融合丰富语义和全局传播的谣言检测方法

Rumor detection method integrating rich semantics and global propagation

关键词:  谣言检测  图神经网络  语义特征  全局传播  外部知识  注意力机制

Key words:  rumor detection  graph neural network  semantic feature  global propagation  external knowledge  attention mechanism

摘要:

现有谣言检测方法主要依赖文本语义特征和网络传播特征,但以短文本为主的源推文易导致语义特征不足,且用于提取传播特征的传播树易产生海量数据。为了解决上述问题,提出一种融合丰富语义和全局传播的谣言检测方法——多视图图神经网络。该模型利用源文本获取结构语义关系,借助外部知识提取潜在语义关系,通过源推文及其响应用户得到用户之间的全局传播关系,最后通过注意力融合机制自动学习不同视图的特征权重,实现信息自适应融合,提升谣言检测准确率。其中,采用Word2Vec对源推文的内容语义进行补充。实验结果表明:利用源文本、外部知识和响应用户分别构图,可有效捕获丰富语义信息和简洁全局传播关系；在公共数据集Twitter15和Twitter16上的性能优于一系列基线模型,准确率分别是90.2%和90.8%。结合消融实验分析,所提方法能够全面捕获源推文的丰富语义特征,且简洁有效地获取其全局传播关系,从而提高谣言检测的准确率。

Abstract:

Existing rumor detection methods mainly rely on text semantic features and network propagation features, but the source tweets dominated by short texts can easily lead to insufficient semantic features, and the propagation tree used to extract propagation features can generate a large amount of data. To solve these problems, we proposed a rumor detection method, namely multi-view graph neural network, which integrated rich semantics and global propagation. This model used source texts to get structural semantic relationships, utilized external knowledge to extract potential semantic relationships, and got the global propagation relationship among users by source tweets and their response users. Finally, it automatically learned the feature weights of different views through the attention fusion mechanism, achieving adaptive information fusion and improving the accuracy of rumor detection. Besides, it adopted Word2Vec to supplement the content semantics of source tweets. Experimental results show that using source texts, external knowledge, and response users to construct graphs, respectively, can effectively capture rich semantic information and concise global propagation relationships. The model outperforms a series of baseline models on the public datasets Twitter15 and Twitter16, with the accuracy rates of 90.2% and 90.8%, respectively. The analysis results from the ablation experiment show that the proposed method can comprehensively capture rich semantic features of the source tweets and effectively obtain the global propagation relationship in a concise manner, so as to improve the accuracy of rumor detection.