ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Abuse Detection in Conversational AI

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7388–7403 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7388 ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Abuse Detection in Conversational AI Amanda Cercas Curry1 and Gavin Abercrombie1 and Verena Rieser1,2 1The Interaction Lab, Heriot-Watt University, Edinburgh 2Alana AI {ac293, g.abercrombie, v.t.rieser}@hw.ac.uk Abstract We present the first English corpus study on abusive language towards three conversational AI systems gathered ‘in the wild’: an open- domain social bot, a rule-based chatbot, and a task-based system. To account for the com- plexity of the task, we take a more ‘nuanced’ approach where our ConvAI dataset reflects fine-grained notions of abuse, as well as views from multiple expert annotators. We find that the distribution of abuse is vastly different compared to other commonly used datasets, with more sexually tinted aggression towards the virtual persona of these systems. Finally, we report results from bench-marking existing models against this data. Unsurprisingly, we find that there is substantial room for improve- ment with F1 scores below 90%. Warning: This paper contains examples of language that some people may find offensive or upsetting. 1 Introduction Abusive language detection has received extensive attention for social media, (see e.g. Vidgen et al., 2020a), but far less within the context of conversa- tional systems. As argued by UNESCO (West et al., 2019), detection and mitigation of abuse towards these (often anthropomorphised) AI systems is im- portant in order to avoid reinforcement of negative gender stereotypes. Following this report, several recent works have investigated possible abuse miti- gation strategies (Cercas Curry and Rieser, 2018, 2019; Chin and Yi, 2019; Ma et al., 2019). How- ever, the results of these studies are non-conclusive as they are not performed with live systems nor with real users – mainly because of the absence of reliable abuse detection tools. The majority of currently deployed systems use simple keyword spotting techniques, (e.g. Ram et al., 2018; Khatri et al., 2018), which tend to produce a high num- ber of false positives, such as cases in which the user expresses frustration, or use of profanities for emphasis, as well as false negatives, e.g. missing out on subtler forms of abuse (Han and Tsvetkov, 2020). Recently, Dinan et al. (2019); Xu et al. (2020) released an abuse detection tool trained on Wikipedia comments and crowd-sourced adversar- ial user prompts (the latter are not freely available). Whereas in this work, • We show that the distribution of abuse towards conversational systems is vastly different com- pared to other commonly used datasets, with more than half the instances containing sex- ism or sexual harassment. • We develop and release a detailed annotation scheme with the help of experts. • We use this scheme to annotate a corpus of 20k ratings on >6k samples (ca. 2k from each system), which we call ConvAbuse. We crit- ically discuss and experiment with different labelling methods for this task. We also re- lease a subset of 4k examples and their expert …

Content truncated to 500 words