Automatic generation of probabilistic relationships for improving schema matching

Schema matching is the problem of finding relationships among concepts across data sources that are heterogeneous in format and in structure. Starting from the “hidden meaning” associated with schema labels (i.e. class/attribute names), it is possible to discover lexical relationships among the elements of different schemata. In this work, we propose an automatic method aimed at discovering probabilistic lexical relationships in the environment of data integration “on the fly”. Our method is based on a probabilistic lexical annotation technique, which automatically associates one or more meanings with schema elements w.r.t. a thesaurus/lexical resource. However, the accuracy of automatic lexical annotation methods on real-world schemata suffers from the abundance of non-dictionary words such as compound nouns and abbreviations. We address this problem by including a method to perform schema label normalization which increases the number of comparable labels. From the annotated schemata, we derive the probabilistic lexical relationships to be collected in the Probabilistic Common Thesaurus. The method is applied within the MOMIS data integration system but can easily be generalized to other data integration systems.