Adapting probability-transitions in DP matching processing for an oral task-oriented dialogue

A system is presented which has been developed to evaluate the introduction of voice technologies in air-traffic controller training. The dialogue system is designed to replace the pseudopilot, a human playing the role of the pilot during training exercises and communicating both with the student controller and with the air-traffic simulator which updates the radar image on a screen. The knowledge representation which renders the cooperation of different knowledge sources possible is described. The approach makes use of pragmatic knowledge to predict a sublanguage which dynamically limits the recognition search space and therefore improves the accuracy. Recognition experiments in a task-simulation environment indicate that the combined use of both dynamic probabilities and dialogue strategies allows the system to obtain performance equal to 96.5% at the sentences level