Multiclass Detection of Misogynistic Bangla Text from YouTube using Transformer-Based Models

Social networks are a powerful tool that helps people stay connected and informed. However, these allow users to hide behind fake profiles, making it easier for some individuals to harass and bully other people, specially women without fear of real consequences. For this reason, misogyny on social media has become a significant issue. These platforms originally intended to encourage free expression, now acting as havens for harassment and hate crimes against women. In Bangladesh, misogynistic comments towards women are particularly widespread on various platforms such as YouTube, Facebook, Instagram, TikTok, and Twitter. To address this problem, we presented a specialized Bengali-language dataset based on YouTube, which has three key categories of misogynistic content: stereotypes, hostility, and discrimination. And to analyze and classify the misogynistic contents we propose to applying advanced natural language processing models such as BanglaBERT, Electra, mBERT, XLM-RoBERTa, and DistilBERT. Among these, BanglaBERT demonstrated the best performance, achieving an accuracy of 69%, demonstrating how well it addresses the difficulties presented by low-resource languages to detect misogyny contents existing on the platform like YouTube. In addition to outperforming other approaches, our aimed methodology makes a significant contribution to creating a safer online environment for all users by detecting and reducing toxic language toward women.