Cross-lingual Induction of Selectional Preferences with Bilingual Vector Spaces

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 921–929, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Cross-lingual Induction of Selectional Preferences with Bilingual Vector Spaces Yves Peirsman QLVL, University of Leuven Research Foundation – Flanders (FWO) yves.peirsman@arts.kuleuven.be Sebastian Padó IMS, University of Stuttgart pado@ims.uni-stuttgart.de Abstract We describe a cross-lingual method for the in- duction of selectional preferences for resource- poor languages, where no accurate monolin- gual models are available. The method uses bilingual vector spaces to “translate” foreign language predicate-argument structures into a resource-rich language like English. The only prerequisite for constructing the bilin- gual vector space is a large unparsed corpus in the resource-poor language, although the model can profit from (even noisy) syntactic knowledge. Our experiments show that the cross-lingual predictions correlate well with human ratings, clearly outperforming monolin- gual baseline models. 1 Introduction Selectional preferences capture the empirical observa- tion that not all words are equally good arguments to a given verb in a particular argument position (Wilks, 1975; Resnik, 1996). For instance, the subjects of the English verb to shoot are generally people, while the direct objects can be people or animals. This is reflected in speakers’ intuitions. Table 1 shows that the combination the hunter shot the deer is judged more plausible than the deer shot the hunter. Selec- tional preferences do not only play an important role in human sentence processing (McRae et al., 1998), but are also helpful for NLP tasks like word sense disambiguation (McCarthy and Carroll, 2003) and semantic role labeling (Gildea and Jurafsky, 2002). Computational models of selectional preferences predict such plausibilities for triples of a predicate p, an argument position a, and a head word h, such as Predicate Relation Noun Plausibility shoot subject hunter 6.9 shoot object hunter 2.8 shoot subject deer 1.0 shoot object deer 6.4 Table 1: Predicate-relation-noun triples with human plau- sibility judgments on a 7-point scale (McRae et al., 1998) (shoot,object,hunter). All recent models take a two- step approach: (1), they extract all triples (p, a, h) from a large corpus; (2), they apply some type of generalization to make predictions for unseen items. Clearly, the accuracy of these models relies crucially on the quality and coverage of the extracted triples, and thus on the syntactic analysis of the corpus. Un- fortunately, corpora that are both large enough and have a very good syntactic analysis are only available for a handful of Western and Asian languages, which leaves all other languages without reliable selectional preference models. In this paper, we propose a cross-lingual knowl- edge transfer approach to this problem: We automat- ically translate triples (p, a, h) from resource-poor languages into English, where large and high-quality parsed corpora are available and we can compute a reliable plausibility estimate. The translations are extracted from a bilingual semantic space, which can be constructed via bootstrapping from large unparsed corpora in the two languages, without the need for parallel corpora or bilingual lexical resources. Structure …

Content truncated to 500 words