ExDoRA: Enhancing transferability of large language models for depression detection using free-text explanations

Y.H.P.P Priyadarshana^*Zilu LiangIan Piumarta

• Kyoto University of Advanced Science (KUAS), Kyoto, Japan

Few-shot prompting in large language models (LLMs) significantly improves performance across various tasks, including both in-domain and previously unseen natural language tasks, by learning from limited in-context examples. How these examples enhance transferability and contribute to achieving state-of-the-art (SOTA) performance in downstream tasks remains unclear. To address this, we propose ExDoRA, a novel LLM transferability framework designed to clarify the selection of the most relevant examples using synthetic free-text explanations. Our novel hybrid method ranks LLM-generated explanations by selecting the most semantically relevant examples closest to the input query while balancing diversity. The top-ranked explanations, along with few-shot examples, are then used to enhance LLMs' knowledge transfer in multi-party conversational modelling for previously unseen depression detection tasks. Evaluations using the IMHI corpus demonstrate that ExDoRA consistently produces high-quality free-text explanations. Extensive experiments on depression detection tasks, including depressed utterance classification (DUC) and depressed speaker identification (DSI), show that ExDoRA achieves SOTA performance. The evaluation results indicate significant improvements, with up to 20.59% in recall for DUC and 21.58% in F1 scores for DSI, using 5-shot examples with top-ranked explanations in the RSDD and eRisk 18 T2 corpora. These findings underscore ExDoRA's potential as an effective screening tool for digital mental health applications.