Entity at SemEval-2021 Task 5: Weakly Supervised Token Labelling for Toxic Spans Detection

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 935–940 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 935 Entity at SemEval-2021 Task 5: Weakly Supervised Token Labelling for Toxic Spans Detection Vaibhav Jain New Delhi, India vaibhav29498@gmail.com Mina Naghshnejad Oakland, CA, USA mina.naghshnejad@gmail.com Abstract Detection of toxic spans - detecting toxicity of contents in the granularity of tokens - is crucial for effective moderation of online discussions. The baseline approach for this problem using the transformer model is to add a token classi- fication head to the language model and fine- tune the layers with the token labeled dataset. One of the limitations of such a baseline ap- proach is the scarcity of labeled data. To im- prove the results, We studied leveraging ex- isting public datasets for a related but differ- ent task of entire comment/sentence classifica- tion. We propose two approaches: the first ap- proach fine-tunes transformer models that are pre-trained on sentence classification samples. In the second approach, we perform weak su- pervision with soft attention to learn token level labels from sentence labels. Our experi- ments show improvements in the F1 score over the baseline approach. The implementation has been released publicly.1 1 Introduction The growth of social media platforms has led to an increase in hate speech and abusive language in on- line communities, primarily due to the anonymity provided on such platforms (Mollas et al., 2020). Since manual moderation is not feasible for the gigantic amount of textual data, automated tox- icity detection has received significant attention with numerous datasets being released in recent years (Pavlopoulos et al., 2020). However, most of the existing work on toxicity detection labels the entire comment as toxic or non-toxic and does not provide information about which specific part of the comment is toxic. In practice, human moder- ators (e.g., news portals moderators) can benefit from information on which character indices of the 1https://github.com/vaibhav29498/Toxi c-Spans-Detection part of the comment that is toxic instead of just a system-generated unexplained toxicity score per post. Designing models that can accurately locate toxic spans within a text is thus a crucial step to- wards successful semi-automated moderation. This is challenging because of the scarcity of datasets that are labeled on a segment or token level. This paper explains our approaches for the SemEval-2021 Task 5 which requires us to identify the character offsets for the toxic spans within a comment (Pavlopoulos et al., 2021). We explore possible techniques for improving the results of the vanilla transformer model by leveraging several available public datasets. 2 Related Work and Background Sequential adaptation Using pre-trained lan- guage models has recently proved to be effective for language understanding tasks (Phang et al., 2018). The supplementary training is particularly beneficial when labeled data is scarce (Phang et al., 2018). A popular transfer learning technique in Natural Language Processing (NLP) has been to pre-train sentence encoder neural networks, such as BERT (Devlin et al., 2019), on unsupervised tasks and then fine-tune the …

Content truncated to 500 words