Benefits and challenges of real-time uncertainty detection and adaptation in a spoken dialogue computer tutor

摘要: We evaluate the performance of a spoken dialogue system that provides substantive dynamic responses to automatically detected user affective states. then present detailed error analysis reveals challenges for real-time affect detection and adaptation. This research is situated in tutoring domain, where student tutor. Our adaptive detects uncertainty each turn via model combines machine learning approach with hedging phrase heuristics; learned uses acoustic-prosodic lexical features extracted from speech signal, as well features. The varies its content based on automatic correctness labels turn. controlled experimental evaluation shows yields higher global than two non-adaptive control systems, but difference only significant subset students. indicates noisy labeling major bottleneck, yielding fewer expected adaptations thus lower performance. However, percentage received adaptation correlates over all Moreover, when accurately recognized adapted to, local significantly improved.