Multi-Class Cyberbullying Detection of Bengali Social Media Texts Using A Hybrid Model

Cyberbullying is the act of a person by employing technology to shame or harass other people. Because bullying on online platforms spreads quickly to larger viewers, it can sometimes get more awful and studies reveal that this kind of activity happens a lot on social media platforms like Facebook and Twitter. Furthermore, generating an improved result is tough because of the unique nature of the Bengali language, a shortage of reliable records, and limited preliminary processing techniques. This study presents a deep learning-based hybrid method with an attention mechanism to identify bully expression in Bangla text by employing a multi-class dataset containing four classes such as political, religious, sexual, and non-bullying feelings. The proposed framework architecture combines Convolutional Neural Network(CNN) and bidirectional LSTM layers, resulting in a remarkable testing performance on the multi-class dataset. The research attempts to maximize performance and compare the suggested model with the initial approaches. Our recommended CNN-BiLSTM with additive attention framework improves cyberbullying identification with 0.86 F1 scores and 85.18% accuracy on a labeled dataset of 42,036 Facebook comments. This study enhances our comprehension of bully expression in the Bangla language, facilitating the development of more sophisticated natural language processing applications in languages with limited resources.