=Paper=
{{Paper
|id=Vol-1766/om2016_Tpaper2
|storemode=property
|title=Analysing top-level and domain ontology alignments from matching systems
|pdfUrl=https://ceur-ws.org/Vol-1766/om2016_Tpaper2.pdf
|volume=Vol-1766
|authors=Daniela Schmidt,Cássia Trojahn,Renata Vieira
|dblpUrl=https://dblp.org/rec/conf/semweb/SchmidtSV16
}}
==Analysing top-level and domain ontology alignments from matching systems==
https://ceur-ws.org/Vol-1766/om2016_Tpaper2.pdf
Analysing Top-level and Domain Ontology Alignments
from Matching Systems
Daniela Schmidt∗ , Cassia Trojahn† , Renata Vieira∗
∗ Pontifical Catholic University of Rio Grande do Sul (Brazil)
daniela.schmidt@acad.pucrs.br, renata.vieira@pucrs.br
† Université de Toulouse 2 & IRIT (France)
cassia.trojahn@irit.fr
Abstract. Top-level ontologies play an important role in the construction and
integration of domain ontologies, providing a well-founded reference model that
can be shared across knowledge domains. While most efforts in ontology match-
ing have been particularly dedicated to domain ontologies, the problem of match-
ing domain and top-level ontologies has been addressed to a lesser extent. This
is a challenging task, specially due to the different levels of abstraction of these
ontologies. In this paper, we present a comprehensive analysis of the alignments
between one domain ontology from the OAEI Conference track and three well
known top-level ontologies (DOLCE, GFO and SUMO), as generated by a set
of matching tools. A discussion of the problem is presented on the basis of the
alignments generated by the tools, compared to the analysis of three evaluators.
This study provides insights for improving matching tools to better deal with this
particular task.
1 Introduction
Guarino [5] classifies ontologies according to their “level of generality”: (i) top-level
ontologies describe very general concepts (e.g., space, time, object, etc.), which are in-
dependent of a particular problem or domain. These ontologies, also named upper or
foundational ontologies [16], are usually equipped with a rich axiomatic layer; (ii) do-
main ontologies and task ontologies that describe, respectively, the entities and other
information related to a generic domain (e.g., biology or aeronautic), or a generic task
or activity (e.g., diagnosis) by specializing the concepts represented in top-level ontolo-
gies; and finally (iii) application ontologies, which describe the roles played by domain
entities when performing an activity (which are, respectively, described by domain and
activity ontologies). While the rich semantics and formalization of top-level ontolo-
gies are important requirements for ontology design [11], they act as well as semantic
bridges supporting very broad semantic interoperability between ontologies [9,10]. In
that sense, they play as well a key role in ontology matching.
However, most efforts in ontology matching have been particularly dedicated to
domain ontologies and the problem of matching domain and top-level ontologies has
been addressed to a lesser extent. This problem poses different challenges in the field, in
particular due to the different levels of abstraction of these ontologies. This is a complex
task, even manually, that requires to deeply identify the semantic context of concepts.
It involves going beyond the frontiers of the knowledge encoded in the ontologies and,
in particular, the identification of subsumption relations. The latter is largely neglect by
most matchers. In fact, when having different levels of abstraction it might be the case
that the matching process is rather capable of identify subsumption correspondences
than equivalence, since the top ontology has concepts at a higher level. Approaches
dealing with this task are mostly based on manual matching [1,12].
This paper tackles the problem of matching domain and top-level ontologies in a
different way. We aim at evaluating how a set of available matching tools, applying
different matching strategies, performs in this task. Even though they were not exactly
developed for that purpose, their output might help us to investigate the problem. We
chose three well-known top-level ontologies (DOLCE, GFO, and SUMO) and one do-
main ontology from the OAEI Conference data set. Nine matching tools have been used
in our experiments. Qualitative and quantitative analyses are based on the point of view
of three evaluators at each generated alignment. The aim is to provide an analysis of the
alignments provided by the tools for the task of aligning ontologies with different levels
of abstraction as well as to discuss our insights on the topic and to provide directions
for future improvements.
The rest of the paper is organised as follows. §2 introduces top-level ontologies and
discusses related work. §3 presents the material and methods used in the experiments,
the results and discussion. Finally, §4 concludes the paper and presents future work.
