CORE: Context-Aware Open Relation Extraction with Factorization Machines

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1763–1773, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. CORE: Context-Aware Open Relation Extraction with Factorization Machines Fabio Petroni Sapienza University of Rome Rome, Italy petroni@dis.uniroma1.it Luciano Del Corro Max Planck Institute for Informatics Saarbr¨ucken, Germany delcorro@mpi-inf.mpg.de Rainer Gemulla University of Mannheim Mannheim, Germany rgemulla@uni-mannheim.de Abstract We propose CORE, a novel matrix fac- torization model that leverages contextual information for open relation extraction. Our model is based on factorization ma- chines and integrates facts from various sources, such as knowledge bases or open information extractors, as well as the con- text in which these facts have been ob- served. We argue that integrating contex- tual information—such as metadata about extraction sources, lexical context, or type information—significantly improves pre- diction performance. Open information extractors, for example, may produce ex- tractions that are unspecific or ambiguous when taken out of context. Our experi- mental study on a large real-world dataset indicates that CORE has significantly bet- ter prediction performance than state-of- the-art approaches when contextual infor- mation is available. 1 Introduction Open relation extraction (open RE) is the task of extracting new facts for a potentially unbounded set of relations from various sources such as knowledge bases or natural language text. The task is closely related to targeted information ex- traction (IE), which aims to populate a knowledge base (KB) with new facts for the KB’s relations, such as wasBornIn(Sepp Herberger, Mannheim). Existing methods either reason within the KB it- self (Franz et al., 2009; Nickel et al., 2011; Dru- mond et al., 2012) or leverage large text corpora to learn patterns that are indicative of KB rela- tions (Mintz et al., 2009; Surdeanu et al., 2012; Min et al., 2013). In both cases, targeted IE meth- ods are inherently limited to an (often small) set of predefined relations, i.e., they are not “open”. The open RE task is also related to open infor- mation extraction (open IE) (Banko et al., 2007; Del Corro and Gemulla, 2013), which extracts large amounts of surface relations and their ar- guments from natural language text; e.g., “critiz- ices”(“Dante”, “Catholic Church”).1 Although open IE is a domain-independent approach, the ex- tracted surface relations are purely syntactic and often ambiguous or noisy. Moreover, open IE methods usually do not “predict” facts that have not been explicitly observed in the input data. Open RE combines the above tasks by predicting new facts for an open set of relations. The key challenge in open RE is to reason jointly over the universal schema consisting of KB relations and surface relations (Riedel et al., 2013). A number of matrix or tensor factorization mod- els have recently been proposed in the context of relation extraction (Nickel et al., 2012; Riedel et al., 2013; Huang et al., 2014; Chang et al., 2014). These models use the available data to learn la- tent semantic representations of entities (or entity pairs) and relations in a domain-independent way; the latent …

Content truncated to 500 words