METHODOLOGY FOR AUTOMATED ASSESSMENT OF THEMATIC RELATEDNESS OF SCIENTIFIC PUBLICATIONS USING SEMANTIC ANALYSIS

Abstract

Relevance. Automated assessment of thematic relatedness between doctoral theses and profiles of potential experts is needed for transparent and reproducible selection of official reviewers and composition of one -time specialized councils; the open NAQA.Svr data make such assessment technically feasible. Object of research: methods and tools for constructing semantic profiles of theses and researchers and for ranking them in a shared vector space. Purpose of the article. To develop and empirically validate a methodology for automated assessment of thematic relatedness of scientific publications based on semantic analysis. Research results. We assembled a corpus of 259 theses, 662 researcher profiles, and 3345 publications; texts were normalized, and publication titles were standardized using a large language model. Semantic representations were obtained with an OpenAI model. We compared two variants of thesis profiling (full description; keywords only) and two variants of researcher profiling (by publications; by keywords). Quality was evaluated against actual reviewer assignments using the metrics NDCG, Hit@k, and NDCG-lift. The best results were consistently achieved by the combination “thesis by keywords” and “researcher by publications,” which yields the highest top-rank accuracy and the largest gain over random selection. Conclusions. Combining a keyword-based thesis profile with a publication-aggregated researcher profile provides the best balance of accuracy and robustness and is practically suitable for preliminary reviewer selection. The proposed methodology is reproducible, scalable, and aligned with open data; promising directions include incorporating temporal weights of publications, handling bilingual terminol ogy, and accounting for procedural constraints in final assignments.