The popularity of microblogging websites makes them important for information dissemination. The diffusion of large volumes of fake or unverified information could emerge and spread producing damage. Due to the ever-increasing volume of data and the nature of complex diffusion, automatic rumor detection is a very challenging task. Supervised classification and other approaches have been widely used to identify rumors in social media posts. However, despite achieving competitive results, only a few studies have delved into the nature of the problem itself in order to identify key empirical factors that allow defining both the baseline models and their performance. In this work, we learn discriminative features from tweets content and propagation trees by following their sequential propagation structure. To do this we study the performance of a number of architectures based on recursive neural networks conditioning for rumor detection. In addition, to ingest tweets into each network, we study the effect of two different word embeddings schemes: Glove and Google news skip-grams. Results on the Twitter 16 dataset show that model performance depends on many empirical factors and that some specific experimental configurations consistently drive to better results.
展开▼
