首页|基于异构图分层学习的细粒度多文档摘要抽取

基于异构图分层学习的细粒度多文档摘要抽取

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
多文档摘要抽取的目标是在多个文档中提取共有关键信息，其对简洁性的要求高于单文档摘要抽取。现有的多文档摘要抽取方法通常在句子级别进行建模，容易引入较多的冗余信息。为了解决上述问题，提出一种基于异构图分层学习的多文档摘要抽取框架，通过层次化构建单词层级图和子句层级图来有效建模语义关系和结构关系。针对单词层级图和子句层级图这2个异构图的学习问题，设计具有不同层次更新机制的两层学习层来降低学习多种结构关系的难度。在单词层级图学习层，提出交替更新机制更新不同的粒度节点，以单词节点为载体通过图注意网络进行语义信息传递;在子句层级图学习层，提出两阶段分步学习更新机制聚合多种结构关系，第一阶段聚合同构关系，第二阶段基于注意力聚合异构关系。实验结果表明，与抽取式基准模型相比，该框架在Multi-news数据集上取得了显著的性能提升，ROUGE-1、ROUGE-2和ROUGE-L分别提高0。88、0。23和2。27，消融实验结果也验证了两层学习层及其层次更新机制的有效性。
Fine-Grained Multi-Document Summarization Extraction Based on Heterogeneous Graph Hierarchical Learning
The objective of multi-document summarization extraction is to extract common key information from multi-documents.Multi-document summarization extraction requires higher simplicity than single-document summarization extraction.The existing multi-document summarization extraction method typically has a performance bottleneck at sentence-level modeling,which is easy to introduce more redundant information.This study proposes a multi-document summarization extraction framework based on heterogeneous graph hierarchical learning to construct word-and subsentence-level subgraphs to effectively model semantic and structural relationships.The framework solves the difficulty of learning multiple structural relationships by proposing two learning layers with different updating mechanisms to learn these two heterogeneous graphs.First,in the learning layer of the word hierarchy graph,an alternate updating mechanism is proposed to update nodes of different granularity,and semantic information is transmitted through the graph attention network with the word node as the carrier.At the learning level of clause hierarchy graph,a two-stage stepwise-learning updating mechanism is proposed to aggregate various structural relationships.The first stage aggregates isomorphic relationships,and the second stage aggregates isomeristic relationships based on attention aggregation.The experimental results show that compared with the extractive benchmark model,the proposed framework achieves significant improvement in Multi-news dataset.ROUGE-1,ROUGE-2,and ROUGE-L increased by 0.88,0.23,and 2.27,respectively.The ablation experiment results also demonstrate the effectiveness of the two learning layers and the hierarchical update mechanism.

extractive multi-document summarizationfine-grained modelingheterogeneous graphhierarchical learningsemantic relationstructural relation

翁裕源、许柏炎、蔡瑞初

展开 >

广东工业大学计算机学院,广东 广州 510006

抽取式多文档摘要 细粒度建模 异构图 分层学习 语义关系 结构关系

科技创新2030—"新一代人工智能"重大项目国家优秀青年科学基金国家自然科学基金国家自然科学基金国家自然科学基金

2021ZD011150162122022618760436197605262206064

2024

10.19678/j.issn.1000-3428.0066794

计算机工程
华东计算技术研究所　上海市计算机学会

计算机工程

CSTPCD北大核心
影响因子：0.581
ISSN：1000-3428
年,卷(期)：2024.50(3)
  • 35