| 摘要: |
| 随着互联网发展和社交网络的大流行，虚假新闻和谣言能够迅速的传播并对公共产生巨大的负面影响。有效的识别出虚假新闻对阻断虚假新闻的快速传播有重要意义。由于新闻中除文字还常常包含图片等多模态的内容，近年来的工作多着重于研究利用深度神经网络提取文本和视觉的特征表示进行虚假新闻检测的方法。然而，却忽略了对新闻中所含的事实和知识的验证。在早期的工作中，研究人员通常将新闻中的实体和关系，与知识库中的实体关系进行比对，计算出新闻的可信度。这类工作基于简单的比对验证，往往难以覆盖海量信息内容的识别。但其利用外部知识库进行计算的思想仍有借鉴意义。我们提出一种知识融合的深度神经网络方法，同时提取新闻中的知识、文本、及视觉三个方面的特征表示，用于虚假新闻识别任务。具体来说，在我们的方法中，我们首先利用大型的开放式知识图谱来计算新闻中实体之间的关系，得到新闻的实体关系矩阵，并用卷积神经网络提取关系矩阵中的特征，作为新闻的知识特征表示。接着，将知识特征与文本和视觉特征相融合，进行虚假新闻的检测。我们的方法在两个大型的多模态数据集上进行验证。数据集分别是从微博采集的中文数据集和从推特上采集的英文数据集。丰富的实验证明了我们算法的有效性。在这两个数据集上，我们的方法都得到了最优的虚假新闻检测结果。 |
| 关键词: 虚假新闻识别 知识图谱 深度神经网络 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2023.06.11 |
| 投稿时间：2020-12-01修订日期：2021-02-24 |
| 基金项目:国家自然科学基金项目青年科学基金项目 |
|
| Multi-Domain Fake News Detection Fusing Knowledge Information |
|
WANG Xi, HAN Jizhong
|
| (Institute of Information Engineering，Chinese Academy of Sciences) |
| Abstract: |
| Fake news and rumors can spread rapidly with the development of Internet and the popularity of social network, and cause a huge negative impact on the public. Fake news detection is very important to blocking the spread of fake news. Since the wildly used of multi-media content in addition to text in posts, recent approaches had paid more attention to how to extract the feature representations of text and vision by using deep neural network. However, they ignored the verification of facts and knowledge information contained in news. In the early fact-checking articles, researchers usually matched the entity and relationship in news with the entity relationship in knowledge base, and calculated the credibility of news. This kind of works were based on simple comparison and verification, which were hard to cover large scale multi-media data. However, the idea of using external knowledge for calculation still is worth learning from. In this paper, we propose a knowledge fusion based deep neural network method, which extracts three domains features such as textual, visual and knowledge information in news for fake news detection task. Specifically, we first use the large open knowledge graphs to obtain the relationship between entities in news, and get the entity relationship matrix of news. Then we obtain the knowledge feature representation of news by extracting the features of the relation matrix with convolution neural network. Finally, knowledge features can be integrated with text and visual features to detect fake news. Our method is validated on two large multi-modal datasets: The Chinese dataset collected from Weibo and the English dataset collected from Twitter. Extensive experiments show the significant improvement of using our al-gorithm, and our model achieved the state-of-the-art performance on both two datasets. |
| Key words: Fake news detection, Knowledge graph, Deep neural network |
