=Paper=
{{Paper
|id=Vol-516/paper-3
|storemode=property
|title=Refining Ontologies by Pattern-Based Completion
|pdfUrl=https://ceur-ws.org/Vol-516/pap19.pdf
|volume=Vol-516
|dblpUrl=https://dblp.org/rec/conf/semweb/NikitinaRB09
}}
==Refining Ontologies by Pattern-Based Completion==
https://ceur-ws.org/Vol-516/pap19.pdf
Refining Ontologies by Pattern-Based Completion
Nadejda Nikitina and Sebastian Rudolph and Sebastian Blohm
Institute AIFB, University of Karlsruhe
D-76128 Karlsruhe, Germany
{nikitina, rudolph, blohm}@aifb.uni-karlsruhe.de
Abstract. Constructing richly axiomatized ontologies for real-world
knowledge-intensive applications is a time-consuming and difficult task. For
this reason, the future relevance of ontologies in practice depends on the
availability of advanced semi-automatic methods for ontology learning and
refinement. In this paper we propose a method to enrich ontologies with
complex axiomatic information by completing partial instantiations of ontology
design patterns.
Keywords: Ontology design patterns, ontology refinement.
1 Introduction
Richly axiomatized ontologies are essential for powerful, knowledge-intensive
applications, since they allow the application of advanced reasoning. However,
ontology modeling and construction is a difficult and time-consuming task.
Considering the fact that real world applications require large-scale knowledge bases,
there is a need for automatic methods to support ontology construction.
Unfortunately, automatically constructed ontologies tend to lack expressivity, e.g.,
axiomatic information about transitivity or symmetry of relations.1 For example, the
relations “before” and “after” found in DBPedia2 are not specified to be transitive
while transitivity would clearly be an expected characteristic of these relations. This
kind of information is difficult to acquire from unstructured resources, since it often
does not explicitly occur in text, but can only be deduced using external information
sources. However, we think that exploiting additional information sources can help to
bring forward the ontology enrichment.
As pointed out in [1], typical conceptual patterns arise during the design of
ontologies for different domains and different tasks. Ontology design patterns [2] –
modeling solutions to solve a recurrent ontology design problem – were introduced to
support the reuse of formalized knowledge. We consider the ontology design patterns
as a potential source to be exploited for ontology refinement.
The usefulness of ontology design patterns in semi-automatic ontology
construction has already been demonstrated in [3], where they have been used to put
automatically constructed ontology elements in context by extending the ontology
1 Relations are also referred to as properties in related literature
2 http://dbpedia.org/page/Angela_Merkel
29
with abstract concepts and relations. In this paper, we aim at the refinement of
ontology’s axiomatization using highly axiomatized knowledge contained in ontology
design patterns. The key idea of the proposed approach is to search for components
within an ontology which partially instantiate a given ontology design pattern. In this
way, potential missing ontology elements can be identified.
If we consider the previously mentioned relations “before” and “after” contained in
DBPedia as well as the ontology design pattern “precedence” introduced in [4] and if
we assume the ontology part including these two relations to be a partial instantiation
of the given ontology design pattern, we see that there are three axioms missing in
DBPedia – the axioms expressing the inverseness of the relations “before” and “after”
and their transitivity.
The ability to automatically recognize partial instantiations of an ontology design
pattern would therefore allow for checking an ontology for potential missing elements
and based on the outcome to automatically generate a list of suggestions for a
refinement. In this way, using frequently occurring and richly axiomatized ontology
design patterns as input could help to add a considerable amount of axioms to a
sparsely axiomatized ontology.
In order to automatically recognize an ontology design pattern by the means of an
algorithm, a set of indicative features of this pattern is required. We identify the
instantiations of ontology design patterns by their structure and the meaning of their
elements expressed by axioms and lexical characteristics of each element. Our
matching algorithm is based on these types of features. In this paper, we presume an
extended kind of ontology design patterns which contain additional lexical
information. In the following, we are going to use the expression ontology pattern or
pattern instead of ontology design pattern.
Our method is mainly independent from the employed concrete ontology
representation language. However, we presume that the underlying ontology
representation language of concerned ontologies supports complex axiomatizations.
The remainder of this paper is organized as follows: The next Section describes
research work related to this paper. Our algorithm is introduced in Section 3. Section
4 summarizes and gives an outlook to further research.
