Improved Divide-and-Conquer Approach to Abstractive Summarization of Scientific Papers
Xin Shen, Wai Lam · 2022 4th International Conference on Natural Language Processing (ICNLP) · 2022
Recently, a simple and effective method called Divide ANd ConquER (DANCER) is proposed for automated abstractive summarization of scientific papers. However, we point out that DANCER still has several weaknesses. The most prominent one is that DANCER relies on ROUGE tool for the Section-to-Summary (SEC2SuMM) alignment. We address the weaknesses and propose a novel framework to improve DANCER. Our method includes a delicate training objective, which learns the SEC2SuMM alignment, and the section-level summarizer in a joint way. We investigate our framework on two publicly available scientific paper summarization benchmarks: PubMed and arXiv. Our method achieves significant improvements over DANCER and several competitive baselines.