Non-Task-Oriented Attentive Listening Dialogue System using Transfer Learning

Abstract

With the development of information and communication technology, practical dialogue systems that can talk with humans are attracting a great deal of attention. Dialogue systems are broadly divided into two types: task-oriented and non-task-oriented. The latter is aimed at dialogue itself such as chat, and basically it is necessary for the user to provide a topic, there is a problem that the motivation for dialogue is reduced due to boredom and familiarity. Therefore, a non-task-oriented attentive listening dialogue system is proposed to improve the problem. In the system, a new method is adopted that combines a response generation model with BERT as an encoder and Transformer as a decoder and transfer learning. In-depth questions and empathy responses are generated to realize listening dialogue by learning the dialogue data acquired using the Twitter API. The process of increasing vocabulary was reduced and response times were shortened and natural dialogue became possible by learning the huge amount of dialogue data collected from Twitter. In addition, by using transfer learning, it became possible to control the response style, also the problem of reduced dialogue motivation caused by poor topic was improved.