引入主题节点的异构图舆情摘要方法  

Heterogeneous Graph Opinion Summary Method Which Introduced Topic Nodes

作　　者：宝日彤 曾淼瑞 孙海春 BAO Ri-tong;ZENG Miao-rui;SUN Hai-chun(Department of Information Network Security,People s Public Security University of China,Beijing 100038,China;Key Laboratory of Security Technology&Risk Assessment,Beijing 100026,China)

机构地区：[1]中国人民公安大学信息网络安全学院,北京100038 [2]安全防范技术与风险评估公安部重点实验室,北京100026

出　　处：《科学技术与工程》2024年第23期9965-9972,共8页Science Technology and Engineering

基　　金：公安部技术研究计划项目(2020JSYJC22);中央高校基本科研业务费专项资金(2022JKF02015)。

摘　　要：微博等社交软件承载着网民对社会舆论事件的不同观点,如何在海量主题评论中识别出有价值的信息已经成为重要课题。提出了一种基于异构图的舆情摘要方法,有效提取热点事件的主流观点,便于引导化解互联网舆情危机。针对多文档摘要任务中难以捕捉跨文档语义关系的难点问题,将主题节点引入评论句子图从而挖掘出输入文档间的潜在语义关联。具体地,抽取评论的主题并构建包含主题节点的异构图模型,利用图注意力机制进行不同粒度节点语义信息的交互,最后结合最大边界相关算法进行候选摘要句子的抽取。实验结果显示,改进模型在英文通用Multi-News数据集上Rouge1、Rouge2、,RougeL分数分别提升了0.46%、0.46%、0.48%;与已有Textrank、Sumpip等热点模型对比,在自制微博评论数据集上该模型性能达到最好。Social applications such as microblogging carry different views of internet users on social opinion events,and how to identify valuable information in the massive amount of thematic comments has become an important issue.An opinion summarization method based on heterogeneous graphs was proposed,which effectively extracted the prevailing viewpoints of hot public opinion events to facilitate the guidance of resolving internet public opinion crises.In order to address the challenging problem of capturing cross-document semantic relationships in the multi-document summarization task,topic nodes were introduced into the comment sentence graph to mine the potential semantic associations among the input documents.Specifically,the topics of comments were extracted to construct a heterogeneous graph model where graph attention mechanism was used to interact with the semantic information of nodes at different granularities,and finally,the maximum bounded correlation algorithm was combined to extract candidate summary sentences.The results show that the improved model improves the Rouge1,Rouge2,and RougeL scores by 0.46%,0.46%,and 0.48%on the English general Multi-News dataset respectively.Comparing with the existing hotspot models such as Textrank,Sumpip and so on,the model achieves the best performance on the self-made microblog comment dataset.

关 键 词：多文档摘要 舆情摘要 主题节点 图注意力机制 微博评论摘要 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]