Unsupervised graph extraction for improvement of multi-domain task-oriented dialogue modelling

This paper proposes a dialogue graph extraction method based on a textual corpus. The dialogue graph displays the main conversational topics for each turn of the dialogue and defines transitions between them. A node in this graph is represents the main conversational topic for the dialogue turn and the edge stands for a possible transition to one of the adjacent topics. The proposed method of dialogue graph extraction is based on a set of clustering algorithms accompanied by a set of heuristics. It is applied to variable size corpora from different domains. We provide visualizations of resulting dialogue graphs for a number of datasets and visualizations aid understanding of the underlying latent dialogue structure in each corpus. As a proof of the concept, we integrate the graph features by a sequence-to-sequence model and improve the current baseline on MultiWOZ 2 response generation task by more than 15%.