=Paper=
{{Paper
|id=Vol-315/paper-3
|storemode=property
|title=Three-Layer OWL Ontology Design
|pdfUrl=https://ceur-ws.org/Vol-315/paper3.pdf
|volume=Vol-315
|dblpUrl=https://dblp.org/rec/conf/kcap/DumontierV07
}}
==Three-Layer OWL Ontology Design==
https://ceur-ws.org/Vol-315/paper3.pdf
Three-Layer OWL Ontology Design
Michel Dumontier1,2, Natalia Villanueva-Rosales1
1
School of Computer Science, 2Department of Biology,
Carleton University, 1125 Colonel By Drive, Ottawa, Ontario, K1S5B6 Canada
michel_dumontier@carleton.ca, nvillanu@scs.carleton.ca
Abstract. Ontology reuse is often predicated on agreement to the semantics
wholly defined within an ontology document. But when faced with overly
constrained semantics that one might partially reject, might one instead prefer
reduced ontological commitment? In this paper, we describe how three layered
ontology design promotes maximal reuse of domain ontologies by separating
taxonomically organized domain terminology from i) disjointess, ii) complex
expressions that define a world view, and iii) application specific requirements
that impose a specific data model for data exchange and document validation.
Keywords: Ontologies, OWL, ontology design, normalization, reuse.
1 Introduction
A major goal of the Semantic Web is to add explicit semantics to web content [1],
which can be realized by using ontologies to describe and relate objects with formal,
logic-based representations. OWL, the Web Ontology Language, is recommended by
the W3C for building Semantic Web ontologies [2]. OWL-DL is a highly expressive
variant of OWL and is based on description logics (DL), a subset of First Order Logic
that allows description of complex concepts from simpler ones with an emphasis on
decidability of reasoning tasks. While the theoretic foundations of DL are well
developed [3], there exists an enormous gap for best practices in designing OWL
ontologies that enable maximal reuse for unexpected application scenarios.
In line with RDF/XML, OWL entities are named by a Universal Resource
Identifiers (URI). A key and controversial issue is that despite any intent of the URI
minter, meaning of a semantic web entity is locally defined by the document it is
found in, opening the possibility that there can be multiple meanings for any URI [4].
Importantly, semantic web users who want to instantiate an OWL class but disagree
with the logical restrictions placed over it in an ontology, could devise or choose the
semantics they agree with. Users might prefer this option rather than minting their
own URI because these then requires mapping to other URIs, currently a semi-
automated and time consuming process. In order to provide ontologies that can be
maximally reused and also reduce the possibility of rogue documents with potentially
inconsistent meanings, we propose a three layer design for OWL ontologies with
incremental ontological commitment. This proposal thus has consequences for the
normalization, modularization, mapping and evolution of OWL ontologies.
2 Three-layer ontology design pattern
We propose a three-layer ontology design of classes/properties and of instances (Fig.
1). The first layer (primitive) consists of classes/properties forming taxonomic trees in
which a single parent may be asserted. The second layer (complex) captures domain
knowledge in the form of necessary or necessary and sufficient conditions. The top
layer (application) is composed of additional restrictions that effectively describe a
particular data model. Similarly, instance assertions are also divided into three
increasingly restricted layers. Instances object assertions (primitive) are separated
from more sophisticated expressions (complex) using specific OWL language
constructs. Finally, axioms that effectively impose more specific assumptions (unique
name assumption, closed world assumption) are maintained in a separate layer for
data model validation. Each layer is now described in greater detail.
App1 Restrictions App2 Restrictions Assumptions
Complex Complex
Class/Property Primitive Instance Primitive
Fig. 1. Three-layer ontology design for classes/properties (solid box) and instances
(dashed box).
2.1 Class/Property Taxonomic Layer
The first layer contains the declaration of domain terminology and relations in the
form of OWL entities (owl:Class; owl:ObjectProperty; owl:DatatypeProperty) that
are annotated with a human readable label (rdfs:label) and associated with a clear,
concise and precise definition (rdfs:comment). We adhere to Alan Rector’s
recommendations on normalization [5], in that each entity is asserted to have exactly
one parent, thus forming primitive trees. Since the design of primitive trees consisting
of “self standing concepts” is acknowledged to be ambiguous and difficult [5], we
relax the criteria of forming disjoint trees, thereby avoiding potential inconsistencies
arising from disjointness. This omission is particularly relevant when reasoning with
additional layers having defined necessary and sufficient conditions such that
instances are realized into multiple classes. For instance, although a scatterplot and a
line graph are different types of statistical graphs 1 , the assertion of disjointness of
these classes in a graph ontology would preclude the discovery of a line graph from a
scatterplot that has lines drawn between the data points. Thus, these primitive layers
enable users to benefit from taxonomically organized declarations of domain
terminology without having to further commit to semantics that might lead to
inconsistencies with additional knowledge. Moreover, it allows users to exchange
data identified by class membership and related with the provided basic relations.
1 http://ontology.dumontierlab.com/statistical-graph-primitive
2.2 Complex layer
The complex layer refines the primitive layer by imposing restrictions such as
necessary or necessary and sufficient conditions beyond the asserted subsumption.
