• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学 ›› 2022, Vol. 44 ›› Issue (09): 1702-1710.

• 人工智能与数据挖掘 • 上一篇    

融合评论的多任务联合谣言检测方法

王繁1,2,郭军军1,余正涛1,2   

  1. (1.昆明理工大学信息工程与自动化学院,云南 昆明 650500;
    2.昆明理工大学云南省人工智能重点实验室,云南 昆明 650500)
  • 收稿日期:2020-12-23 修回日期:2021-04-12 接受日期:2022-09-25 出版日期:2022-09-25 发布日期:2022-09-25
  • 基金资助:
    国家重点研发计划(2018YFC0830101,2018YFC0830100);国家自然科学基金(61866020,61972186,61762056,61472168);云南省科技厅面上项目(2019FB082)

A multi-task joint rumor detection method combining comments

WANG Fan1,2,GUO Jun-jun1,YU Zheng-tao1,2   

  1. (1.Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500;
    2.Yunnan Key Laboratory of Aritficial Intelligence,Kunming University of Science and Technology,Kunming 650500,China)
  • Received:2020-12-23 Revised:2021-04-12 Accepted:2022-09-25 Online:2022-09-25 Published:2022-09-25

摘要: 目前,针对微博领域的谣言检测方法主要基于微博正文,同时辅以用户评论特征、传播特征等信息进行判定。然而已有方法没有考虑用户评论质量会直接影响谣言检测的性能,质量低的评论甚至会引入无用甚至负面的特征,进而对谣言检测的性能带来更大的影响。针对该问题,基于用户评论和谣言检测的关联性,首次提出一种考虑评论有效性,并基于多任务联合学习的谣言检测方法。首先将谣言检测作为主任务,用户评论相关性检测为辅助任务;然后采用门控机制和注意力机制过滤和选择有效的用户评论特征;最后基于自主构建的3万条疫情微博谣言数据集进行实验。实验结果表明,对用户评论进行筛选不仅可以提升谣言检测性能,还能对用户评论质量进行判定。

关键词: 谣言检测, 联合学习, 用户评论, 评论有效性

Abstract: At present, the rumor detection method for microblog field is mainly based on the microblog text itself, supplemented by information such as user comment characteristics and propagation characteristics. However, the current methods ignore the quality of user comments that may directly affect the performance of rumor detection and introduce useless or even negative features, exerting an impact on the performance of detection. In response to this problem, based on the relevance of user comments and rumor detection, a rumor detection algorithm that considers the effectiveness of comments is proposed. It considers the effectiveness of microblog comments while determining rumors, and rumor detection is implemented based on the the multi-task joint learning method. Firstly, rumor detection is taken as the main task, and user comment correlation detection is taken as the auxiliary task. Secondly, the gating mechanism and the attention mechanism are used to filter and select effective user comment features. Finally, experiments on the self-constructed dataset with 30,000 epidemic microblog rumors show that the screening of user comments can not only improve the performance of rumor detection, but also realize the judgment of the quality of user comments.

Key words: rumor detection, joint learning, user comment, comment validity