ACE-GoT: Affective and Contextual Expansion via Graph-of-Thought for Emotional Support Conversations

Emotional Support Conversation (ESC) systems are required not only to generate fluent responses but also to exhibit both empathy and interpretable reasoning. While existing Chain-of-Thought (CoT) approaches have improved model transparency, their linear and static structures are insufficient to capture the multifaceted nature of emotions and the complexity of situational contexts. Consequently, they struggle to adapt support strategies dynamically. In this paper, we propose ACE-GoT (Affective and Contextual Expansion via Graph-of-Thought), an emotional support framework built upon the GoT paradigm. ACE-GoT consists of six modules that jointly construct a graph-structured reasoning process. Among them, the Affective Knowledge Expansion Module and the Semantic Relational Inference Module are the core components, enabling affective commonsense inference and modeling semantic connections between events and emotions. These modules explicitly expand the underlying emotional context, providing richer contextual cues for downstream decision-making. The remaining modules support this expansion and organize the reasoning graph for coherent integration. Furthermore, the framework incorporates strategy planning grounded in Helping Skills Theory to enhance response adaptability. Experimental results on multiple ESC benchmarks demonstrate that, without further fine-tuning, ACE-GoT significantly outperforms baselines in the G-Eval framework, which has been shown to correlate with human preferences in prior studies. Notably, ACE-GoT achieves a 3.08-point improvement in ROUGE-L over the strong ESCoT_LLaMA baseline. Human evaluation further confirm the superiority of our approach. This work demonstrates the effectiveness of the ACE-GoT framework in advancing the quality, generalizability, and interpretability of emotional support dialogues.