@inproceedings{steed-etal-2022-upstream,
    title = "{U}pstream {M}itigation {I}s \textit{ {N}ot} {A}ll {Y}ou {N}eed: {T}esting the {B}ias {T}ransfer {H}ypothesis in {P}re-{T}rained {L}anguage {M}odels",
    author = "Steed, Ryan  and
      Panda, Swetasudha  and
      Kobren, Ari  and
      Wick, Michael",
    editor = "Muresan, Smaranda  and
      Nakov, Preslav  and
      Villavicencio, Aline",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/moar-dois/2022.acl-long.247/",
    doi = "10.18653/v1/2022.acl-long.247",
    pages = "3524--3542",
    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 \textit{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 \textit{before} fine-tuning does little to mitigate the classifier{'}s discriminatory behavior \textit{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."
}