Towards Transfer Learning of Link Specifications

Over the last years, link discovery frameworks have been employed successfully to create links between knowledge bases. Consequently, repositories of high-quality link specifications have been created and made available on the Web. The basic question underlying this work is the following: Can the specifications in these repositories be reused to ease the detection of link specifications between unlinked knowledge bases? In this paper, we address this question by presenting a formal transfer learning framework that allows detecting existing specifications that can be used as templates for specifying links between previously unlinked knowledge bases. We discuss both the advantages and the limitations of such an approach for determining link specifications. We evaluate our approach on a variety of link specifications from several domains and show that the detection of accurate link specifications for use as templates can be achieved with high reliability.