Error Handling in Spoken Dialogue Systems August 28-31, 2003 Château d'Oex, Vaud, Switzerland

Using Grammatical Analysis to Detect Misrecognitions

Teresa Zollo

Department of Computer Science State University of New York College at Geneseo, USA

In systems that use grammatical analysis rather than concept spotting to accomplish natural language understanding, the presence or absence of the top-level constituent "turn" can be used to reliably detect whether the user’s speech was misrecognized. In this paper, a description of the structure of wellformed spoken turns in practical human-computer dialogue is given. We explain how that description of turns can be encoded the context-free grammar rules used by a parser, and how the result of the parser’s analysis can be used as a basis for detecting misrecognitions. We provide the results of an evaluation of this error detection strategy in the TRIPS-Pacifica domain showing 92.1%accuracy in classifying speech recognition hypotheses as correct or erroneous, an improvement of 18.2 percentage points above the majority-class baseline.

Full Paper

Bibliographic reference.  Zollo, Teresa (2003): "Using grammatical analysis to detect misrecognitions", In EHSD-2003, 83-88.