MIPT-NSU-UTMN at SemEval-2021 Task 5: Ensembling Learning with Pre-trained Language Models for Toxic Spans Detection

Abstract

This paper describes our system for SemEval-2021 Task 5 on Toxic Spans Detection. We developed ensemble models using BERT-based neural architectures and post-processing to combine tokens into spans. We evaluated several pre-trained language models using various ensemble techniques for toxic span identification and achieved sizable improvements over our baseline fine-tuned BERT models. Finally, our system obtained a F1-score of 67.55% on test data.

arXiv commentAccepted at SemEval-2021 Workshop, ACL-IJCNLP 2021

References in corpus (7)

• Efficient Estimation of Word Representations in Vector Space
• Decoupled Weight Decay Regularization
• Fine-Grained Analysis of Propaganda in News Articles
• Toxic Language Detection in Social Media for Brazilian Portuguese: New Dataset and Multilingual Analysis
• Deep Learning Models for Multilingual Hate Speech Detection
• Context-Aware Attention for Understanding Twitter Abuse
• SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval)