What Did You Learn To Hate? A Topic-Oriented Analysis of Generalization in Hate Speech Detection

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 3495–3508 May 2-6, 2023 ©2023 Association for Computational Linguistics What Did You Learn To Hate? A Topic-Oriented Analysis of Generalization in Hate Speech Detection Tom Bourgeade1, Patricia Chiril3, Farah Benamara1,2, Véronique Moriceau1 (1) IRIT, Université de Toulouse, CNRS, Toulouse INP, UT3, Toulouse, France {firstname.lastname}@irit.fr (2) IPAL, CNRS-NUS-ASTAR, Singapore (3) University of Chicago, Chicago, IL, USA pchiril@uchicago.edu Abstract Warning: This paper includes messages that may contain instances of vulgarity, degrading terms, or hate speech, which may be offensive or upsetting to some readers. Hate speech has unfortunately become a signif- icant phenomenon on social media platforms, and it can cover various topics (misogyny, sex- ism, racism, xenophobia, etc.) and targets (e.g., black people, women). Various hate speech de- tection datasets have been proposed, some an- notated for specific topics, and others for hate- ful speech in general. In either case, they often employ different annotation guidelines, which can lead to inconsistencies, even in datasets focusing on the same topics. This can cause issues in models trying to generalize across more data and more topics in order to improve detection accuracy. In this paper, we propose, for the first time, a topic-oriented approach to study generalization across popular hate speech datasets. We first perform a comparative analy- sis of the performances of Transformer-based models in capturing topic-generic and topic- specific knowledge when trained on different datasets. We then propose a novel, simple yet effective approach to study more precisely which topics are best captured in implicit mani- festations of hate, showing that selecting com- binations of datasets with better out-of-domain topical coverage improves the reliability of au- tomatic hate speech detection. 1 Introduction On social media and other online communication platforms, hate speech (HS hereafter) is found in many forms, from textual harassment to threats, targeting an individual or group (e.g., black peo- ple, women), based on some characteristics, such as race, color, ethnicity, gender, sexual orientation, nationality, religion, etc, which we refer to as top- ics. While the broad nature of these topics of HS is generally understood (Erjavec and Kovaˇciˇc, 2012), determining whether a social media post is a man- ifestation of HS (and if so, to which topic(s) it belongs to) is not a trivial task for humans and au- tomated machine-learning systems. Indeed, the lat- ter often require large quantities of annotated data, which in the case of HS, is made difficult due to (1) the absence of a comprehensive definition of these topics; but also, (2) the potential internal biases and subjectivity present in annotators and/or anno- tation guidelines employed to construct annotated datasets (Schmidt and Wiegand, 2017; Fortuna and Nunes, 2018; Vidgen et al., 2019; Fortuna et al., 2020). To palliate these issues, cross-dataset evalua- tion has become an active line of research aiming at studying the generalization capabilities of HS detection systems to unseen data during training (Talat et al., 2018; Ludwig et al., 2022; Toraman et al., 2022). …

Content truncated to 500 words