Deep Attention Model with Multiple Features for Rumor Identification

With the rapidly development of social networks and advances in natural language processing (NLP) techniques, rumors are extremely common and pose potential threats to community. In recent years, massive efforts are working on detecting rumors by using various techniques like simply investigating the content of texts, exploring the abnormality of propagation. However, these techniques are not ready to fully tackling this emerging threats due to the dynamic variations of rumors in a period of time. In this paper, we observed that the user feedback provides a clean signal for determining the trend of rumors, thus we combine the text content and the improved representation of network topology to characterize the dynamic features of rumors in a period of time. In detection, we employ a deep attention model with proposed features for spotting the minor differences between legitimate news and rumors. Experimental results show that our approach give an accuracy more than 94.7% in detecting rumors and outperforms previous approaches. Our studies also give a new insight that user interactions could be working as an important asset in rumor identification.