Towards Reliable and Empathetic Depression-Diagnosis-Oriented Chats

TL;DR

This work introduces SEO, a two-component ontology framework for depression-diagnosis chats that unifies Symptom-related TOD and Empathy-related chit-chat into a cohesive Task-Oriented Chat (TOC) system. By annotating the D$^4$ dataset with SEO and developing Res-Track for dialogue state tracking, the approach enables dynamic intent transitions and emotionally supportive, diagnostically reliable interactions. Experimental results across intent prediction, DST, response and summary generation, and risk classification show significant gains over the baseline, with human evaluators noting improved fluency, coherence, and empathy. The framework advances digital mental health tooling by delivering more accurate depression assessment while maintaining empathetic engagement, with careful attention to ethical considerations and limitations in empathy quality and evaluation metrics.

Abstract

Chatbots can serve as a viable tool for preliminary depression diagnosis via interactive conversations with potential patients. Nevertheless, the blend of task-oriented and chit-chat in diagnosis-related dialogues necessitates professional expertise and empathy. Such unique requirements challenge traditional dialogue frameworks geared towards single optimization goals. To address this, we propose an innovative ontology definition and generation framework tailored explicitly for depression diagnosis dialogues, combining the reliability of task-oriented conversations with the appeal of empathy-related chit-chat. We further apply the framework to D$^4$, the only existing public dialogue dataset on depression diagnosis-oriented chats. Exhaustive experimental results indicate significant improvements in task completion and emotional support generation in depression diagnosis, fostering a more comprehensive approach to task-oriented chat dialogue system development and its applications in digital mental health.

Figures (4)

• Figure 1: Task-Oriented Chat in Depression Diagnosis, complex intents with flexible intent transition, simultaneously enveloping TOD and Chit-chat in single turn.
• Figure 2: SEO Intent Distribution of D$^4$. The length along the circumference serves as an indicator of the proportional representation of the intent. Note that a 3-level intent hierarchy is maintained, the complete intent list is provided in Appendix \ref{['tab:ontology']}.
• Figure 3: Intent Transition Tendency. The height represents the proportion of dialogues at this core symptoms and empathy strategies mentioned in §\ref{['subsec:ontology definition']}
• Figure 4: Symptom Intent Selection Tendency. The value on the vertical axis indicates the core depressive symptoms mentioned in §\ref{['subsec:ontology definition']}