Modeling vs encoding for the Semantic Web
Editors: Krzysztof Janowicz, Pennsylvania State University, USA; Pascal Hitzler, Wright State University, USA
Solicited reviews: Thomas Lukasiewicz, Oxford University, UK; Giancarlo Guizzardi, Federal University of Espírito Santo, Brazil
Open review: Pascal Hitzler, Wright State University, USA
Institute for Geoinformatics (ifgi), University of Münster, Weselerstr. 253, D-48151 Münster, Germany
Abstract. The Semantic Web emphasizes encoding over modeling. It is built on the premise that ontology engineers can say
something useful about the semantics of vocabularies by expressing themselves in an encoding language for automated rea-
soning. This assumption has never been systematically tested and the shortage of documented successful applications of Se-
mantic Web ontologies suggests it is wrong. Rather than blaming OWL and its expressiveness (in whatever flavor) for this
state of affairs, we should improve the modeling techniques with which OWL code is produced. I propose, therefore, to sepa-
rate the concern of modeling from that of encoding, as it is customary for database or user interface design. Modeling seman-
tics is a design task, encoding it is an implementation. Ontology research, for applications in the Semantic Web or elsewhere,
should produce languages for both. Ontology modeling languages primarily support ontological distinctions and secondarily
(where possible and necessary) translation to encoding languages.
Keywords: Ontology modeling and encoding, semantic engineering, expressiveness, algebra, functional languages, Haskell
The Semantic Web transcends all previous at-
tempts at enriching data with explicit semantics. Yet,
the modeling languages brought to the task are
weaker than those for conceptual modeling at
smaller scales, such as databases or user interfaces.
As a consequence, the Semantic Web rests on the
premise that it is possible to produce and understand
ontologies in OWL, using editors like Protégé1 .
Many of those who have tried this doubt the premise,
especially if they have also done other kinds of con-
ceptual modeling. Their experience in over three
decades of conceptual modeling does not support
the conclusion that description logic statements
adorned with syntactic sugar and design patterns are
sufficient (or even necessary) to capture what people
mean when they use a vocabulary.
1 Witness the online guide to Protégé: “The Protégé platform
supports two main ways of modelling ontologies – frame-based
and OWL” (http://protege.stanford.edu/doc/owl/getting-started.
Modeling semantics is a design task, encoding it
is an implementation. With the former we explore
how to constrain human and machine interpretations
of vocabularies, with the latter we support auto-
mated reasoning. Expressiveness is an essential cri-
terion for the former, decidability for the latter. Mix-
ing the two concerns is harmful for both tasks, but
routinely done. While there are complementary ap-
proaches to encode semantics (for example, through
machine learning), the scope of this note is limited
to conceptual modeling.
2. Can we support our own goals?
Architects do not start a design by constructing
geometric figures, database administrators do not
start a project by creating relational tables, and user
interface designers do not model interfaces in a user
interface toolkit. Each of these fields has its model-
ing languages and environments, allowing, for ex-
Semantic Web 1 (2010) 1–5
1570-0844/10/$27.50 © 2010 – IOS Press and the authors. All rights reserved
ample, to sketch a building, draw diagrams, or com-
What does the Semantic Web offer in support of
design? How does it assist modelers in choosing
ontological distinctions, testing their implications,
exploring their varieties, experimenting with appli-
Ontology engineers can choose from informal or
semi-formal techniques, such as twenty question
games, sorting tasks, concept mapping, or document
analysis2. They can follow best practice3 and get
advice on pitfalls to avoid4. The more formally in-
clined designers can use methods like OntoClean ,
which ask key questions, for example, about identity
Yet, the gap between informal knowledge elicita-
tion techniques, design patterns, and design methods
on the one hand, and useful, tested OWL axioms on
the other often remains too wide to jump across
without breaking a leg or two. Evidence for this
comes in the form of typical problems found in
− confusing instance-of with subclass-of;
− confusing part-of with subclass-of;
− leaving the range of an OWL property unspeci-
− introducing concepts and properties that are not
sufficiently distinguished from others (a.k.a.
These and other well-known problems may just
be attributed to sloppiness in modeling. However, if
it is too easy to be sloppy without noticing it, the
Semantic Web will have a serious quality and repu-
tation problem. Also, some of these problems occur
in prominent spots, such as Protégé’s Guide to Cre-
ating Your First Ontology5, which teaches us, for
example, to model Côtes d’Or region as a class (!)
and furthermore as a subclass-of Bourgogne region.
