Databases Reference
In-Depth Information
some recent challenges directly refer to this issue [
Shvaiko and Euzenat 2008
]. The
selection of a schema matcher may be guided by the results that it obtains using
some benchmarking tools. In addition, a few recent works have been proposed
to automatize this matcher selection. We describe each of them in the rest of this
section.
6.1
AHP
Authors of
Malgorzata et al.
[
2006
] have proposed to select a relevant and suit-
able matcher for ontology matching. They have used Analytic Hierarchical Process
(AHP) to fulfil this goal. They first define characteristics of the matching process
divided into six categories (inputs, approach, usage, output, documentation and
costs). Users then fill in a requirements questionnaire to set priorities for each
defined characteristic. Finally, AHP is applied with these priorities and it outputs
the most suitable matcher according to user requirements. This approach has two
drawbacks: (1) there is no experiment demonstrating its effectiveness and (2) cur-
rently there does not exist a listing of all characteristics for all matching tools. Thus,
the user would have to manually fill in these characteristics.
6.2
RiMOM
RiMOM [
Li et al. 2009
] is a multiple strategy dynamic ontology matching system.
Different matching strategies are applied to a specific type of ontology information.
Based on the features of the ontologies to be matched, RiMOM selects the best
strategy (or strategy combination) to apply. When loading the ontologies, the tool
also compute three feature factors. The underlining idea is that if two ontologies
share similar feature factors, then the strategies that use these factors should be
given a high weight when computing similarity values. For instance, if the
label
meaningful
factor is low, then the
Wordnet-based strategy
will not be used. Each
strategy produces a set of correspondences, and all sets are finally aggregated using a
linear interpolation method. A last strategy dealing with ontology structure is finally
performed to confirm discovered correspondences and to deduce new ones. Contrary
to other approaches, RiMOM does not rely on machine learning techniques. It is
quite similar to the AHP work by selecting an appropriate matcher based on input's
features. RiMOM participated to the 2009 OAEI campaign [
Zhang et al. 2009
].
Results show that the tool performed well in different tracks (anatomy, benchmark,
instance matching). For instance, it achieves F-measures above 75% for all datasets
in the instance matching track.
