=Paper=
{{Paper
|id=Vol-156/paper-13
|storemode=property
|title=FalconAO: Aligning Ontologies with Falcon
|pdfUrl=https://ceur-ws.org/Vol-156/paper13.pdf
|volume=Vol-156
|dblpUrl=https://dblp.org/rec/conf/kcap/JianHCQ05
}}
==FalconAO: Aligning Ontologies with Falcon==
https://ceur-ws.org/Vol-156/paper13.pdf
Falcon-AO: Aligning Ontologies with Falcon
Ningsheng Jian, Wei Hu, Gong Cheng, Yuzhong Qu
Department of Computer Science and Engineering
Southeast University
Nanjing 210096, P. R. China
{nsjian, whu, gcheng, yzqu}@seu.edu.cn
ABSTRACT As an infrastructure for semantic web applications, Fal-
Falcon-AO is an automatic tool for aligning ontologies. con is a vision of our research group. It will provide
There are two matchers integrated in Falcon-AO: one is enabling technologies for finding, aligning and learning
a matcher based on linguistic matching for ontologies, ontologies, and ultimately for capturing knowledge by
called LMO; the other is a matcher based on graph an ontology-driven approach. It is still under develop-
matching for ontologies, called GMO. In Falcon-AO, ment in our group. As a component of Falcon, Falcon-
GMO takes the alignments generated by LMO as exter- AO is an automatic tool for aligning ontologies. It is
nal input and outputs additional alignments. Reliable dedicated to aligning web ontologies expressed in OWL
alignments are gained through LMO as well as GMO DL [5]. There are two matchers integrated in current
according to the concept of reliability. The reliabil- version of Falcon-AO (version 0.3). One is a matcher
ity is obtained by observing the linguistic comparability based on linguistic matching for ontologies, called LMO,
and structural comparability of the two ontologies be- and the other one is a matcher based on graph matching
ing compared. We have performed Falcon-AO on tests for ontologies, called GMO.
provided by OAEI 2005 campaign and got some prelim-
inary results. In this paper, we describe the architec- 1.1 Linguistic Matching for Ontologies
ture and techniques of Falcon-AO in brief and present As is known, linguistic matching plays an important
our results in more details. Finally, comments about role in matching process. Generally, linguistic similar-
test cases and lessons learnt from the campaign will be ity between two entities relies on their names, labels,
presented. comments and some other descriptions.
LMO combines two different approaches to gain linguis-
Categories and Subject Descriptors tic similarities: one is based on lexical comparison; the
D.2.12 [Software]: Interoperability; I.2.6 [Artificial other is based on statistic analysis.
Intelligence]: Knowledge Representation Formalisms
and Methods; I.5.3 [Pattern Recognition]: Cluster- In lexical comparison, we calculate the edit distance [4]
ing—Similarity measures between names of two entities and use the following
function to capture the string similarity (denoted by
General Terms SS ):
Experimentation, Measurement
ed
Keywords SS = 1/e |s1.len+s2.len−ed| (1)
Semantic Web, Ontology Alignment, Mapping, Match-
ing, Similarity Measurement Where ed denotes the edit distance between s1 and s2 ;
s1.len and s2.len denote the length of the input strings
1. PRESENTATION OF THE SYSTEM s1 and s2, respectively.
In statistic analysis, we use the algorithm of VSM [6]
(Vector Space Model) in our implementation. Given a
collection of documents, we denote N the number of
unique terms in the collection. In VSM, we represent
each document as a vector in an N -dimensional space.
The components of the vector are the term weights as-
signed to that document by the term weighting function
for each of the N unique terms. Clearly, most of these
85
are going to be 0, since only a few of the N terms actu- K-Cap 2005 Workshop on Integrating Ontologies 1 .
ally appear in any given document. In our scenario, we
construct a virtual document for each of the ontology The main idea of GMO is as follows. Similarity of two
entities (classes, properties and instances). The virtual entities from two ontologies comes from the accumula-
document of an entity consists of ”bag of terms” ex- tion of similarities of involved statements (triples) tak-
tracted from the entity’s names, labels and comments ing the two entities as the same role (subject, predicate,
as well as the ones from all neighbors of this entity. The object) in the triples, while the similarity of two state-
term weighting functions are defined as follows: ments comes from the accumulation of similarities of
involved entities of the same role in the two statements
being compared.
T ermW eighting = T F ∗ IDF (2) Usually, GMO takes a set of matched entity pairs, which
t are typically found previously by other approaches, as
TF = (3)
T external mapping input in the matching process, and
1 D outputs additional matched entity pairs by comparing
IDF = ∗ (1 + log2 ) (4)
2 d the structural similarity.
