CSECU-DSG at SemEval-2021 Task 5: Leveraging Ensemble of Sequence Tagging Models for Toxic Spans Detection

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 990–994 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 990 CSECU-DSG at SemEval-2021 Task 5: Leveraging Ensemble of Sequence Tagging Models for Toxic Spans Detection Tashin Hossain, Jannatun Naim, Fareen Tasneem, Radiathun Tasnia, and Abu Nowshed Chy Department of Computer Science and Engineering University of Chittagong, Chattogram-4331, Bangladesh {tashin.hossain.cu,jannatun.naim.cu,fareen.tasneem, radia.tasnia.cu}@gmail.com and nowshed@cu.ac.bd Abstract The upsurge of prolific blogging and mi- croblogging platforms enabled the abusers to spread negativity and threats greater than ever. Detecting the toxic portions substantially aids to moderate or exclude the abusive parts for maintaining sound online platforms. This pa- per describes our participation in the SemEval 2021 toxic span detection task. The task re- quires detecting spans that convey toxic re- marks from the given text. We explore an ensemble of sequence labeling models includ- ing the BiLSTM-CRF, spaCy NER model with custom toxic tags, and fine-tuned BERT model to identify the toxic spans. Finally, a major- ity voting based fusion method is used to de- termine the unified toxic spans. Experimental results depict the competitive performance of our model among the participants. 1 Introduction Social media being a key factor in the world dy- namics and toxicity in user-generated contents is a real threat. Threats and hatred instigated in posts and blogs implants fear in users’ minds and pre- vents them from sharing their creative thoughts, valuable opinions to critical information. Some- times it leads to severe mental trauma and fatalities. Hence, it is a formidable task to precisely detect toxicity in comments and posts to be able to mod- erate those portions and provide the users a safe online platform to express themselves. Toxic span detection is a process where the spe- cific toxic segment of a text is detected instead of detecting the whole text as toxic. The goal of this task is to eradicate the vagueness that is present in simple toxic text classification models and help the moderator to precisely moderate the toxic portions instead of the whole post. To elucidate the task, two examples are presented in Table 1. The first four authors have equal contributions. Text#1: How fucking stupid are you? Span: [4, 5, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17] Text#2: What a sociopathic and parasitic leader we have. Span: [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 23, 24, 25, 26, 27, 28, 29, 30, 31] Table 1: Example of sample texts with toxic spans. Here, the “fucking stupid” portion of Text#1 is toxic and is attacking the personality of the second person, so the indices of this portion are included in the toxic span. In Text#2, the “sociopathic and parasitic” fragment is used as a toxic adjective to describe the leader in that context. Consequently, the indices of this fragment are incorporated in the span. We need to detect such spans accurately to remove toxicity from user content and preserve the safe and sound flow …

Content truncated to 500 words