A comparative analysis of methods for automatic summarization of scientific texts in Russian

Abstract

Automatic summarization of scientific texts remains a pressing problem in natural language processing, particularly for Russian-language corpora, where the application of modern methods is limited by the lack of specialized models and datasets. This paper conducts a systematic comparative analysis of extractive and abstractive summarization models using a Russian-language multimodal dataset of scientific articles. A comprehensive set of metrics was used to assess quality: classical statistical metrics (ROUGE, TF-IDF) and modern neural semantic metrics (BERTScore, cosine measure between embeddings). The results show that extractive methods (TextRank, LexRank) provide high lexical accuracy (ROUGE-1 F1 = 0.38; TF-IDF = 0.87), while neural network abstractive models (rut5_base_sum_gazeta, mbart_ru_sum_gazeta) demonstrate superiority in semantic metrics (BERTScore F1 = 0.70; cosine similarity = 0.79), approaching the level of human summarization. The study suggests that the choice of model should be determined by the specific task: extractive methods are preferable for formalized texts, while abstractive ones provide a more natural and flexible presentation. doi 10.54708/19926502_2026_30211264