The overuse of is-a relation
The is-a relation is often overused by ontology developers. Also, the is-a relation sometimes is confused with other relations. This article explains why it may happen and how to avoid that.
Ontology developers tend to overuse the is-a (or subsumption) relation, and ontology development environments support this tendency. It is so easy in protégé to place a class which is related to another one as the subclass of that one. If the relation is different from the is-a, then it takes considerably more efforts to define how actually those classes are related.
2. The is-a relation
The importance of the is-a relation lies in the fact that is-a links between classes form a hierarchy (or in some cases, a lattice) of classes, a backbone of an ontology. The hierarchical organization of entities, virtually always by is-a relations, enables inheritance of properties [Brachman, 1983]. The is-a relation can be easily confused with a number of other relations, i.e. instance-of, member-of, part-of [Brachman, 1983]. Taht is why it is important to strictly follow the agreed in the research community explicit definition of the semantic meaning of the is-a relation. A suit of of biomedical ontologies OBO uses the RO (Relations Ontology) defines this relation as follows:
C is-a C’ if and only if: given any c that instantiates C at a time t, c instantiates C’ at t.
P is-a P’ if and only if: that given any p that instantiates P, then p instantiates P’.
The relation is transitive, reflexive, and anti-symmetric. This means that if A is-a B, and B is-a C holds; then A is-a C holds too. The relation A is-a A holds and if A is-a B holds then B is-a A does not hold.
If this definition does not hold for the classes C and C’, than an ontology developer needs to identify by what other relation the classes C and C’ should be linked.
let us consider an example from MGED ontology (MO) v1.1.9 (note that MO has been significantly updated since v1.1.9):
‘growth condition’ (“a description of the conditions used to grow organisms or parts of the organism. This includes isolated environments such as cultures and open environments such as field studies.”) has subclass ‘water’ (“Water consumed by or enveloping the organism that the biosource is derived from”).
In this example, the is-a relation between the classes ‘growth condition’ and ‘water’ is used incorrectly. The class ‘water’ is not a sub-class of the class ‘growth condition’, because there are many instances of portions of water which are not used for growing. In fact, there is no any instance of a portion of water which is a growth condition. However the is-a relation will be a valid one between the classes ‘growth condition’ and ‘water growth condition’ (“a description of the conditions used to grow organisms or parts of the organism where water is consumed by or enveloping the organism that the biosource is derived from”).
3. The part-of relation
The next after the is-a relation most used relation is the part-of relation. RO ( the relations ontology) defines this relation as follows:
C part-of C’ if and only if: given any c that instantiates C at a time t, there is some c’ such that c’ instantiates C’ at time t, and c *part-of* c’ at t.
P part-of P’ if and only if: given any p that instantiates P at a time t, there is some p’ such that p’ instantiates P’ at time t, and p *part_of* p’ at t. (Here *part-of* is the instance-level part-relation.)
As with the is-a relation, the part-of relation is considered to be transitive, reflexive, and anti-symmetric.
For a number of applications such a definition of the part-of relation may be over-simplistic. GALEN defines has-grain relation [Rector et al., 2002], DOLCE (a Descriptive Ontology for Linguistic and Cognitive Engineering) distinguish five different parthood relations [Guarino et al., 1996], the Foundational Model of Anatomy (FMA) defines containment relations [Mejino et al., 2003]. Also, a parthood relation between i.e. physical objects has different semantic meaning compare with a parthood relation between information content entities. Currently this distinction is not well defined, and the IAO (an ontology of information artefacts) project aims to provide a solution.
The relation part-of should not be confused with the relation member-of between an element and a set to which it belongs to (defined in IAO), and with the relation member-of-organization (defined in OBI, the ontology of biomedical investigations).
The is-a and part-of relations can be confused.
Novice ontology developers are often puzzled by how to correctly assert a part-of relation between classes, i.e. an arm is a part of a body. They may be puzzled by this, because firstly one needs to define both of the classes. It is not obvious a subclass of what class the class ‘arm’ is. A standard solution would be to define the classes ‘body’ and ‘body part’, where a body part can not exist by itself but needs a body to be a part of it. After the definition of the classes, it is strait-forward to assert an appropriate relation between them.
4. Other relations
An ontology can and should include a number of different relations, i.e. has-quality, has-participant, is-about. However, care should be taken that relations are logically defined and the assertion of the relations follows those definitions.
Sometimes it is hard to accurately define relations between classes. The basic rule is that if one can not clearly define the distinction between some entities, it maybe better not to distinguish them at all. For example, the class ‘material entity’ subsumes BFO (basic foundational ontology) derived ‘object’, ‘fiat object part’, and ‘object aggregate’, because the three level theory of granularity in BFO is inadequate for complex biological use cases.
Ontology developers should resist the attraction and ease of overusing is-a relations. RO and other ontologies provide a rich and powerful set of relations. The more relations are used in an ontology, the more types of queries are supported by such an ontology.
About the author
Larisa N. Soldatova is RC UK Fellow at the Computer Science Department, Aberystwyth University
Ron Brachman. What IS-A is and IS-A isn’t, Computer, 0018-9162: 30-36, 1983
N. Guarino, S. Pribbenow, andL. Vieu. Model-ingpartsandwholes. Data&KnowledgeEngineering,20(3):257–258,1996.
A. Rector, J. Rogers, A. Roberts, and C. Wroe. Scale and context: Issues in ontologies to link health and bio-informatics. In Proceedings of the AMIA 2002 Anual Symposium, pages 642–646, 2002.
J. L. V. Mejino, A. V. Agoncillo, K. L. Rickard, and C. Rosse. Representing complexity in part-whole relationships within the foundational model of anatomy. In Proceedings of the American Medical Informatics Association Fall Symposium, 2003.
Tags: Larisa Soldatova