Multi-Class Cyberbullying Analysis in Bengali: A Comparative Study of BERT-Based Deep Learning Models

The rise of social media has increased online harassment and bullying, which is a big challenge for individuals and society. Their detection and addressing require advanced tools, especially for under-resourced languages like Bengali, which lack datasets and processing tools. Cyberbullying analysis, generally used to understand opinions and emotions, can also be applied to detect social media bullying. This paper investigates transformer based models for the detection of Cyberbullying related content in Bengali and puts special focus on the BanglaBERT model. In this regard, a dataset consisting of 42,036 Bengali social media comments is used, which was created by combining two publicly available datasets: "Multi-class Bangla social media comments" and "Bangla online comments dataset". The dataset consists of four classes: Political, Religious, Sexual, and Not Bully. The results show that the BanglaBERT model performs best, achieving 90% accuracy, and 91% F1-score, outperforming all other BERT models. This research has given importance to the pre-trained models designed for a low resource language and set a new benchmark for detecting social media bullying in the Bengali language. Overall, it highlights the potential of advanced NLP models to address the pressing issue of online harassment efficiently.