International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 08 | Aug 2025
p-ISSN: 2395-0072
www.irjet.net
Scientific Paper Summarizer Using BERT, Sci-BERT, and BART Mr.Nikhil Mane1, Dr. Rudragoud S. Patil 2, Prof. Shubhada S. Kulkarni3 1 Department of Computer Science and Engineering, Gogte Institute of Technology,
Belagavi– 590008, Karnataka, India
2Department of Computer Science and Engineering, Gogte Institute of Technology,
Belagavi– 590008, Karnataka, India
3Department of Computer Science and Engineering, Gogte Institute of Technology,
Belagavi– 590008, Karnataka, India ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The surge in scientific publications has made it
sections such as Introduction, Methodology, Results, and Conclusion each serve distinct purposes. Treating a paper as a single block of text often results in summaries that fail to capture section-specific context and key findings.
increasingly difficult for researchers to quickly extract essential insights from large volumes of academic literature. This paper proposes a hybrid, section-wise summarization system that merges the precision of extractive models—BERT and SciBERT—with the fluency of the abstractive model BART. The approach first segments input documents, whether PDF or plain text, into standard research sections such as Abstract, Methodology, and Conclusion. Extractive models capture the most relevant and domain-specific content, while BART rephrases the information into clear, well-structured summaries. The framework’s performance is evaluated using ROUGE-1, ROUGE-2, and ROUGE-L metrics. Results indicate that BERT and SciBERT provide strong factual accuracy, with SciBERT performing better on specialized terminology, whereas BART delivers superior readability and narrative flow. Overall, the hybrid approach strikes an effective balance between technical accuracy and linguistic clarity, offering a practical tool for academic and research-oriented summarization.
To address this limitation, this study presents a sectionwise summarization framework that integrates the strengths of both extractive and abstractive strategies. BERT and SciBERT are employed for extractive summarization to ensure factual precision and domain-specific relevance, while BART is used for abstractive summarization to enhance readability and coherence. By processing each section independently, the system preserves structural integrity while providing summaries that are both technically accurate and accessible to a wider audience. The proposed system is designed to accept research papers in PDF or plain text format, automatically detect standard sections, and generate high-quality summaries for each. Performance is evaluated using ROUGE-1, ROUGE-2, and ROUGE-L metrics, enabling a comprehensive comparison of model effectiveness. Experimental results demonstrate that extractive models excel in factual accuracy, with SciBERT performing better on specialized terminology, while BART consistently produces more fluent and engaging summaries. This combination offers a balanced solution that meets the needs of both technical experts and general readers.
Keywords: Text Summarization, BERT, SciBERT, BART, Natural Language Processing, Extractive Summarization, Abstractive Summarization, ROUGE Metrics.
1.INTRODUCTION The rapid expansion of digital libraries and open-access repositories has led to an unprecedented surge in the volume of scientific literature published each year. While this growth enriches the academic landscape, it also creates significant challenges for researchers, educators, and professionals who must navigate vast amounts of information to extract relevant insights. Traditional literature review methods are often time-consuming, requiring meticulous reading of lengthy, complex, and highly technical documents.
2.PROBLEM STATEMENT The volume of scholarly publications is growing at an unprecedented pace, making it increasingly difficult for researchers, academicians, and industry professionals to stay updated with the latest developments in their fields. Reading and comprehending full-length research papers is a timeintensive process, especially when working under tight deadlines for literature reviews, project planning, or academic referencing.
Automatic text summarization has emerged as a promising solution to this problem by condensing lengthy research articles into concise, informative summaries. However, many existing summarization approaches overlook the structured nature of academic writing, where
© 2025, IRJET
|
Impact Factor value: 8.315
Existing summarization tools offer partial relief but suffer from notable limitations. Many rely solely on extractive methods, which directly select sentences from the source text without rephrasing or improving readability. Such
|
ISO 9001:2008 Certified Journal
|
Page 499