WLV-RIT at SemEval-2021 Task 5: A Neural Transformer Framework for Detecting Toxic Spans

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 833–840 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 833 WLV-RIT at SemEval-2021 Task 5: A Neural Transformer Framework for Detecting Toxic Spans Tharindu Ranasinghe1, Diptanu Sarkar2, Marcos Zampieri2, Alexander Ororbia2 1University of Wolverhampton,UK 2Rochester Institute of Technology, USA T.D.RanasingheHettiarachchige@wlv.ac.uk Abstract In recent years, the widespread use of social media has led to an increase in the generation of toxic and offensive content on online plat- forms. In response, social media platforms have worked on developing automatic detec- tion methods and employing human modera- tors to cope with this deluge of offensive con- tent. While various state-of-the-art statistical models have been applied to detect toxic posts, there are only a few studies that focus on de- tecting the words or expressions that make a post offensive. This motivates the organization of the SemEval-2021 Task 5: Toxic Spans De- tection competition, which has provided par- ticipants with a dataset containing toxic spans annotation in English posts. In this paper, we present the WLV-RIT entry for the SemEval- 2021 Task 5. Our best performing neural trans- former model achieves an 0.68 F1-Score. Fur- thermore, we develop an open-source frame- work for multilingual detection of offensive spans, i.e., MUDES, based on neural trans- formers that detect toxic spans in texts. 1 Introduction The widespread adoption and use of social media has led to a drastic increase in the generation of abusive and profane content on the web. To counter this deluge of negative content, social media com- panies and government institutions have turned to developing and applying computational mod- els that can identify the various forms of offensive content online such as aggression (Kumar et al., 2018, 2020), cyber-bullying (Rosa et al., 2019), and hate speech (Ridenhour et al., 2020). Prior work has either designed methods for identifying conversations that are likely to go awry (Zhang WARNING: This paper contains text excerpts and words that are offensive in nature. et al., 2018; Chang et al., 2020) or detecting of- fensive content and labelling posts at the instances level – this has been the focus in the recent shared tasks like HASOC at FIRE 2019 (Mandl et al., 2019a) and FIRE 2020 (Mandl et al., 2020), Ger- mEval 2019 Task 2 (Struß et al., 2019), TRAC (Kumar et al., 2018, 2020), HatEval (Basile et al., 2019a), OffensEval at SemEval-2019 (Zampieri et al., 2019b) and SemEval-2020 (Zampieri et al., 2020). With respect to identifying offensive language in conversations, comments, and posts, noticeable progress has been made with a variety of large, annotated datasets made available in recent years (Pitenis et al., 2020; Rosenthal et al., 2020). The identification of the particular text spans that make a post offensive, however, has been mostly ne- glected (Mathew et al., 2021) as current state-of- the-art offensive language identification models flag the entire post or comment but do not actually highlight the offensive parts. The pressing need for toxic span detection models to assist human …

Content truncated to 500 words