Exploring the Repetition Problem in Empathetic Dialogue Generative Model

Abstract

We explore the task of generating an empathetic response of dialogue agents. Recent empathetic dialogue agents have shown their abilities to generate empathetic responses equipped with additional modules, such as mixture-of-experts, emotion mimicry, or external knowledge. However, we find even these agents prefer to generate generic and repetitive responses like "I am sorry to hear that ...", which is the empathy statement in the task of empathetic dialogue generation. From the analysis, we reveal that this tendency is attributed to the nature of EMPATHETICDIALOGUES dataset. Through the experiment, we confirm that only reducing repetition problem in the empathetic dialogue generation is not optimized solution.