Content representation for microblog rumor detection

Abstract

In recent years, various social network applications have emerged to meet users demand of social activity. As the biggest Chinese Microblog platform, Sina Weibo not only provides users with a lot of information, but also promotes the diffusion spread of rumors which generated huge negative social impacts. To quickly detect rumors from Sina Weibo, many research works focus on social attributes in social network. However, content play an important role in rumor diffusion, and it was ignored in many research works. In this paper, we use two different text representations, bag of words model and neural network language model, to generate text vectors from rumor contents. Furthermore, we compared performance of two text representations in rumor detection by using some state-of-the-art classification algorithms. From the experiments in 10, 000 Sina Weibo posts, we found that the best classification accuracy of bag of words model is over 90%, and the best classification accuracy of neural network language model is over 60%. It indicates that words of posts aremore useful than semantic context vectors representation in rumor detection.