SkoltechNLP at SemEval-2021 Task 5: Leveraging Sentence-level Pre-training for Toxic Span Detection

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 927–934 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 927 SkoltechNLP at SemEval-2021 Task 5: Leveraging Sentence-level Pre-training for Toxic Span Detection David Dale‡, Igor Markov‡,†, Varvara Logacheva‡, Olga Kozlova†, Nikita Semenov†, and Alexander Panchenko‡ ‡Skolkovo Institute of Science and Technology, Moscow, Russia †Mobile TeleSystems (MTS), Moscow, Russia {d.dale,igor.markov,v.logacheva,a.panchenko}@skoltech.ru {oskozlo9,nikita.semenov}@mts.ru Abstract This work describes the participation of the Skoltech NLP group team (Sk) in the Toxic Spans Detection task at SemEval-2021. The goal of the task is to identify the most toxic fragments of a given sentence, which is a bi- nary sequence tagging problem. We show that fine-tuning a RoBERTa model for this prob- lem is a strong baseline. This baseline can be further improved by pre-training the RoBERTa model on a large dataset labeled for toxicity at the sentence level. While our solution scored among the top 20% participating models, it is only 2 points below the best result. This sug- gests the viability of our approach. 1 Introduction Toxicity and offensive content is a major concern for many platforms on the Internet. Therefore, the task of toxicity detection has attracted much atten- tion in the NLP community (Wulczyn et al., 2017; Hosseini et al., 2017; Dixon et al., 2018). Until recently, the majority of research on toxicity fo- cused on classifying entire user messages as toxic or safe. However, the surge of work on text de- toxification, i.e., editing of text to keep its content and remove toxicity (Nogueira dos Santos et al., 2018; Tran et al., 2020), suggests that localizing toxicity within a sentence is also useful. If we know which words of a sentence are toxic, it is easier to “fix” this sentence by removing or replac- ing them with non-toxic synonyms. Mathew et al. (2020) make human labelers annotate the spans as rationales for classifying a comment as hateful, offensive, or normal. They show that using such spans when training a toxicity classifier improves its accuracy and explainability and reduces unin- tended bias towards toxicity targets. This year the SemEval hosts the first competition on toxic spans detection, namely, SemEval-2021 Task 51 (Pavlopoulos et al., 2021). It provides training, development, and test data for English. As far as we know, it is the first attempt to explicitly formulate toxicity detection as sequence labeling instead of classification of sentences. Multiple NLP tasks recently benefited from transfer learning — transfer of probability distri- butions learned on some task to another model solving a different task. The most common ex- ample of transfer learning is the use of embeddings and language models pre-trained on unlabeled data (e.g. ELMo (Peters et al., 2018), BERT (Devlin et al., 2019) and its variations, T5 (Raffel et al., 2020), etc.) on other tasks (e.g. He et al. (2020); Wang et al. (2020) inter alia use pre-trained BERT models to perform tasks from the GLUE bench- mark (Wang et al., 2018)). Word-level toxicity classification can be for- mulated …

Content truncated to 500 words