Design considerations for knowledge source representations of a stochastically-based natural language understanding component

We describe a stochastic component for spoken natural language understanding in an application for train travel information retrieval in French. The aim is to discuss the design considerations for representing knowledge sources appropriately in such a parsing component. The development focuses on the design of a stochastic model topology that is optimally adapted in quality and complexity to the task model and the available training data. Another important issue concerns the iterative semantic labeling of large data amounts used for the component training, The parser has been evaluated on both corrected and uncorrected speech recognizer output transcriptions. (C) 1999 Elsevier Science B.V. All rights reserved.