REBL: Entity Linking at Scale
Abstract (link to article)
REBL is an extension of the Radboud Entity Linker (REL) for Batch Entity Linking. REBL is developed after encountering unforeseen issues when trying to link the large MS MARCO v2 web document collection with REL. In this paper we discuss the issues we ran into and our solutions to mitigate them. REBL makes it easier to isolate the GPU heavy operations from the CPU heavy operations, by separating the mention detection stage from the candidate selection and entity disambiguation stages. By improving the entity disambiguation module we were able to lower the time needed for linking documents by an order of magnitude. The code for REBL is publicly available on GitHub.