Leveraging RF and BERT for Superior Hate Speech Detection in the Bengali Language

The present research utilises the BHSSP dataset, which has 50,000 labelled instances, to use machine learning and deep learning models to offer a hybrid architecture for Bengali hate speech identification. A robust preprocessing pipeline was implemented to address noise, including the removal of duplicates, punctuation, digits, stop words, and emojis. Bangla GloVe embeddings were used for enhanced feature extraction, while models such as Random Forest (RF), Decision Tree (DT), XGBoost (XGB), and Support Vector Machine (SVM) were compared with deep learning approaches like CNN, LSTM, GRU, and BERT. Additionally, hybrid models RF+BERT and GRU+BERT were introduced. Results demonstrate the superiority of BERT-based models, with GRU+BERT achieving an F1-Score of 0.99, effectively capturing the complexities of Bengali hate speech. This work highlights the efficacy of advanced embeddings and hybrid architectures in low-resource linguistic settings and provides a foundation for broader applications in hate speech detection for Bengali and similar languages.