Detecting Culture-specific Tags for News Videos through Multimodal Embedding
Chun-Yu Tsai, John R. Kender · 2017
Many videos on the Web about international events are maintained in different countries, and some come with text descriptions from different cultural points of view. We introduce a new task-detecting culture-specific tags for news videos: given video keyframes and culture information, assign the most relevant tags with cultural preferences. We approach this problem by mapping visual and multilingual textual features into a joint latent space by reliable visual cues, by our proposed two-view pair-pair embedding and three-view embedding, through various canonical correlation analyses variants (Canonical Correlation Analysis, Deep Canonical Correlation Analysis, eneralize Canonical Correlation Analysis). For human-interest international events such as epidemics and transportation disasters, we proof that, for the same news event, tags of videos are significantly different in different cultures.