ogy reasoning therefore plays a central role in both the development of high-quality ontologies and the deployment of ontologies in applications.

In spite of the complexity of reasoning with OWL ontologies, highly optimized DL reasoning systems (such as FaCT++, owl.man.ac.uk/factplusplus/, Racer, www.racer-systems.com/, and Pellet, pellet.owldl.com/) have proved effective in practice; the availability of such systems was one of the key motivations for the W3C to base OWL on a DL. State-of-the-art ontology-development tools (such as SWOOP, code.google. com/p/swoop/, Protégé 4, and TopBraid Composer, www.topbraidcomposer. com) use DL reasoners to give feedback to developers about the logical implications of their designs. This feedback typically includes warnings about inconsistencies and synonyms.

An inconsistent (sometimes called “unsatisfiable”) class is one for which its description is “overconstrained,” with the result that it can never have instances. This inconsistency is typically an unintended consequence of the design (why introduce a name for a class that can never have instances?) and may be due to subtle interactions among axioms. It is therefore useful to be able to detect such classes and bring them to the attention of the ontology engineer. For example, during the recent development of an OWL ontology at NASA’s Jet Propulsion Laboratory, the class “OceanCrustLayer” was found to be inconsistent. Engineers discovered (with the help of debugging tools) that this was the result of its being defined as both a region and a layer, one (a layer) a 2D object and the other (a region) a 3D object. The inconsistency thus highlighted a fundamental error in the ontology’s design.

It is also possible that the descriptions in an ontology mean that two classes necessarily have exactly the same set of instances; that is, they are alternative names for the same class. Having multiple names for the same class may be desirable in some situations (such as to capture the fact that “myocardial infarction” and “heart attack” are the same thing). However, multiple names could also be the inadvertent result of interactions among descriptions or of basic errors by the ontology designer; it is therefore use-

Reliability and
correctness
are particularly
important when
ontology-based
systems are used
in safety-critical
applications; in
those involving
medicine, for
example, incorrect
reasoning could
adversely affect
patient care.

ful to be able to alert developers to the presence of such synonyms.

In addition to checking for inconsistencies and synonyms, ontology-development tools usually check for implicit subsumption relationships, updating the class hierarchy accordingly. This automated updating is also a useful design aid, allowing ontology developers to focus on class descriptions, leaving the computation of the class hierarchy to the reasoner; it can also be used by developers to check if the hierarchy induced by the class descriptions is consistent with their expert intuition. The two may not be consistent when, for example, errors in the ontology result in unexpected subsumption inferences or “underconstrained” class descriptions result in expected inferences not being found. Not finding expected inferences is common, as it is easy to inadvertently omit axioms that express “obvious” information. For example, an ontology engineer may expect the class of patients with a fracture of both the tibia and the fibula to be a subClassOf “ patient with multiple fractures”; however, this relationship may not hold if the ontology doesn’t include (explicitly or implicitly) the information that the tibia and fibula are different bones. Failure to find this subsumption relationship should prompt the engineer to add the missing DisjointClasses axiom.

Reasoning is also important when ontologies are deployed in applications, when it is needed to answer standard data-retrieval queries, and to answer conceptual queries about the structure of the domain. For example, biologists use ontologies (such as the Gene Ontology, or GO, and the Biological Pathways Exchange ontology, or BioPAX) to annotate (Web-accessible) data from gene-sequencing experiments, making it possible to answer complex queries (such as “What DNA-binding products interact with insulin receptors?”). Answering requires a reasoner to not only identify individuals that are (perhaps only implicitly) instances of DNA-binding products and of insulin receptors but to identify which pairs of individuals are related (perhaps only implicitly) via the interacts With property.

Finally, in order to maximize the benefit of reasoning services, tools should be able to explain inferences; without explanations, developers may