UFI4ER: An Utterance-Level Feature Dynamic Interaction Model for Cognition-Enhanced Empathetic Response Generation

In multi-turn empathetic dialogues, the implicit features such as cognition, affection, and behavior, expressed in the utterances, are not static. Instead, they naturally engage in dynamic interactions throughout the conversation, facilitating the emergence and development of empathy. However, existing works primarily focus on capturing dialogue-level features, disregarding the sequential structure of dialogues and failing to perceive the dynamic interactions of utterance-level features. Additionally, aligning phrase-level commonsense knowledge with the context poses challenges. To address these limitations, we propose an utterance-level feature dynamic interaction model for cognition-enhanced empathetic response generation. We construct a two-stage graph attention network that integrates event chains to enhance cognitive understanding and leverages the distinctive structure of the dialogue along with the dynamic interactions of utterance-level features to simulate the progression of empathy. Furthermore, we incorporate contextual commonsense knowledge to enhance the understanding of the context. Additionally, we have designed our proposed method as a prompt template to guide the Large Language Models (LLMs) in generating more empathetic responses. Experimental results on EmpatheticDialogues demonstrate the superiority of our approach over baselines in both automatic and manual evaluations.