CovSumm: an unsupervised transformer-cum-graph-based hybrid document summarization model for CORD-19.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Biomedical Informatics

Background:

• The COVID-19 pandemic has led to an exponential increase in scientific publications, causing significant information overload for researchers and medical professionals.
• Staying updated with the latest COVID-19 research is crucial but challenging due to the sheer volume of published studies.

Purpose of the Study:

• To address the information overload in COVID-19 scientific literature by developing an effective single-document summarization model.
• To present and evaluate a novel hybrid unsupervised approach named CovSumm on the CORD-19 dataset.

Main Methods:

• A hybrid text summarization model, CovSumm, was developed, combining two extractive approaches: GenCompareSum (transformer-based) and TextRank (graph-based).
• The model was tested on 840 scientific papers from the CORD-19 dataset, specifically those published between January 1, 2021, and December 31, 2021.
• Sentence ranking for summary generation was based on the combined scores from the GenCompareSum and TextRank components.

Main Results:

• The CovSumm model achieved state-of-the-art performance on the CORD-19 dataset, as measured by the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) scores.
• The model obtained ROUGE-1 scores of 40.14%, ROUGE-2 scores of 13.25%, and ROUGE-L scores of 36.32%.

Conclusions:

• The proposed hybrid unsupervised approach, CovSumm, demonstrates superior performance compared to existing methods for single-document summarization of COVID-19 literature.
• CovSumm offers a promising solution for efficiently summarizing large volumes of scientific text, aiding researchers in staying current with critical findings.