《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (12): 3820-3828.DOI: 10.11772/j.issn.1001-9081.2024121751

• 人工智能 • 上一篇    下一篇

基于多图神经网络和图对比学习的科学文献摘要模型

赵红燕, 郭力华, 刘春霞, 王日云   

  1. 太原科技大学 计算机科学与技术学院,太原 030024
  • 收稿日期:2024-12-12 修回日期:2025-03-02 接受日期:2025-03-04 发布日期:2025-03-13 出版日期:2025-12-10
  • 通讯作者: 刘春霞
  • 作者简介:赵红燕(1977—),女,山西运城人,副教授,博士,CCF会员,主要研究方向:自然语言处理、智能信息处理
    郭力华(1999—),男,山西运城人,硕士研究生,主要研究方向:长文档摘要生成
    刘春霞(1977—),女,山西大同人,副教授,硕士,CCF会员,主要研究方向:智能软件工程技术
    王日云(1998—),男,山西大同人,硕士研究生,主要研究方向:知识图谱推理。

Scientific document summarization model based on multi-graph neural network and graph contrastive learning

Hongyan ZHAO, Lihua GUO, Chunxia LIU, Riyun WANG   

  1. College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan Shanxi 030024,China
  • Received:2024-12-12 Revised:2025-03-02 Accepted:2025-03-04 Online:2025-03-13 Published:2025-12-10
  • Contact: Chunxia LIU
  • About author:ZHAO Hongyan, born in 1977, Ph. D., associate professor. Her research interests include natural language processing, intelligent information processing.
    GUO Lihua, born in 1999, M. S. candidate. His research interests include long document summarization generation.
    LIU Chunxia, born in 1977, M. S., associate professor. Her research interests include intelligent software engineering technology.
    WANG Riyun, born in 1998, M. S. candidate. His research interests include knowledge graph reasoning.
  • Supported by:
    Surface Program of Natural Science Foundation of Shanxi Province(202203021211199);Open Foundation of Key Laboratory of Shanxi Province(CICIP2022004);Doctor Research Startup Foundation of Taiyuan University of Science and Technology(20212075)

摘要:

长文档摘要生成面临句间关系的捕捉、长距离依赖及文档信息的高效编码与提取等难题,一直是自然语言处理领域的一个难点任务。同时,科学文献通常包含多个章节和段落,具有复杂的层次结构,使科学文献的摘要生成任务更具挑战性。针对以上问题,提出一种基于多图神经网络(GNN)和图对比学习(GCL)的科学文献摘要模型(MGCSum)。首先,对于输入的文档,通过同构GNN和异构GNN分别建模句内与句间关系,以生成初始句子表示;其次,将这些句子表示馈送到一个多头超图注意网络(HGAT),并在其中利用自注意机制充分捕捉节点和边之间的关系,从而进一步更新和学习句间的表示;再次,引入GCL模块增强全局主题感知,从而提升句子表示的语义一致性和区分度;最后,采用多层感知器(MLP)和归一化层计算一个得分,用于判断句子是否应被选为摘要。在PubMed和ArXiv数据集上的实验结果表明,MGCSum模型的表现优于多数基线模型。具体地,在PubMed数据集上,MGCSum模型的ROUGE-1、ROUGE-2和ROUGE-L分别达到了48.97%、23.15%和44.09%,相比现有的先进模型HAESum (Hierarchical Attention graph for Extractive document Summarization)分别提高了0.20、0.71和0.26个百分点。可见,通过结合多GNN和GCL,MGCSum模型能够更有效地捕捉文献的层次结构信息和句间关系,提升了摘要生成的准确性和语义一致性,展现了它在科学文献摘要生成任务中的优势。

关键词: 科学文献摘要, 抽取式摘要, 图神经网络, 超图注意网络, 图对比学习

Abstract:

Long document summarization generation faces challenges such as capturing inter-sentence relationships, long-range dependencies, and efficient encoding and extraction of document information, which has always been a difficult task in the field of natural language processing. At the same time, scientific documents, characterized by multiple chapters and paragraphs with complex hierarchical structures, further increases the difficulty of the summarization task of scientific documents. To address these issues, a scientific document Summarization model based on Multi-Graph Neural Network (GNN) and Graph Contrastive Learning (GCL) (MGCSum) was proposed. Firstly, for a given input document, intra-sentence and inter-sentence relationships were modeled by using homogeneous GNN and heterogeneous GNN, respectively, so as to generate initial sentence representations. Secondly, these sentence representations were fed into a multi-head HyperGraph ATtention network (HGAT), where self-attention mechanisms were used to fully capture relationships between nodes and edges, thereby updating and learning inter-sentence representations. Thirdly, a GCL module was introduced to enhance global topic awareness, thereby improving the semantic consistency and discriminability of sentence representations. Finally, a Multi-Layer Perceptron (MLP) and a normalization layer were applied to calculate a score for determining whether a sentence should be selected for summarization. Experimental results on the PubMed and ArXiv datasets indicate that the MGCSum outperforms most baseline models. Specifically, on the PubMed dataset, MGCSum achieves the ROUGE-1, ROUGE-2, and ROUGE-L of 48.97%, 23.15%, and 44.09%, respectively, with improvements of 0.20, 0.71, and 0.26 percentage points, respectively, compared to the existing state-of-the-art model HAESum (Hierarchical Attention graph for Extractive document Summarization). It can be seen that by integrating multi-GNN and GCL, MGCSum captures hierarchical structural information and inter-sentence relationships more effectively, enhancing the accuracy and semantic consistency of summarization, and demonstrating its advantages in scientific document summarization tasks.

Key words: scientific document summarization, extractive summarization, Graph Neural Network (GNN), HyperGraph ATtention network (HGAT), Graph Contrastive Learning (GCL)

中图分类号: