Enabling Representation Learning in Ontology-Driven Conceptual Modeling Using Graph Neural Networks

Conceptual Models (CMs) are essential for information systems engineering since they provide explicit and detailed representations of the subject domains at hand. Ontology-driven conceptual modeling (ODCM) languages provide primitives for articulating these domain notions based on the ontological categories put forth by upper-level (or foundational) ontologies. Many existing CMs have been created using ontologically-neutral languages (e.g., UML, ER). Connecting these models to ontological categories would provide better support for meaning negotiation, semantic interoperability, and complexity management. However, given the sheer size of this legacy base, manual stereotyping is a prohibitive task. This paper addresses this problem by proposing an approach based on Graph Neural Networks towards automating the task of stereotyping UML class diagrams with the meta-classes offered by the ODCM language OntoUML. Since these meta-classes (stereotypes) represent ontological distinctions put forth by a foundational ontology, this task is equivalent to ontological category prediction for these classes. To enable this approach, we propose a strategy for representing CM vector embeddings that preserve the model elements’ structure and ontological categorization. Finally, we present an evaluation that shows convincing learning of OntoUML model node embeddings used for OntoUML stereotype prediction.