ViHOS: Hate Speech Spans Detection for Vietnamese

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 652–669 May 2-6, 2023 ©2023 Association for Computational Linguistics ViHOS: Hate Speech Spans Detection for Vietnamese Phu Gia Hoang, Canh Duc Luu, Khanh Quoc Tran, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen University of Information Technology, Ho Chi Minh City, Vietnam Vietnam National University, Ho Chi Minh City, Vietnam {19520215, 19521272}@gm.uit.edu.vn {khanhtq, kietnv, ngannlt}@uit.edu.vn Abstract The rise in hateful and offensive language di- rected at other users is one of the adverse side effects of the increased use of social networking platforms. This could make it difficult for hu- man moderators to review tagged comments fil- tered by classification systems. To help address this issue, we present the ViHOS (Vietnamese Hate and Offensive Spans) dataset, the first human-annotated corpus containing 26k spans on 11k comments. We also provide definitions of hateful and offensive spans in Vietnamese comments as well as detailed annotation guide- lines. Besides, we conduct experiments with various state-of-the-art models. Specifically, XLM-RLarge achieved the best F1-scores in Single span detection and All spans detection, while PhoBERTLarge obtained the highest in Multiple spans detection. Finally, our error analysis demonstrates the difficulties in detect- ing specific types of spans in our data for future research. Our dataset is released on GitHub1. Disclaimer: This paper contains real com- ments that could be considered profane, of- fensive, or abusive. 1 Introduction Social networking sites have been widely used all over the world. Here, users can easily share their thoughts, connect with others, or earn money by selling items, creating content, and so on. Since these sites are universally accepted, many extreme users misuse comment functions to abuse other individuals or parties with hate and offensive lan- guage. Consequently, it has been proved that these types of speech could harm other users’ health (Mo- han et al., 2017; Anjum et al., 2018). Sometimes these behaviors can be considered cyberbullying, cyber threats, or online harassment. However, current studies are mainly about clas- sifying comments as a whole with binary labels 1https://github.com/phusroyal/ViHOS Figure 1: An example of the aid of spans detection for traditional foul language tagging system can provide additional insightful knowledge about tagged comments for human moderators. (Zampieri et al., 2019; Nguyen et al., 2021b) or multiple labeling schemes of abusive behaviors (Davidson et al., 2017; Founta et al., 2018; Mathur et al., 2018). These efforts are made to aid hu- man moderators, who need to review a massive number of online tagged comments that violate their community standards. However, a system that can highlight the spans that make a comment hateful or offensive can be more advantageous to human moderators who frequently deal with long and tedious comments and prefer explana- tions over a system-generated unexplained tag per comment. Furthermore, in some cases, using high- lighted spans and moderators’ context knowledge, they can take some actions to stop cyberbullying or online harassment. Nevertheless, there is only a study on toxic spans, SemEval-2021 Task 5: Toxic Spans Detection …

Content truncated to 500 words