Acquisition Of Chinese Verb Selectional Preference Information

The semantic selectional restriction of verbs towards their arguments has always been a quite focused problem in the natural language processing field. In the past, the researchers' work on the problem of verb semantic selectional restriction mainly limited on the level of sentences. They focus on the semantic selectional restriction problem of verbs towards their arguments under the condition that verbs are taken as the center predicates in sentences, while lack a deep research on the semantic preference problem of verbs appearing in compound structure. In this article, the semantic selectional reference problem in Chinese compound structure is mainly researched.NV compound structure ,which is a compound structure with size of two, and has a noun as its modifier and a verb nomination as its center word, is common in Chinese. In this article, the semantic selectional restriction problem in NV compound structure is illustrated as classifying problem. And two semantic selectional restriction model is introduced: one is a classifying model based on automatic lexical model; the other is a classifying model based on attribute words' similarity. This article's new ideas mainly embody in the following points: a)In this article, automatically acquired lexical models are taken as the classifying characteristics of the classifier, using the positive and negative lexical models together to tag the semantic relations. And in the end the automatically acquired lexical model is compared with manual lexical model, proving that automatic lexical model is better than other models. b)the semantic similarity between two words based on attribute words calculation is introduced, and used as semantic selectional classifying. Attribute words are used to describe every aspects of concept definition. The more common attributes two concepts share, the more similar the two concepts may turn out to be. By translating the Miller-Charles word pairs standard into a Chinese version, we compare the result of attribute words model with the results calculated with HowNet and the results from juxtaposition words model, we could see that the attribute words' similarity model has a better result.