Advancing XAI: new properties to broaden semantic-based explanations of black-box learning models

Pavel Kosov, Nahla El Kadhi, Cecilia Zanni-Merk, Latafat A. Gardashova

Abstract: For a long time, experts in different areas where Artificial Intelligence (AI) is widely applied have been requesting more clarity for the decisions made by AI. DARPA came up with a new framework for eXplainable AI (XAI) where the system exploits an explainable model to provide explanations through the explanation interface to users based on the level of their expertise. Later, ontologies were integrated in various ways, which paved the way for clearer explanations. Provided enough Expert Knowledge, ontologies can be a potent tool in XAI. Based on the ideas of Bellucci et al., their explainable system provides comprehensive explanations based on ”visible” properties found in images by Machine Learning (ML) models and described via ontologies. However, we believe that any property, not only visible ones, can be used to explore the data. New ”explanatory” properties are proposed to be used for explanations. Our system exploits ML models and more user-oriented Expert Knowledge using a wider range of properties for objects to build a more profound XAI.