DeepHateDetect: A Comparative Study of Traditional, Sequential, Transformer, and Large Language Models for Hate Speech Detection in Bengali using Explainable AI

The spread of hate speech on the internet has been associated with an increase in violent acts committed against minority groups all across the world. Nowadays, usage of social media is at its peak, and so is hate speech on online social media. Almost every continent has reported incidents. Most of the world’s population uses Facebook alone, and many people increasingly converse on social media. This study used a private dataset collected from different social media platforms. This study aims to discover an efficient and optimal method by using various models that can detect Bengali hate speech on a different text corpus. The overall procedure includes analyzing the collected data. Labeling, cleaning, vectorization, and further data set processing. Several traditional machine learning (ML) models implemented with advanced grid search. In addition, several sequential and transformer-based models were used. Also, a large language model was applied using quantization. Among these different types of models, LLaMA 3-8B gained the highest accuracy of 0.99. Furthermore, the predictability of LLaMA 3-8B has been demonstrated using explainable AI with the LIME framework.