Quantifying hot topic dynamics in scientific literature: An information-theoretical approach

Abstract

Understanding the internal structure of scientific discourse is essential for tracking the evolution of research topics and their conceptual interdependencies. However, existing approaches such as dynamic topic modeling and neural topic models often fail to capture fine-grained semantic shifts among known concepts, or require substantial computational resources. Co-occurrence networks offer a more interpretable alternative, but typically rely on correlation-based weights that lack metric properties, preventing rigorous temporal comparison and topological interpretation. To address this gap, we introduce a metric-based framework for analyzing the evolving structure of concept networks in the scientific literature. Using 10,370 research articles (2010-2023) on international security from JSTOR and PORTICO, we compute the normalized variation of information (NVI) distances to construct annual concept networks with a well-defined geometric structure. We then quantify semantic change using velocity matrices and extract major trends using Minimum Spanning Tree (MST) analysis. Our results reveal that conceptual shifts are concentrated in temporally localized hubs and are not driven by co-occurrence frequency alone, but by contextual information and shared uncertainty between concept distributions. By introducing a scalable, interpretable, and mathematically grounded approach to tracking concept dynamics, this study contributes new tools for topic evolution analysis and offers insight into the structural organization and reconfiguration of knowledge over time.