- From: <larry.eshelman@philips.com>
- Date: Wed, 11 Sep 2002 16:42:34 -0400
- To: www-webont-wg@w3.org
Here is my part-whole example for the how-to-do document. This is in the form of a general how-to-do guide for expressing part-whole relations in OWL, and includes a discussion of some of the issues (with some valuable input from Guus). I plan to include Guus's wine-region example as part of this section of the document. If anyone has any other part-whole examples, please send them to me, and I'll include them in the write-up, when I add Guus's. --Larry ------------------------------------------------------------------------ PART-WHOLE RELATIONS Although part-whole relations play an central role in many domains, there is no consensus on how part-whole relations should be modeled. Of course, everyone agrees that part-whole relations should not be mixed with the kind-of (subsumption) hierarchy. Although this may seem obvious to those familiar with ontologies, there are examples where this still occurs -- e.g., some controlled medical terminologies. Beyond this, there is one principle that is generally accepted: the part-whole relation is a transitive relationship. One can infer, for example, that since the finger is part of the hand and the hand is part of the human body, then the finger is part of the human body. On the surface, however, it appears that transitivity does not always hold. The classic example is the following: an arm is part of a musician, the musician is part of an orchestra, but it would be odd to conclude that the arm is part of the orchestra. However, this example is not completely convincing, since an arm is not part of a musician per se, but of a human (body). "Musician" is a role of a human. A more convincing example is provided by Rogers and Rector: a part-of an artery is part-of the thigh, and the thigh is part-of the leg, but a part-of an artery is not normally thought of as part-of the leg. Many authors (including Rogers and Rector) have argued that the solution to such examples is to distinguish different kinds of part-whole relations, and then to not mix them. Winston, Chaffin and Herrmann, for example, proposed six different types of part-whole relations (specializations): (1) component-integral_object (2) member-collection (3) portion-mass (4) stuff-object (5) feature-activity (6) place-area. Rogers and Rector block the a-part-of-an-artery-is-part-of-the-leg inference by distinguishing two specializations: segment-of and component-of. (Rogers and Rector treat "segment-of" as a subspecialization of the portion-mass specialization.) Thus, a segment-of an artery is a component-of the thigh, and the thigh is a segment-of the leg, but a segment-of an artery is not a segment-of the leg. The inference is blocked by not allowing transitivity across two different specializations of part-whole. Although most authors agree that something like this notion of part-whole specialization is needed, there is no agreement as to what the specialization categories should be. It should be pointed out that sometimes transitivity can lead to examples that seem linguisticly odd, even though they are confined to a part-whole specialization and are conceptually correct. For example, a fingernail is part of the finger, which is part of a hand, which is part of an arm, although it does seem somewhat odd to say that a fingernail is part of an arm. Part of the problem here is that what we mean by "arm" is somewhat ambiguous. We may mean the extremity attached to the shoulder, or we may mean the extremity between the shoulder and the hand. In light of this overview, one can conclude that part-whole relationships can be safely modeled if the following three principles are followed: (1) Keep kind-of and part-of hierarchies strictly separate. (2) Treat part-of as a transitive relation (i.e., a property that is transitive). (3) Create specialized part-of relations to reflect different uses of part-of, allowing transitivity only within a specialization. Finally, it should be noted that although following these principles should lead to a model that doesn't make any false inferences, it will not necessarily allow all the inferences that one may ideally want to model. In particular, fully modeled part-of relations may need to allow limited inheritance of certain properties -- whereas a model based on the above principles will not. For example, in some domains such as human anatomy, functions of some substructures may be inheritable via the component-of relation, but only up to a certain level: The Islets of Langerhans are components of the pancreas and they are what secrete insulin; yet, we attribute this function to the whole pancreas, although we don't carry this attribution to the gastrointestinal system (Rogers and Rector). Inheritance (or propagation) in part-whole relationships is very complex. There can be standard inheritance (like with subclass) -- e.g., the owner of the whole is also owner of the part. There can be upward propagation -- e.g., if the part if faulty, the whole is faulty. And there can be systematic relationships -- e.g., the weight of the whole is the sum of the weight of its parts. Modeling such sophisticated part-whole relationships is very difficult, and is beyond the scope of this document. EXAMPLES As an illustration of how to model part-of relations in OWL, let us start with the component-of specialization of the part-whole relationship. We will specify that all the classes of things that are to be related in this way are subclasses of the Component class. (This is a very flat class. They may also be subclasses of other classes that have more structure.) The purpose of making them subclasses of Component is to enable us to specify what kinds of things can be related via the component-of relation. For example, in a human anatomy ontology, the class Heart would be declared to be a subclass of both the class Organ and the class Component. The Component class would have the properties hasComponent and componentOf, which are then inherited by Heart. The AorticValve would also be declared to be a subclass of the class Component as well as the HeartValve class, which in turn is a subclass of Organ. Then AorticValve would be linked via the property componentOf to Heart. The class AorticValveCusp would be declared to be a subclass of Component and Organ and linked via the property componentOf to AorticValve. Because the property componentOf is transitive, one would be able to infer that the AorticValveCusp is a component-of the Heart. Since there may be other specializations of part-whole in which we are interested, we make the component-of specialization a subproperty of a general part-of property. However, since we don't want transitivity to carry over to other specializations, we specify that the subproperty is transitive, not its parent property. (This is different from the wine-region example where we want transitivity across different types of subareas, each of which is a subproperty.) Below is a general scheme for modeling the component-of relationship for any kind of component, be they parts of the body or mechanical widgets. The only aspect not discussed so far is the class IntegralWhole. In the human anatomy ontology, the HumanBody class would be a subclass of the IntegralWhole class since it would not have any componentOf property, but only the hasComponent property. Classes: widget:Component rdf:type rdfs:Class. widget:IntegralWhole rdf:type rdfs:Class. Properties: widget:hasPart rdf:type rdfs:Property. widget:PartOf owl:inverseOf widget:hasPart widget:hasComponent rdf:SubPropertyOf widget:hasPart widget:hasComponent rdf:type rdfs:TransitiveProperty. widget:hasComponent owl:allValuesFrom widget:Component. widget:componentOf owl:inverseOf widget:hasComponent. widget:componentOf owl:allValuesFrom widget:Component widget:IntegralWhole. widget:componentOf owl:cardinality 1. References Winston, Chaffin and Herrmann. A Taxonomy of Part-Whole Relations. Cognitive Science, 11: 417-444, 1987. [Seminal article on part-whole relations] Artale, A, Franconi, E., Guarino, N. and Luca, P. Part-Whole Relations in Object-Centered Systems: An Overview. Data and KnowledgeEngineering 20(3): 347-383, 1966. <http://citeseer.nj.nec.com/artale96partwhole.html> [Good overview of various approaches to part-whole relations] Rogers, Jeremy and Rector, Alan. GALEN's Model of Parts and Wholes: Experience and Comparisions. <http://www.amia.org/pubs/symposia/D200050.PDF> [Description logic based approach to part-whole relations.] Schulz, S., Romacker, M. and Hahn, U.. Part-Whole Reasoning in Medical Ontologies Revisited. <http://www.amia.org/pubs/symposia/D005124.PDF> [Another description logic based approach to part-whole relations] Philips Research USA **My opinions do not reflect those of my employer**
Received on Wednesday, 11 September 2002 16:40:23 UTC