HITSZQ at SemEval-2023 Task 10: Category-aware Sexism Detection Model with Self-training Strategy

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 934–940 July 13-14, 2023 ©2023 Association for Computational Linguistics HITSZQ at SemEval-2023 Task 10: Category-aware Sexism Detection Model with Self-training Strategy Ziyi Yao1, Heyan Chai1, Jinhao Cui1, Siyu Tang1 and Qing Liao1,2∗ 1 Harbin Institute of Technology, Shenzhen, China 2 Peng Cheng Laboratory, Shenzhen, China {yaoziyi,chaiheyan,cuijinhao,tangsiyu999}@stu.hit.edu.cn, liaoqing@hit.edu.cn Abstract This paper describes our system used in the SemEval-2023 Task 10 Explainable Detection of Online Sexism (EDOS). Specifically, we par- ticipated in subtask B: a 4-class sexism classifi- cation task, and subtask C: a more fine-grained (11-class) sexism classification task, where it is necessary to predict the category of sexism. We treat these two subtasks as one multi-label hierarchical text classification problem and pro- pose an integrated sexism detection model for improving the performance of the sexism de- tection task. More concretely, we use the pre- trained BERT model to encode the text and class label and a hierarchy-relevant structure encoder is employed to model the relationship between classes of subtasks B and C. Addi- tionally, a self-training strategy is designed to alleviate the imbalanced problem of distribu- tion classes. Extensive experiments on subtasks B and C demonstrate the effectiveness of our proposed approach. 1 Introduction Sexism is especially rampant online due to the anonymity of the Internet, making the online com- munity unfriendly towards women (Mosca et al., 2021; Abburi et al., 2020; García-Baena et al., 2022). Detecting sexism on social media is mean- ingful in building a more harmonious Internet world. Sexism detection (Ahuir et al., 2022; del Arco et al., 2022) can be considered as a particu- lar sentiment classification task, applied to predict the categories (the detailed categories are shown in Figure 1) of prejudice or discrimination based on a person’s gender in the whole sentence. For example, the sentence “As Roosh said having dis- cussions with woman is a waste of time anyway.” is a sexist comment, which belongs to a kind of descriptive attack expressing derogation against women. ∗ Corresponding Author Figure 1: Taxonomic hierarchy structure of the three subtasks. SemEval-2023 Task 10 (Kirk et al., 2023) (Ex- plainable Detection of Online Sexism, EDOS), is an online sexism detection and classification task. The organizers provide a labeled dataset and two unlabeled datasets for training the three subtasks. The labeled dataset consists of 14,000 annotated samples of which 3398 are sexist. For the sexist entries, as shown in Figure 1, fine-grained labels of the 4 categories (subtask B) and 11 vectors (subtask C) are also provided. Subtask A is a binary classification task aimed at predicting whether a post is sexist or not. Subtask B and subtask C classify the sexist posts into more fine-grained categories. There are four categories for subtask B, which are threats, derogation, ani- mosity, and prejudiced discussions. For subtask C, there are eleven target vectors. The hierarchical structure of the three tasks is shown in Figure 1. One obvious characteristic of subtasks B and C is that categories in subtask …

Content truncated to 500 words