Multilingual and Multi-Aspect Hate Speech Analysis

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4675–4684, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4675 Multilingual and Multi-Aspect Hate Speech Analysis Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, Dit-Yan Yeung Department of Computer Science and Engineering The Hong Kong University of Science and Technology nousidhoum@cse.ust.hk, zlinai@connect.ust.hk, hzhangal@cse.ust.hk, yqsong@cse.ust.hk, dyyeung@cse.ust.hk Abstract Current research on hate speech analysis is typically oriented towards monolingual and single classification tasks. In this paper, we present a new multilingual multi-aspect hate speech analysis dataset and use it to test the current state-of-the-art multilingual multitask learning approaches. We evaluate our dataset in various classification settings, then we dis- cuss how to leverage our annotations in order to improve hate speech detection and classifi- cation in general. 1 Introduction With the expanding amount of text data generated on different social media platforms, current fil- ters are insufficient to prevent the spread of hate speech. Most internet users involved in a study conducted by the Pew Research Center report hav- ing been subjected to offensive name calling on- line or witnessed someone being physically threat- ened or harassed online.1 Additionally, Amnesty International within Element AI have lately re- ported that many women politicians and journal- ists are assaulted every 30 seconds on Twitter.2 This is despite the Twitter policy condemning the promotion of violence against people on the basis of race, ethnicity, national origin, sexual orienta- tion, gender identity, religious affiliation, age, dis- ability, or serious disease.3 Hate speech may not represent the general opinion, yet it promotes the dehumanization of people who are typically from minority groups (Soral et al., 2017; Martin et al., 2012) and can incite hate crime (Ross et al., 2017). Moreover, although people of various linguistic backgrounds are exposed to hate speech (Waseem 1http://www.pewinternet.org/2017/07/11/online- harassment-2017/ 2https://www.amnesty.org.uk/press-releases/women- abused-twitter-every-30-seconds-new-study 3https://help.twitter.com/en/rules-and-policies/hateful- conduct-policy et al., 2017; Ross et al., 2017), English is still at the center of existing work on toxic language anal- ysis. Recently, some research studies have been conducted on languages such as German (Kratzke, 2017), Arabic (Albadi et al., 2018), and Ital- ian (Sanguinetti et al., 2018). However, such stud- ies usually use monolingual corpora and do not contrast, or examine the correlations between on- line hate speech in different languages. On the other hand, tasks involving more than one lan- guage such as the hatEval task4, which covers En- glish and Spanish, include only separate classifica- tion tasks, namely (a) women and immigrants as target groups, (b) individual or generic hate and, (c) aggressive or non-aggressive hate speech. Treating hate speech classification as a binary task may not be enough to inspect the motivation and the behavior of the users promoting it and, how people would react to it. For instance, the hateful tweets presented in Figure 1 show toxicity directed towards different targets, with or without using slurs, and generating several types of reac- tions. We believe that, in order to …

Content truncated to 500 words