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

Area of Science:

• Computational linguistics
• Mental health informatics
• Artificial intelligence in healthcare

Background:

• Mental health disorders like anxiety and depression are a global concern, necessitating early detection.
• Traditional diagnostic methods face accessibility and stigma challenges.
• AI and NLP show promise for automated mental health detection via social media text, but research is limited in low-resource languages.

Purpose of the Study:

• To develop and evaluate a deep learning framework for mental health classification (anxiety, depression, neutral) in Urdu.
• To create the first large-scale, publicly available Urdu dataset for mental health classification.
• To establish performance benchmarks for Urdu mental health detection using AI.

Main Methods:

• Creation of a novel annotated dataset of 36,000 Urdu tweets, classified into anxiety, depression, and neutral categories.
• Systematic translation, manual annotation, and automated labeling with rigorous quality validation.
• Adaptation and evaluation of three deep learning architectures: CNN+BiLSTM, CNN+BiGRU, and UrduBERT.

Main Results:

• UrduBERT achieved the highest accuracy (81.71%), outperforming CNN+BiLSTM (79.08%) and CNN+BiGRU (78.25%).
• UrduBERT demonstrated superior precision, recall, and F1-scores across all classification categories.
• Transformer-based models like UrduBERT excel in capturing linguistic nuances in morphologically rich languages like Urdu.

Conclusions:

• The proposed deep learning framework, particularly UrduBERT, offers a robust solution for mental health classification in Urdu.
• This work bridges a critical gap in computational mental health research for low-resource languages.
• The framework supports scalable, culturally sensitive mental health monitoring and early detection in underrepresented communities.