2 Related Work
Semi-automatic ontology construction and refinement has been addressed by
several approaches relying on different types of data sources. However, only few of
them aim at the acquisition of complex axioms going beyond the modeling
capabilities of RDFS.
There is a range of methods exploiting the information contained within natural
language texts in order to acquire additional axioms (for an overview see [5]). [6]
proposes a method for an axiomatization of glossaries such as WordNet based on
parsing and converting of natural language descriptions into formal definitions. [7]
also aims at the acquisition of complex axioms by the means of deep syntactic
analysis of natural language definitions.
30
There is a range of approaches relying on multiple data sources such as [8] which
aims at the acquisition of a particular type of axioms, namely disjointness axioms, by
gathering syntactic and semantic evidence from different data sources. RELExO [9]
combines learning complex class descriptions from textual definitions with the FCA-
based technique of relational exploration in order to clarify the subclass relationship
of concepts of an ontology. It generates hypotheses about class extension
relationships which cannot be deduced or denied using the axioms already contained
in the ontology. Then, it looks for counterexamples in the set of instances contained in
the ontology and, if none could be found, it asks the expert to provide a
counterexample or to approve the suggested hypothesis. RoLExO [10] relies on the
same type of user interaction and hypothesis verification, but generates hypotheses
about complex domain-range restrictions. A method proposed in [11] is another
example of extracting hypothetical domain axioms based on a given set of entities.
These approaches are complementary to ours, since they rely on other sources of
information to acquire complex axioms.
Blomqvist [3] proposes a framework for pattern-based semi-automatic ontology
construction and refinement. This work focuses on the refinement of ’lightweight’
ontologies concerning the logical complexity and expressiveness, which are not
intended to obtain a rich axiomatization. For this reason, ontology patterns are not
used to enrich the ontology with complex axioms, but to put the automatically learnt
ontology elements into context by connecting them with the more general concepts
and relations of the pattern.
To the best of our knowledge we are the first to address the general use of ontology
design patterns for semi-automatic enrichment of ontologies with complex axioms.
3 Matching Ontologies and Ontology Patterns
The proposed method is based on ontology matching. Matching of ontologies has
been widely covered in literature. An overview of the existing approaches can be
found in [12]. We use a modified ontology matching technique due to the specific
requirements for matching ontology patterns with ontologies. The main particularity
of pattern matching is the high average level of abstractness characteristical for the
concepts of a pattern. The concepts contained in a pattern are usually abstract enough
to match many different concepts in an ontology. Therefore, relations are often the
major indicators for a pattern instantiation. Especially lexical information about
relations is essential for a better performance of the matching algorithm. Thus, we
consider it useful to invest additional effort to a-priori enrich patterns with lexical
information. Our algorithm is designed to exploit provided additional lexical
information.
Before presenting the algorithm, we state the underlying criteria for a high
likelihood of pattern realization by an ontology part. Thereby, we reduce the problem
of identifying partial instantiations of a pattern to the problem of identifying complete
pattern instantiations. We rely on the following set of criteria:
31
A part O of an ontoology ONTOLOGY is a potential instantiation of the considered
c
ontology pattern P, if its structure can be matched completely with the structure of P
in a way that
1. Each concept CP contained in P has exactly one corresponding conccept CO in
O, which is equuivalent to CP or a subconcept of it;3
2. Each relation RP contained in P has exactly one corresponding relaation RO in
O, so that doomain and range concepts of RO are the correespondents
(according to 1) of the domain and range concepts of RP and RO impllies RP;
3. Each axiom AP of P can be deduced from ONTOLOGY when its conncepts and
relations are reeplaced by their correspondents according to 1 and 2.
Fig. 1. Matching of an onntology part and a pattern
Figure 1 shows how w the content ontology pattern “Participation” is matchhed with a
part of an example onttology according to the stated criteria. The concept “S Student” is
a specific “Object”, thet concept “Lecture” is a specific “Event”, and thus they
correspond as requiredd by condition 1. Relations “hasParticipant” and “hasA Attendee”
as well as “participateesIn” and “attends” are expressed by synonymous exxpressions
and their domain and range
r concepts correspond to each other (condition 1). For this
reason they imply eachh other and correspond to each other as required by coondition 2.
