Intrinsic and Extrinsic Evaluation of an Automatic User Disengagement Detector for an Uncertainty-Adaptive Spoken Dialogue System

We present a model for detecting user disengagement during spoken dialogue interactions. Intrinsic evaluation of our model (i.e., with respect to a gold standard) yields results on par with prior work. However, since our goal is immediate implementation in a system that already detects and adapts to user uncertainty, we go further than prior work and present an extrinsic evaluation of our model (i.e., with respect to the real-world task). Correlation analyses show crucially that our automatic disengagement labels correlate with system performance in the same way as the gold standard (manual) labels, while regression analyses show that detecting user disengagement adds value over and above detecting only user uncertainty when modeling performance. Our results suggest that automatically detecting and adapting to user disengagement has the potential to significantly improve performance even in the presence of noise, when compared with only adapting to one affective state or ignoring affect entirely.