We explore the extent to which misrecognitions and corrections in spoken dialogue systems can be predicted from information about other turn categories in the preceding or following context. Features including whether or not subsequent turns represent `aware' sites, in which users first become aware that the system has misheard them, or corrections, and whether or not prior turns represent `aware' sites or are misrecognized are used to identify following turns as potential correction sites and to detect whether previous turns were in fact misrecognized. This represents a new phase in our ongoing work identifying corrections and misrecognitions to improve the performance of spoken dialogue systems.
Cite as: Hirschberg, J., Litman, D., Swerts, M. (2001) Detecting misrecognitions and corrections in spoken dialogue systems from `aware' sites. Proc. ITRW on Prosody in Speech Recognition and Understanding, paper 12
@inproceedings{hirschberg01_prosody,
author={Julia Hirschberg and Diane Litman and Marc Swerts},
title={{Detecting misrecognitions and corrections in spoken dialogue systems from `aware' sites}},
year=2001,
booktitle={Proc. ITRW on Prosody in Speech Recognition and Understanding},
pages={paper 12}
}