Enhancing Proactive Emotional Support Dialogue System with Look-forward Strategy Planning

Anonymous

Abstract: Effective emotional support (ES) is crucial to preventing severe mental health issues amid widespread mental disorders and limited access to psychological counseling. However, current emotional support conversations are limited by their simplistic single-turn interactions and lack the capability for multi-turn, look-forward strategy planning, which impedes accurately identifying users' emotional states. Additionally, ground-truth-based evaluation metrics fall short in practically assessing supportiveness and empathy in realistic dialogues. In this paper, we introduce a proactive emotional support conversational system (ProESC) to address these issues. Utilizing a small pre-trained language model, we enable the anticipation of future support strategy sequences as simulation hints, guiding LLMs in generating emotionally supportive responses and training with goal-oriented rewards. For pragmatic user feedback assessment, we employ a GPT-4 based user simulator to represent vulnerable users in need of support, evaluating responses with multi-faceted metrics. Extensive experiments demonstrate that our model surpasses competitive baselines in both strategy planning and dialogue generation, offering a more nuanced and effective approach to emotional support.

Paper Type: long

Research Area: Dialogue and Interactive Systems

Contribution Types: Model analysis & interpretability, Approaches to low-resource settings

Languages Studied: English