New techniques for speech understanding

A standard semantic analyzer accepting dictated sentences belonging to a grammar is used in a vocal information retrieving system. The problems of managing sentences not in the grammar and errors from the speech recognizer are addressed. Three different methods are proposed, based, respectively, on alignment techniques, conceptual categories (CCs), and neural networks. To test and compare the three approaches, several benchmarks were run. Each benchmark is specific for a kind of query sentence category: in-grammar with recognition errors, out-of-grammar with no recognition errors, etc. It is concluded that the CCs extraction phase combined with the neural approach for evaluating logical rules leads to a solid speech understanding system.<>