NTU_NLP at SemEval-2020 Task 12: Identifying Offensive Tweets Using Hierarchical Multi-Task Learning Approach

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 2105–2110 Barcelona, Spain (Online), December 12, 2020. 2105 NTU_NLP at SemEval-2020 Task 12: Identifying Offensive Tweets Using Hierarchical Multi-Task Learning Approach Po-Chun Chen1, Hen-Hsen Huang2, Hsin-Hsi Chen1 1Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan 2Department of Computer Science, National Chengchi University, Taipei, Taiwan pcchen@nlg.csie.ntu.edu.tw, hhhuang@nccu.edu.tw, hhchen@ntu.edu.tw Abstract This paper presents our hierarchical multi-task learning (HMTL) and multi-task learning (MTL) approaches for improving the text encoder in Sub-tasks A, B, and C of Multilingual Offensive Language Identification in Social Media (SemEval-2020 Task 12). We show that using the MTL approach can greatly improve the performance of complex problems, i.e. Sub-tasks B and C. Coupled with a hierarchical approach, the performances are further improved. Overall, our best model, HMTL outperforms the baseline model by 3% and 2% of Macro F-score in Sub-tasks B and C of OffensEval 2020, respectively. 1 Introduction Multilingual Offensive Language Identification in Social Media (OffensEval 2020) hosted by (Zampieri et al., 2020) is a popular competition attracting many teams. The task is based on Offensive Language Identification Dataset (OLID) 1.0 and a new dataset SOLID. OffensEval 2020 consists of three subtasks. In Sub-task A, the goal is to identify if a given tweet is offensive or non-offensive. In Sub-task B, the focus is to identify if the offensive content in a tweet is targeted or untargeted. In Sub-task C, systems have to detect the type of a target in an offensive tweet. The purpose of this paper is to investigate the multi-task learning (MTL) approach to the OffensEval tasks. We propose the MTL model, which is based on the popular text encoder BERT (Devlin et al., 2018) with the MTL architecture. To leverage the hierarchical nature of the three subtasks in OffensEval 2020, we further propose the HMTL model, which is based on the BERT with a hierarchical MTL ar- chitecture. This paper is organized as follows. We summarize the related work in Section 2. Section 3 introduces the dataset, the preprocessing procedure, and the architectures of our models. Our results in OffensEval 2019 and OffensEval 2020 are reported and discussed in Section 4. Finally, Section 5 concludes this paper. 2 Related Work The overview paper of OffensEval 2019 (Zampieri et al., 2019b) shows that 70% of the teams that participated last year use the deep learning models, and 8% of them use the BERT model in Sub-task A. Many of the top-ranked teams use the BERT model. The NULI team (Liu et al., 2019), which wins the first place in Sub-task A, proposes a BERT-based model and an LSTM-based model. Their work shows that the pre-trained BERT model has a good competition result in the task. The vradivchev_anikolov team (Radivchev and Nikolov, 2019) wins the first place and the second place in Sub-tasks C and A, respectively. They explore several models, among which BERT has the highest performance. The UM- This work is licensed under a Creative Commons Attribution 4.0 International Licence. …

Content truncated to 500 words