On the shape of semantic space - what can we infer from large-scale statistical properties of texts?

Abstract

[eng] The large amount of digitized linguistic data opens up the unique possibility of using the methodology of complex systems to understand high-level human cognitive processes. Two such issues are i) the way we categorize the continuous space of real-world features into discrete concepts, and ii) the way we use language to copy a line a thought from one brain to another. In this work I address both questions by formulating a simple text generation model which reproduces the three major characteristic large-scale statistical laws of human language streams, namely Zipf’s law, Heaps’ law and Burstiness. Furthermore, the generation itself can be described as a random walk on a scale-free, highly clustered and low dimensional complex network, suggesting that this class of networks is appropriate as a minimal model of the semantic space. Entangling the global characteristics of the semantic space is an inevitable step towards analyzing texts as trajectories in such a space, with promising applications such as author or style identification, personal disorder diagnosis, or the evolution of cultural traits mirrored by text production characteristics.