Learning class-to-class selectional preferences

摘要

Selectional preference learningmethods have usually focused on wordto-class relations, e.g., a verb selects asits subject a given nominal class. Thispapers extends previous statisticalmodels to class-to-class preferences,and presents a model that learnsselectional preferences for classes ofverbs. The motivation is twofold:different senses of a verb may havedifferent preferences, and some classesof verbs can share preferences. Themodel is tested on a word sensedisambiguation task which usessubject-verb and object-verbrelationships extracted from a smallsense-disambiguated corpus.