Forschungs- & Entwicklungsinformationsdienst der Gemeinschaft - CORDIS

Modularization of ontologies

The increasing awareness of the benefits of ontologies for information processing in open and weakly structured environments has lead to the creation of a number of such ontologies for real-world domains. In complex domains such as medicine these ontologies can contain thousands of concepts. Examples of such large ontologies are the NCI cancer ontology with about 27.500 and the Gene ontology with about 22.000 concepts. Other examples can be found in the area of e-commerce where product classification such as the UNSPSC or the NAICS contain thousands of product categories.

While being useful for many applications, the size and the monolithic nature of these ontologies cause new problems that affect different steps of the ontology life cycle.
Maintenance: Ontologies that contain thousands of concepts cannot be created and maintained by a single person. The broad coverage of such large ontologies normally requires a team of experts. In many cases these experts will be located in different organizations and will work on the same ontology in parallel. An example for such a situation is the gene ontology that is maintained by a consortium of experts.

Publication: Large ontologies are mostly created to provide a standard model of a domain to be used by developers of individual solutions within that domain. While existing large ontologies often try cover a complete domain, the providers of individual solutions are often only interested in a specific part of the overall domain. The UNSPSC classification, for example, contains categories for all kinds of products and services while the developers of an online computer shop will only be interested in those categories related to computer hardware and software.

Validation: The nature of ontologies as reference models for a domain requires a high degree of quality of the respective model. Representing a consensus model, it is also important to have proposed models validated by different experts. In the case of large ontologies it is often difficult if not impossible to understand the model as a whole due to cognitive limits. What is missing is an abstracted view on the overall model and its structure as well as the possibility to focus the inspection on a specific aspect.

Processing: On a technical level, very large ontologies cause serious scalability problems. The complexity of reasoning about ontologies is well known to be critical even for smaller ontologies. In the presence of ontologies like the NCI cancer ontology, not only reasoning engines but also modelling and visualization tools reach their limits. Currently, there is no OWL-based modelling tool that can provide convenient modelling support for ontologies of the size of the NCI ontology.

All these problems are a result of the fact that the ontology as a whole is too large to handle. Most problems would disappear if the overall model consists of a set of coherent modules about a certain subtopic that can be used independently of the other modules, while still containing information about its relation to these other modules.

- In distributed development, experts could be responsible for an single module and maintain it independently of other modules thus reducing revision problems.

- Users of an ontology could use a subset of the overall ontology by selecting a set of relevant modules. While only having to deal with this relevant part, the relations to other parts of the model are still available through the global structure.

- Validation of a large ontologies could be done based on single modules that are easier to understand. Being related to a certain subtopic, it will be easier to judge the completeness and consistency of the model. Validated modules could be published early while other parts of the ontology are still under development.

- The existence of modules will enable the use of software tools that are not able to handle the complete ontology. In the case of modelling and visualization tools, the different modules could be loaded one by one and processed individually. For reasoning tasks we could make use of parallel architectures where reasoners work on single modules and exchange partial results.

In other areas, e.g. software engineering, these problems are tackled by partitioning monolithic entities into sets of meaningful and mostly self-contained modules. We have developed a similar approach for ontologies. We have designed a method for automatically partitioning large ontologies into smaller modules based on the structure of the class hierarchy. We have shown that the structure-based method performs surprisingly well on real-world ontologies. This claim is supported by experiments carried out on realworld ontologies including SUMO and the NCI cancer ontology. The results of these experiments are available online at

Verwandte Informationen

Reported by

Vrije Universiteit Amsterdam
De Boelelaan 1081a
1081HV Amsterdam
See on map
Folgen Sie uns auf: RSS Facebook Twitter YouTube Verwaltet vom Amt für Veröffentlichungen der EU Nach oben