A combined method for discovering short-term affect-based response rules for spoken tutorial dialog

A good tutoring system should be able to detect and respond to subtle changes in the affective state of the learner, as a way to motivate and encourage the student, thereby improving the learning outcomes. This responsiveness should also operate at the sub-second timescale, as with some human tutors. Modeling this ability is, however, a challenge. This paper presents a combined method for the discovery of the rules governing such real-time responsiveness. This method uses both machine-learning and perceptual techniques, both with and without reference to internal states. This method is illustrated with the problem of choosing supportive acknowledgments in memory-reinforcing quiz dialogs. A wizard-of-oz experiment showed that users prefer a tutorial system based on responsive rules to one that chooses acknowledgments at random.