Empathetic Dialogue Generation Model with Multi-Source Emotion and Intent Perception

Empathy plays a crucial role in improving the comfort of human-machine interactions. To achieve empathetic responses, a deep understanding of the user's emotional state and psychological intent is necessary. However, most of present research faces two significant issues: 1) perceiving the user's emotional state through a relatively simplistic approach, 2) linking various additional information, like user intent, into the model's input introduces more noise, which limits the model's ability to understand user intent effectively. Therefore, this paper proposes an empathetic dialogue generation model that incorporates multi-source emotion perception with intent perception. Concerning multi-source emotion perception, the model predicts the speaker's emotions by focusing on the overall contextual semantics, specific emotional words, and inferring potential emotional states of the speaker using accumulated commonsense. Concerning intent perception, to acquire the most pertinent reason information related to the user's expressed intent, incorporate special identifiers into the input sequence within the commonsense knowledge graph. We evaluated our approach on the publicly available EmpatheticDialogue dataset. The experimental results demonstrate that our approach surpasses the majority of mainstream baselines and excels in generating responses that are both more empathetic and contextually relevant.