Using Classes As Property Values

Natasha Noy, Stanford University
Last revised: May 3d, 2004

[Note: This document will be a part of a larger document that will provide an introduction and overview of all ontology design patterns produced by the Semantic Web Best Practices and Deployment Working Group]

Comments, suggestions and modifications suggested by Pat Hayes shown in yellow. Comments are in monospace type, suggested modifications to text in normal font. Comments written May 14 2004

General issue

When and how to use classes as values for properties? Why only values? Why not as having properties applied to them i.e. more generally as individuals? What are advantages and drawbacks? What are different approaches and workarounds? Which solutions are in OWL DL (and hence OWL Lite) and which solutions are in OWL Full but not in OWL DL require us to move outside OWL DL?

Use case example

Suppose we have a set of books about animals and want to annotate each book with its subject, which is a particular species class? That is, does it have to be a species? of animal (or animals) that it talks about. This emphasis on books is unfortunate since it gets the issue mixed up with issues of symbol vs. symbolised, class vs. concept, etc.. Further, we want to be able to say infer? that a book about african lions is also a book about lions (For example, when retrieving all books about lions from a repository, we want books that are annotated as books about african lions to be included in the results).

More specifically, consider two book examples: (1) "Lions: Life in the Pride", which is a book that "presents an introduction to lions describing their physical characteristics, habitat, young, food, predators, and relationship to people"; and (2) "The African Lion," which "describes the physical characteristics, habitat, and behavior of the" african lions. We would like to specify that the first book described the animal species of Lions, and the second describes a species of African Lion. However, we also want to retrieve the second book when a query is about Lions, not just African Lions.

We consider animals classes of animals? to be subjects of the books and would like to use the Dublin Core property dc:subject for this annotation.

Other use case scenarios

This issue arises in general when we have a hierarchy of concepts and would like to use it as a terminology to annotate other classes or individuals. Terminology?? This reads oddly to me, and seems on a tangent. The classes as individuals issue arises in many cases which have nothing particularly to do with terminology. Consider using a hierarchy of different genre genres? to annotate music CDs, or linking classes or individuals in an ontology to the corresponding concepts in a standard reference terminology (e.g., UMLS is such standard reference terminology for many medical applications).

Notations

In all the figures below, ovals represent classes and rectangles represent individuals. The orange color signifies classes or individuals that are specific to a particular approach.

Approaches

Approach 1: using classes directly as values

In the first approach, we can simply use classes from the subject hierarchy as values for properties (in our example, as values for the dc:subject property). We can define a class BookAboutAnimals of all books about animals. Suggest omit this as it plays no role in the example. For simplicity, we omit classes for other animals, such as mammals and felines.

Here is a definition of an individual (a specific book that we are annotating) that is an instance of the BookAboutAnimals class with the corresponding subject (for simplicity, we assume that each book discusses only one class of animals):

:LionsLifeInThePrideBook
      a       :BookAboutAnimals ;
      rdfs:seeAlso <http://isbn.nu/0736809643> ;
      :bookTitle "Lions: Life in the Pride" ;
      dc:subject :Lion .

The book "The African Lion" will be represented as:

:TheAfricanLionBook
      a       :BookAboutAnimals ;
      rdfs:seeAlso <http://isbn.nu/089686328X> ;
      :bookTitle "The African Lion" ;
      dc:subject :AfricanLion  .

:AfricanLion
      a       owl:Class;
      rdfs:subClassOf :Lion .

Considerations when choosing approach 1:

• The resulting ontology is in OWL Full, but not in OWL DL.
• This approach is probably the most succinct and intuitive among all the approaches proposed here.
• Applications using this representation can directly access the information needed to infer that Lion (the subject of the LionsLifeInThePrideBook individual) is a subclass of Animal and that AfricanLion (the subject of the TheAfricanLionBook individual) is a subclass of Lion. And also, for example, that if Joe is an African Lion then Joe is one of the things that the book is about.
• If we need to restrict the range of values for the dc:subject property for the BookAboutAnimals class to the class Animal and its subclasses, we will need to create another class (a metaclass) of classes that will have the class Animal and its subclasses as instances:
  • The class AnimalClass is a class, which is a subclass of owl:Class (In OWL, any Any class that has other classes as instances must be a subclass of owl:Class) You could add that this also goes outside the DL case, and also that in RDFS it is not required to be an rdfs:subClassOf owl:Class; this is an artifact of the special OWL extensional semantic conditions. Or just add the qualification suggested. .
    

    :AnimalClass
          a       owl:Class ;
          rdfs:subClassOf owl:Class .

