Directions for NLP Practices Applied to Online Hate Speech Detection

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11794 - 11805 December 7-11, 2022 ©2022 Association for Computational Linguistics Directions for NLP Practices Applied to Online Hate Speech Detection Paula Fortuna, Mónica Domínguez NLP Group Pompeu Fabra University Barcelona, Spain first.last@upf.edu Leo Wanner ICREA and Pompeu Fabra University Barcelona, Spain leo.wanner@upf.edu Zeerak Talat Simon Fraser University Burnaby, Vancouver zeerak_talat@sfu.ca Abstract Addressing hate speech in online spaces has been conceptualized as a classification task that uses Natural Language Processing (NLP) tech- niques. Through this conceptualization, the hate speech detection task has relied on com- mon conventions and practices from NLP. For instance, inter-annotator agreement is concep- tualized as a way to measure dataset quality and certain metrics and benchmarks are used to assure model generalization. However, hate speech is a deeply complex and situated con- cept that eludes such static and disembodied practices. In this position paper, we critically reflect on these methodologies for hate speech detection, we argue that many conventions in NLP are poorly suited for the problem and en- courage researchers to develop methods that are more appropriate for the task. 1 Introduction Online hate speech is the cause for growing con- cern, due to its social impacts (Soral et al., 2018). In response, automatic detection of hate speech has become a popular research area in Natural Lan- guage Processing (NLP), with a substantial number of papers at leading conferences, targeted work- shops and shared tasks dedicated to it. Simultane- ously, content moderation infrastructures that rely on machine learning technologies have been pro- posed to address the propagation of online harms. The increase in research interest to hate speech detection has spurred on a growth and variety in annotated resources for the task created within the academy and industry. However, at the same time, critical work on hate speech detection has found that there are significant challenges related to the published research outcome with respect to the construction of data (Vidgen and Derczynski, 2021; Fortuna et al., 2020), model generalizability (Nozza, 2021; Fortuna et al., 2021), and socially bi- ased effects of models (Talat et al., 2018; Davidson et al., 2019). These challenges indicate that despite a techno-optimist attitude from the NLP commu- nity to the task (Talat et al., 2021), the acceleration in research on the topic has impeded a maturation of the basic conventions, principles and practices which this research is based on. Hence, it seems timely to assess whether the NLP conventions and practices that have been followed so far are indeed appropriate for the classification of hate speech. In contrast to prior work, which has sought to address specific shortcomings of machine learning algorithms for hate speech detection (e.g., Dixon et al., 2018), our work presents a comprehensive analysis of the inadequacies of some of NLP’s methodological conventions for algorithmic hate speech detection. That is, in this position paper we reflect on contemporary methodological challenges that arise from applying common NLP methodolo- gies to hate speech detection. More specifically, …

Content truncated to 500 words