Term Definitions Help Hypernymy Detection

Proceedings of the 7th Joint Conference on Lexical and Computational Semantics (*SEM), pages 203–213 New Orleans, June 5-6, 2018. c⃝2018 Association for Computational Linguistics Term Definitions Help Hypernymy Detection Wenpeng Yin and Dan Roth University of Pennsylvania {wenpeng,danroth}@seas.upenn.edu Abstract Existing methods of hypernymy detection mainly rely on statistics over a big corpus, ei- ther mining some co-occurring patterns like “animals such as cats” or embedding words of interest into context-aware vectors. These approaches are therefore limited by the avail- ability of a large enough corpus that can cover all terms of interest and provide suffi- cient contextual information to represent their meaning. In this work, we propose a new paradigm, HYPERDEF, for hypernymy de- tection – expressing word meaning by en- coding word definitions, along with context driven representation. This has two main benefits: (i) Definitional sentences express (sense-specific) corpus-independent meanings of words, hence definition-driven approaches enable strong generalization – once trained, the model is expected to work well in open- domain testbeds; (ii) Global context from a large corpus and definitions provide comple- mentary information for words. Consequently, our model, HYPERDEF, once trained on task- agnostic data, gets state-of-the-art results in multiple benchmarks1. 1 Introduction Language understanding applications like textual entailment (Dagan et al., 2013), question answer- ing (Saxena et al., 2007) and relation extraction (Mintz et al., 2009), benefit from the identification of lexical entailment relations. Lexical inference encompasses several semantic relations, with hy- pernymy being one of the prevalent (Roller et al., 2014; Shwartz et al., 2016), an i.e., “Is-A” relation that holds for a pair of terms2 (x, y) for specific terms’ senses. Two families of approaches have been studied for identifying term hypernymy. (i) Pattern match- 1cogcomp.org/page/publication_view/836 2This paper uses “term” to refer to any words or phrases. ing exploits patterns such as “animals such as cats” to indicate a hypernymy relation from “cat” to “an- imal” (Hearst, 1992; Snow et al., 2004). How- ever, it requires the co-occurrence of the two terms in the same sentence, which limits the recall of this method; (ii) Term representation learning de- pends on a vector embedding of each term, where each entry in the vector expresses an explicit con- text feature (Baroni et al., 2012a; Roller and Erk, 2016; Shwartz et al., 2017) or a latent semantic (Fu et al., 2014; Vulic and Mrksic, 2017; Glavas and Ponzetto, 2017). Both approaches hinge on acquiring context- aware term meaning in a large corpus. The gen- eralization of these corpus-based representation learning paradigms, however, is limited due to the domain specificity of the training data. For ex- ample, an IT corpus hardly mentions “apple” as a fruit. Furthermore, the surrounding context of a term may not convey subtle differences in term meaning – “he” and “she” have highly similar con- text that may not reveal the important difference between them. Moreover, rare words are poorly expressed by their sparse global context and, more generally, these methods would not generalize to the low resource language setting. Humans can easily determine …

Content truncated to 500 words