OWL ontologies cannot be expected to be di-
rectly written or understood by modelers, because
OWL is optimized to support machine reasoning,
not human thought. The Semantic Web is today at a
stage of maturity that databases had passed in the
1970’s and user interfaces in the 1980’s, when they
2 see http://www.semanticgrid.org/presentations/ontologies-
tutorial/GGFpart4.ppt for an excellent overview
3 for example, http://www.w3.org/2001/sw/BestPractices/OEP
abandoned using a single paradigm for encoding and
modeling. Relational algebra for database encoding
and logical devices for user interfaces have been
complemented by conceptual modeling languages
like entity-relationship or state-transition diagrams,
and subsequently by much more elaborate design
A likely consequence of the relatively immature
state of modeling support is that today’s Semantic
Web contains assertions, whose implications have
never been understood by anybody, and which may
have been tested for satisfiability at best, but not for
correctness or relevance.
In the face of this situation, some commonly en-
countered claims about the goals of the Semantic
Web appear rather bold. For example, in a classical
paper introducing OWL, it has been said that
“ontologies are expected to be used to provide
structured vocabularies that explicate the relation-
ships between different terms, allowing intelligent
agents (and humans) to interpret their meaning flexi-
bly yet unambiguously” .
The goal of unambiguous interpretation is a for-
midable one, and variants of the idea that ontologies
contain “precisely defined meanings” are propa-
gated throughout the Semantic Web literature and in
countless project proposals. A recent paper co-
authored by the Semantic Web’s father, Tim Bern-
ers-Lee, even carries it forward to linked data,
“meaning is explicitly defined”
according to the authors . Claims like these, if not
taken with a very large grain of salt, vastly overstate
the achievable goals of the Semantic Web and create
expectations that are bound to be disappointed, at
least with the currently available modeling support.
In the rest of this note, I will first suggest that the
Semantic Web community should adopt a more
modest engineering view of semantics. Then, I will
argue why OWL is too weak for modeling. Finally, I
will propose to use and develop modeling languages
to complement today’s encoding languages.
3. An engineering view of semantics
Neither linguists nor philosophers have so far
been able to define meaning as an object of scien-
tific study in a way that would capture what people
mean when they use vocabularies. Thus, specifying
particular “meanings”, or targeting unambiguous
interpretations, rests on shaky grounds, no matter
W. Kuhn / Modeling vs encoding for the Semantic Web2
how it is attempted. Yet, while philosophers figure
out what meaning really means, information scien-
tists and engineers can use ontologies pragmatically
to constrain interpretations.
Ontology engineers have recommended striving
for minimal ontological commitments , rather
than for any kind of completeness in ontological
specifications. In this spirit, I have recently pro-
posed a pragmatic view of concepts and their speci-
fications . It treats meaning as a process to be
constrained, rather than as an object to be defined.
As the saying “words don’t mean, people do” ex-
presses, it is people who mean something when they
use a vocabulary, rather than the words having a
fixed meaning. Such a process view of meaning can
be captured by the threefold view of concepts in the
meaning triangle (Fig. 1), involving people using
− words to
− express conceptualizations and to
− refer to something in the world.
For example, the community of English speakers
uses the word “star” to refer to shapes like that in
the bottom-right corner of Fig. 1. Note that a
speaker’s or listener’s idea of a star may not exactly
match the instance referred to.
If one adopts such a triadic view of concepts, on-
tologies do not need to specify “meanings”, much
less the existence of something in reality. They sim-
ply provide humans or machines with constraints on
how to apply and interpret vocabularies. These con-
straints support reasoning, while not removing all
The Semantic Web does not need to and probably
cannot raise the bar higher than this semantic engi-
neering goal. Traditional formal semantics is com-
patible with it, as long as truth is not put before
meaning. Truth is a consequence of meaning, just as
much as it is a cause, as the cyclic nature of the tri-
angle implies. Truth conditions do not “define”
meaning, but constrain interpretations of vocabular-
ies to the ones that make sentences “true”. For natu-
ral or informal technical languages, where there are
no logical truth criteria, this is equivalent to sen-
tences being correctly interpreted in the language
community. For example, English speakers need to
be able to correctly interpret the word “star” when it
is used to refer to the bottom-right shape in Fig. 1.
