Hierarchical Data Augmentation for Rumor Verification on Twitter

Unlimited dissemination of rumors in social media has a tremendous negative impact on our society. To address this issue, many rumor verification models have been proposed and achieved reasonable verification performance. However, the imbalanced data distribution between samples heavily limit the further prosperity of the deep learning-based models. To alleviate challenges, we propose a novel hierarchical data augmentation method for the rumor verification task (termed as HDA-RV), which consists two data augmentation methods (tweet-level and thread-level data augmentation). Tweet-level data augmentation simulates the noise of text information in social media and thread-level data augmentation corresponds to the noise of the propagation structure in social networks. Experiments on the PHEME dataset show that our method can effectively alleviate the problem of data imbalance.