Compositional adaptation of explanations in textual case-based reasoning

Abstract

When problem solving systems are deployed in real life, it is usually not enough to provide only a solution without any explanation. Users need an explanation in order to trust the system’s decisions. At the same time, explanations may also function internally in the system’s own reasoning process. One way to come up with an explanation for a new problem is to adapt an explanation from a similar problem encountered earlier, which is the idea behind the case-based explanation approach introduced by [29]. The original approach relies on manual construction of cases with explanations, which is difficult to scale up. In earlier work, therefore, we developed a system for automatic acquisition of cases with explanations from textual reports, including retrieval and adaptation of such cases [32,33]. In this paper, we improve the adaptation method by combining explanations from more than one case, which we call compositional adaptation. The method is evaluated on an incident analysis task where the goal is to identify the root causes of a transportation incident, explaining it in terms of the information contained in the incident description. The evaluation results show that the proposed approach increases both the recall and the precision of the system.