Transferring Large Language Models for Depression Detection through Multi-Party Conversation Analysis

Due to the complexity of traditional fine-tuning methods, prompt-based engineering methods, such as prompt tuning (PT), have recently received increasing attention for casting various downstream tasks into a large language model (LLM) format by prepending soft tunable embeddings to input sequences while keeping the majority of the LLM frozen [1]. On the other hand, the discourse analysis of text-based multi-party conversations (MPCs) has been used to obtain valuable insights like speaker emotion recognition [2]. Although most prior work considers indomain knowledge transfer using PT, much work remains to be done for PT-based out-of-domain (OOD) knowledge transfer such as adapting a LLM's knowledge about MPC modelling to a semantically dissimilar task such as depression detection. This study aims to detect human depression in MPCs by leveraging an LLM's knowledge about MPC modelling.