Listening-Oriented Response Generation by Exploiting User Responses

Abstract

A dialog system is a system that enables humans and machines to communicate in natural language. Dialogue systems are largely divided into a task-oriented dialogue system and a chat-oriented dialogue system. The task-oriented dialogue system has a goal to accomplish a specific task, such as plane ticket booking, restaurant information providing, or bus route information providing. The task-oriented dialogue system has proven its usefulness in the industry in narrow domains but is not adaptable for providing natural conversation experience to humans in open-domain situations. On the other hand, the chat-oriented dialogue system aims to have a natural conversation with the user. The chat-oriented dialogue system does not have a specific task to accomplish but has a goal to act like a human.

Although listening to a conversation partner is a key factor in the success of dialogue systems or conversational agents, recent neural conversation systems have no interest in generating listening-oriented responses. In this paper, we conduct research on a dialogue system that provides responses that make users feel listened to. In detail, we propose an end-to-end dialogue system that generates listening-oriented responses which make users reveal themselves and feel positive emotions.

The proposed model uses `self-disclosure' and `positiveness' as listening features and generates responses appropriately to the features. Furthermore, the model infers a user response that will be brought out at the end of the dialogue and uses the inferred user response for generating a system response. By utilizing both listening features and user responses, our model becomes capable of generating listening-oriented responses. In quantitative and qualitative experiments, our model shows that it can generate listening-oriented responses that induce users to disclose themselves and talk positively. The results also show that the model utilizing user responses generates more listening-oriented responses than the models not using user responses.