If the pattern also coontains an axiom declaring “participatesIn” to be thhe inverse
relation of “hasParticiipant”, then according to condition 3 it must be possible to
deduce it from the set of ontology axioms. In this case, an axiom declaring “attends”
to be inverse to “hasAtttendee” would suffice.
The criteria stated above provide the basis for our matching algorithm m. It uses
lexical properties of reelations and concepts to verify whether concepts andd relations
correspond to each other
o as required in 1 and 2. We will briefly desscribe the
matching of ontology elements
e based on lexical properties in the following subsection
s
before discussing the algorithm
a in more detail.
3 Equivalence, subconcep
pt and implication relationships are correspondences establisshed by the
ontology matching metthod introduced later on.
32
3.1 Matching Lexical Properties
The goal of the lexical matching is to determine whether a concept is equivalent to
or a subconcept of a particular concept and a relation is implied by a particular
relation. For this purpose, we use the lexical information contained in ontologies and
rely on the availability of particular lexical information in patterns. In the following,
we describe these kinds of information.
Lexical information contained in ontologies differs in its detailedness and purpose.
We distinguish between a label and a linguistic pattern (LP). A label of an element is
a string used as a name for a concept or a relation whereas a LP is used to recognize
the instances of a concept or a relation in text. LPs can range from simple regular
expressions to more complex structures enriched with different kinds of linguistic
information such as concept’s part-of-speech type. Even though LPs would be very
useful for matching due to their potential richness, the representation of LPs is not
standardized in widely used ontology representation languages such as OWL.
Therefore, we do not consider LPs in our approach and use only the labels of
elements.
For the verification of conditions 1 and 2 using labels, we rely on a list of
synonyms and hyponyms for each pattern concept and a list of synonyms and
troponyms4 for each pattern relation. We match each synonym and hyponym or
troponym with each label of the potentially corresponding pattern element in the
ontology based on string-similarity.
3.2 Matching Algorithm
The matching algorithm in its simplified form can be stated as shown in Fig. 2 and
Fig. 3. The algorithm receives an ontology and an ontology pattern as input and
generates a list of pattern instantiations as output. It first identifies pairs of lexically
matching ontology and pattern elements. Then, to avoid unnecessary computations, it
selects the pattern element, which has the fewest lexical matches in the ontology.
Since the pattern can only be matched as long as all of its elements have a
corresponding element in the ontology, considering only the occurrences of the
pattern element with the fewest number of correspondents assures that the least
number of ontology parts is analyzed. Each occurrence of the selected element is then
analyzed using the recursive procedure growAlignments starting with the given pair
of matched elements.
Due to possible hyponymy or troponymy between the elements of the pattern and
the ontology, several valid alignments are possible. For a particular initial partial
alignment the outcome can differ depending on the order in which elements are
matched. For this reason, the algorithm tries all possible ways to construct an
alignment and gathers all valid alignments.
4 Troponyms are expressions for more restrictive relations
33
Fig. 2. Alignment algorithm
Fig. 3. Recursive procedure growAlignments
It marks the already matched pattern and ontology elements green and the
remaining elements red. For each green pattern element A it calculates the remaining
red neighbors and matches each of them with the remaining red neighbors of the
ontology element corresponding to A. If the lexical matching was successful for a pair
34
of elements, they are included into the current alignment which forms the input for
another run of the described procedure. The resulting alignments are gathered in a set.
After collecting all valid alignments for the currently analyzed pattern occurrence,
axiomatic matching is applied to each alignment to verify that axioms of the ontology
pattern can be deduced from the axioms of the ontology, if concepts and relations in
the pattern axioms are replaced by the concepts and relations of the ontology. For this
task, a state-of-the-art reasoner such as Pellet5 or HermiT6 can be used.
4 Ontology Refinement Based on Partial Pattern Instantiations
The algorithm presented above can be used to find partial pattern instantiations by
separating pattern elements into obligatory and optional elements and applying the
algorithm to the set of obligatory pattern elements. Assuming the availability of a set
of richly axiomatized patterns containing additional sets of synonyms and hyponyms
for each concept and relation, the ontology engineer can compose a list of patterns for
the ontology refinement by choosing the whole set at once or selecting some patterns
manually if he or she only needs a particular type of patterns.
