ЛОГIКО-ОНТОЛОГIЧНА РЕКОНСТРУКЦIЯ ДИСКУРСИВНОÏ СТРУКТУРИ НАУКОВОГО ТЕКСТУ ТА ÏÏ РЕАЛIЗАЦIЯ В ШI-СИСТЕМI РЕЦЕНЗУВАННЯ

Context. The growing number of scientific publications and the emergence of tools based on large language models (LLMs) highlight the need for automated verification of the structural quality of scientific texts. Most existing solutions focus on surfacelevel linguistic analysis and do not account for logical-discursive integrity – specifically, whether the text includes a hypothesis, method, results, conclusions, and whether these elements are connected by normative relationships.Objective. The aim of this study is to develop an ontology-driven approach for the formalized verification of scientific text structures by constructing an ontological knowledge graph and evaluating its compliance with a predefined normative model of scientific discourse.Method. A model is proposed based on two interrelated ontologies: “Scientific Publication” (defining node types and their roles) and “Reviewing” (defining logical-discursive requirements). The text is represented as a graph where nodes are formed through semantic markup using an LLM, and connections are verified according to a set of normative rules. A specialized GPT agent capable of dynamically applying ontological knowledge during analysis and review generation is employed for implementation.Results. The model enables automatic detection of discourse structure violations: the absence of key elements, logical discontinuities, substitution of scientific novelty with practical significance, and incorrect interpretation of results. The proposed metrics quantitatively capture the level of structural completeness and consistency. Provided examples of graphs and reviews demonstrate that the system can detect non-obvious, latent logical inconsistencies even in formally complete texts.Conclusions. The scientific novelty of the study lies in introducing the ontological graph as an interpretable model of scientific argumentation, used in tandem with a large language model. The practical significance lies in establishing a foundation for semiautomated reviewing, structural analysis of publications, and academic writing training. The methodology is scalable to other genres of scientific texts and can potentially be integrated into editorial platforms.