A time-scale representation of EEG Higuchi fractal dimension.

As a core track of global technological competition, the academic output in the AI field has grown rapidly over the past decade. Data from the Web of Science shows that a large number of English academic papers in the AI field were published in 2023, covering numerous sub-directions such as machine learning, computer vision, and natural language processing. The rapid accumulation of massive literature not only provides theoretical support for the field’s development but also causes an “information overload” problem: researchers struggle to quickly identify core research hotspots and track topic evolution trajectories, research management departments lack a quantitative basis for formulating R&D policies, and enterprises face difficulties in accurately deploying technological innovation directions. Among various text mining technologies, topic models are the most effective tool to mine implicit topic structures from massive unstructured academic texts, which can help solve the information overload problem. Text mining technology offers an effective solution to this problem. Among its core tools, topic models can mine implicit topic structures from unstructured text. For instance, Topaloğlu et al. applied the Latent Dirichlet Allocation (LDA) model to conduct topic clustering on machine learning research in the intensive care unit field, accurately identifying core technical directions. Yu et al. systematically revealed the topic distribution and hot trends in the AI field based on the LDA model, providing references for research topic selection. However, most current bibliometric studies in the AI field have two limitations: first, over-reliance on a single static model (e.g., LDA) for hotspot identification, which fails to capture the temporal evolution laws of topics, second, dynamic topic models (e.g., DTM) are prone to topic drift when used independently, leading to distorted evolution trajectory analysis. Therefore, constructing a combined model with both “static recognition accuracy” and “dynamic tracking capability” to conduct dynamic topic evolution analysis of English academic papers in the AI field is of great theoretical significance and practical value for revealing the development trends and iteration directions of domain topics. To address the key gaps in current bibliometric research on the AI field, this study focuses on three core objectives: (1) identifying the core research topics in the AI field over the past decade and their static clustering characteristics; (2) exploring the dynamic evolution laws of these core topics, including the classification and characteristics of emerging, declining, and continuously active topics; and (3) clarifying the coupling characteristics between topics, as well as the derivative, fusion, or substitution relationships that occur during their evolution.