Detection of Topic Transition in a Chat Dialogue System

Abstract

The purpose of this study is to develop a system for detect topic transitions in a chat dialogue in order to grasp the context, which enabling to reduce conversation breakdown. Words used in utterances are vectorized using Word2Vec, and the similarity between utterances is calculated by cosine similarity. Then, suppose that the topic transitions when the similarity between utterances becomes low. Using an actual dialogue corpus, we conducted a subject experiment comparing the results of the system extracting topic transitions with the results actually judged by subjects, and found that the topic transitions were captured at a position close to the human judgment. However, the system presumed that there was no topic transition when similar words were used even for different topics. In addition, the accuracy was improved by adding the algorithm to use an intersection of words between distant sentences.