2 Background
2.1 Top-level ontologies
A top-level ontology is a high-level and domain independent ontology. The concepts
expressed are intended to be basic and universal to ensure generality and expressive-
ness for a wide range of domains. It is often characterized as representing common
sense concepts and is limited to concepts which are meta, generic, abstract and philo-
sophical. There are two approaches for the use of top-level ontologies [16], top-down
and bottom-up. The top-down approach uses the ontology as a foundation for deriving
concepts in the domain ontology. In this way, we take the advantage of the knowledge
and experience already expressed in the top-level ontology. In a bottom-up approach,
one usually matches a new or existing domain ontology to the top-level ontology. This
approach represents more challenges since inconsistencies may exist between domain
and top-level ontologies [16]. This paper focuses on the latter approach.
Several top-level ontologies have been proposed in the literature. The reader can
refer to [9] for a review of them. Here, we briefly introduce some well-known and
largely used top-level ontologies which are used further in our evaluation:
– DOLCE [4]: Descriptive Ontology for Linguistic and Cognitive Engineering has been
proposed by Nicola Guarino and his team at LOA (Laboratory for Applied Ontology).
DOLCE is the first module of the WonderWeb Foundational Ontologies Library. The
focus of the DOLCE is to grasp the underlying categories of human cognitive tasks
and the socio-cultural environment. It is an ontology of particulars and include con-
cepts such as abstract quality, abstract region, physical object, process, and so on.
– GFO [6]: General Formal Ontology is a top-level ontology for conceptual modeling
that has been proposed by the Onto-Med Research Group. It includes elaborations
of categories such as objects, processes, time and space, properties, relations, roles,
functions, facts, and situations. The work is in progress on the integration with the
notion of levels of reality in order to more appropriately capture entities in the mate-
rial, mental, and social areas.
– SUMO [13]: Suggested Upper Merged Ontology is an upper level ontology that
has been proposed as a starter document for The Standard Upper Ontology Work-
ing Group, an IEEE working group of collaborators from the fields of engineering,
philosophy, and information science. The SUMO provides definitions for general-
purpose terms and acts as a foundation for more specific domain ontologies. It is
being used for research and applications in search, linguistics and reasoning.
2.2 Related work
In the literature, we see the growing importance of aligning domain and top-level on-
tology. Recently, in [14], correspondences between DBPedia ontology and DOLCE-
Zero [3], a module of DOLCE, are used to identify inconsistent statements in DBPedia.
The authors focus on finding systematic errors or anti-patterns in DBPedia. For this
task, they exploit previously established alignments between the DBpedia ontology and
DOLCE-Zero. They argued that by aligning these ontologies and by combining reason-
ing and clustering of the reasoning results, errors affecting statements can be identified
at a minimal human workload.
In several proposals, alignments between top and domain ontologies are manually
generated. In [12], the authors align a domain ontology describing web services (OWL-
S) with DOLCE, in order to overcome conceptual ambiguity, poor axiomatization, loose
design and narrow scope of the domain ontology. They developed a core ontology of
services to serve as middle level between the foundational and domain ontologies and
used a module for DOLCE called Descriptions and Situations (D&S) previously devel-
oped. The alignment process has been manually done and combined both bottom-up
and top-down approaches. First, they used DOLCE as foundational ontology and ex-
tended it with the D&S module. This basis has been then used for developing the core
ontology of services. Next, they manually aligned OWL-S to the core ontology.
In [1], two domain ontologies of GeoScience (GeoSciML and SWEET) were manu-
ally aligned with DOLCE Lite. The authors discussed about the matter of aligning foun-
dational ontologies with these domain ones as a basis for integrating knowledge in this
specific domain. The aim is to produce a unified ontology in which both GeoSciML and
SWEET are aligned to DOLCE. The alignment process was done in two steps. First,
each domain ontology was individually aligned with DOLCE. Then, both ontologies
were manually aligned to each other.
In [17], a manually generated alignment between a upper and a biomedical ontology
is used for filtering out correspondences at domain level that relate two different kinds
of ontology entities. The matching approach is based on a set of similarity measures and
the use of top-level ontologies as a parameter for better understanding the conceptual
nature of terms within the similarity calculation step. That allows for reducing the pos-
sibility of associations between terms derived from different categories. A set of initial
experiments showed an improvement on the alignment quality when using this kind of
approach. Evaluation of the generated correspondences has been manually done.
