Year: 2025
Type: dissertation
Abstract: This thesis investigates how embeddings from different pre-trained large language models (LLMs) influence the prediction of mental health rating scale scores from free-text input, and examines how well the models' underlying word-and sentence-level attributions align with expert judgments.We compare three LLMs-BERT, RoBERTa, and Mixedbread-in their ability to predict PHQ-9 depression scores and highlight important text components.A multi-level visualization tool was reconstructed and extended to make model predictions more transparent at the word, sentence, and paragraph levels.Our results show that larger, more specialized models like Mixedbread yield stronger correlations with expert ratings, both in paragraph-level predictions and sentence-level contribution rankings.However, alignment at the word level was limited, indicating that further refinement in modeling or annotation strategies may be needed to capture finegrained linguistic cues.This work highlights both the promise and current limitations of using LLMs in clinical contexts, and proposes directions for building interpretable, NLP-based tools to support mental health assessment.
Authors: Broman, Nils, Olsson, Rasmus
Institution: Lund University
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7006375265