OntoConnect

Abstract

Ontology alignment is performed to combine or integrate multiple knowledge bases at the elemental and structural levels. The current state-of-the-art systems use many different approaches to match semantics, syntax, and terminologies of different ontological entities. However, most of the ontology alignment systems depend on domain knowledge, which makes the alignment process domain-specific. To address this challenge, we aim at developing an ontology alignment approach that is independent of domain knowledge. To achieve this goal, an ontology alignment approach is proposed which exploits an unsupervised learning method using a recursive neural network to align classes between different ontologies. In particular, the proposed approach extracts structural information of the classes in ontology to train the unsupervised model. The proposed approach is tested against a reference gold copy of the Anatomy data set in the Ontology Alignment Evaluation Initiative. Our evaluation results show that the proposed unsupervised neural network approach using the meta information of ontological classes yields satisfactory results with a precision of 95.66% and F-measure of 80.26% for a similarity threshold of 0.96 with the 100-dimension input vector. Increasing the input vector dimension to 300 results in improved precision of 97.71% and F-measure of 80.38% with a 0.96 threshold. The significance of the proposed approach is that it can be used for ontology alignment independent of domain expertise and without the need for human intervention.