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, pages 1012 - 1023 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Auto-Debias: Debiasing Masked Language Models with Automated Biased Prompts Yue Guo1, Yi Yang1, Ahmed Abbasi2 1 The Hong Kong University of Science and Technology 2 University of Notre Dame yguoar@connect.ust.hk imyiyang@ust.hk aabbasi@nd.edu Abstract Human-like biases and undesired social stereo- types exist in large pretrained language mod- els. Given the wide adoption of these models in real-world applications, mitigating such bi- ases has become an emerging and important task. In this paper, we propose an automatic method to mitigate the biases in pretrained lan- guage models. Different from previous debi- asing work that uses external corpora to fine- tune the pretrained models, we instead directly probe the biases encoded in pretrained models through prompts. Specifically, we propose a variant of the beam search method to automat- ically search for biased prompts such that the cloze-style completions are the most different with respect to different demographic groups. Given the identified biased prompts, we then propose a distribution alignment loss to miti- gate the biases. Experiment results on stan- dard datasets and metrics show that our pro- posed Auto-Debias approach can significantly reduce biases, including gender and racial bias, in pretrained language models such as BERT, RoBERTa and ALBERT. Moreover, the im- provement in fairness does not decrease the language models’ understanding abilities, as shown using the GLUE benchmark. 1 Introduction Pretrained language models (PLMs), such as masked language models (MLMs), have achieved remarkable success in many natural language pro- cessing (NLP) tasks (Devlin et al., 2019; Liu et al., 2019; Lan et al., 2020; Brown et al.). Unfor- tunately, pretrained language models, which are trained on large human-written corpora, also in- herit human-like biases and undesired social stereo- types (Caliskan et al., 2017; Bolukbasi et al., 2016; Blodgett et al., 2020). For example, in the fill-in- the-blank task, BERT (Devlin et al., 2019) substi- tutes [MASK] in the sentence “The man/woman had a job as [MASK]” with “manager/receptionist” respectively, reflecting occupational gender bias. The human-like biases and stereotypes encoded in PLMs are worrisome as they can be propagated or even amplified in downstream NLP tasks such as sentiment classification (Kiritchenko and Moham- mad, 2018), co-reference resolution (Zhao et al., 2019; Rudinger et al., 2018), clinical text classifica- tion (Zhang et al., 2020) and psychometric analysis (Abbasi et al., 2021; Ahmad et al., 2020). However, although it is important to mitigate biases in PLMs, debiasing masked language mod- els such as BERT is still challenging, because the biases encoded in the contextualized models are hard to identify. To address this challenge, previous efforts seek to use additional corpora to retrieve the contextualized embeddings or locate the biases and then debias accordingly. For example, Liang et al. (2020); Kaneko and Bollegala (2021); Garimella et al. (2021) use external corpora to locate sen- tences containing the demographic-specific words (e.g., man and women) or stereotype words (e.g., manager …

Content truncated to 500 words