Interactive Acquisition of Justifications: Learning “Why” by Being Told “What”

In this paper I describe an approach to automated knowledge acquisition in which users specify desired system behavior by constructing justifications of examples. Justifications are explanations of why example behaviors are appropriate in given situations. I analyze the problem of acquiring justifications, showing how current knowledge acquisition techniques are best suited for asking what-questions while justifications are naturally viewed as answers to why-questions. I sketch a new approach for acquiring justifications that transforms why-questions into what-questions, borrowing the sources of power of existing techniques. In this approach, users construct justifications by selecting facts that specify what is relevant in a situation from a space of facts provided by the elicitation tool. Justifications are then used to create operational mappings from situations to intended outcomes. I show how the approach is applied to two different knowledge acquisition problems: the acquisition of diagnostic strategy and the acquisition of design rationale. I conclude by identifying common characteristics of the two applications and discuss how their design distributes the cognitive load between human and machine.