Unsupervised Learning of Ontology-Linked Selectional Preferences

We present a method for extracting selectional preferences of verbs from unannotated text. These selectional preferences are linked to an ontology (e.g. the hypernym relations found in WordNet), which allows for extending the coverage for unseen valency fillers. For example, if drink vodka is found in the training corpus, a whole WordNet hierarchy is assigned to the verb to drink (drink liquor, drink alcohol, drink beverage, drink substance, etc.), so that when drink gin is seen in a later stage, it is possible to relate the selectional preference drink vodka with drink gin (as gin is a co-hyponym of vodka). This information can be used for word sense disambiguation, prepositional phrase attachment disambiguation, syntactic disambiguation, and other applications within the approach of pattern-based statistical methods combined with knowledge. As an example, we present an application to word sense disambiguation based on the Senseval-2 training text for Spanish. The results of this experiment are similar to those obtained by Resnik for English.