Abstract
A semantic space model provides a framework of semantic representation. In this model, each word is represented by a high-dimensional vector and the degree of semantic similarity between any two words can be easily computed as the cosine of the angle formed by their vectors. Recently, a number of methods have been proposed for constructing semantic spaces, but little has been known about the properties of different semantic spaces, in particular what kinds of semantic relations can be represented by what kinds of semantic spaces. In this study, we constructed fourteen different semantic spaces using three corpora (i.e., Japanese newspaper articles, Japanese novels, and Japanese dictionary), two construction methods (i.e., term frequency (TF) and term cooccurrence (CO)) and three context-window sizes (i.e., article, paragraph, and sentence). We then examined the properties of these spaces by comparing the ability to represent three semantic relations (i.e., coordination⁄synonymy, superordination, and collocation) and their eight subrelations. As a result, we demonstrated that, regardless of construction method and window size, the coordination⁄synonymy relation was better represented by the dictionary-based semantic spaces, but the collocation relation was better represented by the newspaper- and TF-based spaces. We also found that the superordination relation was better represented by the TF-based spaces with paragraphs as a window size, and corpus difference between dictionary and newspaper did not affect the representational ability of superordination. In addition, we investigated the effects of dimensionality reduction by singular value decomposition. The overall result was that the performance in predicting word association was degraded, but the performance of typicality judgment for the coordination⁄synonymy relation was improved by dimentionality reduction.
