A Multi-resolution Mechanism with Multiple Decoders for Empathetic Dialogue Generation

Abstract

To build an interactive human-like conversational agent, one of the key challenges is to learn to generate responses in a more empathic way. When a conversation has a tone of empathy to it, the person conversing with the conversational agent will feel a sense of contentment. Building an empathetic dialogue system results in increased user interaction and thereby more effective communication. To improve the perception and expression of emotional states, the proposed model uses a multi-resolution mechanism that helps to capture the nuances of user emotions effectively. Moreover, it uses separate decoders, which are optimized to respond to specific emotions, and combines the output response representations to generate an empathetic response. Experimental results show that the proposed approach excels the state-of-the-art baselines in empathy, relevance, and fluency.