Harnessing historical corrections to build test collections for named entity disambiguation

Abstract

Matching mentions of persons to the actual persons (the name disambiguation problem) is central for many digital library applications. Scientists have been working on algorithms to create this matching for decades without finding a universal solution. One problem is that test collections for this problem are often small and specific to a certain collection. In this work, we present an approach that can create large test collections from historical metadata with minimal extra cost. We apply this approach to the dblp collection to generate two freely available test collections. One collection focuses on the properties of name-related defects (such as similarities of synonymous names) and one on the evaluation of disambiguation algorithms.