北京邮电大学学报

  • EI核心期刊

北京邮电大学学报 ›› 2023, Vol. 46 ›› Issue (4): 97-102.

• 研究报告 • 上一篇    下一篇

基于层次门控交互融合网络的谣言检测方法

苏兴 ,禹可, 吴晓非   

  1. 北京邮电大学 人工智能学院
  • 收稿日期:2022-07-10 修回日期:2022-09-05 出版日期:2023-08-28 发布日期:2023-08-24
  • 通讯作者: 禹可 E-mail:yuke@bupt.edu.cn
  • 基金资助:
    国家自然科学基金项目

Gated Interactive Fusion Network for Rumor Detection

SU Xing,  YU Ke,  WU Xiaofei   

  1. School of Artificial Intelligence, Beijing University of Posts and Telecommunications
  • Received:2022-07-10 Revised:2022-09-05 Online:2023-08-28 Published:2023-08-24
  • Supported by:
    the National Natural Science Foundation of China

摘要: 针对现有谣言检测方法对多特征做处理时因特征间差异导致特征冲突的问题,提出了一种基于层次门控交互融合网络的谣言检测方法首先,利用一阶门控对原贴和评论的语义特征和情感特征做特征增强,然后,利用二阶门控对增强特征做跨语义特征融合,以解决特征融合时由于不同特征之间的差异引入噪声的问题在公开的Weibo 数据集和自建的 Weibo22 数据集上,所提方法的检测正确率分别为 96.71% 97.36% 。 与检测性能最好的基线方法相比,检测正确率分别提高了 0.84% 1.31% ,训练时间分别减少了 53% 46% 。

关键词: 谣言检测 , 门控网络 , 特征融合

Abstract: To address the issue of feature conflicts caused by differences between features when the existing rumor detection methods deal with multiple features, a hierarchical gated interactive fusion network-based rumor detection method is proposed. First, the first-order gate unit is conducted to obtain the enhanced semantic and sentiment features of original posts and comments. Then, the second-order gate unit is used to perform cross-semantic feature fusion on the enhanced features to solve the problem of introducing noise due to differences between different features during feature fusion. On the public Weibo dataset and the self-built Weibo22 dataset, the detection accuracy of the proposed method is 96.71% and 97.36% , respectively. Compared with the baseline methods with the best detection performance, the detection accuracy of the proposed method is improved by 0.84% and 1.31% , respectively, and the training time is reduced by 53% and 46% , respectively.

Key words: rumor detection , gated network , feature fusion

中图分类号: