Disinformation in Multimedia Annotation

Abstract

Popularity of online videos is increasing at a rapid rate. Not only the users can access these videos online, but they can also upload video content on platforms like YouTube and Myspace. These videos are indexed by user generated multimedia annotation, also known as metadata, which is usually rich contextual information added by users about the content of the videos to facilitate access to their videos. Metadata plays a crucial role in techniques for video search and retrieval. However, this freedom of choosing annotation causes some uploaders to provide additional tags which are not even related to the content of the videos. Therefore, it is essential to verify the relevance of user-generated tags with the content of the video. Given the sheer volume of video content uploaded everyday, manual tag validation can be a highly labor intensive task. In this paper, we propose a method to automatically analyze user generated tags against video content to identify relevance of these tags and to detect irrelevant and misleading metadata for online videos. Our contributions are three-fold: First, we study nature of user-assigned tags and characterize them in two categories-generic and specific tags. Second, we propose a novel hierarchical graph based approach to identify tags which are relevant to content of the video. Third, we present a way to use user-generated comments for multimedia annotation verification. We demonstrate results of our method and evaluation on 300 YouTube videos for three different categories. The results show that we are able to identify relevant tags with average recall of 0.813 and average precision of 0.97.