Stochastic understanding models guided by connectionist dialogue acts detection

We study the use of specific stochastic models for the understanding process in a spoken dialogue system. A previous classification of the user turns in terms of dialogue acts is accomplished by connectionist models to guide the understanding process. Some specific issues are explored, like the multiclass classification problem, the smoothing of models, and the generation of the frames which constitute the input of the dialogue manager. Some experiments using the correct transcription of the user turns and the output of the speech recognizer are presented.