These include more expressive aspects of the OWL-DL language (OWL 1.1)
including disjunction, existential and universal restriction, union, intersection,
complement, and (qualified) cardinality restrictions, transitivity of properties,
reflexivity of properties, etc. The complex layer aims to augment the ontology with
domain knowledge that would lead to meaningful inferences, and includes mappings
to upper level ontologies. Drawing from the statistical graph domain, a Time Series
Graph would necessarily require time intervals as categorical data. Fully defining
necessary and sufficient conditions (equivalentClass) enables realization of class
membership by DL reasoners. For instance, the complex layer of our statistical graph
ontology 2 fully describes a Graph Title, thus titles of a graph are automatically
discovered.
2.3 Application restrictions
The application restriction layer applies highly restrictive constraints and may be used
for the purposes of document validation and application interoperability. This layer
refines primitive classes and their relations, and should extend the complex layer, if
defined. Requirements are described using necessary conditions and if necessary
applying (carefully) extra closure axioms with universal restrictions (only) and
cardinality restrictions (min, max, exactly). For instance, the iGraph ontology 3
specifies the necessary conditions that defines a valid data model to be used with the
iGraph application [6]. Alternatively, one could define new application specific
classes (i.e. iGraphLineGraph) and determine whether instances are realized as class
members from necessary and sufficient conditions. In this way, the specification for
necessary information is completely encoded in the ontology and can be validated
using a DL reasoner.
2.4 Instance layers
Assertions about instances are also placed in separate layers for data
exchange/integration, reasoning and document validation. The instance primitive
layer includes assertions about class membership, object relations and datatype
values. This layer can be seen as a solution for developing a flexible model that can
be used to annotate publicly available data (e.g. from databases) without interest or
need of commitment to more complex logical restrictions. For example, our
instantiated statistical graph ontology 4 asserts that graph is an instance of the class
Graph and this is related to the plot instance of the class Plot class using the hasPart
property defined in our basic relation ontology.
2 http://ontology.dumontierlab.com/statistical-graph-complex
3 http://ontology.dumontierlab.com/igraph
4 http://ontology.dumontierlab.com/igraph-example
A second layer (complex) adds axioms such as existential restrictions to denote
necessary relations between instances (possibly unknown) or (qualified) cardinality
restrictions to identify how many of such relations must to exist. Thus, expressive
OWL constructs such as owl:sameAs, owl:differentFrom may be specified in this
layer.
Finally, the assumptions layer contains axioms that are needed to enforce a specific
data model that could modify the reasoning inferences. For instance, if we want to
assume that the information is complete and we would like to obtain inferences bases
on this assumption, we might want to “close” the knowledge base by asserting closure
axioms (i.e. complex class membership, ABox closure, etc).
3 Discussion
The increasing expressivity in the three-layered ontology design is analogous to the
formulation of more complex descriptions from simpler ones in description logics.
Domain knowledge is first captured as a taxonomic tree of non-disjoint classes which
is asserted in an OWL ontology along with basic relations required to populate the
ontology. Such normalization is an important aspect of ontology design because it
provides front line support for ontology reuse, maintainability and evolution.
Additional knowledge leading to meaningful inferences (e.g. membership inferences
via realization) is declared in a separate layer, thus enabling multiple “world views”,
i.e. ontologies with the same primitive layer, but different complex layers, could
potentially drive to different sets of inferences. Finally, the specification of
application requirements aims to ensure application compatibility and support for
interactions with other data models, e.g. relational databases. In providing multiple
levels of ontological commitment, users may choose to commit to i) a non-disjoint
primitive tree, ii) a logically enriched tree containing additional, explicit domain
knowledge, or iii) a constrained data model. Thus, our approach promotes ontology
reuse, at various levels of ontological commitments, which should be acceptable to a
broad number of requirements. Indeed, since OWL currently only supports document
level imports (owl:imports), a user now has three documents to choose from, rather
than rejecting the ontology altogether due to restrictions placed on domain entities
that they may not agree with. Moreover, he may choose to build a new complex layer.
Ontologies containing taxonomically organized terminology are already used to
manage biomedical terminology (i.e. the Gene Ontology). As more sophisticated
semantic web ontologies (i.e. BioPAX) emerge, we advocate that more restricting
axioms, for specific domain and data models can be described in a separate layer. The
BioPax OWL-DL ontology aims to exchange molecular interactions, complexes and
pathways [7], and places database-like restrictions to enforce single values and
closures. While proteins are routinely referred to using multiple names, instances of
BioPax Protein class are restricted to having at most one name (using datatype
property bp:NAME). This choice clearly reflects a desire to maintain the relational
database model in which name is a single valued field in a table containing protein
entries. Importantly, certain restrictions require closed world reasoning that holds in
the relational model, but does not hold in OWL. It seems unreasonable to outright
reject instantiating the class bp:Protein because of this restriction. Thus, cleanly
separating application restrictions from domain terminology allow users to instantiate
the ontology with the declared classes and relations, but don’t have to commit to
restrictions placed to facilitate (legacy) data exchange. Should the exclusion of role-
based restrictions in the primitive layer lead to inconsistent ontologies using a
complex layer, then procedures to resolve the discrepancy may yet have to be
developed.
