Complex Network Analysis of Distributional Semantic Models

Akira UtsumiThe University of Electro-Communications

Abstract

A number of studies on network analysis have found the small-world and scale-free properties in the network of free word association, which reflects human semantic knowledge. Nevertheless, there have been very few attempts to apply network analysis to distributional semantic models (DSMs), despite the fact that DSMs have been extensively studied as a model of human semantic knowledge. In this paper, therefore, we analyze the small-world and scale-free properties of DSM networks. We demonstrate that DSM networks exhibit the same properties as the word association network. Especially, we show that DSM networks have the distribution of the number of connections that follows the truncated power law, which is also observed in the association network. This result indicates that DSMs provide a plausible model of semantic knowledge. Furthermore, we propose a modified version of Steyvers and Tenenbaum's (2005) growing network model, which involves the processes of semantic differentiation and experiential correlation. This model can better explain different distributions generated by various DSM implementations.

Files

Complex Network Analysis of Distributional Semantic Models (263 KB)



Back to Table of Contents