Empath: Enhancing LLMs’ Empathy in Emotional Support Conversation with Group MoELoRA and Experience RAG

ACL ARR 2025 May Submission1155 Authors

Abstract: This paper aims to enhance the empathetic capabilities of Emotional Support Conversation (ESC) models. To this end, we propose **Empath**, a novel ESC framework under the guidance of a licensed psychotherapist and grounded in Bohart and Greenberg’s Empathy Theory, designed to improve Person, Affective, and Cognitive Empathy. At its core, **Empath** features the Group MoELoRA architecture and Experience RAG. Group MoELoRA personalizes support by tailoring character perspectives and dynamically adjusting support strategies based on emotional and contextual cues, while Experience RAG enriches interactions by aligning seeker concerns with relevant counselor experiences for deeper understanding. To train **Empath**, we introduce EmpathSupport-52k, a large-scale, multi-role, multi-strategy psychological counseling dataset. Extensive experiments demonstrate that **Empath** surpasses baseline models in both empathetic engagement and emotional support effectiveness.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: spoken dialogue systems, applications, conversational modeling

Contribution Types: NLP engineering experiment, Data resources, Data analysis

Languages Studied: English

Submission Number: 1155