Training selection for tuning entity matching

Entity matching is a crucial and difficult task for data integration. An effective solution strategy typically has to combine several techniques and to find suitable settings for critical configuration parameters such as similarity thresholds. Supervised (trainingbased) approaches promise to reduce the manual work for determining (learning) effective strategies for entity matching. However, they critically depend on training data selection which is a difficult problem that has so far mostly been addressed manually by human experts. In this paper we propose a trainingbased framework called STEM for entity matching and present different generic methods for automatically selecting training data to combine and configure several matching techniques. We evaluate the proposed methods for different match tasks and smalland medium-sized training sets.