PingAnLifeInsurance at SemEval-2023 Task 10: Using Multi-Task Learning to Better Detect Online Sexism

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 2188–2192 July 13-14, 2023 ©2023 Association for Computational Linguistics PingAnLifeInsurance at SemEval-2023 Task 10: Using Multi-Task Learning to Better Detect Online Sexism Mengyuan Zhou, Xiaolong Hou, Meizhi Jin, Xiyang Du, Cheng Chen, Lianxin Jiang, Jianyu Li and Zhenggang Wei Ping An Life Insurance Company of China, Ltd. {ZHOUMENGYUAN425, HOUXIAOLONG430, JINMEIZHI005, DUXIYANG037, CHENCHENG498, JIANGLIANXIN769, LIJIANYU002, WEIZHENGGANG617} @pingan.com.cn Abstract This paper describes our system used in the SemEval-2023 Task 10: Towards Explainable Detection of Online Sexism (Kirk et al., 2023). The harmful effects of sexism on the internet have impacted both men and women, yet cur- rent research lacks a fine-grained classification of sexist content. The task involves three hi- erarchical sub-tasks, which we addressed by employing a multitask-learning framework. To further enhance our system’s performance, we pre-trained the roberta-large (Liu et al., 2019b) and deberta-v3-large (He et al., 2021) models on two million unlabeled data, resulting in sig- nificant improvements on sub-tasks A and C. In addition, the multitask-learning approach boosted the performance of our models on sub- tasks A and B. Our system exhibits promising results in achieving explainable detection of on- line sexism, attaining a test f1-score of 0.8746 on sub-task A (ranking 1st on the leaderboard), and ranking 5th on sub-tasks B and C. 1 Introduction The use of sexist language and behavior in online spaces has long been a problem, with harmful ef- fects on those who are targeted. This issue has become even more pronounced in recent years, as the internet has become a platform for social and political discourse, making it easier for sexism to be disseminated on a large scale (Plaza et al., 2023). Automated tools have been developed to address this issue, with many platforms now using machine learning models to detect and assess sexist con- tent at scale (Lopez-Lopez et al., 2021). However, most of these tools only provide generic, high-level classifications for sexist content, without further explanation. This lack of interpret ability can be problematic, as it can make it difficult for users and moderators to understand how automated tools make their decisions. To address this issue, In Task 10, we need to develop English-language models for sexism detec- tion that are not only more accurate, but also more explainable (Kirk et al., 2023). To achieve this, the official provides not only supervised data but also 2 million unlabeled data for sexist content classifi- cations from two prominent online platforms: Gab and Reddit. The task consists of three hierarchical subtasks. The first subtask, Binary Sexism Detection, is a two-class (or binary) classification task in which systems must predict whether a post is sexist or not sexist. The second subtask, Category of Sexism, is designed for posts identified as sexist in the first subtask. Here, systems must predict one of four categories: threats, derogation, animosity, or preju- diced discussions. The final subtask, Fine-grained Vector of Sexism, is also for posts identified as sexist in the first subtask. This subtask requires …

Content truncated to 500 words