In equation (3), t denotes the number of times where Our previous experiments showed that GMO were irre-
one term occurs in a given document and T denotes the placeable when there was little gain from lexical com-
maximum number of times. In equation (4), D denotes parison. In addition, GMO can be integrated with other
the number of documents in collection and d denotes matchers. While using GMO approach to align ontolo-
the number of documents where the given term occurs gies, there should be another component to evaluate
at least once. reliability of alignments generated by GMO.
We can gain the cosine similarity between documents 1.3 Linguistic vs. Structural Comparability
(denoted by DS ) by taking the vectors’ dot product: Given two ontologies to be aligned, GMO always tries
to find all the possible matched entity pairs. However,
how to evaluate the reliability of these matched entity
pairs is still a problem. As mentioned above, another
DS = N · N t (5)
component is needed to select more reliable matched
entity pairs by using other information. In Falcon-AO,
It is worthy of note that there are several preparing we use a simple approach to observe the reliability of
steps before calculating term weights, such as splitting matched entity pairs output by GMO, and select more
words, stemming and removing stop words. reliable matched entity pairs to the users. The approach
is based on the measure of linguistic comparability (LC )
The two methods described above will both take effect and structural comparability (SC ) of two ontologies to
in ontology matching. In our implementation, we com- be aligned.
bine them together, and use the following equation to
calculate the final linguistic similarity. Please note that Given two ontologies O1 , O2 to be aligned, the linguistic
the parameters in the equation comes from our experi- comparability (LC) of O1 and O2 is defined as follows:
ence:
M
LC = p (7)
LinguisticSimilarity = 0.8 ∗ DS + 0.2 ∗ SS (6) NO1 ∗ NO2
Currently, we do not use any lexicons in LMO and it is Where M denotes the number of entity pairs with sim-
certain that the use of lexicons may bring some benefits ilarity larger than c and c is an experience value; NO1
for matching. We plan to take into account using some and NO2 represent the number of named entities in O1
lexicons in later versions. and O2 , respectively.
1.2 Graph Matching for Ontologies The structural comparability is determined through com-
paring the occurrences of built-in properties used in
Another important component in Falcon-AO is GMO,
the two ontologies to be aligned. The built-in prop-
which is based on a graph matching approach for ontolo-
erties are RDF [2], RDFS [1] and OWL [5] built-in vo-
gies. It uses directed bipartite graphs to represent on-
cabularies used as properties in triples (e.g. rdf:type,
tologies and measures the structural similarity between
rdfs:subClassOf and owl:onProperty).
graphs by a new measurement. Details of the approach
1
are described in another paper [3] also presented in the http://km.aifb.uni-karlsruhe.de/ws/intont2005
86
We use VSM method to observe the structural compa- datatypes, data literals and URIs used in both ontolo-
rability. The vectors V1 , V2 represent the frequency of gies. And in Falcon-AO we take the alignments gen-
built-in properties used in O1 and O2 and the element erated by LMO as the other part of external mapping.
vij denotes the number of occurrence of built-in prop- Entities involved in the alignments generated by LMO
erty pj in Oi . The structural comparability of O1 and are set to be external entities and GMO will just output
O2 is the cosine similarity [7] of V1 and V2 : mapping between internal entities.
When the alignments generated by LMO and GMO are
V1 · V2 obtained, Falcon-AO will integrate these alignments by
SC = observing the linguistic comparability and structural
kV1 k kV2 k
Pn comparability, following the rules below:
j=1 v1j ∗ v2j
= qP qP (8)
n n
j=1 v1j ∗ v1j j=1 v2j ∗ v2j 1. We take that linguistic similarity is somewhat more
reliable than structural similarity, and that the
1.4 Implementation alignments generated by LMO are always accepted
LMO and GMO are integrated in Falcon-AO. Align- by Falcon-AO.
ments output by Falcon-AO come from the integration
of alignments generated by LMO and GMO. The archi- 2. When the linguistic comparability is high and the
tecture of Falcon-AO is shown in Figure. 1. structural comparability is low, only alignments
generated by GMO with high similarity are reli-
able and accepted by Falcon-AO.
Ontool gy
Ontool gy 3. If the linguistic comparability is low, all of the
alignments generated by GMO are accepted by
Falcon-AO. In this case, there is no enough infor-
Parser
mation to measure these alignments and we can
only assume that they are reliable.
