A Method for Determining the Utterance-Response Relationship between Dialogues using a Language Model for Construction of Dialogue Corpus using Novels

Abstract

The purpose of this study is to automatically construct a dialogue corpus using utterances by characters in novels, in order to reduce the cost of constructing a Japanese daily dialogue corpus. However, there is a problem with the small number of utterances contained in a single conversation when extracting consecutive utterances in the novel as a single conversation. Therefore, we focus on utterances in which the utterance-response relationship is established even if other text is inserted between them, and propose a method for determining the utterance-response relationship between two utterances using a language model. From the experimental results, we confirmed that the proposed method can determine the utterance-response relationship between two utterances with the average of F1 value of 80.3%.