4. Does OWL support modeling?
Naked OWL code does not convey much insight.
Consider how hard it is to understand what some-
body else’s OWL statements say about a collection
of concepts. OWL editors like Protégé provide class
and property hierarchies as useful overviews, but do
not show how the stated properties interact. Some
context and rationale for the statements may be
guessed from annotations, but the processes in
which concepts are used remain informal and often
invisible. Graph representations, showing classes as
nodes and relations as edges, can be produced with
Protégé plug-ins, but these tend to work only one-
way and show only parts of the story. Even profes-
sional (and costly) development environments
provide only limited and fractioned support for ex-
ploring, communicating, and evaluating – in other
words for modeling before, while, and after encod-
In addition to these usability and understandabil-
ity issues, OWL and the current tools around it are
often not expressive enough for modeling. For ex-
− treat properties and relations as the same,
though these are two rather different ideas in
− limit us to binary relations, though interesting
relations often start out as ternary or more;
− provide a well-defined primitive for taxono-
mies (subclass-of), but not for partonomies and
other formal ontological relations;
− make it hard to encode processes and events,
though these are often essential elements of
OWL’s expressiveness may be sufficient (or even
overkill) for eventual encoding and machine reason-
ing; but human understanding and reasoning require
Fig. 1. The meaning triangle.
W. Kuhn / Modeling vs encoding for the Semantic Web 3
more expressive languages and we do not seem to
understand yet what these should be. Web searches
for “ontology modeling languages”, for example,
lead either to OWL or to UML or to requirements
for better modeling languages with experimental
implementations at best. While it is possible and
valuable to introduce ontological distinctions into
diagrammatic modeling languages like UML ,
automated reasoning support at the modeling stage
requires more formal languages. Note that a call for
more powerful modeling languages does not contra-
dict the idea of lightweight approaches in implemen-
tation. On the contrary, better modeling tends to
produce leaner, simpler encodings.
5. Toward formal modeling languages
Modeling ontologies involves tasks like
− finding out what should be said,
− understanding what has been said,
− conveying that understanding to others,
− checking whether it is what was intended,
− spotting errors in what has been said, and
− testing whether what has been said is relevant
These are human reasoning and communication
challenges that are unlikely to be met by reasoning
and visualization at the encoding level. As Nicola
Guarino proposed in , they require ontological
distinctions built into modeling languages and auto-
mated reasoning support during modeling. What
exactly the ontological distinctions should be re-
mains an important research question. The formal
ontological distinctions proposed by Guarino (essen-
tially, those of OntoClean ) are immensely useful,
but appear difficult to build into modeling languages.
Here, I will propose some distinctions that are al-
ready available in existing languages. Their choice
has been motivated by the work of Joseph Goguen
on algebraic theories  and 25 years of conceptual
modeling applying these ideas, 15 years of which
using the functional language Haskell (http://
haskell.org), into which many of them are built.
Similar ideas and arguments with a larger scope can
be found in .
A fundamental ontological distinction is that be-
tween objects, events, and properties (these labels
vary, but the basic idea remains the same). It is quite
natural for humans to think in terms of objects and
events with properties, rather than just in terms of
predicates. Description logics do not allow for this
distinction. Functional languages offer formal dis-
tinctions, based on kinds of functions. For example,
properties (say, temperature) can be modeled as
functions mapping objects (say, air masses) to val-
ues; events (say, a storm) can be modeled as func-
tions mapping between objects (say, air masses
With a distinction between objects and events
comes the need and opportunity to model the par-
ticipation relation (say, of air masses in storms).
Since participation is fundamental for semantics
(witness the idea of thematic roles), having it as a
modeling primitive would be very useful. Types in
any language provide it. Data types and operation
signatures capture the possible participation of ob-
jects in operations. Thereby, typing also provides a
theory for instantiation, which is an undefined
primitive in description logics: instances of a type
are the individuals who can participate in its opera-
Distinguishing properties from relations is
straightforward through unary functions for proper-
ties and n-ary (Boolean) functions for n-ary relations.
Any interesting ontological relations (such as part-of
or location) can then be specified through equational
axioms, well known from algebraic specifications.
For example, one can specify that an object that was
put into another object is in that object as long as it
has not been taken out again. Function composition
allows for defining semantic constraints involving
such sequences of events, which abound in practice.