For each pattern in this list, the ontology engineer can select the obligatory
elements and the level of accuracy, which is the acceptable extent of pattern
incompleteness, expressed as the number of pattern elements relative to the total
number of pattern elements. The level of accuracy can be set for each pattern or for
the whole refinement process. It allows to limit the required user interaction and at the
same time to influence the matching performance towards a higher recall or a higher
precision. The ontology engineer can also use the default settings. Per default, the
level of accuracy is greater zero, which allows considering potential pattern matches
containing at least one pattern element. The default obligatory elements are the
concepts and relations of each pattern. Axioms are however optional.
After the setup, the algorithm is run for each of the selected patterns. Found
alignments are checked for axiomatic incompatibility with the optional pattern
elements in order to avoid refinement suggestions which result in an inconsistent
ontology. Finally, for each pattern, a list of refinement suggestions is generated and
presented to the ontology engineer, who can select some suggestions for the
integration into the ontology and start the automatic integration process.
During the integration, partial alignments are used to integrate the unmatched
pattern elements into the ontology. Thereby, matched elements themselves are not
integrated, but are replaced in axioms by their correspondents before integrating the
axioms and unmatched pattern elements into the ontology (Table 1). Concepts and
relations missing in the ontology can be optionally renamed by an expert in order to
obtain less general names and in this way to better suit the level of abstraction present
in the ontology.
5 http://clarkparsia.com/pellet/
6 www.hermit-reasoner.com/
35
Table 1. Integration of pattern elements into an ontology.
Type of unmatched Action before inserting the element into the ontology
pattern element
Concept Optional renaming by an expert
Relation Replacement of all matched pattern concepts contained in domain
and range axioms by their correspondents, optional renaming by
an expert
Axiom Replacement of all matched concepts and relations by their
correspondents
5 Feasibility Study
We conducted an experiment on the ontologies contained in the Watson Ontology
repository7 in order to assess the potential of the proposed method. In the experiment,
we used the previously described example consisting of the transitive relations
“before” and “after” to examine how well the proposed method can perform for
axioms involving transitivity and inverseness of relations.
In the experiment, we used only the relation label itself for the lexical matching.
The relations were considered as obligatory pattern elements whereas the axioms
about their transitivity and inverseness were considered to be optional.
We used the Watson Search Engine [13] to identify the ontologies containing an
ObjectProperty definition for at least one of the relations. Thereby, 14 documents
were identified and matched against the pattern with results as displayed in Table 2.
Table 2. Experiment results: matching of the after-before-pattern with ontologies indexed by
the Watson Ontology Search Engine.
Ontology URL Result
morpheus.cs.umbc.edu/aks1/ontosem.owl Inverse only
lists.w3.org/Archives/Public/www-rdf-logic/2003Apr/att- “After” is missing
0009/SUMO.daml
secse.atosorigin.es:10000/ontologies/SUMO.owl “After” is missing
daml.umbc.edu/ontologies/cobra/0.3/daml-time Inverse only
ai.sri.com/daml/ontologies/time/Time.daml Inverse only
cs.umd.edu/~golbeck/daml/slaveOnt.daml No transitivity and
no inverseness
cs.vu.nl/~pmika/owl-s/time-entry-fixed.owl Complete
isi.edu/~pan/damltime/time-entry.owl Complete
pervasive.semanticweb.org/ont/2004/06/time Complete
pervasive.semanticweb.org/ont/dev/time Complete
isi.edu/~pan/damltime/time.owl Complete
mogatu.umbc.edu/ont/2004/01/Time.owl Complete
sweet.jpl.nasa.gov/sweet/time.owl Complete
daml.umbc.edu/ontologies/cobra/0.4/time-basic Complete
7 http://watson.kmi.open.ac.uk/WatsonWUI/
36
Eight of 14 documents resulted in a complete match of the pattern including all
axioms. Three of the ontologies did not include the inverseness axiom, but the
transitivity axioms. Two documents did not contain a definition for the relation
“after”, but a definition for the relation “before” which was defined as transitive. One
document did not contain any of the mentioned axioms. We manually examined the
refined ontologies and found that the performed completions were semantically
justified.
6 Summary and Outlook
In this paper, we presented an algorithm for the identification of ontology pattern
instantiations in ontologies along with a method to transfer complex axioms contained
in ontology design patterns into a target ontology. The results of our experiment
demonstrate the potential of the reuse of formalized knowledge. However, in order to
assess the impact of the method more precisely, we plan a large-scale evaluation
involving a large set of pattern with different characteristics.
