A dependency modelling approach for the management of ontology-based integration systems

Abstract

Ontology-based approaches that formally represent the meaning of information in a system, offer the hope of dealing with semantic heterogeneity when integrating heterogeneous data sources [Halevy 2001, Noy 2004, Wache et al. 2001, Doan and Halevy 2005, Pollock 2002]. While these ontology-based approaches offer significant advantages [Cruz and Xiao 2005, Noy 2004, Wache et al. 2001, Halevy 2005] over traditional approaches (e.g. ETL1), they tend to require semantic mappings to create loose coupling of systems to enable integration. The mappings may serve to relate ontologies to other ontologies (inter-ontology mappings) or to relate ontologies to underlying information sources (e.g. a database). However, when such semantic systems are scaled up, the semantic mappings also need to grow and evolve [Bernstein and Melnik 2007, Velegrakis et al. 2003, Yu and Popa 2005, Halevy et al. 2005, and Topaloglou 2007]. Failure to provide methods to manage and evolve the semantic mappings can make the integration systems brittle. There is currently little research to help identify, manage and evolve semantic mappings when the integration system is evolving [Bernstein and Melnik 2007, Haas 2007, Doan and Halevy 2005, Kondylakis et al. 2009]. The first part of the evolution problem, identifying and managing the mappings that need to evolve, is addressed by the dependency model in this thesis. The dependency model is important in the context of ontology-based data integration because it promises to enhance the scalability of integration systems by allowing them to find which elements of the integration system are impacted when a data source or ontology changes. This thesis has developed an ontology-based domain specific dependency model, a more general dependency metamodel and a tool that can represent and analyse dependencies that occur between mappings, ontologies and databases in an ontology based integration system. The approach has been developed and evaluated using two industrial datasets.