SAPGraph: Structure-aware Extractive Summarization for Scientific Papers with Heterogeneous Graph

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 575–586 November 20–23, 2022. ©2022 Association for Computational Linguistics 575 SAPGraph: Structure-aware Extractive Summarization for Scientific Papers with Heterogeneous Graph Siya Qi∗1 Lei Li†∗1 Yiyang Li1 Jin Jiang1 Dingxin Hu1 Yuze Li1 Yingqi Zhu1 Yanquan Zhou1 Marina Litvak 2 Natalia Vanetik2 1Beijing University of Posts and Telecommunications 2Shamoon College of Engineering {qsy,leili,kenlee,jiangjin}@bupt.edu.cn {hudingxin,lyzbupt,zhuyq,zhouyanquan}@bupt.edu.cn {marinal,natalyav}@sce.ac.il Abstract Scientific paper summarization is always chal- lenging in Natural Language Processing (NLP) since it is hard to collect summaries from such long and complicated text. We observe that previous works tend to extract summaries from the head of the paper, resulting in information incompleteness. In this work, we present SAP- Graph1 to utilize paper structure for solving this problem. SAPGraph is a scientific paper extractive summarization framework based on a structure-aware heterogeneous graph, which models the document into a graph with three kinds of nodes and edges based on structure information of facets and knowledge. Addi- tionally, we provide a large-scale dataset of COVID-19-related papers, CORD-SUM. Ex- periments on CORD-SUM and ArXiv datasets show that SAPGraph generates more compre- hensive and valuable summaries compared to previous works. 1 Introduction In recent years, scientific papers represented by COVID-19-related papers have shown an expand- ing growth in a short period, which produces in- formation overload and makes it difficult for re- searchers to follow. Automatic summarization can help researchers quickly focus on valuable infor- mation in the article and be updated about the lat- est research progress. The goal of automatic sum- marization is to condense a long text into a con- cise summary while retaining essential information. It evolves mainly in two directions: abstractive and extractive methods. Abstractive summariza- tion generates summaries which are rewritten and refined (Lewis et al., 2020; Zhang et al., 2020), while the extractive one selects text segments as summaries (Liu and Lapata, 2019; Nallapati et al., 2017; Zhong et al., 2020; S et al., 2021), which *The first two authors contributed equally. †Corresponding author. 1Available at: https://github.com/cece00/SAPGraph Bacille Calmette-Guérin (BCG) is an attenuated strain of Mycobacterium bovis currently used as a vaccine against tuberculosis. Global distribution and propagation of BCG has contributed to the in vitro evolution of the vaccine strain and is thought to partially account for the different outcomes of BCG vaccine trials. Previous efforts by several molecular techniques effectively identified large sequence polymorphisms among BCG daughter strains, but lacked the resolution to identify smaller changes. In this study, we have used a NimbleGen tiling array for whole genome comparison of 13 BCG strains. Using this approach, in tandem with DNA resequencing, we have identified six novel large sequence polymorphisms including four deletions and two duplications in specific BCG strains. Moreover, we have uncovered various polymorphisms in the phoP-phoR locus. Importantly, these polymorphisms affect genes encoding established virulence factors including cell wall complex lipids, ESX secretion systems, and the …

Content truncated to 500 words