DS at SemEval-2023 Task 10: Explaining Online Sexism using Transformer based Approach

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1102–1106 July 13-14, 2023 ©2023 Association for Computational Linguistics DS at SemEval-2023 Task 10: Explaining Online Sexism using Transformer based approach Madisetty Padmavathi 4WE Tech Solutions Hyderabad, Telangana India padmadatascience0@gmail.com Abstract In this paper, I describe the approach used in the SemEval 2023 - Task 10 Explainable De- tection of Online Sexism (EDOS) competition (Kirk et al., 2023). I use different transformer models, including BERT and RoBERTa which were fine-tuned on the EDOS dataset to clas- sify text into different categories of sexism. I participated in three subtasks: subtask A is to classify given text as either sexist or not, while subtask B is to identify the specific cate- gory of sexism, such as (1) threats, (2) deroga- tion, (3) animosity, (4) prejudiced discussions. Finally, subtask C involves predicting a fine- grained vector representation of sexism, which included information about the severity, target and type of sexism present in the text. The use of transformer models allows the system to learn from the input data and make predictions on unseen text. By fine-tuning the models on the EDOS dataset, the system can improve its performance on the specific task of detecting online sexism. I got the following macro F1 scores: subtask A:77.16, subtask B: 46.11, and subtask C: 30.2. 1 Introduction Sexism refers to any form of discrimination or prej- udice against an individual based on their gender and can manifest in various ways, including ha- rassment, stereotyping, and unequal treatment. It can be directed towards women, as well as indi- viduals who identify as non-binary or gender non- conforming. The intersection of sexism with other forms of discrimination, such as racism or homo- phobia, can result in even greater harm to individ- uals who experience multiple forms of marginal- ization. These actions not only harm individual women but also create a toxic online culture that perpetuates social injustices and asymmetries. On- line sexism also contributes to the underrepresen- tation of women and other marginalized groups in certain online spaces and industries, limiting their opportunities to participate fully in the digital world. Although automated tools are now extensively employed to detect and evaluate sexist content at a large scale, most of these tools merely provide generic and high-level classifications without any additional clarification. This is a common limi- tation of many automated tools that are used to detect and evaluate sexist content online. While these tools can be useful in identifying patterns of sexist language or behavior, they often lack the con- textual understanding necessary to provide more nuanced classifications or explanations. This can be a problem because online sexism can take many different forms, and what may be considered sexist in one context may not be in another. Therefore, it is essential to have tools that can accurately identify and contextualize instances of online sexism, rather than relying solely on high-level classifications. To enhance interpretability, trust, and compre- hension of the decisions made by automated tools, it …

Content truncated to 500 words