Making an Information State-Based Spoken Dialogue System Less Fragile

We describe several architectural features of the Athos agentbased framework for the construction of spoken dialogue systems, and additional features in two application demonstrators, AthosMail and AthosNews, that contribute to the robustness of their operation. These include adaptation of recognition resources to personal and application data, use of a broad-coverage robust parser as an intermediate layer between ASR and logical interpretation building, discourse model preinitialization, sophisticated reasoning-based discourse interpreter capable of resolving complex expressions referring to sets and individuals.