Schema matching is a vital and critical step in many application domains such
as semantic web, integration of web oriented data, e-commerce, schema evolution
and migration, application evolution, data warehousing, database design, XML
message mapping, XML-relational data mapping and schema integration. A schema
matching process takes two schemas as input and returns a set of matching element
pairs which are semantically related to one another between the two schemas
(Madhavan et al., 2001; Melnik
et al., 2002; Do et al., 2002; Bernstein
et al., 2004; Aumueller et al., 2005).
Schema matching is still at large a manual process and hence, is highly
time consuming, error-prone and expensive. Hence, automation of this process
aids in a faster, less labor intensive integration approach which is crucial
for large scale data integration systems. We have concentrated in this
study on matching XML schemas, due to the popularity of XML model, the
large proliferation of XML documents on-line and the emergence of XML
as a defacto standard for sharing data among sources. Though there are
several methods proposed in the literature for the schema matching process,
there is still a large space for improvement in terms of accuracy and
reduction in complexity.
The matching process in general comprises of linguistic matching phase, a structural
matching phase (Madhavan et al., 2001; Do
et al., 2002; Aumueller et al., 2005)
and in some approaches a filtering phase to extract the matching element pairs
from the two schemas (Melnik et al., 2002). The
linguistic matching process identifies linguistic similarity between the elements
that are similar in meaning. The structural matching process identifies similarity
between elements based on their structure i.e., based on the number of similar
attributes or sub elements they have.
In this study, a novel method for identifying element similarities using
a simple path matching algorithm has been proposed.
A brief survey of some of the most successful schema matching approaches
found in the literature will help in understanding the current direction
of research in the schema matching field. All these approaches in general
compute the similarity between elements of the two schemas as a value
in the interval 0 and 1, where, a value of 1 means the elements are identical
and 0 means completely dissimilar.
Cupid (Madhavan et al., 2001) proposes a hybrid
schema matching algorithm to compute the similarity between the schema elements.
The matching algorithm operates on schema trees. It first computes the linguistic
similarity between the schema elements based on morphological normalization,
categorization, string-based techniques and a thesauri look-up. Then it computes
structural similarity between the schema elements. It computes the final similarity
coefficient combining the weighted linguistic similarity measure and weighted
structural similarity measure. The chosen matching pairs are those whose similarity
coefficients are greater than a given threshold.
SF (Melnik et al., 2002) proposes a structural
algorithm to match diverse data structures like data schemas, data instances,
or a combination of both. It first converts the schemas into directed labeled
graphs. It then uses an iterative fix point computation to find the similarities
between the schema elements. The similarity computation relies on the intuition
that elements are similar if their adjacent elements are similar. It also proposes
several filters for extracting the best matching pairs from the resultant candidate
matches returned by the algorithm.
LSD (Doan et al., 2001) uses machine learning
techniques to semi automatically compute semantic mapping between schemas. It
uses a set of learners like Whirl Learner, Naïve Bayesian learner, Name
Matcher using TF/IDF similarity measure and also domain specific learners like
County-Name Recognizer to learn the similarity patterns between the schema elements.
The predictions of the individual learners are then combined using a meta-learner.
The technique is extensible i.e., new learners can be added to it.
COMA (Do et al., 2002) proposes a flexible composite
match approach to combine different match algorithms. It uses multiple schema
level matchers exploiting linguistic, data-type, structural information and
previous matches to perform the matching process. The novelty in present approach
is that, we use a simple path matching algorithm in conjunction with the regular
match approaches, there by improving the accuracy of the matching process.
A HYBRID PATH MATCHING ALGORITHM
The focus of the proposed approach is in the development of a simple path matching
algorithm which is used for identifying candidate match pairs between the schema
elements. The algorithm takes two schemas in the form of schema trees (Fig.
1) as input. First, it extracts the paths from the root of the schema tree
to its various leaf elements. The algorithm treats each path in the schema tree
as a vector of strings. It then matches each path of the source schema tree
with the paths of the target schema tree and returns a path similarity matrix
between the source and target schema tree paths.
The similarity between any two path is a value in the interval 0 and
1, where, 0 means that the paths are totally dissimilar and 1 means the
paths are totally identical. The path similarity, pathsimmat is computed
based on the number of similar elements between the paths. The element
similarity simmat, is the summation of linguistic similarity measure LingSim
computed using linguistic similarity computation techniques like-levenshtein
distance, affix matching anddomain specific dictionaries and structural
similarity measure StructSim, computed using a simple structural similarity
computation technique-based on the number of linguistically similar children
each pair of matched elements have.
||Po, purchase order schema variant 1
The element pairs whose linguistic similarity measure is greater than
a predefined threshold lcutoff, are considered to be similar. The element
similarity measure, simmat thus computed is again a value in the interval
0 and 1, where, 0 means the elements are dissimilar and 1 means the elements
are identical. Then for each path from source schema tree the path with
the maximum similarity measure from the target schema tree is identified
and if this value is greater than a predefined threshold, scutoff the
paths are considered as similar. The complete algorithm is sown in Code
Once similar paths are identified, the element pairs from the matching
path whose similarity measure simmat is greater than a given threshold,
cutoff value is chosen as the candidate matching pairs.
In order to evaluate the performance of this algorithm a prototype implementing
this algorithm was built. The performance of the algorithm has been shown in
terms of precision and recall, a common measure proposed in the literature (Do
et al., 2002) used to test these types of systems. Precision and
Recall are defined as:
||Ratio of the correct matches identified to total number of matches
|| Ratio of the correct matches identified to total number of manually
The data sources were chosen from diverse domains so that the performance
is not domain specific. The characteristics of data sources are shown
in Table 1.
First the optimum values for the tunable parameters lcutoff, scutoff
and cutoff value are obtained. This is done by gradually increasing the
values of lcutoff and scutoff from 0 to 1 and trying to find the precision
and recall of the algorithm for those values. From the experimentation
it was found that for lcutoff = 0.4 and scutoff = 0.1 optimum precision
and recall values resulted. The results are shown in Fig.
2. Similarly by varying the value of cutoff value from 0 to 1 for
the optimum values of lcutoff and scutoff determined earlier, it was found
that for cutoff value = 0.5 the algorithm gave optimum precision and recall
values. The result of this experimentation is shown in Fig.
|| Characteristics of the data sources used in the experimentation
|| Experimentation results of precision and recall for
various values of lcutoff and scutoff
|| Experimentation results for various of cutoff value
||Performance analysis of the algorithm for the different
Then the performance of the algorithm was evaluated for optimum values of the
tunable parameters lcutoff, scutoff and cutoff value that was determined in
the earlier experimentation. The results of this experimentation are shown in
Fig. 4. It can be seen that the precision and recall of the
algorithm ranged between 0.75 to 1.0 which was quite impressive given the performances
of similar approaches in the literature (Do et al.,
2002; Doan et al., 2001; Madhavan
et al., 2001). The snapshots of the prototype that we used for experimentation
purpose are shown in Fig. 5 and 6.
|| Selection of schemas for matching purpose and setting
up the various parameters for operation
|| Match results for the two selected schemas
CONCLUSION AND FUTURE WORK
The algorithm is extremely successful in identifying one to one correspondences
between schema elements. Future directions are working on ways to upgrade
the algorithm to identify other type of correspondences such as one to
many, many to one and many to many correspondences. And working on ways
to use user feedback and earlier match results to further enhance the
performance of the algorithm.