Alignment of Ontologies Using Directed Graphs, Neural Network and Genetic Algorithm in the Web of Data

Abstract

Semantic interoperability remains a central challenge for the Semantic Web, as the growth of structured data increases the need to reconcile heterogeneous ontologies across domains such as education, health, defence, agriculture and biomedicine. This paper proposes a hybrid ontology-alignment architecture that combines directed-graph representations to capture semantic and taxonomic structure, graph neural networks (GNNs) enriched with BioBERT embeddings to learn latent semantic correspondences, and a genetic algorithm to optimise the final selection of matches under structural-consistency constraints. The approach is illustrated on the biomedical domain, where reference ontologies such as SNOMED CT, ICD-10 and LOINC are large, complex and structurally rich, and where aligning hospital data with such reference terminologies supports interoperability and diagnostic assistance. Simulated experiments inspired by the OAEI Anatomy and LargeBio benchmarks indicate that combining the three layers improves the F-measure relative to each layer taken in isolation, primarily by reducing false positives through structural filtering. A component-wise ablation study isolates the contribution of each layer, and the simulated results are positioned, for indicative purposes only, against published OAEI systems. A complementary, non-simulated evaluation on the actual small LargeBio ontology fragments (FMA–NCI, FMA–SNOMED, SNOMED–NCI), scored locally with a MELT-style evaluator, shows the hybrid pipeline reaching the best precision of all matcher variants on every pair but not the best F-measure, since it trades recall away more than it recovers, a real-data finding that qualifies the simulated ablation and points to blocking-width and fitness-weight calibration as concrete next steps. The paper also identifies the remaining methodological steps an official OAEI submission under SEALS/MELT, an equivalent run on Anatomy, statistical validation across multiple seeds, and a dedicated scalability study required to move from illustrative simulation to a fully validated system.