摘要
Abstract
Emotional Support Conversation(ESC)systems are crucial in many applications,such as customer service interactions and auto-mated psychological therapy.However,existing ESC systems typically overlook two important issues:(a)they generally rely on simple emo-tion labels to describe user states,capturing the reasons behind users' emotions less frequently;(b)they often focus on optimizing strategy se-lection,potentially neglecting the relevance and diversity of responses.Therefore,this paper proposes an Emotion Cause enhanced Model With Latent Variable(ECLV).First,the paper introduces commonsense knowledge to supplement emotion-related information,establishing a fine-grained emotional state of the user to provide clues for inferring the emotion causes.Subsequently,continuous latent variables are used to describe the latent features of the input data,and a memory scheme is employed to allow the decoder to directly focus on this information,thereby improving response relevance and diversity.The model surpasses existing baselines in both automatic and human evaluations,with sig-nificant improvements in strategy prediction accuracy(ACC),response relevance(B-2),and diversity(D-2)metrics,validating the mod-el's effectiveness.关键词
情感支持对话/隐变量/情绪原因/常识知识/记忆模式Key words
emotional support conversation/latent variable/emotion cause/commonsense knowledge/memory schema分类
信息技术与安全科学
