Open Relation Extraction and Grounding

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 854–864, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Open Relation Extraction and Grounding Dian Yu, Lifu Huang, Heng Ji Computer Science Department Rensselaer Polytechnic Institute {yud2, huangl7, jih}@rpi.edu Abstract Previous open Relation Extraction (open RE) approaches mainly rely on linguis- tic patterns and constraints to extract im- portant relational triples from large-scale corpora. However, they lack of abilities to cover diverse relation expressions or measure the relative importance of candi- date triples within a sentence. It is also challenging to name the relation type of a relational triple merely based on con- text words, which could limit the useful- ness of open RE in downstream applica- tions. We propose a novel importance- based open RE approach by exploiting the global structure of a dependency tree to extract salient triples. We design an unsu- pervised method to name relation types by grounding relational triples to a large-scale Knowledge Base (KB) schema, leverag- ing KB triples and weighted context words associated with relational triples. Exper- iments on the English Slot Filling 2013 dataset demonstrate that our approach achieves 8.1% higher F-score over state- of-the-art open RE methods. 1 Introduction Open Relation Extraction (open RE) (Banko and Etzioni, 2008) aims at extracting relational triples from an open-domain corpus. Each triple contains two arguments and a phrase which denotes the re- lation between them. In this paper, we focus on discovering relations between entities. Most successful open RE approaches (Fader et al., 2011; Xu et al., 2013; Bovi et al., 2015; Bhutani et al., 2016) extract salient relational triples based on lexical or syntactic patterns. How- ever, such handcrafted or automatically learned patterns are incapable of covering diverse rela- tion expressions (Soderland et al., 2013). Sub- sequently, the shortest path between arguments derived from a dependency tree has been widely applied to generate patterns to capture long- distance and complex relations. However, addi- tional heuristic rules are usually needed to filter out the resulting large number of meaningless pat- terns (Wu and Weld, 2010; Mausam et al., 2012; Bovi et al., 2015). Besides, such flat syntactic structures lack the ability to measure the relative importance of candidate triples in a sentence. For example, the sentence in E1 places particular em- phasis on the relation between “Lucille Clifton” and “1936” which therefore should be retained. E1 “Lucille Clifton, whom he married in 1958, was born in 1936.” We notice that a candidate relational triple is likely to be salient if its two arguments are strongly connected in a dependency tree. Instead of re- lying on patterns to capture important triples, we use an importance-based strategy by exploring the entire dependency tree structure to automatically measure the connection strength of candidate ar- gument pairs. Specifically, we assume that a re- lational triple is important if there is a relatively short random walk-based distance between two relatively important arguments, measured against the entire dependency tree of a given sentence. For each …

Content truncated to 500 words