Bengali Hate Speech Categorization: A Comparison of Hybrid Deep Learning and Transformer-Based Models

The proliferation of hate speech in online social engagement platforms poses significant challenges to digital communication platforms, particularly in linguistically diverse regions such as South Asia. The rapid growth of virtual entertainment has facilitated easier interactions across online platforms, but this has also led to an increase in targeted hate speech. This paper presents a system for categorizing hate speech in Bengali text, leveraging a combination of hybrid deep learning and transform learning techniques. We constructed a dataset of approximately 19,398 Bengali texts from YouTube comments, which were manually classified into six categories: Hate Speech, Religious Hatred, Inciteful, Political Comments, Religious Comments, and Communal Attack. Among hybrid deep learning models, CNN-BiLSTM achieved the highest accuracy of 76%, surpassing both the GRU-LSTM and BiGRU-CNN models. Notably, BanglaBERT excelled as the top-performing Transformer algorithm with 80 % accuracy. This exhibits the efficacy of the hybrid model in categorizing hate speech within Bengali text. We have also evaluated the effectiveness of our proposed approach by comparing with baseline methods.