Run-time discourse processing to supplement incomplete ASR

The discourse processing component of a spoken dialogue system can play any of the following four roles in improving the system´s overall efficacy as a tool: the dialogue manager can prevent/discourage out-of-grammar input by setting up the immediate context so as to force the user to constrain the input speech; the dialogue model can be used to reduce word-error rate by indicating to the speech recognizer (SR) the most probable language model to use; the dialogue model can be used to rescore the n-best output of the SR component; and dialogue management strategies can compensate for weaknesses in the input signal by bringing context information to bear on the task of understanding input speech in the presence of incomplete recognition. It is this fourth role for discourse that is the topic of this paper