A closer approach to ours has been presented in [7,8], where a repository of ontolo-
gies called ROMULUS aims at improving semantic interoperability between founda-
tional ontologies. In order to provide the alignments available in ROMULUS, the au-
thors aligned three foundational ontologies (DOLCE, BFO and GFO) with each other
in a semi-automatic way. The alignment process used seven available matching tools
(H-Match, PROMPT, LogMap, YAM++, HotMatch, Hertuda, Optima). The resulting
manual alignment consists of 35 manual correspondences between DOLCE and GFO,
17 between DOLCE and BFO, and 23 between BFO and GFO. It has been used as a
gold standard for comparison with the output of the tools. However, here we focus on
the alignment of top and domain ontologies.
Analysing the impact of using top ontologies as semantic bridges (as in [17]) has
been done in [10]. A set of algorithms exploiting such semantic bridges are applied and
the authors studied under which circumstances upper ontologies improves traditional
matching approaches that do no exploit them. They developed different algorithms :
one that does not look at the ontology structure; one that looks at the identity and struc-
tural information of concepts to decide when the concepts are related; and another one
aggregating the structural algorithm with another that does not use upper ontologies.
The experiments involved 17 ontologies and 3 top-level ontologies (SUMO, Cyc and
DOLCE) used as bridges for matching domain ontologies. 10 tests cases were designed
and for each, a reference alignment was manually created including only concepts.
These works use top-level ontologies as a resource for producing better domain
ontologies and alignments. Some have used alignments with top-level ontologies that
were manually made and others apply automatic approaches for matching ontologies
of same level or for analysing the impact of using top ontologies as semantic bridges in
the matching process. In fact, the best part of efforts in ontology matching research are
targeted to align same domain ontologies while matching domain ontologies with top-
level ontologies poses different challenges. This paper tackles the problem in a different
way and analyse the behaviour of available matching tools when aligning domain with
top-level ontologies. The analysis is more qualitative than quantitative, therefore it is
based on a reduced data set, one domain ontology against three of the most well known
top-level ontologies available. The experiments are described in the next section.
3 Experiments
3.1 Data set and matchers
OAEI Conference data set. The OAEI Conference data set1 contains 16 ontologies
covering the domain of conference organization. A subset of 21 reference alignments
involving 7 ontologies (Ekaw, Conference, Sigkdd, Iasted, ConfOf, Cmt and Edas) has
been published. We have chosen this data set because it provides expressive ontologies
and is one of the most popular data set in the ontology matching evaluation community
[2]. In the experiments presented below, we have used one ontology (the Conference
1 http://oaei.ontologymatching.org
ontology2 ). This ontology has 60 concepts, 46 object properties and 18 data properties.
Here, we focus on the alignment of concepts.
Top-level ontologies. The top-level ontologies DOLCE Lite, GFO Basic, and SUMO-
OWL were aligned with the Conference ontology:
– DOLCE Lite3 : the lite version is freely available and it is composed by 37 Concepts
and 70 Object properties.
– GFO Basic4 : the basic version is freely available and it is composed by 45 Concepts
and 41 Object properties.
– SUMO5 : the OWL version is freely available and composed by about 4.500 Concepts
and 778 Object properties.
Ontology matching tools. A set of tools, publicly available, from previous OAEI cam-
paigns (not limited to Conference track top participants), and implementing different
matching strategies was selected. Even though they are not exhaustive and were not ex-
actly developed for that purpose, their output might help us to investigate the problem of
aligning domain and top-level ontologies. Aroma6 is a hybrid tool based on association
rules; Falcon-AO7 applies linguistic and structural approaches, as Lily8 , which includes
debugging strategies; LogMap9 applies logical reasoning and repair strategies and its
variant LogMap-Lite is essentially based on string similarities; MaasMatch adopts a
similarity cube and a disambiguation phase as described in [15]; WeSeE-Match10 uses
web search results for improving similarity measures; WikiMatch11 uses Wikipedia as
external knowledge source and YAM++12 applies both linguistic and graph-based ap-
proaches together with machine learning. MaasMatch and YAM++ use WordNet as
background knowledge. All the tools were run with their default configuration settings.
