PUB - Publikationen an der Universität Bielefeld

In the quest to successfully convey internal decisions and actions to interacting humans, it is imperative for Explainable Artificial Intelligence (XAI) systems to integrate explanation models, which are informed by the mechanisms of human-human interactions. Explanation, being a pervasive element in human communication, emerges from an interactive process involving an explainee, who lacks understanding of some matter, and an explainer, who has knowledge to convey to address the explainee's knowledge gap. So far, there is a lack of methodological frameworks to provide deeper insights into this interactive process in which the explainer and explainee adapt to each other to jointly shape the explanation. We propose a novel analytical framework that assumes that explainer and explainee take turns in a dialogue to establish an explanation by performing speech acts, which we refer to as explanatory moves in this work. We present a comprehensive inventory of explanatory moves that is informed by categories discussed in the existing literature, and annotate an existing corpus of explanatory dialogues derived from the online magazine WIRED based on this inventory. As a key element of our approach, we represent explainers and explainees by vectors that represent their behavior in terms of the inventory of explanatory moves as dimensions. We discuss how this representation can be used to predict the proficiency level of explainees as well as how relationships between explainees and explainers can be described in terms of similarity in this vector space. Moreover, we demonstrate how the framework can be used to reveal structural patterns in the way explainees and explainers interact to establish an explanation.