Describir: Automatic induction of language model data for a spoken dialogue system