Learning to Distinguish Hypernyms and Co-Hyponyms

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2249–2259, Dublin, Ireland, August 23-29 2014. Learning to Distinguish Hypernyms and Co-Hyponyms Julie Weeds, Daoud Clarke, Jeremy Reffin, David Weir and Bill Keller Department of Informatics, University of Sussex, Brighton, UK juliewe,D.Clarke,J.P.Reffin,davidw,billk@sussex.ac.uk Abstract This work is concerned with distinguishing different semantic relations which exist between distributionally similar words. We compare a novel approach based on training a linear Support Vector Machine on pairs of feature vectors with state-of-the-art methods based on distributional similarity. We show that the new supervised approach does better even when there is minimal information about the target words in the training data, giving a 15% reduction in error rate over unsupervised approaches. 1 Introduction Over recent years there has been much interest in the field of distributional semantics, drawing on the distributional hypothesis: words that occur in similar contexts tend to have similar meanings (Harris, 1954). There is a large body of work on the use of different similarity measures (Lee, 1999; Weeds and Weir, 2003; Curran, 2004) and many researchers have built thesauri (i.e. lists of “nearest neighbours”) automatically and applied them in a variety of applications, generally with a good deal of success. In early research there was much interest in how these automatically generated thesauri compare with human-constructed gold standards such as WordNet and Roget (Lin, 1998; Kilgarriff and Yallop, 2000). More recently, the focus has tended to shift to building thesauri to alleviate the sparse-data problem. Dis- tributional thesauri have been used in a wide variety of areas including sentiment classification (Bollegala et al., 2011), WSD (Miller et al., 2012; Khapra et al., 2010), textual entailment (Berant et al., 2010), pre- dicting semantic compositionality (Bergsma et al., 2010), acquisition of semantic lexicons (McIntosh, 2010), conversation entailment (Zhang and Chai, 2010), lexical substitution (Szarvas et al., 2013), tax- onomy induction (Fountain and Lapata, 2012), and parser lexicalisation (Rei and Briscoe, 2013). A primary focus of distributional semantics has been on identifying words which are similar to each other. However, semantic similarity encompasses a variety of different lexico-semantic and topical re- lations. Even if we just consider nouns, an automatically generated thesaurus will tend to return a mix of synonyms, antonyms, hyponyms, hypernyms, co-hyponyms, meronyms and other topically related words. A central problem here is that whilst most measures of distributional similarity are symmetric, some of the important semantic relations are not. The hyponymy relation (and converse hypernymy) which forms the ISA backbone of taxonomies and ontologies such as WordNet (Fellbaum, 1989), and determines lexical entailment (Geffet and Dagan, 2005), is asymmetric. On the other hand, the co- hyponymy relation which relates two words unrelated by hyponymy but sharing a (close) hypernym, is symmetric, as are synonymy and antonymy. Table 1 shows the distributionally nearest neighbours of the words cat, animal and dog. In the list for cat we can see 2 hypernyms and 13 co-hyponyms1. 1We read cat in the sense domestic cat rather than big cat, hence tiger …

Content truncated to 500 words