Dynamic Demonstration Retrieval and Cognitive Understanding for Emotional Support Conversation

TL;DR

This work tackles ESC by addressing two core challenges: generating contextually relevant, empathetic responses through dynamic demonstration retrieval and achieving deep cognitive understanding of users’ implicit mental states. It introduces Dynamic Demonstration Retrieval and Cognitive-Aspect Situation Understanding (D^2RCU), integrating in-context learning with persona-aware demonstrations and a COMET-powered cognitive module, fused via a multi-knowledge decoder. The approach leverages ATOMIC relations (Effect, Intent, Need, Want) to enrich cognitive awareness and uses Dense Passage Retrieval to select pertinent demonstrations from the ESC domain. Evaluations on the ESConv dataset show substantial gains (up to 13.79% in the overall score), with human judges noting improvements in fluency, coherence, comforting, and usefulness, and the authors provide public code for reproducibility.

Abstract

Emotional Support Conversation (ESC) systems are pivotal in providing empathetic interactions, aiding users through negative emotional states by understanding and addressing their unique experiences. In this paper, we tackle two key challenges in ESC: enhancing contextually relevant and empathetic response generation through dynamic demonstration retrieval, and advancing cognitive understanding to grasp implicit mental states comprehensively. We introduce Dynamic Demonstration Retrieval and Cognitive-Aspect Situation Understanding (\ourwork), a novel approach that synergizes these elements to improve the quality of support provided in ESCs. By leveraging in-context learning and persona information, we introduce an innovative retrieval mechanism that selects informative and personalized demonstration pairs. We also propose a cognitive understanding module that utilizes four cognitive relationships from the ATOMIC knowledge source to deepen situational awareness of help-seekers' mental states. Our supportive decoder integrates information from diverse knowledge sources, underpinning response generation that is both empathetic and cognitively aware. The effectiveness of \ourwork is demonstrated through extensive automatic and human evaluations, revealing substantial improvements over numerous state-of-the-art models, with up to 13.79\% enhancement in overall performance of ten metrics. Our codes are available for public access to facilitate further research and development.

Figures (5)

• Figure 1: Illustration of an emotional support conversation. For queries posted by help-seeker, knowledge is extracted to support responses through the dynamic demonstration selector and COMET within the proposed D$^2$RCU.
• Figure 2: The overall architecture of our proposed D$^2$RCU, comprises three key components: ①Dynamic Demonstration Selector, ②Cognitive-Aspect Situation Understanding, and ③Multi-Knowledge Fusion Decoder. The details of the Cognitive Understanding process are highlighted by the bottom orange box. Given a user's post, we first obtain prior knowledge through dynamic demonstration selection and COMET commonsense extraction, then understand the user's feeling through the cognition of the current situation, and finally generate responses through a knowledge-aware decoder.
• Figure 3: The Dynamic Demonstration Selector architecture. The retrieval component (top) utilizes a pre-trained DPR to identify relevant responses within the training set, while the demonstration component (down) compiles the most pertinent user-system pairs, informed by the top $s$ results.
• Figure 4: The top-n accuracy of predicted strategies. D$^2$RCU consistently gains best performance.
• Figure 5: Comparative results from our case studies of D$^2$RCU and state-of-the-art PAL.