A multilingual expectations model for contextual utterances in mixed-initiative spoken dialogue

This paper describes a model of generating expectations that are used to improve speech recognition and to resolve elliptical expressions in dialogue context. The algorithm is domain and language independent and part of the dialogue manager. We use the expectation model to weight a speech recognizer’s grammar rules in dialogue context which improves recognition rates significantly as shown in the evaluation. We explain what types of expectations the system can generate and give a classification of system actions based on speech act theory, explain the resolution of elliptical expressions and their interpretation in context, and evaluate the presented algorithm in a multilingual system with English and German speech recognition.