An Entropy-Based Approach to Automatic Detection of Critical Changes in Human-Machine Interaction

This paper introduces and exemplifies an entropy-based approach to automatic detection of critical changes in flow of human-machine interaction. The underlying intuition is that a critical change is a local phenomenon associated with a dialogue fragment that can be detected by examining patterns of interactional entropy by means of segmentation and thresholding. In terms of units of analysis, the notion of (normalized) interactional entropy is introduced at the level of dialogue act types. In order to account for the sequential order of dialogue acts in a given dialogue fragment, interactional entropy is re-evaluated with each ensuing dialogue act. The proposed approach is validated for thirty-six dialogues between children and a conversational human-like robot, recorded in therapeutic settings. Finally, the generalizability of the proposed approach is discussed.