Declarative design of spoken dialogue systems with probabilistic rules

Spoken dialogue systems are instantiated in complex architectures comprising multiple in-terconnected components. These architectures often take the form of pipelines whose com-ponents are essentially black-boxes developed and optimised separately, using ad-hoc spec-ification formats for their inputs and outputs, domain models and parameters. We present in this paper an alternative mod-elling approach, in which the dialogue pro-cessing steps (from understanding to manage-ment and to generation) are all declaratively specified using the same underlying formal-ism. The formalism is based on probabilistic rules operating on a shared belief state. These rules are expressed as structured mapping be-tween state variables and provide a compact, probabilistic encoding for the dialogue pro-cessing models. We argue that this declarative approach yields several advantages in terms of transparency, domain-portability and adap-tivity over traditional black-box architectures. We also describe the implementation and vali-dation of this approach in an integrated archi-tecture for human-robot interaction.