A multi-layer annotated corpus for information extraction in Russian clinical NLP.

Area of Science:

• Natural Language Processing (NLP)
• Bioinformatics
• Computational Linguistics

Background:

• Clinical exome sequencing reports contain vital genetic and phenotypic data.
• Extracting this information automatically is challenging due to unstructured text, especially in Russian.
• Limited annotated corpora exist for Russian genetic report analysis.

Purpose of the Study:

• To introduce GENEXOM, the first multi-level annotated corpus for Russian clinical exome sequencing reports.
• To facilitate automated biomedical information extraction from Russian genetic reports.
• To provide a resource for improving Named Entity Recognition (NER) and Relation Extraction (RE) in this domain.

Main Methods:

• Developed GENEXOM with 5,318 reports (318 authentic, 5,000 synthetic).
• Annotated 16 entity and 7 relation types aligned with genetic standards (HGVS, OMIM, ClinVar, ACMG/AMP).
• Fine-tuned transformer models (RuBERT, RuBioBERT, ModernBERT) for NER and RE tasks.

Main Results:

• Achieved substantial inter-annotator agreement (span-level F1-IAA = 0.83, macro κ = 0.79).
• ModernBERT demonstrated top performance: F1 = 0.88 for NER and F1 = 0.836 for RE.
• The corpus enables robust biomedical information extraction.

Conclusions:

• GENEXOM is a valuable resource for Russian medical NLP and biomedical research.
• It supports critical downstream tasks like variant interpretation and knowledge graph construction.
• The corpus and code are publicly available to advance research.