  • The class Lion (and other subclasses of the Animal class) have AnimalClass as its type:
    

    :Lion
          a       :AnimalClass ;
          rdfs:subClassOf :Animal .

  • The class BookAboutAnimals has AnimalClass as an allValuesFrom¹ range restriction for the dc:subject property:
    

      :BookAboutAnimals
          a       owl:Class ;
          rdfs:subClassOf
                  [ a       owl:Restriction ;
                    owl:allValuesFrom :AnimalClass ;
                    owl:onProperty dc:subject
                  ] .
    	   

OWL code for approach 1

[N3] [RDF/XML abbrev] [Abstract syntax]

Summary for approach 1

Approach 2: Creating a hierarchy of subjects and a parallel set of subject individuals

We can treat the hierarchy of animal species as a hierarchy of subjects, create individuals corresponding to all the subjects and use these individuals as values for the dc:subject property. Thus, we will have, for example, an individual LionSubject that will be an instance of the Lion class.

Why did you use rdf:type as the link? On the face of it, that seems clearly broken, and does not seem necessary. There is a perfectly rational approach which would be like this but using a different property to connect to the classes, say 'ex:referenceClass', thereby avoiding the need to include the subject in the class (and making a clear distinction between this case and your case 4, which currently there is no distinction as far as I can see.) Of course this would not be in DL, but it still makes perfect sense as an option in Full.

As this would help to make clear, many of the problems arise because in DL, the ONLY property that can have a class as a value is rdf:type. (It might be worth saying this explicitly somewhere.) If this restriction is lifted, many options are opened up.

Suggestion: make this modification to this example, and say clearly what the merits of it are but that it is not in DL. Then treat your example 4 as being what you get when you 'squeeze' this style into DL, and so force the subjects to be actual lions. That has the merit of making the disticntion clear and also showing clearly the reason for having to introduce "prototypical lions".

We can then use the LionSubject as the value of the property dc:subject for the LionsLifeInThePrideBook individual:

  :LionSubject
      a       :Lion .

In this case, the definition of the LionsLifeInThePrideBook refers to the LionSubject individual

:LionsLifeInThePrideBook
      a       :BookAboutAnimals ;
      rdfs:seeAlso <http://isbn.nu/0736809643> ;
      :bookTitle "Lions: Life in the Pride" ;
      dc:subject :LionSubject . 

Considerations when choosing approach 2:

• The resulting ontology is in OWL DL
• The class Lion has an instance that is the subject lion. Creating an instance of the Lion class to represent a specific lion at the zoo would be inconsistent with this interpretation. Therefore, we will need to have a different class to serve as type for lions at the zoo. This is wrong. It is not INCONSISTENT with this interpretation: it merely creates a somewhat unusual instance of the class. One could have a class of lions which included imaginary lions (such as the hero of a CSLewis novel) and also an 'archetypical lion', which corresponds to a common NL usage where a class is singularized by a prototype, as in "The African lion is a magnificent beast". No inconsistency here, just a need to be a little careful when selecting from the class.Still, it might be good to give a warning, eg by coninuing with the wording: would create a class containing a mixture or real lions and imaginary or virtual lions.
• In this approach, there is no explicit direct relation (i.e, a relation that a general-purpose reasoner can readily identify) between the LionSubject individual defined above and, for example, a AfricanLionSubject individual, which is an instance of AfricanLion:
  

  :AfricanLionSubject
        a       :AfricanLion .

  An application trying to utilize this relation (for example, to extract books about african lions when asked for books about lions), will need to be aware of this specific approach and know to trace back to the corresponding classes, their subclasses, and respective individuals. A general-purpose reasoner will not be able to use this information directly. Note however that the individual AfricanLionSubject is also an instance of the Lion class. Therefore, if we ask for all books where dc:subject is an instance of the Lion class we will get the books that are annotated with AfricanLionSubject.
• There is a maintenance penalty: We need to maintain a set of instances for all subjects in addition to the hierarchy of subjects and ensure that the two sets—classes representing subjects and corresponding individuals—are consistent with each other (e.g., that they have the same names, etc.)
• Some may consider this approach to be "too messy" for the simple task at hand
• This approach explicitly separates the terminology from the ontology of corresponding concepts
• Defining a range restriction for the dc:subject property for the class BookAboutAnimals is straightforward. We define an allValuesFrom¹ restriction that states that all values of the dc:subject property are instances of the class Animal:
  

      :BookAboutAnimals
        a       owl:Class ;
        rdfs:subClassOf
                [ a       owl:Restriction ;
                  owl:allValuesFrom :Animal ;
                  owl:onProperty dc:subject
                ] .

