A Statistical and Machine Learning Approach for Summarising Computer Science Research Papers

Abstract

Academics, researchers and students usually read a lot of papers for their research or to keep up-to-date with the latest works. The high number of papers available makes the process time-consuming. A solution is to summarise the papers and allow the reader to decide if the papers are relevant to their work and whether they require more attention. A system has been built to generate extractive summaries of computer science research papers. We demonstrate how the intrinsic statistical characteristics of computer science research papers such as the document length or the presence of certain keywords can help train a machine learning classifier model that can achieve state-of-the-art performance. Human and automatic evaluation using ROUGE has been carried out to measure performance. Results show that the proposed model performs better than TextRank and BERT on both human and automatic evaluation. It also does better than BART on human evaluation.