Depression Detection on Social Media Using Multi-Task Learning with BERT and Hierarchical Attention: A DSM-5-Guided Approach

Depression represents a major global health challenge, yet traditional clinical diagnosis faces limitations, including high costs, limited coverage, and low patient willingness. Social media platforms provide new opportunities for early depression screening through user-generated content. However, existing methods often lack systematic integration of clinical knowledge and fail to leverage multi-modal information comprehensively. We propose a DSM-5-guided methodology that systematically maps clinical diagnostic criteria to computable social media features across three modalities: textual semantics (BERT-based deep semantic extraction), behavioral patterns (temporal activity analysis), and topic distributions (LDA-based cognitive bias identification). We design a hierarchical architecture integrating BERT, Bi-LSTM, hierarchical attention, and multi-task learning to capture both character-level and post-level importance while jointly optimizing depression classification, symptom recognition, and severity assessment. Experiments on the WU3D dataset (32,570 users, 2.19 million posts) demonstrate that our model achieves 91.8% F1-score, significantly outperforming baseline methods (BERT: 85.6%, TextCNN: 78.6%, and SVM: 72.1%) and large language models (GPT-4 few-shot: 86.9%). Ablation studies confirm that each component contributes meaningfully with synergistic effects. The model provides interpretable predictions through attention visualization and outputs fine-grained symptom assessments aligned with DSM-5 criteria. With low computational cost (~50 ms inference time), local deployability, and superior privacy protection, our approach offers significant practical value for large-scale mental health screening applications. This work demonstrates that domain-specialized methods with explicit clinical knowledge integration remain highly competitive in the era of general-purpose large language models.