A New Approach for Ontology Generation from Relational Data Design Patterns

Abstract

Knowledge graphs can effectively manage domain knowledge such as entities, properties, relations and events. Recent years have witnessed quite a few successful AI applications with the help of knowledge graphs. High-quality knowledge graphs are often built with a domain ontology, which is known to be a tedious and manual intensive task. Well-designed Relational Databases (RDBs) is an important source for obtaining ontologies and knowledge graphs. For this reason, designing automatic algorithms to generate ontologies from data has been a hot research topic. However, existing methods only consider basic ontology elements directly encoded in the RDB metadata, such as table names and foreign keys. This not only leads to erroneous or meaningless ontology classes, but also overlooks a large portion of hidden relations. As we know, database design patterns (DDP) often contain hidden semantics, which should be reconstructed into the ontology. This paper studies how typical DDPs could be recognized and how to extract ontologies from such DDPs. We focus on the JC3IEDM database, a rigorously-designed RDB in the field of joint operation, which is a widely-accepted industrial standard. Through analysis, we summarize 5 representative relational DDPs and identify their metadata characteristics. Based on these characteristics, we propose a new approach to automatically recognize and classify these patterns. Finally, we derive a rule-based method to generate OWL-format ontologies according to the identified DDPs. Empirically, we verify our approach using the whole JC3IEDM database, which contains more than 200 tables and thousands of data columns. The experimental results show that our method can automatically generate the ontology for this large-scale database, and the resultant ontology has improved quality compared to existing methods.

X. Li and R. Luo—Contributed equally to this work.