Computational analysis and detection of Bengali communal violent speech

Abstract

Communal violence is intensified by the widespread use of cyber hate, leading to aggression and increased conflicts among different religious, ethnic, and social groups, creating a barrier to social harmony. This research focuses on evaluating Bengali textual data sourced from Twitter and Reddit comments. The primary objective of this study is to enhance the accuracy of detecting communal violence-inciting speech. To achieve this, we employed and fine-tuned large language models, specifically the pre-trained BanglaBERT, aiming for a significant improvement over existing detection methodologies. Improving the detection of communal violent speech will help content moderation systems to effectively moderate and remove content linked to communal violence, thereby fostering communal peace in the Bengali-speaking regions.