3.2 Results and discussion
Manual evaluation. For our experiments, we ran each of the above mentioned systems
for the pairs composed by the Conference ontology against each top-level ontology. We
then merge the alignments generated by the matchers, resulting in 28 correspondences
(Table 1), and submitted the resulting merge to the analysis of three evaluators. The
evaluators are researchers that have common-sense knowledge about conferences (the
domain ontology), with a strong background in Computer Science and well-familiarised
2 http://oaei.ontologymatching.org/2015/conference/data/Conference.owl
3 http://www.loa.istc.cnr.it/old/DOLCE.html
4 http://onto.eva.mpg.de/gfo-bio/gfo-bio.owl
5 http://www.adampease.org/OP/SUMO.owl
6 https://exmo.inrialpes.fr/software/aroma/
7 http://ws.nju.edu.cn/falcon-ao/
8 http://cse.seu.edu.cn/people/pwang/lily.htm
9 https://www.cs.ox.ac.uk/isg/tools/LogMap/
10 http://www.ke.tu-darmstadt.de/resources/ontology-matching/wesee-match
11 http://www.ke.tu-darmstadt.de/resources/ontology-matching/wikimatch
12 http://www.lirmm.fr/yam-plus-plus/
with ontology matching. Each of the 28 correspondences (pairs of concepts) were pre-
sented to the evaluators, separately, via an online evaluation form (Figure 1). In this
form, the first concept in the pair denotes the domain concept and the second one
denotes the top concept. For the top concepts, a description (as provided by the top-
level ontology) is presented in the form. Checking the ontologies could be done outside
the evaluation form. Figure 1 shows one example for the pair ‘Abstract’ - ‘Abstract
(DOLCE)’ presented to the evaluators. The evaluators analysed each correspondence
and selected one type of relation – Equivalent, Sub/Super concept, or None – according
to the relation they judged as correct.
Fig. 1. Example of correspondence as shown in the online evaluation form.
A summary of the correspondences generated by the matchers together with the
results of the manual annotation is presented in Table 1. In this table, the first col-
umn presents the concepts of the domain ontology for which one correspondence was
found by at least one matcher. The second column shows the top-level concept that was
aligned with the corresponding domain concept. The concept hierarchy is included for
all concepts. The third column identifies the top-level ontology involved in the align-
ment. The fourth, fifth, and sixth columns are used to show the evaluators judgment
about the pair of concepts. The numbers indicate how many evaluators voted for each
type of correspondence. Finally, the last column summarizes how many tools aligned
the corresponding pairs of concepts.
Regarding the evaluators judgement, there was total agreement among them in 20
(out of 28 correspondences). However, for 14 of them, no relation has been identified
so that half of the automatically aligned concepts were considered neither equivalent
nor subsumed. In 3 cases there was total agreement regarding “Subsumption”, and in 3
cases total agreement for “Equivalence”. From the 8 pairs resulting in a disagreement,
only 2 of those corresponded to a full disagreement. These 2 cases of total disagreement
were discussed among the evaluators, and in one case a total agreement for subsumption
was reconsidered. For the other case, a partial agreement for ‘None’ (no relation) was
achieved. The results in Table 1 correspond to the final agreement.
We note that, regarding the 28 correspondences, only 18 concepts of a total of 60
from the domain ontology participated in a correspondence.
Tools alignment evaluation. The evaluation of the alignments generated by the tools
is based on their precision with respect to the manual analysis.We consider 4 sets of
alignments:
– P1 considers the cases of total agreement, where a correspondence is considered as
correct if it has been marked either as equivalent or subsumed by the evaluators (21
correspondences regardless the type of relation – equivalence, subsumption or none
– where 7 of them correspond to either equivalence or subsumption);
– P2 considers the cases involving both total and partial agreements (28 correspon-
dences regardless the type of relation with 14 corresponding to either equivalence or
subsumption);
– P3 considers only total agreement for equivalences (matchers have generated only
equivalences) (21 correspondences with 3 equivalences);
– P4 considers both total and partial agreements only for equivalences (28 correspon-
dences with 4 equivalences).
