The Unspoken Words: Leveraging Transformer-based Approaches for Detecting Depression from Social Media

Mental health issue is a topic of great concern in the present world, yet in spite of being so widely spoken language in Bengali, there are not many notable works addressing this problem. This paper focuses on detecting and classifying depressive content in Bengali text by leveraging linguistic patterns and emotional cues. A dataset of 3,002 Bengali social media texts has been created to identify signs of depression. The dataset is categorized into two classes: Depressive and Non-Depressive, enabling the detection of depression from social media content. This paper presents a comprehensive evaluation of detecting depression for Bengali text, encompassing five machine learning techniques, three deep learning models, and five transformer-based approaches. Among the evaluated models, Bangla-BERT, our proposed transformer-based method, achieved the highest accuracy of 0.96. The model demonstrates exceptional capability in detecting depression by effectively understanding the contextual nuances of sentences. The findings highlight the potential of transformer-based architectures, particularly Bangla-BERT, in advancing sentiment analysis tasks for low-resource languages (LRLs) like Bengali.