Event

Predicting Content of a User's Utterance in a Dialog

Svetlana Stoyanchev

SUNY, Stony Brook

Thursday, February 26, 2009
2:00-3:30 p.m.

In deployed dialog systems, the system often requests confirmation of user-provided task-relevant concepts. The user utterance following a confirmation prompt may be a simple confirmation or rejection, but may also be a correction or topic change containing concepts not in the system's confirmation language model. For example, in data samples from the Let's Go! bus information dialog system, 18.1% of post-confirmation user utterances contain a concept. These utterances are likely to be misrecognized, causing frustration for the user. Speech recognition performance can be improved by automatically determining whether the user's post-confirmation utterance contains no concept, or a concept of a particular type, and adapting the recognizer's language model accordingly. In this work, we evaluate automatic concept type classification for Let's Go! and its effect on speech recognition.