hub at SemEval-2021 Task 5: Toxic Span Detection Based on Word-Level Classification

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 904–908 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 904 hub at SemEval-2021 Task 5: Toxic Span Detection Based on Word-Level Classification Bo Huang, Yang Bai, Xiaobing Zhou* School of Information Science and Engineering Yunnan University, Yunnan, P.R. China *Corresponding author:zhouxb@ynu.edu.com Abstract This article introduces the system description of the hub team, which explains the related work and experimental results of our team’s participation in SemEval 2021 Task 5: Toxic Spans Detection. The data for this shared task comes from some posts on the Internet. The task goal is to identify the toxic content con- tained in these text data. We need to find the span of the toxic text in the text data as accu- rately as possible. In the same post, the toxic text may be one paragraph or multiple para- graphs. Our team uses a classification scheme based on word-level to accomplish this task. The system we used to submit the results is ALBERT+BILSTM+CRF. The result evalua- tion index of the task submission is the F1 score, and the final score of the prediction re- sult of the test set submitted by our team is 0.6640226029. 1 Introduction and Background From the popularization of the Internet to the first year of the mobile Internet in 2011, the number of online social media and social media users has con- tinued to grow. In the context of the ever-expanding user base, coupled with the free and interactive features of online social media communication. Therefore, online social media has exposed many issues worthy of our attention, such as the lack of communication standards and the out-of-control of information dissemination, which makes the dis- semination of online social media prone to various negative functions (Baccarella et al., 2018). The task of toxic span detection is to detect the span of text with toxic information in the text (Pavlopoulos et al., 2021). The goal of the task is to predict the beginning and ending character positions in the text as accurately as possible. Re- viewing the content in online media can effectively avoid the spread of a series of negative informa- tion such as cyber violence, cyberbullying, and false news. Audits are essential to promote healthy online discussions. However, the content and num- ber of posts in social media are too large, and the manual review method obviously cannot achieve a good effect. Therefore, in combination with the development of modern technology, achieving a semi-automatic audit is the best solution. 2 Related Work There are many different kinds of methods for iden- tifying negative information in social media, but usually, these methods mainly focus on supervised learning. Simple SurfaceFeatures similar to the bag of words model can provide very clear and easy-to-understand information in text processing tasks. The general approach of this method is to merge multiple larger n-grams into a feature set (Nobata et al., 2016; Djuric et al., 2015). Use the artificial neural network method to …

Content truncated to 500 words