Towards Better Misogyny Detection in Bangla: Improved Dataset and Cutting-Edge Model Evaluation

Misogyny is hatred, disdain, or prejudice toward women or girls. It can be verbal abuse, aggression, harassment, or discrimination. In Bengali NLP, the scarcity of well-balanced datasets for misogyny detection is a significant problem. While there has been previous work in this area for the Bangla language, the existing dataset, which only contains 2258 comments, has proven insufficient for robust detection, resulting in suboptimal performance in classification models. To tackle this, we have carefully curated a dataset from different sources, including social media and different websites. This dataset features five categories of misogynistic comments (Discredit, Stereotype & Objectification, Sexual Harassment, Threats of violence, Dominance & Derailing) and one category of Non-misogynistic comments. We evaluated the performance of multiple Machine Learning models such as SVM(Support Vector Machine), Naive Bayes, KNN, Decision Tree, and Random Forest. Bangla BERT base and Bangla BERT Electra, two Deep Learning models, were also used for classification. Additionally, we employed ensemble approaches using Bangla BERT embeddings (Bangla BERT Base and BERT-base-multilingual-cased) with models like LSTM, Bi-LSTM, and CNN+LSTM. Among them, Bi-LSTM with BERT (Bert-base-multilingual-cased) Embeddings achieves the highest accuracy, 91.24%, in classifying misogynistic comments. The new improved dataset has been made publicly available as Ben-Misog to help the Bengali Natural Language Processing community progress.