DGAC: Dialogue Graph Auto Construction Based on Data with a Regular Structure

Abstract

AbstractA scripted graph is a way to represent a dialogue scenario in a dialogue system. It is often used in development when the dialogue has a regular structure. In this paper, we propose a method for recovering or extracting regular structures from the data by building a dialogue graph. The dialogue graphs constructed by our method can be used for a more accurate pre-selection of candidates for response selection models on the MultiWOZ dataset. Quality improvements are demonstrated for various response selection models: statistical, pre-trained, and fine-tuned. Obtained results demonstrate the applicability of our approach to the automatic construction of a dialogue graph for the tasks of creating scenario-driven dialogue assistants and improving response selection models.KeywordsDialogue graph auto constructionDialogue systemGraph neural networkClusteringIntents