Leveraging BanglaBERT: A Transformer-Based Approach for Multiclass Hate Speech Detection in Bangla Social Media

In languages like Bengali, where automatic moderation systems are still in their infancy, the rise of hate speech has become an increasingly pressing matter as social media continues to influence digital communication. The process of detecting hate speech is difficult due to the intricacy of the Bengali script and the lack of high-quality annotated datasets. In this work, we provide a deep learning-based method that uses the Bangla BERT model to precisely detect and categorize hate speech in Bengali literature. To increase model robustness, we employ a comprehensive text preparation approach that includes tokenization, normalization, and management of informal linguistic variations. We explore the effects of important hyperparameters, such as batch size, number of epochs, and learning rate, on model performance by optimizing BanglaBERT on a carefully selected dataset. According to experimental data, our transformer-based model achieves a higher accuracy of 93.21%, recall of 92.11%, precision of 92.82% and F1-score of 92.46%, outperforming conventional machine learning classifiers by a wide margin. Additionally, we thoroughly examine the model’s generalizability across various hate speech types, emphasizing the significance of contextual awareness in Bengali natural language processing (NLP). The knowledge gathered from this study aids in the continuous creation of efficient hate speech detection systems that can support real-time content moderation and provide a more secure and welcoming online space for Bengali-speaking groups.