hhuEDOS at SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS) Binary Sexism Detection (Subtask A)

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1476–1482 July 13-14, 2023 ©2023 Association for Computational Linguistics hhuEDOS at SemEval-2023 Task 10: Explainable Detection of Online Sexism (EDOS) – Binary Sexism Detection (Subtask A) Wiebke Petersen and Diem-Ly Tran and Marion Wroblewitz Heinrich Heine University Düsseldorf, Germany wiebke.petersen@hhu.de, diem-ly.tran@hhu.de, marion.wroblewitz@hhu.de Abstract In this paper, we describe our contribution to the SemEval-2023 Task 10 (Subtask A), a shared task on detecting and predicting sexist language. The dataset consists of labeled sexist and non-sexist data targeted towards women ac- quired from both Reddit and Gab. We present and compare several approaches we experi- mented with and our final submitted model. Additional error analysis is given to recognize challenges we dealt with in our process. A total of 84 teams participated. Our model ranks 55th overall in Subtask A of the shared task. 1 Introduction Online sexism is a consistently growing problem on the internet especially on social media platforms such as Reddit and Twitter. When women are the target, it creates a hostile environment and perpet- uates sexist beliefs and social stigmas (Dehingia, 2020). Investigating the data gathered from these platforms to not only specifically detect sexist con- tent and certain patterns, but rather explain why it is sexist, gives us a better understanding of the impact it has on women, enhances its interpretability, and helps create a much more welcoming environment. This paper describes our group’s approach for SemEval-2023 Task 10-A (Kirk et al., 2023), which deals with binary sexism classification of sexist and non-sexist content. In order to solve this problem, we worked on fine-tuning transformer language models and on using the provided and additional non-labelled data for an semi-supervised extension of the training data set. Additionally, we com- pared our results with traditional machine learning classification methods such as a Random Forest or k-Nearest Neighbours classifier. In view of the fact that the provided data had significantly less sexist entries and recognizing the conflict between sexist data and data that is clearly offensive yet non-sexist, we also experimented with data augmentation to overcome the unbalance in the data. It turns out that our best performing model is a BERT model that is fine-tuned with the manually labeled training data as well as with the automati- cally labeled additional data that was provided by the organizers. We present the task and background in section 3, our approaches in section 4 and our error analysis in section 5, respectively. 2 Related Work The detection of hate speech on online platforms has been gaining popularity within the Natural Lan- guage Processing community in recent years, fo- cusing on tasks in sub-areas such as identifying racist or abusive content (for literature reviews see Poletto et al., 2021; Jahan and Oussalah, 2021). However, fewer studies have concentrated on hate speech specifically targeted at women (e.g., Jha and Mamidi, 2017). Classical machine learning techniques based on manual feature engineering use a range of fea- tures for hate speech detection (for an …

Content truncated to 500 words