In search of optimal human-expert system explanations: Empirical studies of human-human and human-expert system interactions

In this project explanations were studied along a continuum ranging from human-human interactions to human-expert system interactions with the goal of identifying features of successful expert system explanations. The project consisted of five distinct phases or steps: (a) defining what a successful explanation entails, (b) observing human-human explanation and formulating hypotheses about the features of successful explanations, (c) testing hypotheses formulated in step b, (d) extending results to an expert system domain and testing again, and (e) from this empirical data, formulating recommendations for expert system explanation designers. The progressive nature of this study allowed conclusions to be drawn about both human-human and human-expert system interactions and the role explanations play in these exchanges. The most salient conclusion drawn from these studies was that explanatory interactions are complex and explanation success is dependent on more than just features of the explanations involved. Individual differences such as an explanation recipient's initial abilities and their participation level in the interaction influence their understanding and performance as much, if not more so, than explanation features. Consistently subjects' participation level interacted with explanation content level. Individuals who are active participants in interactions with an expert perform better when given explanations with low levels of content, whereas passive participants benefit from explanations with high levels of content. Overall, an active participation level increases performance and understanding in human-human interactions, but this result does not generalize to human-expert system interactions where an active participation style is detrimental to performance. This and other inconsistencies between human-human and human-expert system interactions are discussed as well as the advantages of the research approach employed in this project. Finally, recommendations based on the results of these studies are provided for expert system explanation designers