Learning to Mediate Perceptual Differences in Situated Human-Robot Dialogue