How people talk about each other: Modeling Generalized Intergroup Bias and Emotion

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2496–2506 May 2-6, 2023 ©2023 Association for Computational Linguistics How people talk about each other: Modeling Generalized Intergroup Bias and Emotion Venkata S. Govindarajan1 Katherine Atwell2 Barea Sinno3 Malihe Alikhani2 David I. Beaver1 Junyi Jessy Li1 1 Department of Linguistics, The University of Texas at Austin 2 Department of Computer Sciences, University of Pittsburgh 3 Department of Political Science, Rutgers University venkatasg@utexas.edu, kaa139@pitt.edu, barea.sinno@gmail.com, malihe@pitt.edu, dib@utexas.edu, jessy@utexas.edu Abstract Current studies of bias in NLP rely mainly on identifying (unwanted or negative) bias towards a specific demographic group. While this has led to progress recognizing and mitigating neg- ative bias, and having a clear notion of the tar- geted group is necessary, it is not always prac- tical. In this work we extrapolate to a broader notion of bias, rooted in social science and psy- chology literature. We move towards predicting interpersonal group relationship (IGR) — mod- eling the relationship between the speaker and the target in an utterance—using fine-grained interpersonal emotions as an anchor. We build and release a dataset of English tweets by US Congress members annotated for interpersonal emotion – the first of its kind, and ‘found su- pervision’ for IGR labels; our analyses show that subtle emotional signals are indicative of different biases. While humans can perform better than chance at identifying IGR given an utterance, we show that neural models perform much better; furthermore, a shared encoding between IGR and interpersonal perceived emo- tion enabled performance gains in both tasks. 1 Introduction Currently, most work studying bias in NLP situates bias as negative or pejorative language use towards an individual or group based on traits like race, gen- der, etc (Kaneko and Bollegala, 2019; Sheng et al., 2019; Sap et al., 2020; Webson et al., 2020; Pryzant et al., 2020; Sheng et al., 2020). While these ap- proaches greatly advance our understanding of bias in language and its impact and mitigation in NLP, focusing on specific demographic dimensions or an individual’s intent is limiting and not always prac- tical. Research in psychology and social science suggests a different perspective. Bias can be seen as a relationship between people and groups, situated in context (Van Dijk, 2009); as such, bias refers to differences in behavior (in this case language use) as a result of differences in the relationship between speaker and target. The language we produce is bi- ased in one way or another, whether we intend to or not, and whether that bias is positive, negative, or not clearly associated with any valuation (Beaver and Stanley, 2018). In psychological work on Linguistic Intergroup Bias (Maass, 1999), bias originates from the re- lationship between the speaker and target of an utterance, i.e. their interpersonal dynamics, and manifests later in subtle ways. Consider the utter- ances (tweets) in (1), drawn from our collected data in which the identity of the speaker and target are masked: (1) a. In-group: We stand w @Doe, …

Content truncated to 500 words