Applying Reinforcement Learning and Multi-Generators for Stage Transition in an Emotional Support Dialogue System

Abstract

The use of empathetic dialogue systems has grown recently. However, establishing them for users experiencing mental depression requires more advanced consoling skills. In this paper, a dialogue system based on Emotional Support was developed. The system offers coping strategies through stages designed to address users' distress in long-term conversations. It employs a recurrent-based approach integrated with reinforcement learning for a decision model, which selects a generator from three specialized conditional generation models to generate empathetic responses. Experimental results showed improvements in BLEU, Rouge-L, and Distinct-n metrics compared to the baseline. On average, the system's BLEU score increased by 0.87, Rouge-L by 1.85, Distinct-1 by 0.69, and Distinct-2 by 2.26. As a result, the system generates responses aligned with Emotional Support skills, ultimately comforting the user’s distress .