MetaESC: Enhancing Emotional Support Conversation through Metacognition

Abstract

Emotional Support Conversation (ESConv) aims to alleviate seekers’ emotional distress by employing supportive strategies and responses. However, current approaches for strategy selection primarily rely on dialogue history, overlooking the importance of reflection on the chosen strategy. Additionally, these approaches simply lower the priority of an erroneously selected strategy, lacking the dynamic adjustment of strategy priorities based on the subsequent dialogue context. In this work, we propose a novel model called the Metacognition Control Network (MetaESC). Drawing inspiration from Metacognition theory, MetaESC enables reflective thinking on dialogue scenarios and strategy choices, facilitating improved strategy planning and supportive responses. Concretely, as emotional fluctuations play a crucial role in determining supporters’ choice of support strategies, we integrate the emotional fluctuations of seekers with a Large Language Model, enabling comprehensive strategy reflection. Moreover, by considering the conversational context, we utilize a multi-head attention mechanism to update the user’s state and dynamically adjust the priority of the selected strategy during the ongoing interaction. Extensive experimentation and empirical studies validate the effectiveness of MetaESC in generating emotion-supportive conversations, outperforming baseline models.