The ontology matching or alignment task identifies inconsistencies among concepts, relationships, and instances in two different ontologies and then resolves correspondence relationships. The authors present a survey of ontology matching techniques implemented in different tools, focusing on scalability and quality issues when matching large-scale ontologies such as the Systematized Nomenclature of Medicine--Clinical Terms (SNOMED CT).

Ontology matching involves three steps: (1) identifying the inconsistencies in different regions of both source and target ontologies, (2) mapping the correspondences (so-called matchers), and (3) resolving the mappings. To achieve scalability and ensure high- quality mappings, ontology matching tools should be efficient in terms of identifying inconsistencies and processing the mapping. The authors provide different scalability techniques for the first two steps and different quality assurance strategies for the last step; they compare many tools according to these techniques.

Partitioning, data indexing, and structural indexing methods reduce the search space of inconsistent concept or relationship alignments. Ontology partitioning approaches such as graph- or logic-based ones break down both source and target ontologies into subontologies based on some criteria (for example, topic).

To reduce processing time, subontology regions derived from ontology partitioning are processed in parallel by n multiple matching processes. Each matcher identifies the correspondences and inconsistencies based on similarity measures of multidimensional features. The sequential execution of a pairwise similarity measure may not be a scalable approach.

To enhance the quality of inconsistent matching among ontology entities or relationships, human intervention (a pay-as-you-go approach) can repair remaining inconsistencies, as opposed to fully automated repair techniques.

The paper compares many matching tools on a large biomedical source. According to the results, a handful of tools show good precision but most show low recall values. The average processing time seems to be drastically different across tools, so scalability remains unclear. It is similarly unclear whether the performance scores are fully generalizable to other domain ontologies.

Project managers and graduate students can use this helpful survey when selecting alignment tools, to review possible techniques and issues.