Existng
i
Mapping LMO
Falcon-AO is implemented in Java. The implemented
External
process can be outlined as follows:
GMO Mapping
1. Input two ontologies and parse them.
Alg
i nmentn
Itegrato
in
2. Run LMO and obtain matched entity pairs.
OutputAlgnment
i s
3. Calculate linguistic comparability and structural
comparability.
Figure 1: System Architecture
4. In the case that linguistic comparability is below
a very low threshold (e.g. 0.01) and the structural
Due to heterogeneous ways in expressing semantics and comparability of them is also low, we take that
the inference capability brought from ontology languages, these ontologies are quite different and Falcon-AO
two ontologies being matched may need to be coordi- exits with no alignment.
nated by removing some redundant axioms from it or 5. Set external entities of the ontologies according to
adding some inferred axioms. So coordination actions the matched entity pairs generated by LMO.
should be taken before using GMO approach. We have
integrated several coordination rules in Falcon-AO. Our 6. Input matched entity pairs generated by LMO into
Parser component based on Jena 2 has the functionality GMO and form external mapping for GMO. In the
of coordinating ontology models. current version of Falcon-AO, all the individuals
of ontologies are specified as external entities and
As is known, given external mapping as input, GMO their similarities are computed by LMO.
can find additional mapping. The external mapping is
made of two parts: one is the existing mapping pre- 7. Run GMO and obtain matched entity pairs.
assigned by the system; the other comes from another
8. Integrate the alignments generated by LMO and
matcher. The existing mapping is the mapping be-
GMO following the rules described above.
tween built-in vocabularies of web ontology languages,
2
http://jena.sourceforge.net/ 9. Exit with alignments as output.
87
1.5 Adaptations Made for the Contest For test 201, where each of the local name of class
For anatomy test, FMA 3 ontology and OpenGALEN 4 and property is replaced by a random one, LMO can
ontology are not OWL DL. In order to make effec- still find some matched classes and properties due to
tive use of descriptions of entities, we have manually the sameness of their labels or comments. With these
found some annotation properties and inputted them matched entity pairs as feed, GMO performs well.
into LMO. With the help of these annotation proper-
ties, Falcon-AO can find about 500 more matched en- In test 202, names of classes and properties are dis-
tity pairs in addition to other 2000 matched entity pairs turbed and their comments are suppressed. LMO can
found by a simple version of LMO. only find little mapping. Meanwhile, Falcon-AO still
performs not bad by running GMO. In this test, we find
2. RESULTS that it is hard to distinguish many properties purely by
In this section we will present the results of alignment the structure of the ontology, since they have the same
experiments on OAEI 2005 campaign. All the align- domains and ranges, and never used in other part of the
ments output by Falcon-AO are based on the same pa- ontologies. Falcon-AO failed to find correct mapping of
rameters. these properties, which makes the result not so well as
test 201.
2.1 Systematic Benchmark Test In test 203, LMO is able to find all the matched en-
We divide all the benchmark tests 5 into five groups: tity pairs. Therefore, it just takes Falcon-AO several
test 101-104, test 201-210, test 221-247, test 248-266 seconds to find all alignments.
and test 301-304. We will report the results of align-
ment experiments on these five groups in correspon- For tests 204 and 208 with naming conventions, both
dence. The full results on all tests are listed in section the linguistic comparability and structural comparabil-
6.3. ity are high. The outputs of the integration of LMO
and GMO are well.
2.1.1 Test 101–104
In tests 101, 103 and 104, the source ontologies contain For the synonym tests 205 and 209, due to the fact that
classes and properties with exactly the same names as no thesaurus is used in our tool, LMO performs not
those in the reference ontologies. LMO can easily get all so well. There are some errors in the outputs of LMO.
the matched entity pairs, and GMO takes little effect. With these errors feed to GMO, GMO failed to perform
well. As a result, the outputs of Falcon-AO may be
In test 102, the linguistic comparability of the two on- weaker than the outputs of using GMO independently.
tologies is nearly zero and the structural comparability
is low as well. So it could be concluded that the two In tests 206, 207 and 210, ontologies to be aligned are
ontologies to be aligned are quite different. Falcon-AO expressed in different languages. Falcon-AO does not
exits with no alignment. have a specific matcher that uses a dictionary for word
translation. However, because of their high structural
The average performance on test 101-104 is shown be- comparability, GMO in Falcon-AO performs not bad on
low: these tests.
