Statistical Language Models for Dialogue Systems

Abstract. The aim of this overview4 is to describe major approaches and trends used for statistical language modelling in automatic speech recognition (ASR) modules of dialogue systems. We do not attempt to provide a thorough and comprehensive overview of all the techniques applied to language modelling in general but concentrate on those that are especially beneficent when implemented in a dialogue system. The paper consists of three major parts. The first one is be devoted to ba-sic principles and approaches to statistical language modeling in speech technology (Bayes rule, N-gram approach). The second one, that can be regarded as an intermediary one, puts special emphasis on those refine-ments of a general model that are vital for dialogue systems (class and cache models, domain adaptation). The last part of this paper deals with ad-hoc statistical language models for dialogue systems (LMDS) and so-lutions to specific problems in that domain (state-dependent models, language model data induction).