基于图模型的多文档摘要生成算法  

Multi-document Summary Generation Algorithm Based on Graph Model

作　　者：张云纯 张琨[1] 徐济铭 袁卫平 蔡颖 高雅 ZHANG Yunchun;ZHANG Kun;XU Jiming;YUAN Weiping;CAI Ying;GAO Ya(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China;Department of Internet Information,Jiangsu Branch of National Computer Network and Information Security Management Center,Nanjing 210019,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094 [2]国家计算机网络与信息安全管理中心江苏分中心互联网信息处,南京210019

出　　处：《计算机工程与应用》2020年第16期124-131,共8页Computer Engineering and Applications

基　　金：南京理工大学自主科研专项(No.30918012204);江苏省研究生科研创新计划(No.KYCX19_0307)。

摘　　要：提出一种基于图模型的多文档摘要生成算法,对海外大量新闻文档进行主题划分,并提取每个主题的摘要。利用传统的基于图模型方法得到的摘要,其冗余度较高,亦不能够充分考虑新闻文本时效性强、主题明确的特征。在文本特征向量化方面,引入了热度系数,改进了传统的TF-IDF算法。在主题的划分方面,采用基于密度的两阶段聚类方法,改进了传统的基于K-Means进行聚类的方法的不足,同时对文本进行更明确、更具层次性的主题划分。在摘要抽取方面,为句子设计了符合新闻文本特征的重要度计算公式。实验结果表明,基于图模型的自动文本摘要生成算法的效果优于传统算法。A multi-document text summarization algorithm based on graph model is proposed to divide a large number of overseas news documents into themes and extract the abstracts of each theme.Abstract generated by the traditional method of abstract generation based on graph model has high redundancy and fails to fully consider the timeliness and clear theme of news text.In the aspect of text feature vectorization,exponential attenuation coefficient is introduced to improve the traditional TF-IDF algorithm.In terms of theme classification,the density-based fast clustering method is adopted,which improves the shortcomings of the traditional K-Means clustering method.At the same time,the two-stage text clustering is used to divide the text into more explicit and hierarchical themes.In the aspect of abstract extraction,this algorithm designs a formula for sentence significance which conforms to the characteristics of news text.Experimental results show that the improved algorithm is superior to the traditional algorithm.

关 键 词：文本聚类 自动摘要 图模型 多特征融合 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]