PAT: A Personality-Driven Augmentation and Transfer Learning Framework for Depression Detection on Social Media

ACL ARR 2026 January Submission10244 Authors

Keywords: Depression Detection, Transfer Learning, Data Augmentation, Contrastive Learning, Interpretability, Psychology, Personality Theories.

Abstract:

Depression is a prevalent mental disorder affecting millions worldwide, with early detection crucial for effective intervention. While existing methods have achieved remarkable results in automated depression detection, they face two key limitations: (1) Weak optimization of post encoders: Relying solely on coarse user-level supervision signals prevents models from capturing depressive cues within individual posts; (2) Lack of interpretability: Current frameworks cannot substantiate their outputs with granular evidence, which undermines their trustworthiness. To address these challenges, we propose PAT, a Personality-driven Augmentation and Transfer learning framework.
PAT first optimizes the post encoder on the post-level depression detection and then transfers it to the user-level task. To fully utilize the scarce post-level data and enhance encoding performance, PAT also introduces personality-driven augmentation and fine-grained contrastive learning.
Extensive experiments demonstrate that PAT significantly outperforms existing baselines. Moreover, PAT provides comprehensive interpretability by delivering user-level predictions, tracing post-level mental-state trajectories, and highlighting key symptoms, thereby offering valuable diagnostic evidence for clinical practice.

Paper Type: Long

Research Area: Clinical and Biomedical Applications

Research Area Keywords: Clinical and biomedical language models, mental health, clinical decision support.

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings

Languages Studied: Chinese, English.

Submission Number: 10244