A large-scale annotated Urdu corpus and deep learning benchmark for mental health classification

Millions of people worldwide are affected by mental health disorders, particularly anxiety and depression, which pose a growing health concern globally. Early detection and intervention are critical for improving treatment outcomes. Conventional clinical evaluation methods face significant barriers including limited accessibility, social stigma, and diagnostic delays. Recent advances in artificial intelligence and natural language processing have demonstrated promising capabilities for automated mental health detection through text analysis of social media. However, most existing research focuses exclusively on high-resource languages such as English, leaving a critical gap for low-resource languages. Our study addresses this limitation by developing a comprehensive deep learning framework for mental health classification in Urdu. More than 230 million people speak Urdu, but it is severely underrepresented in computational mental health research. This research provides the first large-scale, publicly released Urdu dataset specifically designed for multi-class anxiety, depression, and neutral classification, and establishes performance benchmarks for future work. We created and validated a novel annotated dataset of 36,000 Urdu tweets classified into anxiety, depression, and neutral categories through systematic translation, manual annotation, and automated labeling with rigorous quality validation. Three established deep learning architectures were adapted and systematically evaluated in this study. The CNN+BiLSTM model achieved an accuracy of 79.08%, while the CNN+BiGRU model attained 78.25%. Among all evaluated approaches, UrduBERT delivered the best performance with an accuracy of 81.71%, along with superior precision, recall, and F1-score values across all categories. These findings highlight the advantage of transformer-based contextual representations in modeling complex linguistic structures in morphologically rich languages. Overall, the proposed framework provides a strong basis for scalable and culturally sensitive mental health monitoring systems, with the potential to support early detection of at-risk individuals in linguistically underrepresented communities.