Abstract

Computing confidence scores for applications, such as dialogue system, information retrieving and extraction, is an active research area. However, its focus has been primarily on computing word-, concept-, or utterance-level confidences. Motivated by the need from sophisticated dialogue systems for more effective dialogs, we generalize the confidence annotation to all the subtrees, the first effort in this line of research. The other contribution of this work is that we incorporated novel long distance features to address challenges in computing multi-level confidence scores. Using Conditional Maximum Entropy (CME) classifier with all the selected features, we reached an annotation error rate of 26.0 % in the SWBD corpus, compared with a subtree error rate of 41.91%, a closely related benchmark with the Charniak parser from (Kahn et al., 2005). 1

Citations