Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5517 - 5529 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Improving Generalizability in Implicitly Abusive Language Detection with Concept Activation Vectors Isar Nejadgholi, Kathleen C. Fraser, and Svetlana Kiritchenko National Research Council Canada Ottawa, Canada {Isar.Nejadgholi,Kathleen.Fraser,Svetlana.Kiritchenko}@nrc-cnrc.gc.ca Abstract Robustness of machine learning models on ever-changing real-world data is critical, es- pecially for applications affecting human well- being such as content moderation. New kinds of abusive language continually emerge in on- line discussions in response to current events (e.g., COVID-19), and the deployed abuse de- tection systems should be updated regularly to remain accurate. In this paper, we show that general abusive language classifiers tend to be fairly reliable in detecting out-of-domain ex- plicitly abusive utterances but fail to detect new types of more subtle, implicit abuse. Next, we propose an interpretability technique, based on the Testing Concept Activation Vector (TCAV) method from computer vision, to quantify the sensitivity of a trained model to the human- defined concepts of explicit and implicit abu- sive language, and use that to explain the gen- eralizability of the model on new data, in this case, COVID-related anti-Asian hate speech. Extending this technique, we introduce a novel metric, Degree of Explicitness, for a single in- stance and show that the new metric is bene- ficial in suggesting out-of-domain unlabeled examples to effectively enrich the training data with informative, implicitly abusive texts. 1 Introduction When machine learning models are deployed in the real world, they must be constantly monitored for their robustness to new and changing input data. One area where this is particularly important is in abusive language detection (Schmidt and Wie- gand, 2017; Fortuna and Nunes, 2018; Nakov et al., 2021; Vidgen and Derczynski, 2020). The con- tent of online conversation is constantly changing in response to political and social events. New categories of abusive language emerge, encompass- ing topics and vocabularies unknown to previously trained classifiers. Here, we tackle three main ques- tions: How can a human user formalize new, rele- vant topics or concepts in text? How do we quantify the sensitivity of a trained classifier to these new concepts as they emerge? And how do we update the classifier so that it remains reliable? As a case study, we consider the rise of COVID- related anti-Asian racism on social media. The COVID-19 pandemic represented an entirely new and unexpected situation, generating new vocab- ulary (COVID-19, coronavirus, social distancing, masking), new topics of conversation (dealing with isolation, working from home), and – unfortunately – new and renewed instances of hate speech directed towards Asian communities. We imagine the case of an abusive language detection algorithm which had been deployed prior to the pandemic: what are the new types of abusive language that have emerged with the recent pandemic? To what extent can deployed classifiers generalize to this new data, and how can they be adapted? Although social events can spark off a specific type of …

Content truncated to 500 words