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

Using Knowledge-Based Scores for Identifying Best Speech Recognition Hypothesis

Iryna Gurevych, Robert Porzel

European Media Lab GmbH, Heidelberg, Germany

The paper presents the evaluation of a knowledge-based scoring method applied to the problem of identifying the best speech recognition hypothesis (SRH) in a functioning multimodal dialogue system. The competing SRHs are evaluated in terms of their semantic coherence using the high-level domain knowledge encoded in the ontology. We conducted an annotation experiment and showed that humans can reliably select the best SRH in a given N-best list (agreement 95.35%). The knowledge-based method identifies correctly 88.07% of the best SRHs (given the baseline 63.91%), which is also an improvement over the automatic speech recognizer (ASR) (83.88% accuracy).

Full Paper

Bibliographic reference.  Gurevych, Iryna / Porzel, Robert (2003): "Using knowledge-based scores for identifying best speech recognition hypothesis", In EHSD-2003, 77-81.