Keyword classification trees for speech understanding systems

• Main Author: Kuhn, Roland
• Other Authors: Mori, Renato De (advisor)
• Format: Others
• Language: en
• Published: McGill University 1993
• Subjects: Computer Science.
• Online Access: http://digitool.Library.McGill.CA:80/R/?func=dbin-jump-full&object_id=41238

Speech understanding systems try to extract meaning from one or several word sequences hypotheses generated by a speech recognizer. Designers of these systems rely increasingly on robust matchers to perform this task; a robust matcher processes semantically important word islands rather than attempting to parse the entire word sequence. This thesis describes a robust matcher for speech understanding whose rules are learned automatically from training data. Learning is carried out by a new set of algorithms involving a new data structure, the Keyword Classification Tree (KCT). By eliminating the need to handcode and debug a large number of rules, this approach facilitates rapid construction of a speech understanding system. Furthermore, the rules learned by a KCT, which depend on a very small number of words in each utterance, are highly resistant to errors by the speaker or by the speech recognizer. The thesis discusses a speech understanding system built at the Centre de Recherche Informatique de Montreal that incorporates this robust matcher, using the DARPA-sponsored Air Travel Information System (ATIS) task as training corpus and testbed.