Precision Recall F-Measure Time The average performance on test 201-210 is described
Average 1.0 1.0 1.0 5s below:
2.1.2 Test 201–210 Precision Recall F-Measure Time
We find that each pair of ontologies of these ten tests has Average 0.96 0.95 0.95 63s
high structural comparability, which means that each
pair of the ontologies to be aligned is quite similar in
structure. Our previous experiments showed that GMO
performed well on these tests even without any addi- 2.1.3 Test 221–247
tional external mapping input. In most tests, LMO just In these tests, the linguistic comparability of each pair
finds a small part of all the matched entity pairs, the of ontologies to be aligned is very high. Most of the
rest are generated by GMO. Since GMO runs slower alignments are found by LMO and GMO takes little
than LMO, it takes Falcon-AO more time to find all effect. So, it only takes Falcon-AO a few seconds to
matched entity pairs. align them.
3
http://sig.biostr.washington.edu/projects/fm/
4
http://www.opengalen.org/ As is shown below, the average performance on these
5
http://oaei.inrialpes.fr/2005/benchmarks/ tests are perfect.
88
Precision Recall F-Measure Time 3. GENERAL COMMENTS
Average 0.99 1.0 0.99 4s In this section, we will summarize some features of Falcon-
AO and the improvement in our future work, some com-
ments about test cases will also be presented.
2.1.4 Test 248–266
These fifteen tests are the most difficult ones in all 3.1 Comments on the Results
benchmark tests, since both their linguistic compara- Our Falcon-AO performs well on benchmark tests 101-
bility and structural comparability are low. In the case 104, 201-210 and 221-247, and the results of test 301-304
that the linguistic comparability between two given on- are moderate, but on test 248-266, Falcon-AO doesn’t
tologies is very low, Falcon-AO would not call any match- perform so well. According to the results on these
ers. However, in these tests, there are still some individ- test cases, we can see the strengths and weaknesses of
uals with the same names, which increase the linguistic Falcon-AO:
comparability. So Falcon-AO will still run GMO inte-
grated with LMO. Strengths
According to the experimental results, Falcon-AO per-
Since the ontology pairs to be aligned are quite different forms well when the structures of the ontologies to be
both in linguistics and in structure, our outputs are not aligned are similar to each other or there is much lexi-
good (with average F-Measure 0.63). Indeed, in some cal similarity between the two ontologies. Particularly,
cases, it is really hard to determine the exact mapping. Falcon-AO performs well when the two ontologies have
For these tests, the time for aligning relies on the size very little lexical similarity but high structural compa-
of two ontologies. rability.
Precision Recall F-Measure Time Weaknesses
Average 0.71 0.60 0.63 60s When there is little common vocabulary between the
ontologies and in the meanwhile the structures of the
ontologies are quite different, Falcon-AO can hardly
2.1.5 Real Ontologies Test 301–304 find the exact mapping. Furthermore, GMO could not
In these tests, each pair of ontologies has high linguistic process very large ontologies, which means that while
comparability but low structural comparability. This aligning very large ontologies, Falcon-AO cannot use
indicates that the outputs of Falcon-AO mainly come their structural information.
from the outputs of LMO. Alignments with high simi-
larity generated by GMO matcher are also reliable and 3.2 Improvement of Falcon-AO
these matched entity pairs should also be output by From the experiments we have learnt some lessons and
Falcon-AO. The average performance on these four tests plan to make improvements in the later versions. The
is presented below: following three improvements should be taken into ac-
count.
Precision Recall F-Measure Time
Average 0.93 0.81 0.86 20s
1. While expressing the same thing, people may use
synonyms and even different languages. Therefore,
it is necessary to use lexicons to match ontologies.
2.2 Blind Tests
Blind tests consist of two groups: directory test 6 and 2. The current version of Falcon-AO did not support
anatomy test 7 , and they are all real world cases. many-to-many mapping. The functionality of find-
ing many-to-many mapping will be included in the
Directory later version of Falcon-AO.
We have got the alignment results on directory test by
using the same set of parameters as the ones for bench- 3. Currently, the measure of linguistic comparability
mark test. and structural comparability of ontologies are still
simple and an improvement will be considered.
Anatomy
Falcon-AO detects that the FMA ontology and Open- 3.3 Comments on the Test Cases
GALEN ontology in anatomy test are so large that our The proposed test cases covered a large portion of dis-
GMO could not process them. Therefore, our alignment crepancies occurring of ontologies while aligning ontolo-
result of anatomy test comes only from a simple version gies. Doing experiments on these test cases is help-
of LMO. ful to improving the alignment algorithm and system.
6
http://oaei.inrialpes.fr/2005/directory/ However, there are few tests on real world ontologies in
7
http://oaei.inrialpes.fr/2005/anatomy/ benchmark tests.
