Focused Entailment Graphs for Open IE Propositions

Proceedings of the Eighteenth Conference on Computational Language Learning, pages 87–97, Baltimore, Maryland USA, June 26-27 2014. c⃝2014 Association for Computational Linguistics Focused Entailment Graphs for Open IE Propositions Omer Levy† Ido Dagan† Jacob Goldberger§ † Computer Science Department § Faculty of Engineering Bar-Ilan University Ramat-Gan, Israel {omerlevy,dagan,goldbej}@{cs,cs,eng}.biu.ac.il Abstract Open IE methods extract structured propo- sitions from text. However, these propo- sitions are neither consolidated nor gen- eralized, and querying them may lead to insufficient or redundant information. This work suggests an approach to or- ganize open IE propositions using entail- ment graphs. The entailment relation uni- fies equivalent propositions and induces a specific-to-general structure. We create a large dataset of gold-standard proposition entailment graphs, and provide a novel algorithm for automatically constructing them. Our analysis shows that predicate entailment is extremely context-sensitive, and that current lexical-semantic resources do not capture many of the lexical infer- ences induced by proposition entailment. 1 Introduction Open information extraction (open IE) extracts natural language propositions from text without pre-defined schemas as in supervised relation ex- traction (Etzioni et al., 2008). These proposi- tions represent predicate-argument structures as tuples of natural language strings. Open IE en- ables knowledge search by aggregating billions of propositions from the web1. It may also be per- ceived as capturing an unsupervised knowledge representation schema, complementing supervised knowledge bases such as Freebase (Bollacker et al., 2008), as suggested by Riedel et al (2013). However, language variability obstructs open IE from becoming a viable knowledge representation framework. As it does not consolidate natural lan- guage expressions, querying a database of open IE propositions may lead to either insufficient or re- dundant information. As an illustrative example, 1See demo: openie.cs.washington.edu querying the demo (footnote 1) for the generally equivalent relieves headache or treats headache returns two different lists of entities; out of the top few results, the only answers these queries seem to agree on are caffeine and sex. This is a major drawback relative to supervised knowledge rep- resentations, which map natural language expres- sions to structured formal representations, such as treatments in Freebase. In this work, we investigate an approach for or- ganizing and consolidating open IE propositions using the novel notion of proposition entailment graphs (see Figure 1) – graphs in which each node represents a proposition and each directed edge reflects an entailment relation, in the spirit of textual entailment (Dagan et al., 2013). En- tailment provides an effective structure for ag- gregating natural-language based information; it merges semantically equivalent propositions into cliques, and induces specification-generalization edges between them. For example, (aspirin, elim- inate, headache) entails, and is more specific than, (headache, respond to, painkiller). We thus propose the task of constructing an entailment graph over a set of open IE proposi- tions (Section 3), which is closely related to Be- rant et al’s work (2012) who introduced predicate entailment graphs. In contrast, our work explores propositions, which are essentially predicates in- stantiated with arguments, and thus semantically richer. We provide a dataset of 30 such graphs, …

Content truncated to 500 words