Introducing CAD: the Contextual Abuse Dataset

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2289–2303 June 6–11, 2021. ©2021 Association for Computational Linguistics 2289 Introducing CAD: the Contextual Abuse Dataset Bertie Vidgen†, Dong Nguyen†⋆, Helen Margetts†‡, Patricia Rossini¶, Rebekah Tromble†§ † The Alan Turing Institute; ⋆ Utrecht University; ‡ University of Oxford; ¶ University of Liverpool; § George Washington University bvidgen@turing.ac.uk Abstract Online abuse can inflict harm on users and communities, making online spaces unsafe and toxic. Progress in automatically detecting and classifying abusive content is often held back by the lack of high quality and detailed datasets. We introduce a new dataset of primar- ily English Reddit entries which addresses sev- eral limitations of prior work. It (1) contains six conceptually distinct primary categories as well as secondary categories, (2) has labels annotated in the context of the conversation thread, (3) contains rationales and (4) uses an expert-driven group-adjudication process for high quality annotations. We report several baseline models to benchmark the work of fu- ture researchers. The annotated dataset, anno- tation guidelines, models and code are freely available. 1 Introduction Social media platforms have enabled unprece- dented connectivity, communication and interac- tion for their users. However, they often harbour harmful content such as abuse and hate, inflicting myriad harms on online users (Waseem and Hovy, 2016; Schmidt and Wiegand, 2017a; Fortuna and Nunes, 2018; Vidgen et al., 2019c). Automated techniques for detecting and classifying such con- tent increasingly play an important role in moder- ating online spaces. Detecting and classifying online abuse is a com- plex and nuanced task which, despite many ad- vances in the power and availability of computa- tional tools, has proven remarkably difficult (Vid- gen et al., 2019a; Wiegand et al., 2019; Schmidt and Wiegand, 2017b; Waseem et al., 2017). As Jurgens et al. (2019) argued in a recent review, research has ‘struggled to move beyond the most obvious tasks in abuse detection.’ One of the biggest bar- riers to creating higher performing, more robust, nuanced and generalisable classification systems is the lack of clearly annotated, large and detailed training datasets. However, creating such datasets is time-consuming, complicated and expensive, and requires a mix of both social and computational ex- pertise. We present a new annotated dataset of ∼25,000 Reddit entries. It contains four innovations that address limitations of previous labelled abuse datasets. First, we present a taxonomy with six conceptually distinct primary categories (Identity-directed, Person-directed, Affiliation- directed, Counter Speech, Non-hateful Slurs and Neutral). We also provide salient subcategories, such as whether personal abuse is directed at a person in the conversation thread or to someone outside it. This taxonomy offers greater coverage and granularity of abuse than previous work. Each entry can be assigned to multiple primary and/or secondary categories (Section 3). Second, we annotate content in context, by which we mean that each entry is annotated in the context of the conversational thread it is part of. Every annotation has a label for whether contex- tual …

Content truncated to 500 words