Table 2 presents the precision of each tool (average of the results for the 3 pairs
of ontologies). Here we have a total of 49 correspondences to be analysed, since more
than one matcher may indicate a correspondence for the same pair. While some tools
were able to generate alignments between Conference and the three top-level ontolo-
gies (LogMap, LogMapLite and YAM++), other systems have generated alignments for
only one pair of ontologies (Conference-DOLCE for Aroma and Conference-GFO for
Falcon-AO). Moreover, some systems were not able to generate any alignment (Lily,
WeSeE and WikiMatch) and some only generate incorrect ones (Falcon-AO).
For those systems generating non empty alignments, MaasMatch and YAM++ were
able to generate more correspondences than the other systems (with LogMap and its
variant coming just behind). These 2 systems use WordNet in their matching approaches.
This background knowledge resource is a source of lexical relations and can potentially
be exploited for finding other relations than equivalence. This can explain the fact that
these systems find more alignments. Their best results were obtained for P2 (however,
the best results for this set have been obtained by LogMap). Contrary to what would be
expected, these systems (and all others, in fact) were not able to generate subsumption
(even though some have been designed to). They generated only equivalences, even
when they were in fact subsumptions.
Looking to the different sets, in P1 , LogMap, LogMapLite and MaasMatch out-
performed YAM++. In P2 , LogMap achieves the best results followed by MaasMatch.
When only equivalence (P3 ) is considered, the numbers drop for some matchers. When
relaxing to both partial and total agreements the results drop even more (P4 ). Some
matchers are doing equivalence consistently (LogMapLite), whereas others are also
indicating correspondences which were in fact considered subsumption by the judges
(LogMap, MaasMatch, YAM++), so that P3 and P4 decrease. Moreover, precision is low
if we compare the results when the same systems are matching domain ontologies13 .
13 http://oaei.ontologymatching.org/2015/conference/eval.html
Table 1. Union of the correspondences found by the tools.
Manual
Conference Ontology Top-Level Ontology Ontologies ≡ w None Tools
Conference document/Conference particular/abstract DOLCE Lite 3 5
contribution/Written contribution/Regular Entity/abstract SUMO 3 3
contribution/Extended abstract/Abstract Individual/Abstract GFO Basic 3 5
Entity/object/artifact/
Person/Committee member/Chair SUMO 3 3
furniture/seat/chair
particular/spatio-temporal-
particular/endurant/non-
Person/Conference applicant DOLCE Lite 3 1
physical-endurant/
non-physical-object
Conference document particular DOLCE Lite 3 1
particular/spatio-temporal-
particular/endurant/physical- DOLCE Lite 1 2 1
Conference part
endurant/feature/relevant-part
particular/abstract/region DOLCE Lite 3 1
particular/spatio-temporal-
particular/perdurant/stative/ DOLCE Lite 1 2 1
Conference proceedings
process
Individual/Concrete/
Processual Structure/ GFO Basic 1 2 1
Process
Conference/Conference volume particular DOLCE Lite 2 1 1
Conference document/Conference
particular/abstract/region/
contribution/Written contribution/ DOLCE Lite 3 1
abstract-region
Regular contribution/Extended abstract
Entity/physical/object/agent/
Organization SUMO 3 2
group/organization
Entity/physical/object/agent/
SUMO 3 1
Organizer group/organization
Entity/physical/object/agent/
SUMO 3 1
organism
Conference document/Conference
Entity/physical/object/artifact/
contribution/Written contribution/ SUMO 3 3
paper
Regular contribution/Paper
Person Individual GFO Basic 1 2 1
Entity/physical/content bearing
Conference document/Conference physical/VisualContentBearing SUMO 3 3
contribution/Poster Object/PrintedSheet/Poster
Individual/Property GFO Basic 3 1
Conference document/Conference particular/abstract/proposition DOLCE Lite 2 1 1
contribution/Presentation Individual/Concrete/Processual
GFO Basic 3 1
Structure/Occurrent/Event
Entity/physical/agent/
Publisher SUMO 3 3
commercial-agent/publisher
particular/spatio-temporal-
Person/Conference applicant/ particular/endurant/physical- DOLCE Lite 3 1
Registeered applicant endurant/physical-object
Entity GFO Basic 3 1
particular/abstract/region/
DOLCE Lite 3 1
Topic temporal-region/time-interval
Category/Concept GFO Basic 1 2 1
particular/spatio-temporal-
particular/quality/physical- DOLCE Lite 3 1
Conference part/Workshop
quality/spatial-location q
Entity/physical/object/region/
geographic- area/LocalizablePlace/ SUMO 3 3
stationary artifact/workshop
Total of correspondences found by the tools: 49
Table 2. Precision of each system considering their complete set of alignments.
