TEG-REP: A corpus of Textual Entailment Graphs based on Relation Extraction Patterns

TEG-REP: A Corpus of Textual Entailment Graphs based on Relation Extraction Patterns Kathrin Eichler, Feiyu Xu, Hans Uszkoreit, Leonhard Hennig, Sebastian Krause German Research Center for Artificial Intelligence Alt-Moabit 91c, Berlin, Germany {kathrin.eichler,feiyu,uszkoreit,leonhard.hennig,sebastian.krause}@dfki.de Abstract The task of relation extraction is to recognize and extract relations between entities or concepts in texts. Dependency parse trees have become a popular source for discovering extraction patterns, which encode the grammatical relations among the phrases that jointly express relation instances. State-of-the-art weakly supervised approaches to relation extraction typically extract thousands of unique patterns only potentially expressing the target relation. Among these patterns, some are semantically equivalent, but differ in their morphological, lexical-semantic or syntactic form. Some express a relation that entails the target relation. We propose a new approach to structuring extraction patterns by utilizing entailment graphs, hierarchical structures representing entailment relations, and present a novel resource of gold-standard entailment graphs based on a set of patterns automatically acquired using distant supervision. We describe the methodology used for creating the dataset and present statistics of the resource as well as an analysis of inference types underlying the entailment decisions. Keywords: Entailment Graphs, Relation Extraction, Textual Entailment 1. Introduction The task of relation extraction (RE) is to recognize and extract relations between entities or concepts in texts. De- pendency parse trees have become a popular source for dis- covering extraction patterns, which encode the grammatical relations among the phrases that jointly express relation in- stances. In rule-based RE methods, the patterns are directly applied to extract relation mentions from parsed sentences of free texts (e.g., Yangarber et al. (2000), Alfonseca et al. (2012)). Other methods treat RE as a classification or sequence-labeling problem, but even for those techniques parse tree patterns have proven useful as key classification features (e.g., Zelenko et al. (2003), Bunescu and Mooney (2005)). In order to circumvent manual annotation work needed for supervised learning, recent work in RE concen- trates on weakly supervised learning, for example based on techniques of distant supervision (Mintz et al., 2009; Krause et al., 2012). These utilize extensive volumes of pre-existing knowledge for partially labeling large volumes of data, resulting in large numbers of unique candidate pat- terns acquired from suspected mentions of relation instances. Among these patterns, some are semantically equivalent, but differ in their morphological, lexical-semantic or syntac- tic form. Some express a relation that entails the target relation or that is entailed by the target relation. Others are semantically unrelated to the target relation. The ba- sic assumption made in this work is that patterns are truly reliable if they express a relation that semantically entails the target relation (Romano et al., 2006). This includes all patterns that express the target relation explicitly or a se- mantically equivalent relation. As an example, the pattern “org|BUYER bought org|ACQUIRED” can be considered to be semantically equivalent to the pattern “org|BUYER pur- chased org|ACQUIRED”, whereas “per|SPOUSE divorced per|SPOUSE” is not semantically equivalent to “per|SPOUSE married per|SPOUSE”, but entails a marriage relation. We propose a new approach …

Content truncated to 500 words