Estimating semantic relation of Japanese noun phrases based on k-nearest neighbor estimation

The noun phrase `NP `no' NP', that consists of two noun phrases NPs connected by an adnominal particles `no', is frequently used in Japanese sentences. The surface structure of this pattern is simple, but it has various semantic structures. For such noun phrases, there has been a grammar proposed, where their semantic structures can be gotten systematically from their syntactic structures. This grammar fractionates noun phrases into four syntactic categories (CN,T,RN,EN). As a result, the syntactic structures in this grammar can be mapped into each semantic structure. But in case that syntactic structure is `T `no' CN', it is necessary to infer semantic relations between T and CN, which don't appear in the surface structure. In this paper, one method of pattern recognization (Bayes decision rule with k-Nearest Neighbor estimation, which is nonparametric estimation of a probability density) is applied to estimating semantic relations of a noun phrase whose syntactic category is `T `no' CN'. As a result of the experiment, the accuracy is about 78%.