Abstract

Abstract: We studied the problem of record normalization overa set of matching records that refer to the same real-world entity. Wepresented three levels of normalization granularities (record-level,field-level and value component level) and two forms of normalization (typicalnormalization and complete normalization).We propose three levels ofgranularities for record normalization along with methods to constructnormalized records according to them. We propose a comprehensive framework forsystematic construction of normalized records. Our framework is flexible andallows new strategies to be added with ease. To our knowledge, this is thefirst piece of work to propose such a detailed framework. We propose andcompare a range of normalization strategies, from frequency, length, centroidand feature-based to more complex ones that utilize result merging models frominformation retrieval, such as (weighted) Borda. We introduce a number ofheuristic rules to mine desirable value components from a field. We use them toconstruct the normalized value for the field. We perform empirical studies onpublication records. The experimental results show that the proposedweighted-Borda-based approach significantly outperforms the baselineapproaches.