We are developing a knowledge-based system capable of understanding oral task-oriented dialogs and process pseudo-natural sublanguages (with few syntactic restrictions) with large vocabularies (several thousand words) in a multi-speaker environment. Information Centers provide a wide range of potential applications, in as much as they deal with the general public, i.e. with a large number of unfrequent and untrained speakers. In this paper, we define on the one hand, the various knowledge sources necessary for understanding and managing natural task-oriented dialogs, and on the other hand, the different types of dynamic information involved in the processing of such dialogs. We then present and comment the architecture of a knowledge-based system that could efficiently operate on multiple knowledge sources and data.