Large-Scale Knowledge Graph Identification using PSL Extended Abstract

The web is a vast repository of knowledge, but automatically extracting that knowledge, at scale, has proven to be a formidable challenge. A number of recent evaluation efforts have focused on automatic knowledge base population (Ji, Grishman, and Dang 2011; Artiles and Mayfield 2012), and many well-known broad domain and open information extraction systems exist, including the Never-Ending Language Learning (NELL) project (Carlson et al. 2010), OpenIE (Etzioni et al. 2008), and efforts at Google (Pasca et al. 2006), which use a variety of techniques to extract new knowledge, in the form of facts, from the web. These facts are interrelated, and hence, recently this extracted knowledge has been referred to as a knowledge graph (Singhal 2012). Unfortunately, most web-scale extraction systems do not take advantage of the rich dependencies found in the knowledge graph; instead approaches consider extractions independently, relying on simple heuristics to enforce consistency. Recent work demonstrates that reasoning jointly is a promising approach to improving the knowledge graph. (Jiang, Lowd, and Dou 2012) choose candidate facts for inclusion in a knowledge base with a joint approach using Markov Logic Networks (MLNs) (Richardson and Domingos 2006). Jiang et al. provide a straightforward codification of ontological relations and candidate facts found in a knowledge base as rules in first-order logic and use MLNs to formulate a probabilistic model. However, due to the combinatorial explosion of Boolean assignments to random variables, inference and learning in MLNs pose intractable optimization problems. Jiang et al. limit the candidate facts they consider, restricting their dataset to a 2-hop neighborhood around each fact, and use a sampling approach to inference, estimating marginals using MC-SAT. Despite these approximations, their work demonstrate the utility of joint reasoning in comparison to a baseline that considers each fact independently. Our work builds on the foundation of Jiang, Lowd, and Dou by providing a richer model for knowledge bases and vastly improving scalability. Our method transforms the noisy output of an information extraction system, a we de-