Enhancing Hate Speech Detection in Bengali Language Using Machine Learning and Deep Learning Algorithms

Freedom of expression on social media has led to numerous cases of cyberbullying, offensive content, and hate comments that can have devastating effects on online users or any individual. Natural language processing is essential for understanding and interpreting human languages, converting them into formats that machines can analyze to extract information and recognize contextual significance. Our proposed study focuses on the Bengali language, an interesting and challenging endeavor due to its rich cultural heritage and deep cultural significance. We utilized both machine learning (ML) and deep learning (DL) algorithms. Initially, four ML models were applied using three different feature extraction techniques, achieving an accuracy of 90% in RF, 89% in SVM, 87.74% in DT, and 86.6% in LR. These traditional ML models faced limitations in capturing the contextual and semantic meanings of the Bengali language. By employing fastText embedding, we carefully handled contextual complexities, and also the detection performance in DL models is improved, with an accuracy of 95% in Bi-LSTM and 93% in the CNN model. Our proposed methodology exceeded other approaches, demonstrating effectiveness in identifying hate or toxic comments and contributing to the development of automated systems to detect abusive content on social networks.