Explore publications.

The growing prevalence of mental health issues, particularly depression and suicidal thoughts, points to the need to develop automated tools capable of detecting such sentiments from online communication. This study addresses some critical challenges by introducing a novel sentiment analysis framework for Bangla text, aimed at classifying content into non-depressive, depressive, and suicidal categories. We propose a hybrid deep learning model leveraging the strengths of transformer-based architectures, designed to manage long textual sequences effectively, a critical aspect in the context of natural language processing. Our model integrates RoBERTa (Robustly Optimised BERT Pre-Training Approach) with a Self-Attention Network (SAN), creating a synergistic framework for nuanced sentiment detection in Bangla social media posts, comments, and articles. This method addresses the scarcity of Bangla specific datasets by utilising a dataset curated for the study. The results demonstrate the superiority of our model, achieving an accuracy of 82.58%, alongside precision, recall, and F1-scores of 82%. Subsequently, it emphasises the potential for the proposed model to support early identification of mental health concerns, thereby enabling timely interventions and contributing to mental health awareness and prevention efforts. In the future, deploying the model as a real-time chatbot or browser extension could scan Bangla social media for depressive, non-depressive, and suicidal content and alert professionals to the risk factors.