89
4. CONCLUSION 6.1 Link to the System and Parameters File
While aligning real ontologies, linguistic matching plays Falcon-AO can be found at
an important role in matching process. Therefore, we http://xobjects.seu.edu.cn/project/falcon/download.html.
integrate our GMO with LMO in Falcon-AO. From the
experiments, we found that, Falcon-AO performed well 6.2 Link to the Set of Provided Alignments
on most of benchmark tests. It is also worthy of note Results presented in this paper are available at
that most of benchmark tests came from artificially al- http://xobjects.seu.edu.cn/project/falcon/results/falcon.zip.
tered ontologies, and more real world ontologies are ex-
pected to be included in benchmark tests. 6.3 Matrix of Results
Runtime Environment: Tests were run on a PC run-
Acknowledgments ning Windows XP with an Intel Pentium 4 2.4 GHz
This work is supported in part by National Key Ba- processor and 512M memory.
sic Research and Development Program of China un-
der Grant 2003CB317004, and in part by the NSF of
Jiangsu Province, China, under Grant BK2003001. We
are grateful to other members in our team for their dis-
cussion and comments that helped strengthen our work.
We also would like to thank OAEI 2005 campaign for
providing test cases.
5. REFERENCES
[1] D. Brickley and R. Guha (eds). RDF Vocabulary
Description Language 1.0: RDF Schema. W3C
Recommendation, 10 February 2004. Latest
version available at
http://www.w3.org/TR/rdf-schema/
[2] P. Hayes (ed.). RDF Semantics. W3C
Recommendation 10 February 2004. Latest
version is available at
http://www.w3.org/TR/rdf-mt/
[3] W. Hu, N. Jian, Y. Qu and Y. Wang. GMO: A
Graph Matching for Ontologies. Submitted to
K-Cap workshop on Integrating Ontologies
(2005). Now is available at
http://xobjects.seu.edu.cn/project/falcon/GMO.pdf
[4] V. Levenshtein. Binary Codes Capable of
Correcting Deletions, Insertions, and Reversals.
Soviet Physics - Doklady 10 (1966) 707-710
[5] P. Patel-Schneider, P. Hayes, I. Horrocks (eds.).
OWL Web Ontology Language Semantics and
Abstract Syntax. W3C Recommendation 10
February 2004. Latest version is available at
http://www.w3.org/TR/owl-semantics/
[6] V. Raghavan and S. Wong. A Critical Analysis of
Vector Space Model for Information Retrieval.
Journal of the American Society for Information
Science 37(5) (1986) 279-287.
[7] G. Salton. Automatic Text Processing.
Addison-Wesley Longman Publishing Co., Inc.
(1989) 613-620.
6. RAW RESULTS
Information about our project can be found at
http://xobjects.seu.edu.cn/project/falcon/falcon.html, and our
tool is available now.
90
No. Precision Recall Time
101 1.0 1.0 4s
102 NaN NaN 6s
103 1.0 1.0 4s
104 1.0 1.0 4s
201 0.98 0.98 105s
202 0.87 0.87 140s
203 1.0 1.0 4s
204 1.0 1.0 22s
205 0.88 0.87 55s
206 1.0 0.99 51s
207 1.0 0.99 51s
208 1.0 1.0 34s
209 0.86 0.86 102s
210 0.97 0.96 68s
221 1.0 1.0 4s
222 1.0 1.0 4s
223 1.0 1.0 4s
224 1.0 1.0 4s
225 1.0 1.0 4s
228 1.0 1.0 3s
230 0.94 1.0 4s
231 1.0 1.0 4s
232 1.0 1.0 4s
233 1.0 1.0 3s
236 1.0 1.0 3s
237 1.0 1.0 4s
238 0.99 0.99 4s
239 0.97 1.0 3s
240 0.97 1.0 4s
241 1.0 1.0 3s
246 0.97 1.0 3s
247 0.94 0.97 3s
248 0.84 0.82 100s
249 0.86 0.86 114s
250 0.77 0.70 7s
251 0.69 0.69 166s
252 0.67 0.67 119s
253 0.86 0.85 80s
254 1.0 0.27 4s
257 0.70 0.64 4s
258 0.70 0.70 162s
259 0.68 0.68 113s
260 0.52 0.48 7s
261 0.50 0.48 8s
262 0.89 0.24 4s
265 0.48 0.45 7s
266 0.50 0.48 8s
301 0.96 0.80 18s
302 0.97 0.67 3s
303 0.80 0.82 39s
304 0.97 0.96 18s
91