Abstract

Depression affects individuals globally, necessitating efficient diagnostic tools. This study introduces an advanced unsupervised hybrid approach, that automatically converts binary-labelled depression datasets into multi-class datasets by integrating a rule-based system with Large Language Models (LLMs). The rule-based system employs the Beck Depression Inventory-II tool that can be used to classify depression levels based on predefined scoring rules, and these rules are segregated into clusters based on score ranges from 0-3. LLMs employ fine-tuned large language Model Meta AI2 (LLaMa2) to generate domain-specific embedding from social media posts. By harnessing LLMs’ contextual understanding, both BDI rules and social media posts are embedded, thereafter cosine similarity is applied to calculate semantic similarities. Based on the similarity score, each post is assigned to the most similar BDI cluster, with the highest similarity score, creating a refined multiclass depression cluster. To evaluate clustering effectiveness, the silhouette score was computed, yielding an average score of 0.45, indicating moderate clustering quality. Additionally, 30% of the binary depression dataset was manually labelled by clinical experts. The Normalized Mutual Information (NMI) score of 0.53 further validated the method, showing strong alignment between the generated clusters and expert-labelled data. This approach enhances depression severity classification, providing a scalable, efficient, and accurate tool for researchers and practitioners.

Authors

Divya Dewangan¹, Smita Selot², Sreejit Panicker³
Shri Shankaracharya Technical Campus, India^1,2, Techment Technology, India ³

Keywords

Large Language Model, Multiclass Depression Dataset, Beck Depression Inventory Tool, Semantic Similarity