System P1 P2 P3 P4
Aroma 0/2 0 1/3 .33 0/2 0 0/3 0
Falcon-AO 0/1 0 0/1 0 0/1 0 0/1 0
Lily - - - -
LopMap 3/9 .33 5/11 .55 3/9 .33 3/11 .27
LogMapLite 3/9 .33 3/9 .33 3/9 .33 3/9 .33
MaasMatch 3/10 .30 5/12 .42 0/10 0 1/12 .08
WeSeE-Match - - - -
WikiMatch - - - -
YAM++ 3/11 .27 5/13 .38 2/11 .18 2/13 .15
Total 12/42 .29 19/49 .39 8/42 .19 9/49 .18
Table 3 shows the overall precision of aligned concepts for each pair of ontolo-
gies (based on the union of generated alignments). As expected, the best precision is
achieved for P2 (for the pairs involving GFO). However, if we consider only equiva-
lences (P3 and P4 ), the best precision was achieved with SUMO. We also observe that
more correspondences have been generated involving DOLCE concepts (12 pairs), but
it corresponds to the lower precision across the different sets.
Table 3. Precision of the alignment union (considering all systems).
Pair of Ontologies P1 P2 P3 P4
Conference - DOLCE Lite 1/8 .13 5/12 .42 0/8 0 0/12 0
Conference - GFO Basic 2/4 .50 5/7 .71 0/4 0 1/7 .14
Conference - SUMO 4/9 .44 4/9 .44 3/9 .33 3/9 .33
Total 7/21 .33 14/28 .50 3/21 .14 4/28 .14
Another simpler way to look at the quality of the alignments generated by the tools
is presented in Table 4. It summarizes the correspondences considered correct by at
least 1, 2 or by all 3 evaluators. The table also indicates the number of times a relation of
equivalence found by the matchers were considered equivalence or subsumption by the
evaluators. It shows that 14 out 28 correspondences made by the tools were considered
as equivalent or subsumed by at least 1 evaluator. Total agreement happened in 7 of
these cases (after discussion on the cases of total disagreement).
Table 4. Number of correct correspondences according to the evaluators analysis.
at least 1 judge at least two judges three judges
w 10 6 4
≡ 4 3 3
w+≡ 14 9 7
Discussion. Regarding the qualitative analysis of the alignments, we observe that the
systems found various correspondences between concepts with the same term (“Ab-
stract”, “Chair”, “Paper”, “Workshop”, “Organization”, “Poster” and “Publisher”). This
is quite expected as all tools are based on some string-based matching strategy. How-
ever, many of them were considered as having no correspondence by the evaluators
(“Abstract”, “Chair”, “Paper”, “Workshop”). Among these concepts, the most common
aligned one was ‘Abstract’ involving DOLCE and GFO (5 tools) and SUMO (3 tools).
Some other concepts were aligned by three or two different tools, but most concepts
were aligned just by one. The other correspondences provided by the tools which were
considered no correspondent by all the evaluators can be found in Table 1.
There were correspondences with the same term which were considered equivalent
by the evaluators :
– “Organization” in the top-level ontology is defined as: a group of people with a com-
mon purpose or function in a corporate or similar institution, the same as in the
conference domain.
– “Poster” is defined as: a printed sheet intended to be posted on a horizontal surface,
so as to make the information it displays visible to passers by.
– “Publisher” in the top-level ontology refers to: some service that includes the publi-
cation of texts, so as in the conference domain.
The 3 cases above were SUMO concepts. The concept “Organization” was aligned
by 2 systems and the others by 3. For some concepts, all evaluators considered that
there was a correspondence but selected subsumption instead of equivalence :
– Organizer and organism: The first concept refers to people who organizes confer-
ences and the second refers to a living individual, then, the concept “Organizer” was
considered as subsumed by “Organism”.
