Human Versus Machine Intelligence in Natural Language Generation

This research article presents a comparative study assessing Natural Language Generation (NLG) models, specifically focusing on the English version of GPT-2 and human-generated texts. The study employs a complexity science framework to analyze the stochastic processes underlying the texts produced by both sources. It outlines a methodological approach that consists of two phases: an analysis phase and a synthesis phase. The analysis phase utilizes Multifractal Detrended Fluctuation Analysis, Recurrence Quantification Analysis, Zipf’s law, and approximate entropy to characterize long-term correlations and regularities in the texts. The synthesis phase constructs synthetic text descriptors for machine learning applications, demonstrating the grouping tendencies of different text types. The findings reveal significant trends in long-range correlations and recurrences, contributing to a deeper understanding of NLG systems and their capabilities. The research aims to enhance text classification, fake news detection, and plagiarism detection methodologies, providing insights into the complexities of human and machine-generated language.