OWL code for approach 2

[N3] [RDF/XML abbrev] [Abstract syntax]

Summary of approach 2

This approach results in an OWL DL ontology and may be a good one to use if staying in OWL DL is important. The approach has a potential disadvantage of having actual subject values be unrelated to one another and hence not allowing a general-purpose reasoner to relate books with a subject "lion" to books with a subject "african lions", for example. You need to maintain consistency between the set of classes representing subjects and the set of corresponding individuals. If the subject hierarchy is not solely a terminology (e.g., you need to represent specific animals), you will need to create a separate class hierarchy for that.

Approach 3: using a property other than rdfs:subclassOf to organize the subject hierarchy

There is something weird about this, since it makes no reference to classes as values at all. So why are you even talking about it?

I'd suggest making this the last 'appraoch', almost as an afterthought, along the lines of :of course, you can just ignore the class heirarchy altogether and describe your own heirarchy. That is ALWAYS an option, after all, and it is one very reliable strategy for staying in the DL space, at an obvious cost.

We can create a single class Subject and make all the subjects to be individuals that are instances of this class Subject:

  :LionSubject
      a       :Subject ;
      :parentSubject :AnimalSubject .

We can then create explicit relations between different subjects, which will re-create the hierarchy for animals that we have in mind. While we create our own property parentSubject, we can also use the corresponding properties from the SKOS-Core 1.0 schema, which is an RDF schema for representing thesauri and similar types of knowledge organization systems. So, for example parentSubject is similar to skos:broader. The SKOS schema provides a rich vocabulary for handling subject hierarchy, with additional properties such as skos:narrower, skos:related, and so on.

 :parentSubject
      a       owl:TransitiveProperty , owl:ObjectProperty ;
      rdfs:domain :Subject ;
      rdfs:range :Subject ;
      rdfs:seeAlso <http://www.w3.org/2004/02/skos/core/broader> .

  :AfricanLionSubject
      a       :Subject ;
      :parentSubject :LionSubject .

Considerations when choosing Approach 3

• The resulting ontology is in OWL DL
• Most DL reasoners will be able to infer transitive relations between subjects Really? Is that statement generally accepted? , such as the fact that AnimalSubject is a parentSubject of LionSubject
• The resulting hierarchy of subjects is not related to or dependent on the class hierarchy representing the same topics (in this case, animals).
• Some may consider a approach having two parallel hierarchies of essentially the same data to be too complicated and difficult to maintain for the simple task at hand This comment seems to be a meta-opinion and rather beside the point. Surely it is for the reader to decide this.
• This approach explicitly separates the terminology from the ontology of corresponding concepts
• Defining a range restriction for the dc:subject property for the class BookAboutAnimals is straightforward. We restrict the values of the dc:subject property to the instances of the class Subject:

      :BookAboutAnimals
        a       owl:Class ;
        rdfs:subClassOf
                [ a       owl:Restriction ;
                  owl:allValuesFrom :Subject ;
                  owl:onProperty dc:subject
                ] .

• There is a serious maintenance penalty: We need to maintain a set of instances for all subjects in addition to the hierarchy of subjects and ensure that the two sets—classes representing subjects and corresponding individuals and values for the parentSubject property—are consistent with each other Why must they be consistent?

OWL code for approach 3

[N3] [RDF/XML abbrev][Abstract syntax]

Summary of Approach 3

This approach may be a good one to use if staying within OWL DL is important. It also allows you to use a DL reasoner to infer transitive relationships between subjects. It does carry the penalty of having two parallel "hierarchies."

Approach 4: using members of a class as values for the property

This seems to me to be identical to # 2.

We can approximate the interpretation that we used in the previous approaches by using unspecified members of a class rather than the class itself as property values. We can define the class BookAboutAnimals as a class of books where the subject is some animal. Correspondingly, a BookAboutLions class will be a class of books where a subject is some (unidentified) lion:

For example, we can define the class BookAboutLions as follows:

:BookAboutLions
      a       owl:Class ;
      owl:equivalentClass
              [ a       owl:Class ;
                owl:intersectionOf ([ a       owl:Restriction ;
                                              owl:onProperty dc:subject ;
                                              owl:someValuesFrom :Lion
                                    ] 
                                    :Book)
              ] .

