Article outline

The development of electronic medical records, clinical information systems, and digital libraries has resulted in the creation of large volumes of medical texts, much of which are presented in an unstructured form. Automatic analysis of such information requires not only detecting references to diseases, medications, anatomical structures, and other medical entities, but also correctly interpreting them. This raises the challenge of linking biomedical terms to a knowledge base, which is essential for building medical information systems, analyzing scientific publications, searching for clinical information, and supporting medical decision-making systems. The most widely used knowledge base for standardizing medical terminology is the Unified Medical Language System (UMLS), which comprises millions of concepts and tens of millions of synonyms. However, the large vocabulary, the presence of numerous synonyms, the homonymy of medical terms, and the limited coverage of the Russian language significantly complicate the task of automatic entity linking.

The goal of this work is to develop a method for linking biomedical terms to UMLS concepts, ensuring high-quality processing of both English- and Russian-language medical texts. To achieve this goal, an approach combining semantic search, automatic translation, and additional re-ranking of identified candidates is proposed.

The BioNNE-L dataset, prepared as part of the BioASQ 2025 competition and based on the NEREL-BIO corpus, was used as the experimental base. The dataset includes three tracks: English, Russian, and multilingual, allowing us to evaluate the method's performance in both monolingual and multilingual problem settings.

The proposed method consists of several sequential steps. During the knowledge base preparation stage, terms, concepts, and their semantic types are extracted from UMLS. Embeddings are then pre-calculated for all terms and a FAISS index is constructed, enabling rapid search for the closest candidates among millions of records. To obtain vector representations, specialized models are used: SapBERT for English and BERGAMOT for Russian and multilingual tracks. When processing Russian-language terms, automatic translation into English is additionally performed, allowing for more comprehensive English-language coverage of UMLS. After searching for the closest candidates, they are re-ranked using cosine similarity, Jaccard coefficient, Levenshtein distance, and the semantic type of the entity. The final step is aggregation of the found terms by UMLS concept identifiers using the LogSumExp function.

Experimental evaluation was conducted using the Accuracy@1, Accuracy@5, and Mean Reciprocal Rank metrics. The proposed method demonstrated competitive results in all three competition tracks. Accuracy@1 was 0.68 for the multilingual track, 0.69 for English, and 0.71 for Russian. These results are comparable with the best solutions submitted by BioNNE-L competition participants.

An error analysis was also conducted to identify the main causes of the decline in linking quality. For the Russian language, the greatest challenges stem from the limited coverage of UMLS terms in Russian, meaning the correct concept is not always present among the returned candidates. For English, the main problem is the homonymy of medical terms and abbreviations, where a single mention may correspond to several different concepts. A separate analysis revealed that medical abbreviations represent the most challenging category, especially when there is no exact match with a term in the knowledge base.

The results confirm the effectiveness of the proposed approach for linking biomedical terms to UMLS concepts. The developed method can be used in the creation of medical information systems and medical text analysis systems. Promising areas for further work include expanding the Russian-language portion of the knowledge base and taking into account the context of the mention when selecting the most appropriate concept.