Quantification and identification of authorial writing style through higher-order text network modeling and analysis

Abstract

Determining the true author of anonymized texts has important applications ranging from text classification and information extraction to forensic investigations. Despite substantial progress, current authorship identification solutions are limited to extracting straightforward semantic relationships in writing styles, lacking consideration for higher-order features among multiple vocabulary, phrases, or sentences in language structure. Here, we propose a novel approach based on hypernetwork theory to encode higher-order text features into a unified text hyper-network and investigate whether the hyper-order topological features of the text hyper-network contribute to revealing the author's stylistic preferences. Our results indicate that metrics of the text hyper-network, such as hyperdegree, average shortest path length, and intermittency, can capture more information about the author's writing styles. More importantly, in the author identification task of 170 novels, our method accurately distinguished the authorship of 81% of the novels, surpassing the accuracy of the method of using paired word relationships. This further highlights the importance of higher-order features in text analysis, beyond mere pairwise interactions of words.