Generating Commonsense Explanations with Answer Set Programming

Abstract

In this thesis, we explore the notion of commonsense explanation in the context of Artifcial Intelligence by extending the formalism of Answer Set Programming (ASP) with formal annotations. To this aim, we defne the concept of support graphs to account for the multiple explanations for each model of a logic program, and we provide diferent operations to flter irrelevant information from the graphs. These defnitions are implemented in a tool called xclingo that additionally allows the specifcation of natural language, commonsense explanations. xclingo obtains the support graphs via an ASP meta-encoding that is proved to be correct. We study diferent examples in the context of ASP such as planning, problem-solving, or diagnosis, among others, and we analyze the efect of alternative annotations for the same scenario, illustrating the need for explanation design. Additionally, we address the generation of non-technical explanations of Machine Learning models for real users in a pair of problems from other disciplines (Medicine and Pharmacy), covering both symbolic and sub-symbolic learning algorithms.