Enhancing Sexism Detection with Ensemble Language Models

This paper presents an evaluation of various combinations of pre-trained language models for the automatic detection and classification of sexist comments on social media platforms. We experimented with several ensemble techniques, including dropout-based fusion and majority voting, to assess the effectiveness of combining different models. Our study focused on both sexism Detection (binary classification) and sexism category (multi-class classification) scenarios using the EDOS2023 dataset. The best-performing ensemble model achieved a Macro-F1 score of 0.8602 on the binary classification task and 0.7085 on the multiclass classification task, outperforming individual models. These results demonstrate that ensemble approaches can significantly enhance the detection and classification of sexist content, offering a robust solution that leverages the strengths of multiple models.