Enhancing Emotional Support Conversation System via Integrating Mental State-Strategy Reasoning
Yaru Cao, Hongzhi Yu, Fucheng Wan · 2024
As the demand for emotional support continues to grow, strategy-driven generation methods have become widely adopted to enhance the performance of emotional support conversation systems. However, relying solely on strategy names is insufficient to capture the underlying logic of emotional support responses. This paper introduces a Mental State-Strategy Reasoning framework designed to help supporters accurately understand the mental states of help-seekers, leading to the selection of appropriate strategies and the clear identification of each strategy's goals. The framework aims to improve the system's capability in emotional support dialogues. To validate the effectiveness of the Mental State-Strategy Reasoning framework, we applied it to the ESConv dataset, creating an expanded dataset that incorporates this framework. A large language model was trained on the extended dataset, and comparative experiments were conducted in both in-domain and out-of-domain settings. The results demonstrated significant improvements over existing strategy-based models across several metrics, such as BLEU and R-L, indicating that integrating Mental State-Strategy Reasoning can effectively enhance emotional support systems and improve the model's generalization performance.