Networks generated from natural language text

Abstract

The study of large-scale characteristics of graphs that arise in natural language processing is an essential step in finding structural regularities. Structure discovery processes have to be designed with an awareness of these properties. Examining and contrasting the effects of processes that generate graph structures similar to those observed in language data sheds light on the structure of language and its evolution. In this chapter, we examine power-law distributions and small world graphs (SWGs) originating from natural language data. There are several reasons for the special interest in these structures. 1.Power laws appear in many rank-frequency statistics. Furthermore, we can construct graphs with words as nodes and use various rules to introduce edges between words. In many cases, this results in SWGs, which again often have a power-law distribution for their node degrees.2.SWGs appear in many other real world data, like social networks of many kinds, in the link structure of the World Wide Web or in traffic networks. It is interesting to analyze all these networks in more detail to identify similarities and differences.3.From an application-driven view, SWGs allow effective clustering strategies in nearly linear time. Because these clusters are often related to the growth process of the underlying graph, they are often meaningful. In the case of natural language these clusters usually reflect semantic and/or syntactic structures.