MedAI at SemEval-2021 Task 5: Start-to-end Tagging Framework for Toxic Spans Detection

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 258–262 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 258 MedAI at SemEval-2021 Task 5: Start-to-end Tagging Framework for Toxic Spans Detection Zhen Wang1, Hongjie Fan2 ∗, Junfei Liu3 School of Software and Microelectronics, Peking University The Department of Science and Technology, China University of Political Science and Law National Engineering Research Center for Software Engineering, Peking University {wang.zh, liujunfei}@pku.edu.cn , hjfan@cupl.edu.cn Abstract This paper describes the system submitted to SemEval 2021 Task 5: Toxic Spans Detec- tion. The task concerns evaluating systems that detect the spans that make a text toxic when detecting such spans are possible. To ad- dress the possibly multi-span detection prob- lem, we develop a start-to-end tagging frame- work on the top of RoBERTa based language model. Besides, we design a custom loss func- tion which take distance into account. In com- parison to other participating teams, our sys- tem has achieved 69.03% F1 score, which is slight lower (-1.8 and -1.73) than the top 1 (70.83%) and top 2 (70.77%), respectively. 1 Introduction In recent years, social networks and microblog- ging sites’ popularity have increased, attracting more users. With a huge user base, social media will continue to publish a large amount of user- generated content. As the use of social media increased, other undesirable phenomena and be- haviors emerged. Social media users often abuse this freedom to spread abusive or hateful posts or comments. In many cases, the user-generated con- tent is offensive or proactive, and users may have to deal with threats such as cyberattacks or cyberbully- ing, and other undesirable (Warner and Hirschberg 2012). Therefore, the issue of detecting and pos- sibly limiting the spread of toxic post has become increasingly important. Although several toxicity or abusive language de- tection datasets (Wulczyn et al. 2016; Borkan et al. 2019) and models (Borkan et al. 2019; Pavlopou- los et al. 2017; Zampieri et al. 2019) have been released, most of them classify whole comments or documents, and do not identify the spans that make a text toxic. But highlighting such toxic spans can * Corresponding author. assist human moderators (e.g., news portals mod- erators) who often deal with lengthy comments, and who prefer attribution instead of just a system- generated unexplained toxicity score per post. The evaluation of systems that could accurately locate toxic spans within a text is thus a crucial step to- wards successful semi-automated moderation. For this reason, SemEval 2021 set up the task Toxic Spans Detection to detect and extract the spans that make a text toxic, when detecting such spans is possible (Pavlopoulos et al. 2021). To ad- dress the possibly multi-span extraction problem, we develop a start-to-end tagging framework with custom distance loss, which can tag the start and end position of a toxic span. Based on this scheme, we can effectively deal with the multi-span extrac- tion problem. The rest of the paper is organized as follows: Section 2 provides system …

Content truncated to 500 words