Hierarchical Dialogue Management

The dialogue management in a Spoken Dialogue System aims at learning a decision rule that can select the best system action in response to the user goal. However, this task results extremely hard for on-line applications, especially because these systems usually act in a multi-domain setting and require a large amount of training data. Additionally, the environment they have to perform in is generally highly noisy, because of the variability of human speech that causes errors in the transcriptions. As a consequence, the input received by the system is corrupted and the response provided to the user is consequently affected. Moreover, the uncertainty on the user intention generates a large input space, since every possible utterance needs to be considered. Therefore, the computational cost dramatically increases and contributes to the complexity of the optimisation problem. In this work, we try to exploit the similarities between the domains and improve the generalisation capability of the model when a large set of domains is considered. For this purpose, we build a treestructured hierarchy for organising the domains, so that similar domains are clustered together. We show that the approach proposed provides performances that are generally comparable to the most recently used models in the literature. Furthermore, the slight improvements gained in some of the experiments open to other developments for future works.