MANIN: Multi-Attention Neural Interaction Networks for Social Rumor Detection

With the fast-changing development of emerging online media, the accurate detection and timely suppression of rumors in online social networks becomes a vital task for maintaining social security and purifying public networks. Most existing works rely only on monotonous text content and shallow semantic information, and there is a lack of critical attention and mining for potential user relationships formed in the process of rumor spreading. To better improve these problems by employing attention mechanisms, we propose a novel deep model, Multi-Attention Neural Interaction Networks, termed as MANIN, which mainly consists of a self-attention-based encoder for extracting rich semantic representations from texts, a post-comment co-attention mechanism for studying the interaction between the original post and its comments, and a graph attention neural network for mining potential user interactions. We have conducted numerous experiments on real-world datasets and verified their validity, and the results show that the model proposed by us outperforms the baselines with an accuracy rate of 81.9% and can detect rumors at early stages.