A specific instance of this class LionsLifeInThePrideBook would then be defined as follows:

:LionsLifeInThePrideBook
      a       :BookAboutLions ;
      rdfs:seeAlso <http://isbn.nu/0736809643> ;
      :bookTitle "Lions: Life in the Pride" .

Alternatively, MyLionBook can be defined as

:MyLionBook
      a       :Book;
      [ a       owl:Restriction ;
                owl:onProperty dc:subject ;
                owl:someValuesFrom :Lion
      ]

Considerations when choosing Approach 3

• The resulting ontology is in OWL DL
• The interpretation of this pattern is different from interpretations of the other patterns in this note:I disagree. I think it is identical to the only coherent interpretation of #2. the subject of the book is a specific lion -- an individual in the Lion class -- rather than the Lion class itself . If you had meant the new individuals in #2 to be the classes, you should have said so. That would be coherent in RDFS: the classes could belong to themselves. We can consider this lion to be simply a prototypical representative of the class.
• A Description Logic reasoner will be able to classify automatically the individual of MyLionBook that has an instance of a Lion class as its subject as an instance of the class BookAboutLions. It will also be able to classify the class BookAboutLions defined above as a subclass of the class BookAboutAninmals (defined in a similar way).
• This approach is likely to be the easiest approach to use if you would like to employ DL reasoning to classify books based on their subjects.
• There can be a maintenance penalty if you do not need to create classes for books on various subjects anyway.You will either have to define anonymous types for the individuals or create a hierarchy of classes with corresponding subjects. In many cases however, the concepts or individuals that you need to annotate with subjects or other properties that may require classes as values, are independent of the hierarchy of the classes used for annotation. For instance, you can have a hierarchy of diseases and use one of the properties to describe drugs used to treat each disease. In this case, defining a someValuesFrom drugX restriction for each disease will be more natural and will not incur a maintenance penalty. In the example of books and subjects however, the hierarchy of books with specific subjects defined through someValuesFrom restrictions is essentially parallel to the hierarchy of subjects (animals, in this case) itself. This paragraph reads rather confusingly (to me). Why were these techniques using a restriction not available in the previous cases?

OWL code for approach 4

[N3] [RDF/XML abbrev][Abstract syntax]

Summary of Approach 4

This approach can make the most use of DL classifiers. It represents a different interpretation of the subject as being a prototypical instance of a class rather than a whole class of things or a specific subject corresponding to that class. With the understanding that you are using a different interpretation, this approach may be a good one to use if using DL reasoners to classify individuals based on their subjects (or another property used in its place) is important.

Approach 5: using classes as values for annotation properties

This is not an 'approach'; it is a refusal to solve the problem. Another way to stay in OWL DL is to use classes as values for annotation properties:

  dc:subject
      a       owl:AnnotationProperty . 

:LionsLifeInThePrideBook
      a       :BookAboutAnimals ;
      rdfs:seeAlso <http://isbn.nu/0736809643> ;
      :bookTitle "Lions: Life In The Pride" ;
      dc:subject :Lion  .

Considerations when choosing Approach 5

• In OWL DL, annotation properties cannot at the same time be defined either object properties or datatype properties. Thus, if dc:subject (or another property you want to use) is defined elsewhere as an object property or a datatype property, it cannot be used as an annotation property.
• In OWL DL, annotation properties cannot have any restrictions, such as cardinality or domain/range restrictions, or subproperties. Thus, if you use this approach, you cannot specify that each book should have at least one subject for example or that individuals that are instances of BookAboutAnimals must all be subclasses of the Animal class.
• Even if the ontology is in OWL DL (i.e., the annotation properties are used consistent with OWL DL restrictions), DL reasoners will not use the information in annotation properties for reasoning. Quite. So this seems to me to be a non-solution, and that it is misleading to offer it as an alternative. One could just as well suggest that the user uses C++ instead of OWL. Thus, in order to extract a book about lions when queried for animal books, one would need to use special-purpose reasoning that uses annotation properties and can perform reasoning with them.

OWL code for approach 5

[N3] [RDF/XML abbrev] [Abstract syntax]

Summary of Approach 5

This approach allows you to use classes directly as property values while staying in OWL DL. No, it doesn't in any real sense.  However, the properties that will have classes as values will have to be defined as annotations and therefore cannot have any additional restrictions defined on them (and should not be declared as object or datatype properties elsewhere). DL reasoners will not use values of annotation properties.

Notes

[1] Using allValuesFrom restriction for the book's subject implies that instances of the class BookAboutAnimals can have only animals as values for the dc:subject property. If we want to allow other terms to be listed as subjects, a someValuesFrom restriction will be appropriate.