A data-centric architecture for data-driven spoken dialog systems

Data is becoming increasingly crucial for training and (self-) evaluation of spoken dialog systems (SDS). Data is used to train models (e.g. acoustic models) and is 'forgotten'. Data is generated on-line from the different components of the SDS system, e.g. the dialog manager, as well as from the world it is interacting with (e.g. news streams, ambient sensors etc.). The data is used to evaluate and analyze conversational systems both on-line and off-line. We need to be able query such heterogeneous data for further processing. In this paper we present an approach with two novel components: first, an architecture for SDSs that takes a data-centric view, ensuring persistency and consistency of data as it is generated. The architecture is centered around a database that stores dialog data beyond the lifetime of individual dialog sessions, facilitating dialog mining, annotation, and logging. Second, we take advantage of the state-fullness of the data-centric architecture by means of a lightweight, reactive and inference-based dialog manager that itself is stateless. The feasibility of our approach has been validated within a prototype of a phone-based university help-desk application. We detail SDS architecture and dialog management, model, and data representation.