title =. J. Mach. Learn. Res. , month = jan, articleno =. 2020 , issue_date =

    work page 2020

  29. [29]

    Mitigating harm in language models with conditional-likelihood filtration , journal =

    Ngo, Helen and Raterink, Cooper and Ara. Mitigating harm in language models with conditional-likelihood filtration , journal =

    work page

  30. [30]

    Advances in Neural Information Processing Systems , editor=

    Fair Infinitesimal Jackknife: Mitigating the Influence of Biased Training Data Points Without Refitting , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    work page 2022

  31. [31]

    Don`t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models

    Panda, Swetasudha and Kobren, Ari and Wick, Michael and Shen, Qinlan. Don`t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models. Findings of the Association for Computational Linguistics: EMNLP 2022. 2022. doi:10.18653/v1/2022.findings-emnlp.372

    work page doi:10.18653/v1/2022.findings-emnlp.372 2022

  32. [32]

    T iny BERT : Distilling BERT for Natural Language Understanding

    Jiao, Xiaoqi and Yin, Yichun and Shang, Lifeng and Jiang, Xin and Chen, Xiao and Li, Linlin and Wang, Fang and Liu, Qun. T iny BERT : Distilling BERT for Natural Language Understanding. Findings of the Association for Computational Linguistics: EMNLP 2020. 2020. doi:10.18653/v1/2020.findings-emnlp.372

    work page doi:10.18653/v1/2020.findings-emnlp.372 2020

  33. [33]

    OpenAI blog , volume=

    Language models are unsupervised multitask learners , author=. OpenAI blog , volume=

    work page

  34. [34]

    Quantifying Attention Flow in Transformers

    Abnar, Samira and Zuidema, Willem. Quantifying Attention Flow in Transformers. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653/v1/2020.acl-main.385

    work page doi:10.18653/v1/2020.acl-main.385 2020

  35. [35]

    Attention Calibration for Transformer in Neural Machine Translation

    Lu, Yu and Zeng, Jiali and Zhang, Jiajun and Wu, Shuangzhi and Li, Mu. Attention Calibration for Transformer in Neural Machine Translation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021. doi:10.18653/v1/2021.ac...

    work page doi:10.18653/v1/2021.acl-long.103 2021

  36. [36]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    SEAT: Stable and Explainable Attention , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2023 , month=. doi:10.1609/aaai.v37i11.26517 , number=

    work page doi:10.1609/aaai.v37i11.26517 2023