A1918
Title: Integrating LLM sentiment, emotional word scoring, and syntactic dependency in speech for depression detection
Authors: Shun Hin Chan - The Hong Kong University of Science and Technology (Hong Kong) [presenting]
Abstract: Advancing statistical approaches for psychosocial health assessment requires integrating diverse analytic perspectives across speech and text. A unified framework combines Large Language Model (LLM) sentiment analysis, syntactic dependency features, and emotional word scoring to improve depression detection. First, a Beyond Holistic Context approach applies randomized chunking and statistical aggregation of LLM outputs, generating 1,000 severity rating iterations per participant, with summary statistics used as predictors of depression. Second, syntactic features, including mean sentence length reflecting average words per sentence and mean dependency distance derived from dependency grammar to capture grammatical relationships, were extracted from caregiver speech. Third, dimensional sentiment analysis using valence and arousal scores quantified emotional word use, with valence reflecting pleasantness and arousal indicating intensity. These predictors were integrated using stacking, an ensemble technique that statistically combines multiple learners. Results demonstrate that stacking meta-models consistently outperformed individual models, achieving stronger predictive performance. This multi-level statistical framework demonstrates the value of combining LLM sentiment analysis, syntactic dependency analysis, and emotional word scoring in speech, underscoring the potential of multidimensional measurements for robust, scalable psychosocial health evaluation.
