Applying edge-counting semantic similarities to Link Discovery: Scalability and Accuracy

With the growth in number and variety of RDF datasets comes an in- creasing need for both scalable and accurate solutions to support link discovery at instance level within and across these datasets. In contrast to ontology matching, most linking frameworks rely solely on string similarities to this end. The limited use of semantic similarities when linking instances is partly due to the current literature stating that they (1) do not improve the F-measure of instance linking approaches and (2) are impractical to use because they lack time efficiency. We revisit the combination of string and semantic similarities for linking instances. Contrary to the literature, our results suggest that this combination can improve the F-measure achieved by instance linking systems when the combination of the measures is performed by a machine learning approach. To achieve this in- sight, we had to address the scalability of semantic similarities. We hence present a framework for the rapid computation of semantic similarities based on edge counting. This runtime improvement allowed us to run an evaluation of 5 bench- mark datasets. Our results suggest that combining string and semantic similarities can improve the F-measure by up to 6% absolute.