Ontology Matching with Heterogeneous Graph Neural Network

Ontology matching supports data integration and interoperability by aligning semantically corresponding concepts, properties, and relations across ontologies. Many existing approaches rely primarily on node-level semantic information and underutilize contextual structural cues in the ontology graph, which can limit matching accuracy. To address this issue, we propose a relation-aware heterogeneous graph neural network framework. Our method first fine-tune BERT to obtain node representations, then extract relation representations, and construct a heterogeneous graph. Next,iteratively updating and fusing node and relation embeddings to strengthen their mutual interactions. Finally, computing a cross-ontology similarity matrix using cosine similarity between node embeddings and apply the Hungarian algorithm to derive an optimal one-to-one alignment. Evaluating the proposed method on three tasks from the OAEI Bio-ML dataset, and the results show that it achieves substantially higher F1 scores than existing methods on two of the tasks.