Deep Learning Approaches for Detecting and Analyzing Abusive Bangla Comments on Religion in Social Media

Abstract

Increment in hate speech particularly on religious beliefs has been the source of significant concern of concern on the internet social sites. Bangladesh culture and religion have been closely connected, and due to the abuse language against particular religious groups, social cohesion and security is endangered over the Internet. Identification of this harmful material in Bangli is not easy because of the insufficiency of language materials and because of the vague meanings of the contexts. The paper describes a deep learning method, which is automated, to recognize abusive remarks on religion in Bangali. An effective way of learning contextual relations in a text was to include Multi-Head Attention mechanism to a Bi-Directional Long Short-Term Memory (Bi-LSTM) network. The problem of the imbalance of the datasets was dealt with by an adapted Focal Loss that helped to improve detection efficiency. In the collection, there were 24,137 Facebook bangla remarks, and they were classified in the category of Normal and Religious Abuse. The proposed model with data preprocessing and partitioning (72 percent training, 8 percent validation, 20 percent testing) yielded an accuracy of 95 percent, surpassing the traditional machine learning, and other subsequent deep learning models, including CNN-LSTM and regular Bi-LSTM. Comparisons between the models (LSTM, CNN-LSTM, RNN, Logistic Regression, SVM, Random Forest) have been done and it was identified that attention mechanism enhanced the classification of abusive religious content. The results have shown that recurrent neural activity coupled with attention systems is an effective technique of identifying online abuse on religion on low resources languages such as Bengali language (Bangla).