Context Aggregation and Analysis : A Tool for User-Generated Video Verification
O. Papadopoulou, D. Giomelakis, L. Apostolidis, S. Papadopoulos, Y. Kompatsiaris
The uncontrolled dissemination of User-Generated Content (UGC) through social media and video platforms raises increasing concerns about the intentional or unintentional spread of misleading information. As a result, people who are turning to the Internet for their daily news, need tools that help them distinguish between reliable and unreliable content. Here we present the Context Aggregation and Analysis tool, with the aim to facilitate the investigation of the veracity of User-Generated videos (UGVs). The tool collects and calculates a set of verification cues based on the video context, that is the information surrounding the video rather than the video itself, and then creates a verification report. The cues include information about the video and user that posted it, as well as the activity of other users surrounding it (what we call “wisdom of the crowd”), cross-checking with previous cases of fakes (“wisdom of the past”), and employing machine learning systems trained on past cases of real and fake videos (“wisdom of the machine”). We evaluate the tool in two ways: i) we carry out a user study where end users are manually assessing the tool’s features on a set of UGVs from a real-world dataset of news-related videos, and ii) we quantitatively evaluate the automatic verification component of the tool. The tool assisted successfully with the debunking of 132 out of 200 fake videos, the verification of 142 out of 180 real videos and the performance of the classifiers reached an F-score of 0.72.