Comparative Evaluation Of a Stochastic Parser On Semantic And Syntactic-Semantic Labels
University of Ulm, Dept. of Information Technology
This paper deals with the evaluation of a stochastic component for natural language understanding alternatively trained on semantic and syntactic-semantic labels. The parser uses semantically-labeled speech data gathered using the LIMSI-ARISE interactive speech system for train travel information retrieval in French. The study shows that introducing additional and coherent information into the semantic corpus allows to relatively improve the semantic frame accuracy of the parser by up to 16.5%. The more complex models yielding a high number of parameters are justified, as long as they convey significant information.
Attribute-value pairs; Data labeling; Hidden Markov Models; Robustness; Semantic frame; Spontaneous human-machine interaction