Some recent examples of these and other modeling
capabilities of Haskell can be found in .
Modeling roles is much harder. Their anti-rigidity
may be captured through so-called dependent types,
whose instances depend on their values. Haskell
type classes (generalizations over types) offer a use-
ful model without explicit typing. For example, the
types Person and University can be declared as be-
longing to a type class which provides the functions
of enrolling and unenrolling. Students are then mod-
eled by terms like enroll (person, university), rather
than by explicit typing.
Haskell’s most useful support for modeling, how-
ever, stems from the fact that, as a programming
language, it allows for simulation: ontological speci-
fications can be tested through their constructive
model, while and after being developed .
W. Kuhn / Modeling vs encoding for the Semantic Web4
As Einstein is said to have pointed out, every-
thing should be made as simple as possible, but not
simpler. Conceptual modeling in the Semantic Web
has been made to look simpler than it is. This carries
the risk of yet another turn against Artificial Intelli-
gence in some of its latest incarnations (not only the
Semantic Web, but also Linked Data).
Yet, the idea of allowing information sharing with
minimal human intervention at run time is too good
to be discredited. Therefore, our challenges as re-
searchers in this field are to
− promise only what we honestly believe we can
− work hard to achieve what we promised;
− validate what we have achieved;
− improve our theories and tools.
These challenges give us some re-thinking and re-
tooling to do, and the new journal is a welcome
place to report progress. It would be a pity if encod-
ing-biased views about admissible approaches to
ontology engineering precluded alternatives that
support conceptual modeling.
Among the many colleagues who helped shape
these ideas over decades in discussions (without
necessarily agreeing) are Andrew Frank, Joseph
Goguen, Nicola Guarino, Krzysztof Janowicz,
Ronald Langacker, David Mark, and Florian Probst,
as well as the three reviewers.
 Bizer, C., Heath, T., Berners-Lee, T., 2009. Linked Data –
The Story So Far, http://tomheath.com/papers/bizer-heath-
berners-lee-ijswis-linked-data.pdf (preprint of a paper to
appear in: Heath, T., Hepp, M., and Bizer, C. (eds.). Spe-
cial Issue on Linked Data, International Journal on Se-
mantic Web and Information Systems (IJSWIS).
 Burstall, R.M., Goguen, J.A., 1977. Putting theories to-
gether to make specifications, Proceedings of the 5th In-
ternational Joint Conference on Artificial intelligence,
Cambridge, USA: 1045–1058.
 Gruber, T.R., 1995. Toward principles for the design of
ontologies used for knowledge sharing. International Jour-
nal of Human-Computer Studies, 43(5–6): 907–928.
 Guarino, N., 2009. The Ontological Level: Revisiting 30
Years of Knowledge Representation. Conceptual Model-
ling: Foundations and Applications. Essays in Honor of
John Mylopoulos. Edited by A. Borgida, V. Chaudhri, P.
Giorgini, E. Yu. Springer-Verlag: 52–67.
 Guarino, N. and Welty, C., 2002: Evaluating Ontological
Decisions with OntoClean. Communications of the ACM,
 Guizzardi, G., 2005: Ontological Foundations for Struc-
tural Conceptual Models. Telematica Instituut Fundamen-
tal Research Series No. 015.
 Horrocks, I., Patel-Schneider, P.F., and Harmelen, F. van,
2003. From SHIQ and RDF to OWL: The making of a web
ontology language. Journal of Web Semantics, 1(1).
 Kuhn, W., 2009. Semantic Engineering. In G. Navratil
(Ed.): Research Trends in Geographic Information Science.
Springer-Verlag, Lecture Notes in Geoinformation and
 Lüttich, K., T. Mossakowski, and B. Krieg-Brückner,
2005. Ontologies for the Semantic Web in CASL. In J.
Fiadeiro, editor, WADT 2004, Springer-Verlag, LNCS
 Ogden, C.K., Richards, I.A., 1923. The Meaning of
Meaning. Harcourt Brace Jovanovich.
 Ortmann, J., and Kuhn, W., 2010. Affordances as Qualities.
6th International Conference on Formal Ontology in
Information Systems (FOIS 2010), IOS Press: 117–130.
 Raubal, M. and W. Kuhn, 2004. Ontology-Based Task
Simulation. Spatial Cognition and Computation 4(1): 15–
W. Kuhn / Modeling vs encoding for the Semantic Web 5