Integrating multiple knowledge sources for utterance verification in a large vocabulary speech understanding system