Automatic Extraction of Language Models from a Linguistic Knowledge Base

We present an algorithm for the extraction of language models from a semantic network that contains syntactic, semantic and pragmatic knowledge. The use of such language models in acoustic recognition processes results in much better system performance in speed as well as in quality of results. The automatic extraction process guarantees that the created models are always up to date and consistent with the knowledge base. The algorithm can be applied to simple constituents as well as to concepts representing an entire task domain.