Comparison of Parsing and Spotting Approaches for Spoken Dialogue Understanding

Abstract

We have studied the optimal strategies for both LR parsing and spotting. In this report, several parsing and spotting approaches for spoken dialogue understanding are compared and evaluated. Here, a novel phrase spotting approach based on progressive search is proposed for robust understanding. The experimental results show that (1) sentence-level parsing is most powerful but not robust, (2) phrase spotting approach is robust against ill-formed utterances, (3) simple word bigram and word spotting get good word accuracy but do not lead to sentence-level understanding. Furthermore, we explore a hybrid approach where the sentence-level parsing is tried and, if it fails, the phrase spotting is performed.