Semantic Matching of Big Ontologies

Abstract

Ontology matching is an efficient method to establish interoperability among heterogeneous ontologies. Large-scale ontology matching still remains a big challenge for its long time and large memory space consumption. The actual solution to this problem is ontology partitioning which is also challenging. Artificial neural networks are powerful computational models biologically inspired from the human brain, and the way how they learn and process information. Deep learning is a promising avenue of research and an important step toward artificial intelligence, emulating the human brain’s mechanisms especially for extremely complex problems. Deep learning techniques have been particularly successful when dealing with high-dimensional and massive amounts of data. However, they have limited use in ontology matching, particularly in large-scale ontology matching. In this research, we propose three different semantic solutions to deal with the largescale ontology matching challenges without partitioning. (1) The first solution is NeuralOM, a supervised reuse approach based on artificial neural networks. It consists of combining the mappings of the top ranked matching systems by means of a single layer perceptron, to define a matching function that leads to generate a better alignment between ontologies. (2) The second solution is DeepOM, an ontology matching system that we propose to deal with the large-scale heterogeneity problem using deep learning techniques. It consists on creating semantic embeddings for concepts of input ontologies using a reference ontology, and use them to train an auto-encoder in order to learn more accurate and less dimensional representations for concepts on which similarities are computed. (3) The third solution is SemBigOM, the global methodology of this research that combines NeuralOM and DeepOM in order to perfectly and independently achieve the large-scale ontology matching process. It consists on exploiting DeepOM to generate initial mappings that NeuralOM requires as input to be reused so as to output the final matching results. The experimental results of evaluating the proposed solutions on different test cases from the Ontology Alignment Evaluation Initiative, and comparing them with all participant systems of these tracks are very encouraging. They demonstrate the high efficiency of the proposed work to increase the performance of the ontology matching task, and to tackle the large-scale ontology matching issue.