Abstract
Life sciences produce a huge amount of data (e.g., clinical trials, scientific articles) so that integrating and analyzing all the datasets related to a given research question like the correlation between phenotypes and genotypes, is a key element for knowledge discovery. The life sciences community adopted Semantic Web technologies to achieve data integration and interoperability, especially ontologies which are the key technology to represent and share the increasing amount of data on the Web. Indeed, ontologies provide a common domain vocabulary for humans, and formal entity definitions for machines.A large number of biomedical ontologies and terminologies has been developed to represent and annotate various datasets. However, datasets represented with different overlapping ontologies are not interoperable. It is therefore crucial to establish correspondences between the ontologies used; an active area of research known as ontology matching.Original ontology matching methods usually exploit the lexical and structural content of the ontologies to align. These methods are less effective when the ontologies to align are lexically heterogeneous i.e., when equivalent concepts are described with different labels. To overcome this issue, the ontology matching community has turned to the use of external knowledge resources as a semantic bridge between the ontologies to align. This approach arises several new issues mainly: (1) the selection of these background resources, (2) the exploitation of the selected resources to enhance the matching results. Several works have dealt with these issues jointly or separately. In our thesis, we made a systematic review and historical evaluation comparison of state-of-the-art approaches.Ontologies, others than the ones to align, are the most used background knowledge resources. Related works often select a set of complete ontologies as background knowledge, even if, only fragments of the selected ontologies are actually effective for discovering new mappings. We propose a novel BK-based ontology matching approach that selects and builds a knowledge resource with just the right concepts chosen from a set of ontologies. The conducted experiments showed that our BK selection approach improves efficiency without loss of effectiveness.Exploiting background knowledge resources in ontology matching is a double-edged sword: while it may increase recall (i.e., retrieve more correct mappings), it may lower precision (i.e., produce more incorrect mappings). We propose two methods to select the most relevant mappings from the candidate ones: (1) based on a set of rules and (2) with Supervised Machine Learning. We experiment and evaluate our approach in the biomedical domain, thanks to the profusion of knowledge resources in biomedicine (ontologies, terminologies and existing alignments).We evaluated our approach with extensive experiments on two Ontology Alignment Evaluation Initiative (OAEI) benchmarks. Our results confirm the effectiveness and efficiency of our approach and overcome or compete with state-of-the-art matchers exploiting background knowledge resources.