Evaluating Document-Tuned Transformer Representations for Person-level Mental Health Assessment

The study evaluates document-tuned transformers against base-transformers for person-level mental health assessment, revealing that document-tuned models, which are further fine-tuned at the document level, achieve a 13.4% increase in Pearson correlation (p=.015) across two psychological datasets. Robustness tests indicate these models maintain higher accuracy under various perturbations and better capture hedged language, suggesting they are more effective for predicting mental health outcomes. This highlights the importance of model representation choice in enhancing the reliability of AI-driven psychological assessments.