摘要
社交平台谣言检测问题通常以源帖文本,回复文本为谣言检测的判断依据。此外,用户相关数据也利于提高谣言检测准确率。根据文本内容和回复内容呈现的序列特性,个人资料和微博统计数据多维度的无序性,提出基于自注意力的卷积神经网络及用户信誉特征谣言检测方法。该方法利用自注意力和卷积神经网络对源帖以及回复文本进行词级和句子级别的二级编码获取文本语义特征和谣言事件回帖的时序特征,并通过自注意力和最大池化结合用户个人信息及微博统计数据编码用户信誉特征进行谣言检测。在取自微博和推特的两个公开数据集上实验表明:1.结合自注意力的卷积神经网络序列编码优于单一的卷积神经网络;2.用户信誉特征能有效提高谣言检测结果准确率。
The rumor detection problem of social platforms is usually based on the source post text and reply text. In addition,user-related data also helps improve the accuracy of rumor detection. Based on the sequence characteristics of text content and reply content, the multi-dimensional disorder of personal data and microblog statistics, this paper pro-posed a self-attention convolutional neural network and user credit feature rumor detection method. The method adopts self-attention and convolutional neural networks to perform word-level and sentence-level coding on source post and reply texts to obtain text semantic features and temporal features of a rumor event. User credit features and microblog statistics are encoded by self-attention and max pooling through user profiles. Experiments are conducted on two public datasets from Weibo and Twitter, and the results demonstrates that: 1. Convolutional neural network sequence coding combined with self-attention is superior to a single convolutional neural network;2. The user credit feature can effectively improve the accuracy of rumor detection.
作者
柳先觉
徐义春
董方敏
Liu Xianjue;Xu Yichun;Dong Fangmin(College of Computer and Inform ation,China Three Gorges University,Yichang 443002,China)
出处
《信息通信》
2020年第12期39-43,共5页
Information & Communications
关键词
自注意力机制
卷积神经网络
最大池化
用户资料
谣言检测
self-attention mechanism
convolutional neural network
max pooling
user profiles
rumor detection
