Nested Propositions in Open Information Extraction

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 55–64, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Nested Propositions in Open Information Extraction Nikita Bhutani Department of EECS University of Michigan Ann Arbor nbhutani@umich.edu H. V. Jagadish Department of EECS University of Michigan Ann Arbor jag@umich.edu Dragomir Radev Department of EECS University of Michigan Ann Arbor radev@umich.edu Abstract The challenges of Machine Reading and Knowledge Extraction at a web scale re- quire a system capable of extracting diverse information from large, heterogeneous cor- pora. The Open Information Extraction (OIE) paradigm aims at extracting assertions from large corpora without requiring a vocabulary or relation-specific training data. Most sys- tems built on this paradigm extract binary re- lations from arbitrary sentences, ignoring the context under which the assertions are cor- rect and complete. They lack the expres- siveness needed to properly represent and ex- tract complex assertions commonly found in the text. To address the lack of representa- tion power, we propose NESTIE, which uses a nested representation to extract higher-order relations, and complex, interdependent asser- tions. Nesting the extracted propositions al- lows NESTIE to more accurately reflect the meaning of the original sentence. Our ex- perimental study on real-world datasets sug- gests that NESTIE obtains comparable preci- sion with better minimality and informative- ness than existing approaches. NESTIE pro- duces 1.7-1.8 times more minimal extractions and achieves 1.1-1.2 times higher informative- ness than CLAUSIE. 1 Introduction Syntactic analyses produced by syntactic parsers are a long way from representing the full meaning of the sentences parsed. In particular, they cannot support questions like “Who did what to whom?”, “Where did what happen?”. Owing to the large, hetero- geneous corpora available at web scale, traditional approaches to information extraction (Brin, 1998; Agichtein and Gravano, 2000) fail to scale to the millions of relations found on the web. As a re- sponse, the paradigm of Open Information Extrac- tion (OIE) (Banko et al., 2007) has seen a rise in interest as it eliminates the need for domain knowl- edge or relation-specific annotated data. OIE sys- tems use a collection of patterns over the surface form or dependency tree of a sentence to extract propositions of the form (arg1,rel,arg2). However, state-of-the-art OIE systems, REVERB (Fader et al., 2011) and OLLIE (Schmitz et al., 2012) focus on extracting binary assertions and suffer from three key drawbacks. First, lack of expressivity of representation leads to significant information loss for higher-order relations and complex assertions. This results in incomplete, uniformative and inco- herent prepositions. Consider Example 1 in Fig- ure 1. Important contextual information is either ignored or is subsumed in over-specified argument and relation phrases. It is not possible to fix such nuances by post-processing the propositions. This affects downstream applications like Question An- swering (Fader et al., 2014) which rely on correct- ness and completeness of the propositions. Second, natural language frequently includes re- lations presented in a non-canonical form that can- not be captured by a small set …

Content truncated to 500 words