The availability of appropriate and complete ontology patterns is essential for the
effectiveness of our approach. Hence, we are currently working on semi-automatic
methods to acquire useful patterns as well as the necessary lexical information for
each pattern. For the former, we are planning to exploit existing ontologies to identify
frequently co-occurring characteristics of ontology elements and in this way to
identify particularly useful ontology patterns for ontology refinement. For the latter,
we expect existing broad-coverage data sets such as WordNet, BillionTriple-
Challenge8 and DBPedia to be valuable resources. We also intend address the
acquisition of composed relation labels such as followed_by or authorOf, since they
are typical in the existing ontologies and difficult to obtain from the usual grossaries.
For this purpose, we intend to use the existing methods for the extraction of synonyms
and hyponyms based on Harris’ Distributional Hypothesis [14] such as [15].
Since the effectiveness of our approach is highly dependent on the quality of the
lexical matching, we are currently working on the incorporation of disambiguation
techniques as well as matching techniques based on LPs in our lexical matching
approach.
Acknowledgments. This work is supported by the EU FP6 NeOn project
http://www.neon-project.org.
References
1. Gangemi, A.: Ontology design patterns for semantic web content. In: International Semantic
Web Conference, 262–276 (2005)
8 http://vmlion25.deri.ie/
37
2. Presutti, V., Gangemi, A.: Content Ontology Design Patterns as Practical Building Blocks
for Web Ontologies, In: Proc. of the 27th Int. Conf. on Conceptual Modeling, (2008)
3. Blomqvist. Blomqvist, E.: Semi-automatic Ontology Construction based on Patterns. PhD-
Thesis. Linköping University, Department of Computer and Information Science (2009)
4. Presutti, et al.: A Library of Ontology Design Patterns: Reusable Solutions for Collaborative
Design of Networked Ontologies. NeOn D2.5.1 (2008)
5. Buitelaar, P., Cimiano, P.: Ontology Learning and Population: Bridging the Gap between
Text and Knowledge, volume 167 of Frontiers in Artificial Intelligence and Applications.
IOS Press (2008)
6. R. Navigli and P. Velardi.: Ontology enrichment through automatic semantic annotation of
on-line glossaries. In Proc. of the 15th Int. Conf. on Knowledge Engineering and
Knowledge Management (EKAW), LNCS, vol. 4248, pp. 126–140. Springer, Heidelberg
(2006)
7. Völker, J., Hitzler, P., Cimiano, P.: Acquisition of OWL DL axioms from lexical resources.
In: Proc. of the 4th European Semantic Web Conference (ESWC’07) (2007)
8. Haase, P., Völker, J.: Ontology learning and reasoning - dealing with uncertainty and
inconsistency. In da Costa, P.C.G., Laskey, K.B., Laskey, K.J., Pool, M., eds.: Proc. of the
Workshop on Uncertainty Reasoning for the Semantic Web (URSW). 45–55 (2005)
9. Völker, J., Rudolph, S.: Lexico-logical acquisition of OWL DL axioms – An integrated
approach to ontology refinement. In: Proc. of the 6th Int. Conf. on Formal Concept Analysis
(ICFCA'08) (2008)
10.Völker, J., Rudolph, S.: Fostering web intelligence by semi-automatic owl ontology
refinement. In Proceedings of the 7th International Conference on Web Intelligence (WI)
(2008)
11.Baader, F., Ganter, B., Sertkaya, B., Sattler, U.: Completing description logic knowledge
bases using formal concept analysis. In Veloso, M.M., ed.: IJCAI. 230–235 (2007)
12. Shvaiko, P., Euzenat, J.: A survey of schema-based matching approaches. Journal on Data
Semantics, IV. 146–171 (2005)
13.d’Aquin, M., Baldassarre, C., Gridinoc, L., Sabou, M., Angeletou, S., Motta., E.: Watson:
Supporting next generation semantic web applications. In Proc. of the WWW/Internet
conference (2007)
14.Harris, Z.S.: Word. Distributional Structure 10. 146–162 (1954)
15.Suchanek, F. M., Sozio, M., Weikum, G.: SOFIE: a self-organizing framework for
information extraction. In Proc. of the 18th Int. Conf. on World Wide Web (WWW '09)
(2009)
38