RELLY: Inferring Hypernym Relationships Between Relational Phrases

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 971–981, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. RELLY: Inferring Hypernym Relationships Between Relational Phrases Adam Grycner, Gerhard Weikum Max-Planck Institute for Informatics Campus E1.4, 66123 Saarbr¨ucken, Germany {agrycner, weikum} @mpi-inf.mpg.de Jay Pujara, James Foulds, Lise Getoor Computer Science Department University of California, Santa Cruz Santa Cruz, CA 95064 {jpujara,jfoulds,getoor} @soe.ucsc.edu Abstract Relational phrases (e.g., “got married to”) and their hypernyms (e.g., “is a relative of”) are central for many tasks including question answering, open information ex- traction, paraphrasing, and entailment de- tection. This has motivated the develop- ment of several linguistic resources (e.g. DIRT, PATTY, and WiseNet) which sys- tematically collect and organize relational phrases. These resources have demonstra- ble practical benefits, but are each limited due to noise, sparsity, or size. We present a new general-purpose method, RELLY, for constructing a large hypernymy graph of relational phrases with high-quality sub- sumptions using collective probabilistic programming techniques. Our graph in- duction approach integrates small high- precision knowledge bases together with large automatically curated resources, and reasons collectively to combine these re- sources into a consistent graph. Using RELLY, we construct a high-coverage, high-precision hypernymy graph consist- ing of 20K relational phrases and 35K hy- pernymy links. Our evaluation indicates a hypernymy link precision of 78%, and demonstrates the value of this resource for a document-relevance ranking task. 1 Introduction One of the many challenges in natural lan- guage understanding is interpreting the multi- word phrases that denote relationships between entities. Semantically organizing the complex re- lationships between diverse phrases is crucial to applications including question answering, open information extraction, paraphrasing, and entail- ment detection (Yahya et al., 2012; Fader et al., 2011; Madnani et al., 2012; Dagan et al., 2005). For example, a corpus containing the phrase “George Burns was married to Gracie Allen” al- lows us to answer the query “Who was the spouse of George Burns?” However, “Jay Z is in a re- lationship with Beyonc´e” provides insufficient in- formation to determine whether the couple is mar- ried. To capture the knowledge found in text, rela- tional phrases need to be systematically organized with lexical links like synonymy (“married to” and “spouse of”) and hypernymy (“in a relationship” generalizing “married to”). Many projects address the challenge of under- standing relational phrases, but existing linguis- tic resources are often limited to synonymy, suffer from low precision, or have low coverage. Sys- tems such as DIRT (Lin and Pantel, 2001), RE- SOLVER (Yates and Etzioni, 2009), and WiseNet (Moro and Navigli, 2012) have used sophisticated clustering techniques to determine synonymous phrases, but do not provide subsumption informa- tion. The PATTY (Nakashole et al., 2012) project goes beyond clustering and introduces a subsump- tion hierarchy, but suffers from sparsity and con- tains few hypernymy links. The HARPY (Gryc- ner and Weikum, 2014) project extended PATTY, generating 600K hypernymy links, but with low precision. Berant et al. (2011) introduced en- tailment graphs that provided a …

Content truncated to 500 words