Explainable AI: Transformer-based Bangla Hate Speech Detection

People spend a tremendous amount of time on social media sites these days, including Instagram, Twitter, Facebook, and YouTube. Anyone may attack an individual or a community via social media, which reduces human respect. Consequently, we detect the Bangla hate speech in this investigation. A challenging endeavor while dealing with low-resource languages like Bangla is the dearth of publicly accessible datasets. We made use of a publicly accessible dataset [1] that has a problem with data imbalance. We processed the raw data using a variety of preprocessing techniques, then we utilized SMOTE to balance the dataset. To categorize the speech as hate speech or non-hate speech, we used machine learning, deep learning, and fine-tuned BanglaBERT. In the end, we used Explainable AI (XAI) to elucidate predictions from models. The results show that BanglaBERT is a transformer-based model beyond conventional methods of machine learning with a 90.89% accuracy rate. We compared our suggested strategy with existing work, and our suggested approach performs better in terms of accuracy.