Repairing mappings across biomedical ontologies by probabilistic reasoning and belief revision

摘要

The ontology matching approaches identify correspondences among entities across ontologies, and the quality of ontology mappings is crucial for supporting knowledge sharing and reuse on the Semantic Web. In the annual Ontology Evaluation Alignment Initiative (OAEI) competitions, matching large and complex, real-world biomedical ontologies is one of the most challenging endeavors. As matching methods are basically heuristic, wrong mappings often exist in the generated alignments. The general framework of mapping validation collects candidate wrong mappings based on unsatisfiable concepts and adopts the removal strategy to gain the coherence of alignments w.r.t. source ontologies. Although it ensures logical coherence, such repairing does not necessarily guarantee the quality of mappings obtained, i.e., the disposed mappings can be positive and the retained ones can be wrong in terms of the domain knowledge intended in the ontologies per se. This can be demonstrated by the existence of incoherences when the UMLS Metathesaurus® has been used as the basis of reference alignments for OAEI biomedical ontology matching tasks. To address this problem, we propose a novel approach for repairing biomedical ontology mappings by probabilistic reasoning and belief revision techniques, featuring a combination of removal strategy and revision strategy. More concretely, mappings are transformed into probabilistic description logics (PDL) conditional constraints and their weights into probability intervals based on our designed rules. Then, the incoherence checking of mappings is reduced to solving a linear program with the constraints under the PDL semantics. For identified incoherent mappings, instead of simply discarding, we revise them by relaxing their probability intervals until the probabilistic coherence is reached. The evaluation on repairing OAEI biomedical alignments shows that our approach can be effective in retaining correct mappings and removing wrong ones. Moreover, we show that repair systems following the general framework of mapping validation have improved their performance when equipped with our revision module at the repair stage. Furthermore, feeding structural matchers with repaired alignments as seeds shows that the mappings generated by our approach lead to the best structural matching result compared with other repair systems. Being non-aggressive, our approach is suitable for applications like ontology-supported medical information retrieval, semantic annotation and indexing of medical articles, and matchmaking and ranking objects among multiple ontologies.