Social-Group-Agnostic Bias Mitigation via the Stereotype Content Model

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4123–4139 July 9-14, 2023 ©2023 Association for Computational Linguistics Social-Group-Agnostic Bias Mitigation via the Stereotype Content Model Ali Omrani† Alireza S. Ziabari† Charles Yu§ Preni Golazizian† Brendan Kennedy† Mohammad Atari† Heng Ji§ Morteza Dehghani† † University of Southern California § University of Illinois at Urbana-Champaign {aomrani, salkhord, golazizi, btkenned, atari, mdehghan}@usc.edu {ctyu2, hengji}@illinois.edu Abstract Existing bias mitigation methods require social- group-specific word pairs (e.g., “man” – “woman”) for each social attribute (e.g., gen- der), restricting the bias mitigation to only one specified social attribute. Further, this constraint renders such methods impractical and costly for mitigating bias in understud- ied and/or unmarked social groups. We pro- pose that the Stereotype Content Model (SCM) — a theoretical framework developed in so- cial psychology for understanding the content of stereotyping — can help debiasing efforts to become social-group-agnostic by capturing the underlying connection between bias and stereotypes. SCM proposes that the content of stereotypes map to two psychological di- mensions of warmth and competence. Using only pairs of terms for these two dimensions (e.g., warmth: “genuine” – “fake”; compe- tence: “smart” – “stupid”), we perform debi- asing with established methods on both pre- trained word embeddings and large language models. We demonstrate that our social-group- agnostic, SCM-based debiasing technique per- forms comparably to group-specific debiasing on multiple bias benchmarks, but has theoret- ical and practical advantages over existing ap- proaches. 1 Introduction The societal impacts of Natural Language Process- ing (NLP) have stimulated research on measuring and mitigating the unintended social-group biases encoded in language models (Hovy and Spruit, 2016). However, the majority of this important line of work is atheoretical in nature and “fails to engage critically with what constitutes ‘bias’ in the first place” (Blodgett et al., 2020). The bias found in language models is rooted in human bi- ases (Caliskan and Lewis, 2022); thus, to alleviate such biases, we should ground our debiasing ap- proaches in social psychological theories of stereo- typing. These theories can help us shed light on the Group-Specific Group Agnostic (Our Approach) Debiasing Goal Group Specific Terms Gender men,woman Race White,black,asian Gender Race Religion Age Gender Race Religion Age Alg. Alg. Gender Race Religion Age Alg. SCM terms genuine,fake smart,stupid Warmth Competence competence Jewish Middle class Rich warmth Women Christians Whites Gay men Blacks Muslims Hispanics Men Asians Jewish Middle class Rich warmth competence Women Christians Whites Gay men Blacks Muslims Hispanics Men Asians W C W W C C Mitigated Bias General Debiasing Figure 1: Group-specific debiasing vs. our proposed group-agnostic approach. Rather than iteratively debias- ing with respect to each social attribute (e.g., gender or race), embeddings or language models are debiased with respect to warmth and competence, the two dimensions of the Stereotype Content Model (SCM). underlying structure of language-embedded biases rather than attending to ad hoc superficial patterns (Osborne et al., 2022). Although there is a multitude of approaches to bias mitigation (Bolukbasi et al., …

Content truncated to 500 words