A Hybrid Deep Learning Approach to Detect Bangla Social Media Hate Speech

Social media has become an integral part of our day-to-day life. In our activities or posts on social media, the presence of hate speech written in the native language or English has increased significantly. It often leads to the spread of negativity, depression, or even sometimes considered cybercrime. In this paper, a hybrid deep learning approach has been taken to detect Bangla social media hate speech using fastText embedding, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network. A publicly available dataset of 30000 samples has been used, and the proposed hybrid model achieved close to 90% accuracy with significant sensitivity and specificity in Bangla hate speech detection. Several related deep learning approaches were evaluated in this same dataset, but none of them performed better than the proposed model. The hybrid model also showed robustness which made it more suitable for this task.