Comments on RDF Schema specification

Hello all,

In the past couple of weeks, we have performed a case study on extending
RDF Schema according to the intended extensibility mechanism as outlined
in the specification. A first draft of the results can be found at 
http://www.ontoknowledge.org/oil/oil-rdfs.pdf. The following is a quick
summary of our findings and suggested changes to the RDFS specification.

We tried to define the ontology language OIL (www.ontoknowledge.org/oil)
as a extension to RDF Schema. OIL combines the most common representational
primitives from frame-based languages with a formal semantics from description logics. 

We used existing primitives as much as possible while retaining an
unambiguous mapping between the original OIL specification and its RDFS
serialization. The resulting extension of RDFS allows the specification
of domain ontologies that are already partially understandable by 
non-OIL-aware RDFS applications, while OIL-aware applications can fully
benefit of the added features, such as formal semantics and reasoning
support. Such extensions of RDF are a likely future use of RDF (RDF as 
the common core between many different ontology languages). Therefore, we
think this case-study is significant.

We encountered a couple of problems while defining the extension.

First, there was no way to overcome the restriction on the rdfs:subClassOf
statement, i.e. the restriction that no cycles are allowed in the 
subsumption hierarchy. We think that this restriction should be dropped:
without cycles one cannot even represent equivalence between two classes 
- in our view this is an essential modelling primitive for any KR/ontology
language. Moreover, these kinds of constraint significantly add to the
complexity of parsing/validating RDF documents in a way which we think 
would be highly undesirable. This is because they are really semantic 
constraints rather than syntactic ones (they limit the kinds of models that
can be represented), even if the reasoning required in order to detect 
constraint violation is of a very basic kind.

Second, in contrast with RDFS, OIL allows more than one range restriction 
on a property. Although this can be circumvented by defining a dummy 
superclass of all classes in the range restriction, we see no reason for 
this restriction in RDFS. From a modelling point of view, allowing more 
than one range restriction is a much cleaner solution. A set of such range
restrictions should be interpreted as the union of these restrictions, 
analogous to the interpretation of multiple domain restrictions.

We also encountered a couple of peculiarities in the RDFS definition 
itself. The most striking of these is the non-standard object-meta model.
The main problem with this non-standard model is that some properties have
a dual role in the RDFS specification, both at the schema level and 
instance level (cf. Nejdl in http://www.kbs.uni-hannover.de/Arbeiten/
Publikationen/2000/modeling2000/wolpers.pdf). This makes it quite a 
challenge for modelers to understand the RDFS specification. We tried 
to make this distinction clearer in our extensions by using the rdf:type relationship consistently as an object-meta relationship.

Furthermore, the semantics of several relationships are unclear. It is
not obvious that the meaning of a list of domain (or range) restrictions
is the union of the elements. Also, the meaning of the subPropertyOf 
relation with respect to the inheritance of the domain and range 
restrictions is unclear.

To summarize, we suggest the following changes to the RDFS specification
to improve extensibility and ease of use:
 * allow more than one range restriction on Properties;
 * allow cyclic inheritance hierarchy for both Classes and Properties;
 * make the meaning of subProperty inheritance clear;
 * make the meaning of a list of domain or range restrictions clear;
 * make clear whether rdfs:subClassOf can be applied to properties and if 
   so, what the meaning of this is.

We welcome any comments on those subjects and our case study.

Regards,

Jeen Broekstra,
Michel Klein.

-- 
Department of Mathematics and Computer Science
Vrije Universiteit Amsterdam, the Netherlands

Received on Wednesday, 14 June 2000 11:07:15 UTC