Unsupervised Link Discovery Through Knowledge Base Repair

The Linked Data Web has developed into a compendium of partly very large datasets. Devising efficient approaches to compute links between these datasets is thus central to achieve the vision behind the Data Web. Several unsupervised approaches have been developed to achieve this goal. Yet, so far, none of these approaches makes use of the replication of resources across several knowledge bases to improve the accuracy it achieves while linking. In this paper, we present Colibri, an iterative unsupervised approach for link discovery. Colibri allows discovering links between n datasets (n ≥ 2) while improving the quality of the instance data in these datasets. To this end, Colibri combines error detection and correction with unsupervised link discovery. We evaluate our approach on benchmark datasets with respect to the F-score itachieves. Our results suggest that Colibri can significantly improve the results of unsupervised machine-learning approaches for link discovery while correctly detecting erroneous resources.