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

TL;DR

This paper presents an ensemble approach combining BERT-based models and post-processing techniques for toxic span detection, significantly improving performance over baseline models and achieving a F1-score of 67.55%.

Contribution

The study introduces an ensemble framework with post-processing for toxic span detection using pre-trained language models, enhancing accuracy over individual models.

Findings

Ensemble models outperform baseline BERT models.

Post-processing improves span detection accuracy.

Achieved F1-score of 67.55% on test data.

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.