Semantic-Based XAI: Leveraging Ontology Properties to Enhance Explainability

The lack of explainability of machine learning models has been an issue for a long time. Experts in various fields of AI applications have sought explainable and trustworthy systems. DARPA has developed a modern approach to eXplainable AI to overcome this issue. Later, Bellucci et al. developed an Ontology-based Image Classifier by enriching the XAI of DARPA with a new method based on semantic web technologies. Using expert knowledge in the form of ontologies, explanations of their system are based on “visible” properties discovered in images. Based on their ideas, Kosov et al. proposed new “explanatory” properties, which showed the possibility of using other data types than images. However, we believe that the architecture of the previous system has to undergo several changes to be fully adapted to using various forms of data. The new architecture is proposed to improve the explanations of multiple data types, bringing us closer to the initial ideas of DARPA. Our approach provides a new flexible system where any machine learning model can use “explanatory” properties in a new way by giving promising results and adapting the system to process tabular data by providing thorough explanations and laying down a strong foundation for improved explainability of other forms of data. The flexibility of the newly designed system allows it to be adapted to any machine learning model. It can work with any user interface tailored to fit each user's specific needs, mental model, and way of thinking.