Bootstrapping dialogue systems: the contribution of a semantic model of interactional dynamics

We investigate an end-to-end method for automatically inducing task-based dialogue systems from small amounts of unannotated dialogue data. The method combines an incremental, semantic grammar formalism - Dynamic Syntax (DS) and Type Theory with Records (DS-TTR) with Reinforcement Learning (RL), where language generation and dialogue management are treated as one and the same decision problem. The systems thus produced are incremental: dialogues are processed word-by-word, shown in prior work to be essential in supporting more natural, spontaneous dialogue. We hypothesised that given the rich linguistic knowledge present within the grammar, our model should enable a combinatorially large number of interactional variations to be processed, even when the system is trained from only a few dialogues. Our experiments show that our model can process 70% of the Facebook AI bAbI data-set - a set of unannotated dialogues in a ‘restaurant-search’ domain even when trained on only 0.13% of the data-set (5 dialogues). This remarkable generalisation property results from the structural knowledge and constraints present within the grammar, and highlights limitations of recent state-of-the-art systems that are built using machine learning techniques only.