STAR: Strategy-Aware Refinement Module in Multitask Learning for Emotional Support Conversations

ACL ARR 2025 February Submission5167 Authors

Abstract: Providing effective emotional support requires strategic approaches because it is inherently complex and should account for the diverse situations and needs of each individual. The Emotional Support Conversation framework structures interactions into three phases—exploration, comforting, and action—guiding strategy selection for response generation. Although multitask learning has been used to jointly optimize strategy prediction and response generation, it often suffers from task interference, where conflicting learning objectives hinder optimization. To address this, we propose the Strategy-Aware Refinement Module (STAR), which separates and selectively integrates the decoder’s hidden states for strategy prediction and response generation through a gating mechanism. This approach preserves task-specific representations while enabling adaptive information exchange, thereby mitigating interference. Experimental results demonstrate that STAR effectively reduces task interference and achieves state-of-the-art performance in both strategy prediction and supportive response generation.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: task-oriented, conversational modeling

Contribution Types: Model analysis & interpretability, Theory

Languages Studied: English

Submission Number: 5167