Identifying hypernyms in distributional semantic spaces

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 75–79, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Identifying hypernyms in distributional semantic spaces Alessandro Lenci University of Pisa, Dept. of Linguistics via S. Maria 36 I-56126, Pisa, Italy alessandro.lenci@ling.unipi.it Giulia Benotto University of Pisa, Dept. of Linguistics via S. Maria 36 I-56126, Pisa, Italy mezzanine.g@gmail.com Abstract In this paper we apply existing directional similarity measures to identify hypernyms with a state-of-the-art distributional semantic model. We also propose a new directional measure that achieves the best performance in hypernym identification. 1 Introduction and related works Distributional Semantic Models (DSMs) measure the semantic similarity between words with proxim- ity in distributional space. However, semantically similar words in turn differ for the type of relation holding between them: e.g., dog is strongly similar to both animal and cat, but with different types of re- lations. Current DSMs accounts for these facts only partially. While they may correctly place both ani- mal and cat among the nearest distributional neigh- bors of dog, they are not able to characterize the different semantic properties of these relations, for instance the fact that hypernymy is an asymmetric semantic relation, since being a dog entails being an animal, but not the other way round. The purpose of this paper is to explore the possi- bility of identifying hypernyms in DSMs with direc- tional (or asymmetric) similarity measures (Kotler- man et al., 2010). These measures all rely on some variation of the Distributional Inclusion Hypothe- sis, according to which if u is a semantically nar- rower term than v, then a significant number of salient distributional features of u is included in the feature vector of v as well. Since hypernymy is an asymmetric relation and hypernyms are seman- tically broader terms than their hyponyms, then we can predict that directional similarity measures are better suited to identify terms related by the hyper- nymy relation. Automatic identification of hypernyms in corpora is a long-standing research line, but most meth- ods have adopted semi-supervised, pattern-based ap- proaches (Hearst, 1992; Pantel and Pennacchiotti, 2006). Fully unsupervised hypernym identification with DSMs is still a largely open field. Various mod- els to represent hypernyms in vector spaces have recently been proposed (Weeds and Weir, 2003; Weeds et al., 2004; Clarke, 2009), usually grounded on the Distributional Inclusion Hypothesis (for a dif- ferent approach based on representing word mean- ing as “regions” in vector space, see Erk (2009a; 2009b)). The same hypothesis has been adopted by Kotlerman et al. (2010) to identify (substitutable) lexical entailments” . Within the context of the Tex- tual Entailment (TE) paradigm, Zhitomirsky-Geffet and Dagan (2005; 2009) define (substitutable) lex- ical entailment as a relation holding between two words, if there are some contexts in which one of the words can be substituted by the other and the meaning of the original word can be inferred from the new one. Its relevance for TE notwithstanding, this notion of lexical entailment is more general and looser than …

Content truncated to 500 words