Causal-Debias: Unifying Debiasing in Pretrained Language Models and Fine-tuning via Causal Invariant Learning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4227–4241 July 9-14, 2023 ©2023 Association for Computational Linguistics Causal-Debias: Unifying Debiasing in Pretrained Language Models and Fine-tuning via Causal Invariant Learning Fan Zhou1 Yuzhou Mao 1 Liu Yu 1 ∗ Yi Yang 2 Ting Zhong 1,3 1University of Electronic Science and Technology of China 2Hong Kong University of Science and Technology 3Kashi Institute of Electronics and Information Industry fan.zhou@uestc.edu.cn, yuzhou.mao@outlook.com, liu.yu@std.uestc.edu.cn, imyiyang@ust.hk, zhongting@uestc.edu.cn Abstract Demographic biases and social stereotypes are common in pretrained language models (PLMs), and a burgeoning body of literature focuses on removing the unwanted stereotypi- cal associations from PLMs. However, when fine-tuning these bias-mitigated PLMs in down- stream natural language processing (NLP) ap- plications, such as sentiment classification, the unwanted stereotypical associations resurface or even get amplified. Since pretrain&fine-tune is a major paradigm in NLP applications, sepa- rating the debiasing procedure of PLMs from fine-tuning would eventually harm the actual downstream utility. In this paper, we propose a unified debiasing framework Causal-Debias to remove unwanted stereotypical associations in PLMs during fine-tuning. Specifically, Causal- Debias mitigates bias from a causal invariant perspective by leveraging the specific down- stream task to identify bias-relevant and label- relevant factors. We propose that bias-relevant factors are non-causal as they should have lit- tle impact on downstream tasks, while label- relevant factors are causal. We perform inter- ventions on non-causal factors in different de- mographic groups and design an invariant risk minimization loss to mitigate bias while main- taining task performance. Experimental results on three downstream tasks show that our pro- posed method can remarkably reduce unwanted stereotypical associations after PLMs are fine- tuned, while simultaneously minimizing the impact on PLMs and downstream applications. 1 Introduction Pretrained language models (PLMs) have achieved remarkable success in many natural language pro- cessing (NLP) tasks. However, PLMs often en- coded undesired social stereotypes and biases, and thus mitigating such biases has become an emerg- ing and important task (Meade et al., 2022). Prior bias mitigation methods often focus on removing ∗Corresponding author unwanted stereotypical associations in PLMs. For example, some works (Zmigrod et al., 2019) pre- train a language model using original and counter- factual corpus in order to cancel-out biased asso- ciations, some works (Liang et al., 2020) focus on debiasing post-hoc sentence representations, and others (Guo et al., 2022; Cheng et al., 2021) de- sign bias-equalizing objectives to fine-tune PLM’s parameters. woman woman woman woman woman woman woman woman woman woman man man man man man man man man man man family family family family family family family family family family career career career career career career career career career career math math math math math math math math math math art art art art art art art art art art science science science science science science science science science science literature literature literature literature literature literature literature literature literature literature technology technology technology technology technology technology technology technology technology technology dance dance dance dance dance …

Content truncated to 500 words