TRACKING INITIATIVE IN COLLABORATIVE DIALOGUE INTERACTIONS

A processing system which acts as a participant in the collaborative dialogue is first trained using a training algorithm. The training algorithm computes, in conjunction with a current turn of the training dialogue, a task initiative index and a dialogue initiative index for a next turn of the dialogue, using cues observed during the current turn, probability assignments associated with the observed cues, and current values of the task initiative index and the dialogue initiative index. The algorithm then predicts a task initiative holder and a dialogue initiative holder for the next turn of the training dialogue, using the task initiative index and dialogue initiative index computed for the next turn. The probability assignments associated with the observed cues are adjusted if the predicted task and dialogue initiative holders do not agree with corresponding respective actual task and dialogue initiative holders for the next turn. The computing, predicting and adjusting operations are repeated for each of the turns of the training dialogue in order to determine appropriate probability assignments for a given set of cues. The processing system then uses the resulting probability assignments to predict initiative shifts in an actual dialogue in which the system acts as a participant.