RoleCF: Role-oriented coarse-to-fine emotion cause recognition for empathetic response generation

Abstract

In empathetic response generation, reasoning about conversational emotions by recognizing the causes of emotions is a key technique for achieving empathy. However, existing approaches encounter two fundamental limitations. First, they predominantly focus on fine-grained analysis of emotion causes at the token level, neglecting the broader, more comprehensive analysis at the utterance level. Second, these methods fail to consider emotion causes from the perspectives of different roles, resulting in biased emotional inference. To tackle the aforementioned challenges, we propose RoleCF, an innovative framework that aims to improve empathetic response generation by identifying role-oriented emotion causes in a coarse-to-fine-grained manner. Our approach models the extraction of emotion causes from different perspectives by constructing two distinct heterogeneous graphs for the user and the agent, respectively. Emotion cause nodes within each graph are utilized to swiftly capture emotion causes at the utterance level, providing a holistic understanding of the dialogue context. In addition, we employ two role-interaction modules to selectively integrate the most relevant information from the counterpart, thereby enhancing the recognition of fine-grained emotion causes. Guided by the agent’s state in the generation process, our model achieves superior performance on two benchmark datasets. This is supported by both automatic and human evaluations, demonstrating its effectiveness in capturing and leveraging the underlying causes of emotions for response generation.