Picture or it didn't happen: catch the truth for events

摘要

Pictures spreading on the Internet are essential for the authenticity of events. Each day, huge amounts of data are published on social media, and many of them are bound with a picture in order to increase their readability and reliability. On social media, the bound pictures often have little relevance to their context. The thing changes when it comes to events in our daily lives. The events on social media are often bound with the spot shooting. People are more willing to believe the events described by these pictures. Nevertheless, it is nightmare to plow through millions of pictures which contain enormous noises and redundancies on social media. Moreover, in order to attract readers, the dishonest often mislead the public, by spreading sensational sham news with specious pictures. This behave severely destroys the honesty in our society. In this paper, we visualize an event from its bound pictures, based on the consistency between picture and event. First, we extract high reliable pictures for the event, by analyzing the consistency on temporal and textual dimensions. Second, the consistency of pictures is optimized in a related picture graph, in order to push up representative pictures of an event. The experiments on a real dataset verify the effectiveness of our method in most cases, comparing to several up-to-date methods.