Listening to Affected Communities to Define Extreme Speech: Dataset and Experiments

Findings of the Association for Computational Linguistics: ACL 2022, pages 1089 - 1104 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Listening to Affected Communities to Define Extreme Speech: Dataset and Experiments Antonis Maronikolakis1∗ Axel Wisiorek1,2 Leah Nann3 Haris Jabbar1 Sahana Udupa3 Hinrich Schütze1 1CIS, Center for Information and Language Processing 2Center for Digital Humanities 3Institute of Social and Cultural Anthropology LMU Munich ∗antmarakis@cis.lmu.de Abstract Building on current work on multilingual hate speech (e.g., Ousidhoum et al. (2019)) and hate speech reduction (e.g., Sap et al. (2020)), we present XtremeSpeech,1 a new hate speech dataset containing 20,297 social media pas- sages from Brazil, Germany, India and Kenya. The key novelty is that we directly involve the affected communities in collecting and anno- tating the data – as opposed to giving com- panies and governments control over defining and combatting hate speech. This inclusive approach results in datasets more representa- tive of actually occurring online speech and is likely to facilitate the removal of the social media content that marginalized communities view as causing the most harm. Based on XtremeSpeech, we establish novel tasks with accompanying baselines, provide evidence that cross-country training is generally not feasi- ble due to cultural differences between coun- tries and perform an interpretability analysis of BERT’s predictions. 1 Introduction Much effort has been devoted to curating data in the area of hate speech, from foundational work (Waseem and Hovy, 2016; Davidson et al., 2017) to more recent, broader (Sap et al., 2020) as well as multilingual (Ousidhoum et al., 2019) approaches. However, the demographics of those targeted by hate speech and those creating datasets are often quite different. For example, in Founta et al. (2018), 66% of annotators are male and in Sap et al. (2020), 82% are white. This may lead to unwanted bias (e.g., disproportionately labeling African Ameri- can English as hateful (Sap et al., 2019; Davidson et al., 2019a)) and to collection of data that is not representative of the comments directed at target groups; e.g., a white person may not see and not 1Code and data available at https://github.com/ antmarakis/xtremespeech Figure 1: Overview of hate speech data collection. In- stead of querying for data on our own, we work with fact-checkers advocating for targeted communities who collect and label data as they organically come across it. This inclusive approach results in datasets more repre- sentative of online speech the communities are exposed to. See §3.2 for definition of XtremeSpeech labels. have access to hate speech targeting a particular racial group. An example from our dataset is the Kenyan so- cial media post “… We were taught that such horri- ble things can only be found in Luo Nyanza.” The Luo are an ethnic group in Kenya; Nyanza is a Kenyan province. The post is incendiary because it suggests that the Luo are responsible for horri- ble things, insinuating that retaliation against them may be justified. Only a group of people deeply rooted in Kenya can collect such examples and un- derstand their …

Content truncated to 500 words