macech at SemEval-2021 Task 5: Toxic Spans Detection

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 1003–1008 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 1003 macech at SemEval-2021 Task 5: Toxic Spans Detection Maggie Cech SAS Institute margaret.cech98@gmail.com Abstract Toxic language is often present in online fo- rums, especially when politics and other po- larizing topics arise, and can lead to people becoming discouraged from joining or contin- uing conversations. In this paper, I use data consisting of comments with the indices of toxic text labelled to train an RNN to deter- mine which parts of the comments make them toxic, which could aid online moderators. I compare results using both the original dataset and an augmented set, as well as GRU versus LSTM RNN models. 1 Introduction In this digital era we live in, almost everyone is communicating online. As of January 2021, Facebook, YouTube, and WhatsApp each have over 2 billion users, which means many differing viewpoints and perspectives being shared (Statista, 2021). With such a huge exchange of ideas, there is bound to be some toxicity within the comments. Aside from discouraging users to continue with or join conversations, toxic comments can also taint users’ perceptions on news sites (Tenenboim et al., 2019). Thus it is important to moderate online conversations without fully censoring users. While forums typically rely on human moder- ators, with such vast amounts of data coming in, it can be difficult for humans to keep up (Nobata et al., 2016). Advances in deep learning and ma- chine learning is making text processing a viable option to replace, or at least assist, human mod- erators clean up comment sections (Consultants, 2019). Some methods rely on simply classifying whether a comment is toxic or not, but identifying what parts of the text are actually toxic can assist moderators and provide insight into what makes language toxic. The SemEval task 5 aims to evalu- ate systems that detect toxic spans wihtin text using datasets where spans within the comments are la- belled as toxic, differing from previously released datasets where whole comments were labelled as toxic or non-toxic (Pavlopoulos et al., 2021). This is inherently a natural language processing task, similar to text classification and sentiment analysis. This study focuses on training a recurrent neural network to determine the indices of a given string that represent the toxic portions of a com- ment. Recurrent neural networks are classically used for natural language and sequence labelling task, and one could view this task as a form of se- quence labelling. The goal of sequence labelling is, given a sequence as input, assign a sequence of labels. Because recurrent neural networks (RNNs) are flexible in their use of context information and can recognize sequential patterns, they are an at- tractive and commonly used choice in sequence labelling (Graves, 2012). This paper approaches the task at hand with a sequence labelling method- ology, applying an RNN and comparing the use of gated reccurent unit (GRU) and long-short term memory unit …

Content truncated to 500 words