Abstract

The SemEval 2021 task 5: Toxic Spans Detection is a task of identifying considered-toxic spans in text, which provides a valuable, automatic tool for moderating online contents. This paper represents the second-place method for the task, an ensemble of two approaches. While one approach relies on combining different embedding methods to extract diverse semantic and syntactic representations of words in context; the other utilizes extra data with a slightly customized Self-training, a semi-supervised learning technique, for sequence tagging problems. Both of our architectures take advantage of a strong language model, which was fine-tuned on a toxic classification task. Although experimental evidence indicates higher effectiveness of the first approach than the second one, combining them leads to our best results of 70.77 F1-score on the test dataset.

Citation

@inproceedings{nguyen-etal-2021-nlp, title = “{S}-{NLP} at {S}em{E}val-2021 Task 5: An Analysis of Dual Networks for Sequence Tagging”, author = “Nguyen, Viet Anh and Nguyen, Tam Minh and Quang Dao, Huy and Huu Pham, Quang”, booktitle = “Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021)”, month = aug, year = “2021”, address = “Online”, publisher = “Association for Computational Linguistics”, url = “https://aclanthology.org/2021.semeval-1.120”, doi = “10.18653/v1/2021.semeval-1.120”, pages = “888–897”, }