Lorify: A Knowledge Base from Scratch

In this paper we discuss our approach to the task of Cold-Start Knowledge Base Population and the challenges associated with it. We describe our knowledge base system Lorify and each of the components necessary to populate it from unstructured text. The pivotal component for building a large-scale knowledge base is scalable cross-document coreference. We address this with a novel clustering algorithm based on Markov-Chain MonteCarlo, and show that it is capable of scaling to much larger sets of entities than typical algorithms. Finally, we detail the performance of this system on the TAC KBP 2012 evaluation.