Statistical data consists mostly of numerical values, entries of codelists like country codes or acronyms for gender. Such values are typically described according to speci c patterns. In this paper we present a novel approach for instance-based schema matching, where regular expressions are utilized for matching patterns of instance values.
In various domains, e.g. the social sciences, the matching of statistical data is a typical task. Schema elements of statistical data, e.g. rows or columns of a spreadsheet, are named usually by simple and short labels, sometimes even with abbreviated terms. However, the structure and semantics of their instances (e.g. numerical values, entries of codelists) di er in various aspects from text-heavy data. Instances are often described by a speci c syntactical pattern, e.g. dates consist of numerical values divided by periods or slashes or a three-letter code for a geographical area.
For instance-based schema matching [
Schema Matching using Regular Expressions
By utilizing pattern classes our approach considers two schema elements as
a match, if their instances can be expressed via at least one regular
expression of the same pattern class. We de ne multiple pattern classes, which
correspond to a speci c data element, e.g. dates, age groups or geographical codes,
and contain various patterns for describing this data element. For a data
element "date" di erent patterns might be e.g. [
We assume two datasets M and N with their schema elements SM 2 M and SN 2 N . The pattern classes Cx with Cx = f(regex; !)jregex matches x; 0 < ! < 1g contain multiple regular expressions regex describing the statistical data element x of the class. They are accompanied with a weighting !.
For each pattern class Cx, we compute an average weighting for every schema element SM and SN . This average weighting indicates how often instances of the schema element can be expressed by a pattern of the class. Hereby, as soon as an instance can be expressed by a (regex; !) 2 Cx, the value of ! is added to the sum of all weightings, whose regular expressions previously matched another instance from this same schema element, resulting in the nal P0 !. The average is then retrieved by normalizing this sum regarding the total number of instances inside this particular schema element. For each SM , this is avg(SM ) = jInstanPce0s!in SM j . For SN the average is calculated analogously. If this average weight is not 0, the schema element is collected among its average weight in a set. We de ne these sets as Mx and Nx with Mx = f(SM ; avg(SM )g and Nx = f(SN ; avg(SN )g.
The Cartesian product of Mx and Nx is computed and added to M atchesx, in which a triple (SM ; SN ; avg(SM ) avg(SN )) de nes a match between a SM and a SN with the probability of avg(SM ) avg(SN ). Finally, the result set M atchesx contains all matches between two datasets M and N .
Our approach has been implemented in Java using the JENA API. The source code and an executable jar le are available at https://github.com/mazlo/smurf. In rst experiments with real-world statistical data we obtained better results for matching schema elements than other existing matching systems. A detailed evaluation with generic test datasets is currently work-in-progress. We aim to extend our approach to extract patterns from instance values and to generate weightings automatically. Feature extraction from instance values can enhance our approach in computing weightings and in assigning regular expressions to adequate pattern classes.