Upstream Mitigation Is Not All You Need: Testing the Bias Transfer Hypothesis in Pre-Trained Language Models
Open accessThe DOI matches its Crossref record; 14 of the 14 references it cites were found in a registry or an index.
Abstract
A few large, homogenous, pre-trained models undergird many machine learning systems -and often, these models contain harmful stereotypes learned from the internet.We investigate the bias transfer hypothesis: the theory that social biases (such as stereotypes) internalized by large language models during pre-training transfer into harmful task-specific behavior after fine-tuning.For two classification tasks, we find that reducing intrinsic bias with controlled interventions before finetuning does little to mitigate the classifier's discriminatory behavior after fine-tuning.Regression analysis suggests that downstream disparities are better explained by biases in the fine-tuning dataset.Still, pre-training plays a role: simple alterations to co-occurrence rates in the fine-tuning dataset are ineffective when the model has been pre-trained.Our results encourage practitioners to focus more on dataset quality and context-specific harms.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Transformers: State-of-the-Art Natural Language Processing
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations · doi:10.18653/v1/2020.emnlp-demos.6 · Cited by 8,741
Energy and Policy Considerations for Deep Learning in NLP
Strubell, E., Ganesh, A., & McCallum, A. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1355 · Cited by 3,684 · cited 1×
“But massive pre-training datasets and increasingly homogeneous model design come with well-known, immediate social risks beyond the fi-nancial and environmental costs (Strubell et al., 2019; Bender et al., 2021).”
BackgroundOn the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676
Language (Technology) is Power: A Critical Survey of “Bias” in NLP
Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.485 · Cited by 1,050 · cited 2×
“These intrinsic biases are typically associated with representational harms, including stereotyping and denigration (Barocas et al., 2017; Blodgett et al., 2020; Bender et al., 2021).”
BackgroundThe Risk of Racial Bias in Hate Speech Detection
Sap, M., Card, D., Gabriel, S., Choi, Y., & Smith, N. A. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1163 · Cited by 817 · cited 1×
“…and Daumé, 2021), the re-inforcement of traditional gender roles in part-of-speech tagging (Garimella et al., 2019), toxic text generation (Gehman et al., 2020), and censorship of inclusive language and AAVE (Blodgett and O’Connor, 2017; Blodgett et al., 2018; Park et al., 2018; Sap et al., 2019).”
BackgroundOn Measuring Social Biases in Sentence Encoders
May, C., Wang, A., Bordia, S., Bowman, S. R., & Rudinger, R. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1063 · Cited by 461 · cited 1×
“Contextual embeddings—token encodings that are conditional on other, nearby tokens—pose an ongoing challenge for intrinsic bias measurement (May et al., 2019; Kurita et al., 2019; Guo and Caliskan, 2021) and bias mitigation (Liang et al., 2020).”
BackgroundMeasuring Bias in Contextualized Word Representations
Kurita, K., Vyas, N., Pareek, A., Black, A. W., & Tsvetkov, Y. (2019) · Proceedings of the First Workshop on Gender Bias in Natural Language Processing · doi:10.18653/v1/w19-3823 · Cited by 348 · cited 4×
“Contextual embeddings—token encodings that are conditional on other, nearby tokens—pose an ongoing challenge for intrinsic bias measurement (May et al., 2019; Kurita et al., 2019; Guo and Caliskan, 2021) and bias mitigation (Liang et al., 2020).”
BackgroundMethodInfluentialReducing Gender Bias in Abusive Language Detection
Park, J. H., Shin, J., & Fung, P. (2018) · Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d18-1302 · Cited by 323 · cited 1×
“…and Daumé, 2021), the re-inforcement of traditional gender roles in part-of-speech tagging (Garimella et al., 2019), toxic text generation (Gehman et al., 2020), and censorship of inclusive language and AAVE (Blodgett and O’Connor, 2017; Blodgett et al., 2018; Park et al., 2018; Sap et al., 2019).”
BackgroundLarge image datasets: A pyrrhic win for computer vision?
Birhane, A., & Prabhu, V. U. (2021) · 2021 IEEE Winter Conference on Applications of Computer Vision (WACV) · doi:10.1109/wacv48630.2021.00158 · Cited by 257 · cited 1×
“The closest example of this phenomena is the “blood diamond” effect (Birhane and Prabhu, 2021), in which stereotyping and denigration in the pre-training corpora pervade subsequently generated images and language even before the fine-tuning stage (Steed and Caliskan, 2021).”
BackgroundBlodgett, S. L., & O'Connor, B. (2017) · arXiv · doi:10.48550/arxiv.1707.00061 · Cited by 153
Cited by
Bias and Fairness in Large Language Models: A Survey
Gallegos, I. O., Rossi, R. A., Barrow, J., Tanjim, M. M., Kim, S., Dernoncourt, F., Yu, T., Zhang, R., & Ahmed, N. K. (2024) · Computational Linguistics · doi:10.1162/coli_a_00524 · Cited by 1,245
“Several works point out that biases in the embedding space have only weak or inconsistent relationships with biases in downstream tasks (Cabello et al., 2023; Cao et al., 2022; Goldfarb-Tarrant et al., 2021; Orgad & Belinkov, 2022; Orgad et al., 2022; Steed et al., 2022).”
BackgroundAuditing large language models: a three-layered approach
Mökander, J., Schuett, J., Kirk, H. R., & Floridi, L. (2024) · AI and Ethics · doi:10.1007/s43681-023-00289-2 · Cited by 289
AI generates covertly racist decisions about people based on their dialect
Hofmann, V., Kalluri, P. R., Jurafsky, D., & King, S. (2024) · Nature · doi:10.1038/s41586-024-07856-5 · Cited by 233
Assessing Cross-Cultural Alignment between ChatGPT and Human Societies: An Empirical Study
Cao, Y., Zhou, L., Lee, S., Cabello, L., Chen, M., & Hershcovich, D. (2023) · Proceedings of the First Workshop on Cross-Cultural Considerations in NLP (C3NLP) · doi:10.18653/v1/2023.c3nlp-1.7 · Cited by 163
Feng, S., Park, C. Y., Liu, Y., & Tsvetkov, Y. (2023) · Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2023.acl-long.656 · Cited by 149
Metadata
- Authors
- Steed, R., Panda, S., Kobren, A., & Wick, M.
- Venue
- Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
- Published
- 2022
- Pages
- 3524-3542
- OpenAlex
- W4285288507
- Cited by
- 42
- References
- 14
- FWCI
- 10.79
- Type
- conference paper
- Language
- en
- Open access
- hybrid
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Upstream Mitigation Is Not All You Need: Testing the Bias Transfer Hypothesis in Pre-Trained Language Models. https://app.cochat.ai/papers/10.18653/v1/2022.acl-long.247
- The paper
- Steed, R., Panda, S., Kobren, A., & Wick, M. (2022). Upstream Mitigation Is Not All You Need: Testing the Bias Transfer Hypothesis in Pre-Trained Language Models. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). https://doi.org/10.18653/v1/2022.acl-long.247
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.