Variational Autoencoding Dialogue Sub-Structures Using a Novel Hierarchical Annotation Schema

Abstract

This work presents a novel method to extract sub-structures in dialogues for the following genres: human-human task driven, human-human chit-chat, human-machine task driven, and human-machine chit-chat dialogues. The model consists of a novel semi-supervised annotation schema of syntactic features, communicative functions, dialogue policy, sequence expansion and sender information. These labels are then transformed into tuples of three, four and five segments, the tuples are used as features and modelled to learn sub-structures in above mentioned genres of dialogues with sequence-to-sequence variational autoencoders. The results analyse the latent space of generic sub-structures decomposed by PCA and ICA, showing an increase in silhouette scores for clustering of the latent space.