– Presentation and event: The first concept refers to the action of explaining about some
topic for a group of people. The second refers to processual structures comprising a
process. “Presentation” was considered as subsumed by “Event” by the judges.
– Registered applicant and physical-object: The first concept refers to people who ap-
ply and is able to participate in the conference. The main characteristic of the second
concept is that they are endurants with unity and most physical objects change some
of their parts while keeping their identity, they can have therefore temporary parts.
In this case, one “Registered applicant” is a person who in some specific time inter-
val assumes this role, but keeping their identity as person, then, it was considered as
subsumed by “Physical-object”.
– Registered applicant and entity: The first concept was interpreted in the same way as
above. The second concept refers to everything that exists in the broadest sense. In
this case, one “Registered applicant” is something that exists, than, the first concept
was considered as subsumed by “Entity”.
An important aspect is that finding subsumption correspondences is in fact highly
desirable when matching domain and top-level ontologies. Ideally, such a matcher
should try to find the closest super concept. However the matchers we tested in this
experiment were not able to generate subsumption, even if some of them (Aroma, for
instance) are supposed to do so. They generated only equivalences, even when they
were in fact subsumptions. This is however an important distinction. Finally, our anal-
ysis does not take into account the inconsistencies introduced in the merging align-
ments from all tools. In fact, it is contradictory that Conference applicant aligns to
Non physical object and Registered applicant to Physical object, considered that the
latter is a subclass of Conference applicant in the domain ontology. This could be ex-
ploited for further filtering out correspondences. We could as well enrich the set of
manually validated correspondences by introducing simple hierarchical reasoning.
To sum up, although the number of evaluators is relatively small, it allowed us to
establish a first evaluation of available tools on the task. Our study was useful to observe
various questions in the task of matching ontologies of different levels of abstraction :
– there was a small quantity of aligned concepts by the tools in general (in total, 18 of
60 concepts), even considering all concepts provided by the top ontologies;
– there were many produced correspondences which were not considered as correspon-
dences by the specialists, many string matching cases which are usually safe in same
domain correspondences did not apply, according to our study;
– there is a lack of comprehensive evaluation data sets (regarding domain vs. top-level
ontologies) to evaluate the systems, and to overcome that we presented an analysis
of the output generated by current systems;
– knowledge on top level ontologies is highly specialized, it is important that such
evaluation considers an overview of experts in this area;
– both domain and top ontologies may lack further context or documentation that is
appropriate to help identifying the right correspondences;
– manual analysis or correspondences generation by specialists is a hard and expensive
work, in this work we ran experiments on a small set of concepts and this problem
has been reduced; bigger data sets would require more efforts;
– matching strategies for dealing with this task should take advantage of structural
features of the ontologies, background knowledge from external resources targeting
subsumption correspondences, and logical reasoning techniques for guarantee the
consistency of the generated alignments;
– at last, but not least, current tools do not distinguish between subsumption and equiv-
alence correspondences, which in this kind of task is a crucial point, finding the
closest super-concept is quite desirable when aligning to a top-level ontology.
4 Concluding remarks and future work
This paper presented an analysis of the alignments between three top-level ontologies
with one domain ontology as produced by a set of matching tools. Our goal was to anal-
yse the behaviour of these tools, which apply diverse matching techniques, with respect
to this task. We could observe that matching top-level and domain ontologies automati-
cally is an interesting and challenging task. Top-level ontologies focus on the standard-
isation of more general concepts to be easily reused in a large amount of domains. On
the other hand, there are a lot of domain ontologies available in different fields. There-
fore, we claim that it is important to reuse the well-founded knowledge available in
the top-level ontologies together with the domain ontologies to reduce the time of on-
tology modeling, the heterogeneity problem of the knowledge representation, and the
complexity of ontology modeling. Hence the automatic matching process should be an
alternative. Furthermore, top-level ontologies are semantic bridges for helping solving
the heterogeneity between domain ontologies that have to be integrated.
As future work, we plan to run experiments exploiting the whole space of possi-
ble alignments (regarding a data set) and to extend the evaluation taking into account
matching tools participating in more recent OAEI campaigns. We plan as well to in-
volve evaluators experimented in top-level ontologies and with different backgrounds
(Computer Scientists, Philosophers) in the manual evaluation process. We intend also
to exploit background knowledge from external resources (like BabelNet) in order to
improve the results reported here, paying special attention to subsumption relations.
