HamiltonDinggg at SemEval-2021 Task 5: Investigating Toxic Span Detection using RoBERTa Pre-training

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 263–269 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 263 HamiltonDinggg at SemEval-2021 Task 5: Investigating Toxic Span Detection using RoBERTa Pre-training Huiyang Ding School of Information University of Michigan Ann Arbor, MI 48105 huiyangd@umich.edu David Jurgens School of Information University of Michigan Ann Arbor, MI 48105 jurgens@umich.edu Abstract This paper presents our system submission to task 5: Toxic Spans Detection of the SemEval- 2021 competition. The competition aims at detecting the spans that make a toxic span toxic. In this paper, we demonstrate our system for detecting toxic spans, which in- cludes expanding the toxic training set with Local Interpretable Model-Agnostic Explana- tions (LIME), fine-tuning RoBERTa model for detection, and error analysis. We found that feeding the model with an expanded train- ing set using Reddit comments of polarized- toxicity and labeling with LIME on top of logistic regression classification could help RoBERTa more accurately learn to recognize toxic spans. We achieved a span-level F1 score of 0.6715 on the testing phase. Our quantita- tive and qualitative results show that the pre- dictions from our system could be a good sup- plement to the gold training set’s annotations. 1 Introduction Toxic messages remain a small but persistent part of online communications (Fortuna and Nunes, 2018; Jurgens et al., 2019). NLP methods have been developed to identify these comments, often rely- ing on deep-language models (Vidgen et al., 2019). However, the part of the message that is specifi- cally toxic is often unknown. Such information is useful not only for validating and explaining the judgments of models (Carton et al., 2018), but can also be useful for moderators to use when making decisions and working with these models in their deployment (Carton et al., 2020; Liu et al., 2021). This paper describes our model1 and error analysis for SemEval-2021 Task 5: Toxic Spans Detection (Pavlopoulos et al., 2021). Our model uses a deep learning approach to identify which tokens are toxic. The approach 1The code is available at https://github.com/ davidjurgens/offensive-span-detection. is motivated by two strands of prior work show- ing (1) that large language models can effectively serve as sequence-to-sequence (seq2seq) models and (2) that pre-training on a similar task can im- prove downstream performance (Phang et al., 2018; Gururangan et al., 2020). Here, we treat the toxic- span detection tasks as a seq2seq task, where given a sequence of tokens, the model outputs per-token judgments of whether the token is in the toxic span. Given the limited training data for Task 5, we increase our training data by generating a silver-standard set of span judgments from LIME explanations (Ribeiro et al., 2016) fropm a model trained to recognize toxic and non-toxic language. These additional judgments are intended to help the model learn the basic span recognition task and identify general toxic language, before fine-tuning on the Task 5 data. 2 System Description Our core system relies on a standard RoBERTa model (Liu et al., 2019) …

Content truncated to 500 words