In Context: Integrating Domain- and Situation-Specific Knowledge

We describe the role of context models in natural language processing systems and their implementation and evaluation in the SmartKom system. We show that contextual knowledge is needed for an ensemble of tasks, such as lexical and pragmatic disambiguation, decontextualizion of domain and common-sense knowledge that was left implicit by the user and for estimating an overall coherence score that is used in intention recognition. As the successful evaluations show, the implemented context model enables a multi-context system, such as SmartKom, to respond felicitously to contextually underspecified questions. This ability constitutes an important step towards making dialogue systems more intuitively usable and conversational without loosing their reliability and robustness.