A Discriminative Approach to Ontology Mapping

Techniques for automatically performing ontology mapping are vital for many real-world applications. Unfortunately, the problem is difficult because many types of evidence must be integrated to make good alignment decisions, and these decisions are co-dependent. In this paper, we propose a conditional random field (CRF) for ontology mapping which combines probabilistic machine learning and dependencies among the prediction. We integrate multiple sources of evidence using clauses in first-order logic, and learn corresponding weights directly from labeled training data. Our experiments show examples of impressive gains when tested on a commonly used mapping corpus; our method achieves an average of 11% (absolute) improvement in F1 when compared to other systems. We also show that our CRF is capable of generalizing from one mapping domain to another—making our supervised approach applicable for domains that lack labeled training data.