Bengali misogyny identification with deep learning and LIME

Abstract

The increase of misogyny across social media platforms highlights the urgent need to create efficient tools for recognizing and responding to gender-based online abuse. This study explores the complex problem of identifying instances of sexism in the Bengali language, a field that has a limited amount of research conducted due to a lack of financial resources and academic interest. We study the performance of BERT-based architectures, in recognizing misogynistic language by using the capabilities of deep learning models. Our research hypothesis is that enhancing the mBERT model with linguistic and cultural variety by employing multilingual training such as merging Bengali, Hindi, and English data for training will improve the ability to detect misogyny in Bengali, potentially transcending language barriers. We offer two extensive experiments that assess the performance of the models and give insight into the strengths and limits of those models. In addition, we employ LIME to uncover the decision-making processes of the models, enhancing their interpretability. Our results contribute to the development of improved methods for identifying online sexism, offering insights for the creation of safer digital environments. This study lays the groundwork for future research on language-specific nuances and cross-lingual trends in the field of gender-based abuse detection.