Brian P. Byrne

A clustering-based approach to ontology alignment

Ontology alignment is an important problem for the linked data web, as more and more ontologies and ontology instances get published for specific domains such as government and healthcare. A number of (semi-)automated alignment systems have been proposed in recent years. Most combine a set of similarity functions on lexical, semantic and structural features to align ontologies. Although these functions work well in many cases of ontology alignments, they fail to capture alignments when terms or structure varies vastly across ontologies. In this case, one is forced to rely on manual alignment. In this paper, we study whether it is feasible to re-use such expert provided ontology alignments for new alignment tasks. We focus in particular on many-to-one alignments, where the opportunity for re-use is feasible if alignments are stable. Specifically, we define the notion of a cluster as being made of multiple entities in the source ontology S that are mapped to the same entity in the target ontology τ. We test the stability hypothesis that the formed clusters of source ontology are stable across alignments to different target ontologies. If this hypothesis is valid, the clusters of an ontology S, built from an existing alignment with an ontology τ, can be effectively exploited to align S with a new ontology τ′. Evaluation on both manual and automated high-quality alignments show remarkable stability of clusters across ontology alignments in the financial domain and the healthcare and life sciences domain. Experimental evaluation also demonstrates the effectiveness of utilizing the stability of clusters in improving the alignment process in terms of precision and recall.