CL-UZH at SemEval-2023 Task 10: Sexism Detection through Incremental Fine-Tuning and Multi-Task Learning with Label Descriptions

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1562–1572 July 13-14, 2023 ©2023 Association for Computational Linguistics CL-UZH at SemEval-2023 Task 10: Sexism Detection through Incremental Fine-Tuning and Multi-Task Learning with Label Descriptions Janis Goldzycher Department of Computational Linguistics University of Zurich goldzycher@cl.uzh.ch Abstract The widespread popularity of social media has led to an increase in hateful, abusive, and sexist language, motivating methods for the automatic detection of such phenomena. The goal of the SemEval shared task Towards Explainable De- tection of Online Sexism (EDOS 2023) is to detect sexism in English social media posts (subtask A), and to categorize such posts into four coarse-grained sexism categories (subtask B), and eleven fine-grained subcategories (sub- task C). In this paper, we present our submit- ted systems for all three subtasks, based on a multi-task model that has been fine-tuned on a range of related tasks and datasets before be- ing fine-tuned on the specific EDOS subtasks. We implement multi-task learning by formu- lating each task as binary pairwise text classi- fication, where the dataset and label descrip- tions are given along with the input text. The results show clear improvements over a fine- tuned DeBERTa-V3 serving as a baseline lead- ing to F1-scores of 85.9% in subtask A (rank 13/84), 64.8% in subtask B (rank 19/69), and 44.9% in subtask C (26/63).1 OFFENSIVE CONTENT WARNING: This re- port contains some examples of hateful content. This is strictly for the purposes of enabling this research, and we have sought to minimize the number of examples where possible. Please be aware that this content could be offensive and cause you distress. 1 Introduction With social media’s expanding influence, there has been a rising emphasis on addressing the widespread issue of harmful language, especially sexist language (Meyer and Cukier, 2006; Simons, 2015; Das et al., 2020). Automatic content mod- eration and monitoring methods have become in- dispensable due to the sheer amount of posts and 1We make our code publicly available at https://github. com/jagol/CL-UZH-EDOS-2023. comments on social media platforms. However, the deployment of automatic methods has led to a new problem: current approaches to sexism detection rely on transformer-based language models whose inner workings, in spite of model interpretability re- search, generally remain opaque (Sun et al., 2021). This stands in contrast with the need for explain- able and transparent decision processes in content moderation. The EDOS 2023 shared task (Kirk et al., 2023) focuses on the detection (subtask A), and coarse- (subtask B) and fine-grained (subtask C) categoriza- tion of sexism. The purpose of sexism categoriza- tion is to aid the explainability of sexism detection models, where categorization can serve as addi- tional information for why a post was classified as sexist. In this paper, we present our approach for all three subtasks. The annotated data for detecting sexism is scarce compared to other natural lan- guage processing tasks and is often not publicly available. In response to this, we adopt a multi-task learning approach, where we first train …

Content truncated to 500 words