This paper proposes a multimodal interactive approach to improving recognition performance of objects a person indicates to a robot. We considered two phenomena in human-human and human-robot interaction to design the approach: alignment and alignment inhibition. Alignment is a phenomenon that people tend to use the same words or gestures as their interlocutor uses; alignment inhibition is an opposite phenomenon, which people tend to decrease the amount of information in their words and gestures when their interlocutor uses excess information. Based on the phenomena, we designed robotic behavior policies that a robot should use enough information without being excessive to identify objects so that people would use similar information with the robot to refer to those objects, which would contribute to improve recognition performance. To verify our design, we developed a robotic system to recognize the objects to which people referred and conducted an experiment in which we manipulated the redundancy of information used in the confirmation behavior. The results showed that proposed approach improved recognition performance of objects to which referred by people.
