Detecting Misrecognitions and Corrections in Spoken Dialogue Systems from ‘Aware’ Sites

摘要: We explore the extent to which misrecognitions and corrections in spoken dialogue systems can be predicted from information about other turn categories preceding or following context. Features including whether not subsequent turns represent ‘aware’ sites, users first become aware that system has misheard them, corrections, prior sites are misrecognized used identify as potential correction detect previous were fact misrecognized. This represents a new phase our ongoing work identifying improve performance of systems.