Knowledge-Infused Topic Model for Empathetic Dialogue Response

Empathetic dialogue generation aims to produce the responses that reflect an understanding of dialogue content in a multi-turn dialogue system through emotional expression.

Traditional methods focus on enhancing the emotional response prediction or improving the dialogue generation either through the reinforcement learning or based on the topic models.

This study employs a graph neural network to learn the characteristics and behaviors over empathetic dialogue interactions. Such a network is further integrated with recent advancements in topic model scheme to enhance the dialogue topic representation.

This model is additionally fused with the external knowledge to impose relevance in the generated response.

Experimental results demonstrate that the proposed method surpasses some related methods in automatic and human evaluations, showing improvement in terms of empathy, relevance, fluency, text perplexity, generation diversity and emotion classification accuracy.