Ontological Mediation in Explainable Artificial Intelligence: Concept-Based Models and Ontology-Grounded Explanatory Justification

Resumen

Current approaches to Explainable Artificial Intelligence (XAI) have made significant progress in producing explanations for complex machine learning models. However, many remain epistemologically fragile because their explanations lack explicit semantic and ontological grounding, limiting their adequacy in critical domains such as healthcare. We argue that this limitation is fundamentally ontological rather than algorithmic.

Meaningful explainability requires the coordination of two complementary explanatory regimes governed by distinct ontological commitments and validation criteria: a fast sub-symbolic regime that produces concept-bearing representations through neural models, and a slower symbolic regime grounded in ontologies, rules, and formally specified knowledge structures. Learned concepts from sub-symbolic learning architectures capable of generating conceptual representations are therefore not self-sufficient explanatory units; they require ontological mediation and validation.

We present a methodology and illustrate it by a proof of concept integrating Concept Embedding Models (CEMs) with a Semantic Data Dictionary (SDD) enriched by ontologies. Within this framework, ontological artifacts enable semantic consolidation, justification, and ontological unpacking of sub-symbolic explanations. A case study on Alzheimer’s disease diagnosis illustrates how this framework improves the coherence, communicability, and contestability of explanations in high-stakes decision contexts, and we argue that semantic and ontological grounding are necessary conditions for explanation, justification, and governance in intelligent systems.