Topic Modeling of Scientific Texts Using BERTopic (Based on Scientific Abstracts on the Great Patriotic War published in 2014–2023)

Amid the continuous growth of the volume of scientific literature, the development of efficient tools for systematizing and automating the analysis of scientific texts has become a critical challenge. These tools are essential to ensure the accessibility, organization, and information retrieval across various scientific domains. Topic modeling is one of the most effective approaches for automated text analysis since it enables the identification of key themes within large collections of texts, facilitating classification and speeding up the search for relevant information. This is particularly important for analyzing scientific literature and tracking research trends. In this study, the authors consider the capabilities of the BERTopic algorithm for topic modeling of Russian-language scientific texts. The dataset consists of scientific articles on the Great Patriotic War, published in Russian academic journals over the past decade (2014–2023). The relevance of this work arises from the limited research on the application of topic modeling algorithms to scientific texts, especially those that utilize large language models. There is a notable gap in the studies focusing on this approach. Additionally, there is an urgent need in more advanced methods for searching and selecting scholarly sources. This further highlights the importance of the current study. The paper provides a detailed description of the data collection, preprocessing, and topic modeling processes. The results reveal the most prevalent topics in academic works, offering valuable insights for researchers in computational linguistics, as well as for historians and scholars of Russian historiography. The approach to corpus building, combining the use of keyword markers, automated data collection tools, and expert validation, can be applied to other subject areas in similar studies. Topic modeling allowed for the identification of key themes within the corpus, along with the keywords that characterize them, and facilitated the creation of both hierarchical and dynamic representations of these topics. BERTopic’s strengths lie in its use of vector-based text representations from large language models, as well as its extensive tools for visualizing the resulting topics.