LCT-1 at SemEval-2023 Task 10: Pre-training and Multi-task Learning for Sexism Detection and Classification

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1573–1581 July 13-14, 2023 ©2023 Association for Computational Linguistics LCT-1 at SemEval-2023 Task 10: Pre-training and Multi-task Learning for Sexism Detection and Classification Konstantin Chernyshev∗ Ekaterina Garanina∗ Duygu Bayram Qiankun Zheng Lukas Edman University of Groningen {k.chernyshev, e.garanina, d.bayram.1, q.zheng.9}@student.rug.nl j.l.edman@rug.nl Abstract Misogyny and sexism are growing problems in social media. Advances have been made in online sexism detection but the systems are of- ten uninterpretable. SemEval-2023 Task 10 on Explainable Detection of Online Sexism aims at increasing explainability of the sexism de- tection, and our team participated in all the proposed subtasks. Our system is based on further domain-adaptive pre-training (Gururan- gan et al., 2020). Building on the Transformer- based models with the domain adaptation, we compare fine-tuning with multi-task learning and show that each subtask requires a differ- ent system configuration. In our experiments, multi-task learning performs on par with stan- dard fine-tuning for sexism detection and no- ticeably better for coarse-grained sexism clas- sification, while fine-tuning is preferable for fine-grained classification1. 1 Introduction Sexism has been appearing frequently in online spaces in recent years, which not only makes on- line spaces unfriendly but also exacerbates so- cial prejudice and causes serious harm to targeted groups. In order to control and mitigate it, con- siderable efforts have been made to detect online sexism (Fersini et al., 2018a,b; Bhattacharya et al., 2020; Fersini et al., 2020; Rodríguez-Sánchez et al., 2021). SemEval-2023 Task 10 on Explainable De- tection of Online Sexism (EDOS) (Kirk et al., 2023) aims to improve interpretability via flagging sexist content (Task A) and further deciding on a particu- lar type of sexism in the text, including 4-category (Task B) and 11-category (Task C) classification systems. In this paper, we present our participation in all EDOS subtasks. Given the complexity of obtaining annotated data, the research on using additional data via dif- *Equal contribution. 1The source code is available at https://github.com/ lct-rug-2022/edos-2023. ferent training techniques is quite intensive. Gu- rurangan et al. (2020) illustrated the benefits of further pre-training and introduced two methods for it: only using the task data or only using the domain data. These approaches, used individually and in combination, are shown to perform well in domain-specific tasks with relatively low costs of computing resources. Safi Samghabadi et al. (2020) developed a uni- fied end-to-end neural model using a multi-task learning (MTL) approach to address the tasks of aggression and misogyny detection. Lees et al. (2020) pre-trained a BERT-based model on 1 bil- lion comments and fine-tuned the model with multi- lingual toxicity data before fine-tuning it on the tar- get dataset. They both demonstrated that using data from similar tasks and fine-tuning the model with it in a multi-task way could improve the model’s performance. Inspired by the above studies, we explore the impact of further pre-training as well as multi-task learning on the EDOS tasks in our work. Specif- ically, we collect several datasets that have been developed for the …

Content truncated to 500 words