Another issue identified by the BioPAX community involves the validation of
OWL documents meant for data exchange, but whose underlying semantics assume
an open world. We propose that additional layers over data may be used to place
certain assumptions, which can then be used to validate specific data models. This
model supports interoperability with relational database systems with primary key
requirements and whose implementation assumes unique name assumption and the
closed world assumption. The topmost layer of assumptions over the data could also
be used to integrate/exchange knowledge with systems with other models such as
modal logic systems. We expect that continued development of reasoning systems
will make this layer less essential in the future.
Our proposal is compatible with approaches towards the automatic design of
ontology modules [8, 9]. Ontology modules include minimal information such that
importing the module would provide the same answers to arbitrary queries as
importing would have been obtained with the entire ontology. While the taxonomic
layer provides some information about the nature of the entities, the complex layer
provides valuable role-based descriptions from which subsequent modularization
would be most effective for semantic query answering.
This proposal has also interesting considerations for ontology mapping and
ontology evolution. Ontology mapping involves the automated mapping of
ontological entities such that for each concept in ontology A, a corresponding concept
in ontology B is identified (if existing), with the same or similar semantics. Major
issues in ontology mapping include 1) ensuring that the restrictions placed in one
document are equivalent to another and 2) dealing with ontology evolution such that
changes are propagated. Our approach potentially would allow agent-based
negotiations to occur in three steps, beginning with the least expressive to the most
expressive set of axioms about the resources. Thus, agents may determine an
equitable mapping at some less expressive level, thereby avoiding inconsistencies that
preclude a mapping. With respect to ontology evolution, class membership assertions
are expected to be more stable than assertions that restrict classes, therefore this
approach will modularize ontology evolution in different documents, separating the
knowledge that evolves more frequently from the one that does not. This model was
useful in the design of ontologies in the iGraph system. In particular, the flexibility to
separate the application restrictions from the complex layer, allows reusing the latter
when modeling other application requirements without committing to the iGraph
requirements [10]. Future work includes applying this model to other ontology-
centered integration systems.
4 Conclusions
In this paper we described a methodology for a three-layered ontology design that
captures domain knowledge at various levels of granularity so as to maximize
ontology reuse, maintainability and evolution, yet also satisfy application specific
requirements for data exchange. This approach separates the declaration and
taxonomic organization of domain concepts from complex class descriptions that may
be used for enhanced reasoning or for placing constraints on data exchange. We
believe that this three-layered design will be particularly useful for reasoning with
across multiple world views and where there is a need for semantic annotations of
data and integrate/exchange knowledge with systems based on other models such as
relational databases and modal logic systems. Research towards flexible exchange is a
key activity that will facilitate the use of the semantic web in real world applications.
Acknowledgments. We thank the anonymous reviewers for providing constructive feedback
that has improved the quality of the paper. Funding for NVR and MD was with CONACYT
scholarship #150581 and NSERC, respectively. We thank Leo Ferres for providing iGraph
system requirements.
References
1. Berners-Lee, T., Hendler, J., Lassila, O.: The Semantic Web. Scientific American. Vol.
(2001)
2. W3C: OWL Web Ontology Language Guide. In: Smith, M.K., Welty, C., McGuinness, D.L.
(eds.) (2004)
3. Baader, F., Calvanese, D., McGuinness, D., Nardi, D., Schneider, P.: The Description Logic
Handbook: Theory, Implementation and Applications. Cambridge (2003)
4. Parsia, B., F. Patel-Schneider, P.: Meaning and the semantic web. Proceedings of the 13th
international World Wide Web conference on Alternate track papers & posters. ACM
Press, New York, NY, USA (2004)
5. Rector, A.L.: Modularisation of domain ontologies implemented in description logics and
related formalisms including OWL. KC2003. ACM Press, Sanibel Island, FL, USA
(2003)
6. Ferres, L., Verkhogliad, P., Lindgaard, G., Boucher, L., Chretien, A., Lachance, M.:
Improving Accessibility to Statistical Graphs: The inspectGraph System. Ninth
International ACM SIGACCESS Conference on Computers and Accessibility,
Tempe, Arizona (2007)
7. Luciano, J.S.: PAX of mind for pathway researchers. Drug Discov Today. Vol. 10 (2005)
937-942
8. Bernardo Cuenca, G., Ian, H., Yevgeny, K., Ulrike, S.: Just the right amount: extracting
modules from ontologies. Proceedings of the 16th international conference on World
Wide Web. ACM Press, Banff, Alberta, Canada (2007)
9. Grau, B.C., Parsia, B., Sirin, E., Kalyanpur, A.: Modularizing owl ontologies. KCAP, Banff,
Canada (2005)
10. Ferres, L.D., Michel, Villanueva-Rosales, N.: Semantic Query Answering with Time-Series
Graphs. The 3rd International Workshop on Vocabularies, Ontologies and Rules for
The Enterprise (VORTE 2007), Annapolis, USA. (2007)