Combining it with logical reasoning is another aim. Finally, we intend to exploit other
data sets such as the ones available on the BioPortal, which contain manually validated
alignments between biomedical ontologies and the top level ontologies GFO and BFO.
References
1. Brodaric, B., Probst, F.: DOLCE ROCKS: Integrating Geoscience Ontologies with DOLCE.
In: Semantic Scientific Knowledge Integration. pp. 3–8 (2008)
2. Cheatham, M., Hitzler, P.: Conference v2.0: An Uncertain Version of the OAEI Conference
Benchmark. In: Proc. of the 13th Intern. Semantic Web Conference. pp. 33–48 (2014)
3. Gangemi, A., Guarino, N., Masolo, C., Oltramari, A.: Sweetening WORDNET with
DOLCE. AI Magazine 24(3), 13–24 (2003)
4. Gangemi, A., Guarino, N., Masolo, C., Oltramari, A., Schneider, L.: Sweetening Ontologies
with DOLCE. In: Proc. of the 13th Intern. Conf. on Knowledge Engineering and Knowledge
Management. pp. 166–181 (2002)
5. Guarino, N.: Formal Ontology in Information Systems: Proceedings of the 1st Intern. Con-
ference. IOS Press, 1st edn. (1998)
6. Herre, H., Heller, B., Burek, P., Hoehndorf, R., Loebe, F., Michalek, H.: General Formal
Ontology (GFO): A Foundational Ontology Integrating Objects and Processes. In: Basic
Principles, Research Group Ontologies in Medicine) (2007)
7. Khan, Z., Keet, C.M.: The Foundational Ontology Library ROMULUS. In: Proc. of the 3rd
Conf. on Model and Data Engineering. pp. 200–211 (2013)
8. Khan, Z., Keet, C.M.: Toward semantic interoperability with aligned foundational ontologies
in ROMULUS. In: Proc. of the 7th Conf. on Knowledge Capture. pp. 23–26 (2013)
9. Mascardi, V., Cordı̀, V., Rosso, P.: A Comparison of Upper Ontologies. In: Proc. of the 8th
AI*IA/TABOO Joint Workshop on Agents and Industry. pp. 55–64 (2007)
10. Mascardi, V., Locoro, A., Rosso, P.: Automatic Ontology Matching via Upper Ontologies: A
Systematic Evaluation. IEEE Trans. on Knowl. and Data Eng. 22(5), 609–623 (2010)
11. Mika, P., Oberle, D., Gangemi, A., Sabou, M.: Foundations for Service Ontologies: Aligning
OWL-S to Dolce. In: Proc. of the 13th Conf. on World Wide Web. pp. 563–572 (2004)
12. Mika, P., Oberle, D., Gangemi, A., Sabou, M.: Foundations for Service Ontologies: Aligning
OWL-S to DOLCE. In: Proc. of the 13th Conf. on World Wide Web. pp. 563–572 (2004)
13. Niles, I., Pease, A.: Towards a Standard Upper Ontology. In: Proc. of the Intern. Conf. on
Formal Ontology in Information Systems. pp. 2–9 (2001)
14. Paulheim, H., Gangemi, A.: Serving DBpedia with DOLCE - More than Just Adding a
Cherry on Top. In: The Semantic Web, pp. 180–196 (2015)
15. Schadd, F.C., Roos, N.: Coupling of Wordnet Entries for Ontology Mapping Using Virtual
Documents. In: Proc. of the 7th Conf. on Ontology Matching. pp. 25–36 (2012)
16. Semy, S., Pulvermacher, M., Obrst, L.: Toward the use of an upper ontology for U.S. govern-
ment and U.S. military domains: An evaluation. Tech. rep., Sub. to Workshop on Information
Integration on the Web, in conjunction with VLDB-2004 (2004)
17. Silva, V., Campos, M., Silva, J., Cavalcanti, M.: An Approach for the Alignment of Biomed-
ical Ontologies based on Foundational Ontologies. Information and Data Management 2(3),
557–572 (2011)