=Paper=
{{Paper
|id=Vol-1114/Poster_Diallo
|storemode=property
|title=Towards Learning Based Strategy for Improving the Recall of the ServOMap Matching System
|pdfUrl=https://ceur-ws.org/Vol-1114/Poster_Diallo.pdf
|volume=Vol-1114
|dblpUrl=https://dblp.org/rec/conf/swat4ls/DialloK13
}}
==Towards Learning Based Strategy for Improving the Recall of the ServOMap Matching System==
https://ceur-ws.org/Vol-1114/Poster_Diallo.pdf
Towards Learning based Strategy for Improving the
Recall of the ServOMap Matching System
Gayo Diallo1, Amal Kammoun1
1
ERIAS, INSERM U897, Univ. of Bordeaux,
146 rue Leo Saignat, 33076 Bordeaux Cedex France
{firstname.lastname}@isped.u-bordeaux2.fr
Abstract. In order to solve interoperability issues among heterogeneous
knowledge based applications, it is important to find correspondences between
their underlying ontologies. This is the aim of the ServOMap system, a generic
approach for large scale ontologies matching. However, although achieving
good results on the official Ontology Alignment Evaluation Initiative dataset,
ServOMap performance remains to be improved in term of recall. We describe
in this paper a strategy based on Machine Learning technique for improving the
discovery of more possible candidate mappings among input ontology entities.
Keywords: ontology matching, contextual similarity, machine learning,
decision tree
1 Introduction
With the proliferation of semantically annotated data and the increase of
knowledge based applications, one of the key challenges is solving interoperability
issues which may arise due, in particular, to the heterogeneity of underlying used
ontologies. A common way is to establish correspondences between these ontologies
[1]. This is usually done with automated systems when the considered ontologies
contain large number of entities as it used to be in the life sciences domain.
The ServOMap Ontology Matching System [2][3] aims at matching ontologies at
large scale. It has been designed with the purpose of facilitating interoperability
between different applications which are based on heterogeneous knowledge
organization systems (KOS). It relies on Information Retrieval (IR) techniques for
computing similarity between entities. ServOMap was among the top systems for
large scale ontology matching during the 2012 Ontology Alignment Evaluation
Initiative (OAEI) challenge [4]. The system provided high precise mappings for most
of the tracks, reaching 99% some times. However, overall, the provided recall was a
step behind similar tools within the contest. Depending on the application domain,
there should be a balance between optimizing recall or precision. Thus, while we may
be more interested in optimizing recall in certain situations, focusing on more high
precise mappings is a requirement for other contexts. Regarding the lower provided
recall, ServOMap performance is penalized by the intensive use of the lexical
description of entities, even for the contextual based similarity strategy as described in
[5].
With respect to that, the aim of the present research work is investigating a way to
improve the overall recall provided by ServOMap without penalizing the performance
in term of precision. To do so, we rely on the highly accurate candidate mappings
generated in the first step of the ServOMap matching process, the lexical similarity
computing strategy. It consists in the use of IR based exact similarity computing by
exploiting local names and synonym terms of the concepts contained in each input
ontology. The approach of the recall improvement, based on Machine Learning (ML)
strategy, is detailed in the next section.
2 Approach
Our method is based on the fact that our previous evaluations proved that the
candidate mappings set generated by the lexical similarity computing of ServOMap is
highly precise. Let’s call this set of candidates Mexact.. Our assumption is that by
learning the behavior of the couples in Mexact, we are able to generate new possible
couples from their surrounding concepts. To do so, as depicted in figure 1, the ML
based contextual similarity computing has three inputs: the Mexact set and the two
input ontologies O1 and O2.The output is a new set of couples Mcontext.. A classifier is
built from a generated learning set based on the Mexact set.
Fig. 1. Overall process for ML based contextual similarity computing. Mexact is the set of high
accurate candidate mappings obtained using an IR based approach. Mcontext is the set of new
candidate mappings obtained by applying contextual based similarity computing thanks to the
use of the ML strategy.
