Databases Reference
In-Depth Information
a
Input
schemas
S1
Matcher
Execution
(sub-workflow)
Pre-
processing
Combination of
matcher results
Selection of
correspondences
Result
Mapping
S2
General workflow for pairwise schema matching
b
Matcher 1
Matcher 2
Matcher 3
Matcher
Matcher
Matcher
…
Matcher 1
Matcher 2
Matcher 1
Sequential matchers
Parallel (independent)
matchers
Mixed strategy
Principal matcher sub-workflows
Fig. 1.1
General match workflows
of which determines a preliminary set of correspondences. After the execution of the
matcher subworkflow, there are typically several postprocessing steps, particularly
the combination of the individual matcher results and finally the selection of the
correspondences from the combined result.
As indicated in Fig.
1.1
b, the individual matchers may either be executed sequen-
tially, independently (in parallel), or in some mixed fashion. In the sequential
approach, the matchers are not executed independently, but the results of initial
matchers are used as input by subsequent matchers. A common strategy, e.g., used
in Cupid (
Madhavan et al. 2001
), is to first execute a linguistic matcher to compare
the names of schema elements and then use the obtained similarities as input for
structure-based matching. In the parallel matcher strategy, individual matchers are
autonomous and can be independently executed from other matchers. This supports
a high flexibility to select matchers for execution and combination. Furthermore,
these matchers may also physically be executed in parallel, e.g., on multicore or
multiserver hardware. On the other hand, the autonomy of individual matchers may
introduce redundant computations, e.g., of name similarities to be used for structural
matching. The mixed strategy combines sequential and parallel matcher execution
and is thus most complex.
There are different methods to combine match results of individual matchers,
e.g., by performing a union or intersection of the correspondences or by aggregating
individual similarity values, e.g., by calculating a weighted average of the individual
similarities. Similarly, there are different methods to finally select the correspon-
dences. Typically, correspondences need to exceed some predetermined threshold
but may also have to meet additional constraints for improved precision. So it is
reasonable for 1:1 mappings to enforce the so-called stable marriages, i.e., a corre-
spondence
1
0
is the most similar element for
c1
and
vice versa. Some ontology matching systems such as ASMOV enforce additional
constraints regarding is-a relationships (see Sect.
4.2
).
1
0
is only accepted if
c
1
c
c
