niceNLP at SemEval-2023 Task 10: Dual Model Alternate Pseudo-labeling Improves Your Predictions

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 307–311 July 13-14, 2023 ©2023 Association for Computational Linguistics niceNLP at SemEval-2023 Task 10: Dual Model Alternate Pseudo-labeling Improves Your Predictions Yu Chang1 , Yuxi Chen1 , Yanru Zhang1,2∗ 1University of Electronic Science and Technology of China, Chengdu 2Shenzhen Institute for Advanced Study, UESTC {yuchang, yuxi.ch}@std.uestc.edu.cn, yanruzhang@uestc.edu.cn Abstract Sexism is a growing online problem. It harms women who are targeted and makes online spaces inaccessible and unwelcoming. In this paper, we present our approach for Task A of SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS), which aims to per- form binary sexism detection on textual content. To solve this task, we fine-tune the pre-trained model based on several popular natural lan- guage processing methods to improve the gen- eralization ability in the face of different data. According to the experimental results, the effec- tive combination of multiple methods enables our approach to achieve excellent performance gains. 1 Introduction Gender discrimination is any mistreatment or nega- tive feelings directed at women based on their gen- der, or based on their gender in combination with one or more other identity attributes (e.g., Black women, Muslim women, trans women). As the Internet continues to grow, sexism has become a growing online problem. It can be harmful to women who are discriminated against while also posing a threat to the healthy functioning of cy- berspace. Large-scale pre-trained models have been widely used to automate the identification of discriminatory statements. However, these models are often capable of simple labeling and not inter- pretative descriptions. Improving the interpretabil- ity of sexism detection can help to thereby enhance the understanding and behavioral decision-making ability of users and moderators of automated tools. To address this issue, SemEval 2023 - Task 10 sup- ports the development of English-language models for sexism detection that are more accurate as well as explainable, with fine-grained classifications for sexist content from Gab and Reddit. This task proposes a new English sexism detection dataset ∗*Corresponding author based on Gab and Reddit, which contains four fine- grained and eleven more fine-grained category la- bels. In Task A, each sentence was first asked to perform a dichotomous classification of the pres- ence or absence of sexism. And tasks B and C re- quire a finer-grained distinction between sentences with sex discrimination (Kirk et al., 2023). In many past studies, language models called BERT have been used extensively for several tasks (Devlin et al., 2018). BERT models can be pre- trained with a self-supervised approach to gener- ate word/tag or sentence representations that are rich in a priori knowledge (Jin et al., 2020). They can then be specifically fine-tuned for many down- stream tasks, including text classification. More- over, various model optimization methods have been shown to be effective in enhancing the fine- tuning of pre-trained models. Therefore, in this work, we employ the current state-of-the-art pre- training model DeBERTa-v3 to exploit the prior knowledge in its pre-trained resources (He et al., 2021). …

Content truncated to 500 words