Learning class-to-class selectional preferences

Learning class-to-class selectional preferences Eneko Agirre IXA NLP Group University of the Basque Country 649 pk. 20.080 Donostia. Spain. eneko@si.ehu.es David Martinez IXA NLP Group University of the Basque Country 649 pk. 20.080 Donostia. Spain. jibmaird@si.ehu.es Abstract Selectional preference learning methods have usually focused on word- to-class relations, e.g., a verb selects as its subject a given nominal class. This papers extends previous statistical models to class-to-class preferences, and presents a model that learns selectional preferences for classes of verbs. The motivation is twofold: different senses of a verb may have different preferences, and some classes of verbs can share preferences. The model is tested on a word sense disambiguation task which uses subject-verb and object-verb relationships extracted from a small sense-disambiguated corpus. 1 Introduction Previous literature on selectional preference has usually learned preferences for words in the form of classes, e.g., the object of eat is an edible entity. This paper extends previous statistical models to classes of verbs, yielding a relation between classes in a hierarchy, as opposed to a relation between a word and a class. The model is trained using subject-verb and object-verb associations extracted from Semcor, a corpus (Miller et al., 1993) tagged with WordNet word-senses (Miller et al., 1990). The syntactic relations were extracted using the Minipar parser (Lin, 1993). A peculiarity of this exercise is the use of a small sense- disambiguated corpus, in contrast to using a large corpus of ambiguous words. We think that two factors can help alleviate the scarcity of data: the fact that using disambiguated words provides purer data, and the ability to use classes of verbs in the preferences. Nevertheless, the approach can be easily extended to larger, non- disambiguated corpora. We have defined a word sense disambiguation exercise in order to evaluate the extracted preferences, using a sample of words and a sample of documents, both from Semcor. Following this short introduction, section 2 reviews selectional restriction acquisition. Section 3 explains our approach, which is formalized in sections 4 and 5. Next, section 6 shows the results on the WSD experiment. Some of the acquired preferences are analysed in section 7. Finally, some conclusions are drawn and future work is outlined. 2 Selectional preference learning Selectional preferences try to capture the fact that linguistic elements prefer arguments of a certain semantic class, e.g. a verb like ‘eat’ prefers as object edible things, and as subject animate entities, as in, (1) “She was eating an apple”. Selectional preferences get more complex than it might seem: (2) “The acid ate the metal”, (3) “This car eats a lot of gas”, (4) “We ate our savings”, etc. Corpus-based approaches for selectional preference learning extract a number of (e.g. verb/subject) relations from large corpora and use an algorithm to generalize from the set of nouns for each verb separately. Usually, nouns are generalized using classes (concepts) from a lexical knowledge base (e.g. WordNet). Resnik (1992, 1997) defines an information- theoretic measure of the association between a verb and nominal WordNet …

Content truncated to 500 words