A Study of Similarity Measures for Personal Names

Abstract

Name matching is a key component of historical record-linkage systems because a person's name is one of the main fields used to identify an individual across multiple databases. However, typographical errors and alternative spellings make name-matching a non-trivial problem. Therefore, record-linkage systems must rely upon good similarity measures to compare personal names. Knowing which similarity measure to use a priori for a given set of personal names is itself a challenging problem. In this thesis, we present an experimental study to investigate the performance of 76 similarity measures for name matching. We first discuss the various characteristics of personal names. We then evaluate and compare the performance of 76 similarity measures using three different datasets. Our results show that there is no single measure that significantly outperforms all the others. However, there are groups of measures that one should consider first when dealing with a name matching task.