COMPARATIVE EVALUATION MODEL BASED ON EXPERT SYSTEMS FOR ABSTRACTIVE SUMMARIZATION OF SCIENTIFIC TEXTS

Abstract

The growing number of scientific publications requires effective methods of summarizing scientific texts. Traditional extractive approaches do not allow preserving the logical structure, while abstract generalization methods are difficult to process highly structured documents. In this study, a comparative assessment model is proposed to improve the quality of abstract generalization of scientific texts. This study presents a comparative evaluation of BART-base and BART-large models. A corpus of 17,038 documents from arXiv and PubMed was used for fine-tuning. Evaluation with ROUGE and BLEU metrics shows that BART-large outperforms BART-base, achieving ROUGE-1 = 0.2224, ROUGE-2 = 0.0695, and BLEU = 0.29. In contrast, BART-base reached ROUGE-1 = 0.1932 and BLEU = 0.25. BART-large demonstrated better phrase-level coherence.