Memory-Enhanced Emotional Support Conversations with Motivation-Driven Strategy Inference

Abstract

The prevalence of mental disorders is increasing globally, highlighting the need for effective emotional support systems. Emotional Support Conversation (ESC) systems have emerged as a promising solution, providing supportive responses to individuals experiencing emotional distress. However, existing ESC systems face challenges in deducing appropriate support strategies with motivations and modeling complex semantic patterns inherent in support strategies. To tackle these challenges, we propose MAGIC, a M emory-enhanced emotional support conversation system with motivation-driven str A te G y I nferen C e. Considering potential responses in future conversations, we instruct models to deduce motivations for selecting appropriate future strategies from the prior dialogue history, by harnessing the knowledge deducing abilities of Large Language Models (LLMs). These deduced motivations serve as chain-of-thought to steer models in understanding the underlying reasons of strategy inference. Moreover, to capture the intricate human language patterns and knowledge embedded in support strategies, we introduce a strategy memory store to enhance strategy modeling, by disentangling the representations from the same-strategy responses as strategy memory. Experimental results on the ESConv dataset demonstrate that MAGIC significantly outperforms state-of-the-art baselines in both automatic and human evaluations, showcasing its effectiveness in generating empathetic and supportive responses. Our codes are available at Github: https://github.com/Wh2018/MAGIC