This paper deals with RDF (Resource Description Framework). The main point is to present a general model describing when and how to exploit RDF technology. It is suggested that RDF(S) 1 functions best as a means to provide mechanisms for expressing contextual and case-specific information. In other words, RDF(S) is suitable for providing different views into a single extensive ontology, rather than specifying the actual ontology. The ontology "behind" the case-specific RDF(S) is likely to be expressed using some other mechanism than RDF(S).
This paper introduces a simple model on how to exploit RDF(S) in large and heterogeneous environments that include several different applications. The opinion is that RDF(S) has a lot of useful features in describing resources, but also some drawbacks. After the model is presented, the possibilities as well as limitations of RDF(S) are discussed.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission by the author.
Semantic Web Workshop 2001 Hongkong, China
Copyright by the author.
1 RDF(S) refers to combined technologies of RDF and
RDFSchema. Cf. [
XML is one proposed representation format for RDF statements.
Of the large amount of technologies in the XML family at least
XML Namespaces [
This chapter presents the general structure of the proposed model. The motivation is to familiarize the reader with different parts of the model.
Figure 1 presents the overview of the model and illustrates the
role of RDF(S). Unlike in [5, 7, and 19], RDF(S) is not intended
to cover the whole semantic categorization in the environment2. It
is rather intended as a mechanism to provide domain-specific
data related to small-scale tasks. An individual RDF document as
2 Note that in [5, 7, and 19] standard RDF(S) is extended with a
language called OIL (Ontology and Inference Language). Also
DAML (DARPA Agent Markup Language) [
Figure 1 is now examined from right to left. The rightmost section of the picture denotes ontology, the most general description of the environment in question. The next two sections are in the core focus of this paper. RDF schemas are seen as domain-specific validating filters. RDF documents are relatively small pieces of information that are validated against RDF schemas. And finally, users are the ones that utilize RDF(S) as their source of knowledge when working in the environment. Users can be software agents as well as human beings.
Details of the ontology are outside the scope of this particular paper; ontology is treated here as a "black box". It could be implemented for example as a semantic network or a tree structure.
The approach in this paper favors the adoption of one big shared
ontology as opposed to several smaller ones. It is acknowledged
that this shared ontology might expand and become too slow and
complicated to use. Additional limitations include the complexity
and slowness of defining standards needed for one large
heterogeneous ontology [
First one of the problems of the shared ontology approach, the slowness of using a large ontology, can be eliminated with RDF schemas. With RDF schemas it is possible to specialize the users of the ontology to be task-specific experts; they do not have to know every bit of information about the environment. The problems with defining standards for large ontologies are outside the scope of this paper.
The size and magnitude of the environment is a relevant question
within the limits of this paper; how large and heterogeneous is
the environment supposed to be? Instead of concentrating on
domains, disciplines, business branches, etc., the concept of
environment is used here along the following guideline: if two
applications share one or more concepts, they belong in the same
One important property of an ontology is extensibility [
RDF schemas are intended to function in a roughly similar role
than DTD's function for XML documents. Individual RDF
documents are validated against some RDF schema. RDF
Schema specification [
The RDF schema specification [
In this paper properties are nevertheless often contrasted with classes and resources. The reason for this is to differentiate the concepts that get defined from those that participate in defining them. xmlns="http://www.santtusvideos.com/schema/"> xmlns="http://www.santtusvideos.com/schema/"> <director </director> <starring </starring> <length
114 </length> </Description> </RDF>
Martin Scorsese
Robert De Niro
Here the mpg-version of Taxi Driver is presented as an RDF
resource. It has three properties: director, starring and length.
These are specified in individual statements. This document is
validated against the RDF schema of an imaginary web video
service called Santtu's Videos. Here all the statements refer to
the same schema but this is not necessary. Features of a given
resource could be defined in separate schema documents.
The shared ontology approach is favored in this paper over the
multiple ontologies approach [
RDF(S) is intended to provide metadata about web resources that
is both human-readable and machine-understandable [
Individual RDF documents are validated against RDF schemas. RDF documents consist of descriptions, which in turn consist of statements. Each description represents some resource. Each statement represents some feature of the resource that is being described.
In RDF schemas the interrelations among selected resources and properties are defined. RDF documents contain naturally more specific and case-related data than RDF schemas; in RDF documents properties and resources are given values and thereby the ontological system is tied to actual instances of the resources. Following is a simple example of RDF in an XML syntax: <RDF xmlns="http://www.w3.org/1999/02/22-rdf
syntax-ns#">
<Description
about="http://www.santtusvideos.com/
taxidriver.mpg">
xmlns="http://www.santtusvideos.com/schema/">
the users of the model proposed here can consist of software
agents in addition to human beings. For example in FIPA
(Foundation for Intelligent Physical Agents) there is already
work done around the usage of RDF as a content language for
software agents [
The semantic information that software agents use should be
mainly external to the agents themselves [
This semantic information external to the agents is distributed to all the other parts of the proposed model: RDF documents, RDF schemas, and the ontology. The agents should be designed so that they understand one or more RDF schemas. The agents can utilize individual RDF documents as pieces of case-specific information. They can do different things with the documents (and applications/services bound to the documents) according to their internal inference rules and the RDF schema/schemas they are committed to.
Also human users of the environment benefit from this model. The model helps people to understand different parts of the environment and applications appearing in it. For example, if someone decides to introduce a new service or resource into the environment, he can examine the schemas of the existing resources (assuming that the schemas are publicly available for examination).
