What social attitudes about gender does BERT encode? Leveraging insights from psycholinguistics

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6790–6809 July 9-14, 2023 ©2023 Association for Computational Linguistics What social attitudes about gender does BERT encode? Leveraging insights from psycholinguistics Julia Watson1 Barend Beekhuizen2 Suzanne Stevenson1 1Department of Computer Science University of Toronto {jwatson, suzanne}@cs.toronto.edu 2Department of Language Studies University of Toronto, Mississauga barend.beekhuizen@utoronto.ca Abstract Much research has sought to evaluate the de- gree to which large language models reflect social biases. We complement such work with an approach to elucidating the connections be- tween language model predictions and people’s social attitudes. We show how word prefer- ences in a large language model reflect social attitudes about gender, using two datasets from human experiments that found differences in gendered or gender neutral word choices by par- ticipants with differing views on gender (pro- gressive, moderate, or conservative). We find that the language model BERT takes into ac- count factors that shape human lexical choice of such language, but may not weigh those fac- tors in the same way people do. Moreover, we show that BERT’s predictions most resemble responses from participants with moderate to conservative views on gender. Such findings il- luminate how a language model: (1) may differ from people in how it deploys words that sig- nal gender, and (2) may prioritize some social attitudes over others. 1 Introduction Language choices are revealing about speakers’ so- cial attitudes – their (evaluative) beliefs, views, and expectations about social phenomena. If a café advertises “gingerbread people,” instead of “gin- gerbread men” (example adapted from Papineau et al., 2022), people may make inferences about the social views of the café owners based on their avoidance of the traditional masculine term. So- cial attitudes typically surface in less “pointed” but higher stakes scenarios, such as a speaker using the pronoun they to refer to a colleague who identifies as nonbinary, reflecting the speaker’s acceptance of nonbinary identities. Much work on the social knowledge encoded in language technology has focused on evaluating whether models encode stereotypical/harmful asso- ciations (e.g., Caliskan et al., 2017; Rudinger et al., 2018), and if so, removing them to “de-bias” NLP (e.g., Bolukbasi et al., 2016; Zhao et al., 2018). However, social knowledge permeates language (e.g., Nguyen et al., 2021), and what counts as harmful depends on one’s perspective (e.g., Blod- gett et al., 2020). To deal effectively with poten- tially harmful associations in NLP, we need a clear understanding of how social attitudes are linked to the language choices people make, so that we can assess the language choices of our technologies. Here we seek to understand what social atti- tudes a large language model encodes, specifi- cally social attitudes about gender. To address this question, we draw on datasets from two psycholin- guistics studies, both of which included language tasks involving gendered and gender neutral lan- guage choices, and surveys eliciting the same par- ticipants’ social attitudes on gender. By explicitly linking people’s language choices with their social attitudes, …

Content truncated to 500 words