Figure 2 details the process of the approach. The contextual based candidate couples
generation is assimilated to a classification task. Therefore, we need a learning set and
a test set. The classification task consists in classifying couples into a correct or
incorrect class. The first step is to compute features for the learning set. This learning
set is based on the Mexact set (considered as correct couples, labeled as “Yes”) and a
randomly generated incorrect set (labeled as “No”). We compute a set of five
similarity measures (Q-Gram, Levenstein, BlockDistance, Jaccard and Monge-Elkam)
between the concepts of each couple (by considering their local names and
synonyms). We have chosen five different similarities to cope with short and long
strings. We perform then the same process on the randomly incorrect set. The
incorrect set is constituted by couples obtained as follows. For each couple (c1, c2) in
Mexact, we compute the 5 similarity scores for (c1, ancestor(c2)), (ancestor(c1), c2),
(descendant(c1), c2) and (c1, descendant(c2)). The ancestor and descendant functions
retrieve the super-concepts and sub-concepts of a given concept.
Fig. 2. Detailed process for acquiring new candidate mappings based on the use of a very high
precise set of initial candidate mappings.
The next step is to build the classifier. We use a decision tree and the J48 algorithm
implemented within the Weka framework [6]. Then, we look up the surrounding
concepts of the couples in Mexact by comparing respectively their sub-concepts, their
super-concepts and their siblings. We keep all couples having the score si=
f(getScoreSub(), getScoreSup(), getScoreSib()) >⍬, where ⍬ is a chosen threshold
and getScoreSub(), getScoreSup(), getScoreSib() are functions computing
respectively for each possible couple (c1, c2) a score from the sub-concepts couples,
super-concepts and siblings couples.
3 Conclusion
We have briefly described in this poster paper a new strategy for generating
candidates mappings based on a ML based contextual based similarity computing.
The approach has been implemented into the new version of the ServOMap system.
Its evaluation showed an improvement on the recall achieved by the system for most
of the standard dataset provided by the OAEI challenge in its 2013 edition1. However,
the precision is slightly decreased some times. For future work, we have to investigate
further to identify the situations where the ML based contextual mapping is well
adapted.
References
1. P. Shvaiko, J. Euzenat. Ontology Matching: State of the Art and Future Challenges. IEEE
Trans. Knowl. Data Eng. 25(1): 158-176 (2013).
2. G. Diallo, M. Ba. Effective Method for Large Scale Ontology Matching. Proceedings of the
5th International Workshop on Semantic Web Applications and Tools for Life Sciences.
CEUR Workshop Proceedings, vol 952, (2012).
3. M. Ba, G. Diallo. Large scale biomedical ontology matching with ServOMap. IRBM,
Volume 34, Issue 1, pp 56-59, (2013).
4. J.-L. Aguirre, K. Eckert, J. Euzenat, A. Ferrara, W. Robert van Hage, L. Hollink, C.
Meilicke, A. Nikolov, D. Ritze, F. Scharffe, P. Shvaiko, O. Svab-Zamazal, C. Trojahn dos
Santos, E. Jimenez-Ruiz, B. Cuenca Grau, and B. Zapilko. Results of the ontology
alignment evaluation initiative 2012. In Pavel Shvaiko, Jeerome Euzenat, Anastasios
Kementsietsidis, Ming Mao, Natasha Fridman Noy, and Heiner Stuckenschmidt, editors,
OM, volume 946 of CEUR Workshop Proceedings. CEUR-WS.org, (2012).
5. M. Ba, G. Diallo. ServOMap and ServOMap-lt results for OAEI 2012. In Pavel Shvaiko,
Jeerome Euzenat, Anastasios Kementsietsidis, Ming Mao, Natasha Fridman Noy, and
Heiner Stuckenschmidt, editors, OM, volume 946 of CEUR Workshop Proceedings. CEUR-
WS.org, (2012).
6. I. Witten, E. Frank, L. Trigg, M. Hall, G. Holmes, S. Cunningham. Weka: Practical machine
learning tools and techniques with java implementations. Proceedings of the
ICONIP/ANZIIS/ANNES'99 Workshop on Emerging Knowledge Engineering and
Connectionist-Based Information Systems, (1999).
1 http://oaei.ontologymatching.org/2013/results/index.html