Error detection in spoken dialogue systems

In conversation, speakers try to ground what they do together, in order to reach mutual understanding. Some miscommunication problems will inevitably occur, as the speakers will try to minimize their collaborative effort for achieving their goals. This is normally not a problem for human speakers, as they have developed methods for handling such errors. However, in human-computer dialogue, this can lead to serious problems if we don’t learn to handle the specific problems that arise due to the errorprone speech recognition technology. Methods for preventing, predicting, detecting and recover from such problems must therefore be investigated. In this paper, three different approaches to error detection are described and discussed. The first is to detect errors in the recognition result in order to choose an appropriate grounding strategy. The second is to detect previous errors that have been made, based on the users reactions to the system’s grounding acts. The third is to predict possible future errors, based on the first turns of the dialogue.