A Neural Rumor Detection Framework By Incorporating Uncertainty Attention On Social Media Texts

Automatic rumor detection technology has become a very urgent need, as rumors can arise and spread dauntingly fast in social media, which bring unforeseeable and devastating impacts. However, current approaches mainly capture the event semantics or user-based features for rumor detection, but neglect the uncertainty expressions that strongly indicate the unverified nature of a rumor. As a result, these methods perform suboptimal when the topics of being verified rumors are changing wildly. In this paper, we present a neural rumor detection framework, namely NERUD. In NERUD, both uncertainty semantics and the event semantics of a word are represented by the attention mechanisms to generate a rumor representation for rumor detection. Experiments were conducted on the benchmark dataset and the Chinese Rumor Corpus (CRC), and the results showed that our NERUD outperformed state-of-the-art approaches on CRC dataset, and the uncertainty semantics was proven effective on rumor detection task.