If he finds a suitable schema, he can utilize it as a means to exploit the ontology. If no schemas as such work for the developer, there still might be some guidelines or pieces of information in existing schemas that help the developer to get started. Either way, people working in an environment with shared ontology can reduce the amount of work with public RDF schemas and this way avoid re-inventing the wheel.
3.1 RDF and web resources
Again: RDF is intended to provide metadata about web
resources. Different web resources naturally have different
means of categorization. For example, libraries, video stores and
digital phone books use different concepts as metadata [
RDF(S) is in this paper proposed as a means to provide
casespecific information rather than means to constitute the whole
ontology. The reason for this reduces to the question concerning
the ontological status of properties. In [
In other words: Concepts that are definienda in some case or
application form the basic level of concepts for that particular
case. In RDF(S) terminology these would be classes to be
defined. There are two other levels in addition: subordinate and
superordinate level. Definiens (property in RDF(S) terminology)
is at the subordinate level when compared with definiendum.
Depending on the domain, however, relations between these
levels vary (cf. [18 and 16]).
3.3 An example characterizing the
casespecificity of RDF(S)
From the electronic video store's point of view director is a
property that characterizes the resources of a class called movie.
This is fine as long as it is clear that somewhere else director
could appear also as a class that gets defined by some other
properties. An electronic catalog of artists might have director as
a class5. Now directors could have movies that they have directed
as their properties. Just the other way around than in the video
store6. This is illustrated in Figure 2.
When representing the world, the structure of concepts should be
as analogous to reality as possible [
Of course it is possible to introduce all (or at least majority of)
concepts twice; once with the status of definiendum and again
with the status of definiens. However, this is not a desirable
solution. It leads to compatibility problems and violates the
simplicity principle of ontologies. The principle states that there
should be as few ontological commitments in an ontology as
possible [
First the RDF documents column of the Figure 3 is examined. It tells us that in video store the movie Taxi Driver has a property named director with the value "Scorsese". Artist catalog, on the other hand, has the director Martin Scorsese with a property named director that has the value "Taxi Driver". So Taxi Driver and Scorsese both appear once as resources to be defined and once as properties. The same thing concerns the RDF schemas column of the picture. In video store director is a property that belongs in the domain7 of movies. In artist catalog the situation is contrary. 7 For the sake of simplicity only one constraint property is presented here. Besides domain, also range is a useful property to be exploited in RDF schemas. In the schema of video store, for example, there could be a range constraint property named person attached to the director property. This would mean that the value of the director property is always a member of the class person. Furthermore, only one domain for each property is presented. If necessary, though, director could have other domains besides movies. It could be attached to TV-series, theatre plays, etc.
A phenomenon closely related to this is observed in [
One remark here could be that since RDF is intended for describing web resources, the movie Taxi Driver (at least in mpg-format as in the code example presented earlier) is more appropriate candidate for a web resource than the director Martin Scorsese. That is because Martin Scorsese can not appear in a format distributed in the Internet unlike Taxi Driver.
Ontologically speaking, however, the movie Taxi Driver is not the same entity as the mpg-version of it distributed in the net. It is rather an abstract thing that has different instances. Compared with object-oriented programming, the movie Taxi Driver would be a class and the copies of that movie (for example the mpgversion distributed in the electronic video store) in turn instances of the class.
The expressive power of RDF(S) does not necessarily complete
all the parts that are needed for expressing a semantic
description of some system 8. An ontology independent of the
domain-specific details of its usage is needed. There should be
an "isolated basic backbone" of ontology that is independent of
any case-specific details [
What RDF(S) technology can do, however, is to provide means to access an ontology characterizing some environment – no matter how large or heterogeneous – in many ways.
5.1 Rethinking the properties
The main problem of RDF(S) presented in this paper is the
division of the concepts into properties and classes. The answer
proposed to this problem is the usage of an external ontology in
addition to the RDF(S). From the ontology's point of view the
usage of concepts in different RDF(S) is based on roles [
At first sight this might seem wise. More carefully examined,
however, this leads to a situation not preferable to defining every
concept twice; once as a property and once as a class. The
number of properties would be doubled as shown in figure 4;
has_directed and directed_by would both exist between a movie
and its director even though they have the same information
content. This would again violate the simplicity principle of
ontologies [
However, introducing rules and restrictions that cope with
limitations of RDF at a higher level does not seem feasible. For
one thing, this again violates the simplicity principle of
ontologies [
If all the concepts and their interrelations are such that they are found in the ontology it could be possible to make that deduction. Possible conflicts should however try to be avoided. If some concept is a property (definiens) in one schema and a resource to be defined (definiendum) in another, does that have any impact on the ontology? If it does, which one of the schemas determines the "ontological" location of the concept in question. If it does not, how it is possible to construct any hierarchy in the ontology (since there would be nothing in addition to the schemas)? On the other hand, if there are some general relations or attributes at the ontological level that are not visible in RDF(S), the deduction is not possible.
Clearly a deduction in the other direction is not possible. There is no way of knowing how the concepts in the ontologies are used and grouped in different RDF(S). This means that it is not possible to deduce all imaginable RDF(S) just by examining the ontology. And this is due to the proposed case-specific nature of RDF(S).
Acknowledgements. Thanks to Joose Niemistö and Johannes Gröhn for their help on this article.