SmartVideoRanking: Video Search by Mining Emotions from Time-Synchronized Comments

摘要

Many people search for and watch videos on videosharing Web sites, where users input a query and rank videos onthe basis of metrics such as view count and rating. However, it isnot always easy to find the desired video with a conventionalsearch. One approach that enables users to more intuitivelysearch for videos they desire is to index them according toemotions. Previous studies have used several predefined emotioncategories, such as "fear" and "funny", for this purpose. However, viewers' emotions tend to be more diverse and specific. In thispaper, we dynamically detect emotions in accordance with aninput query and implement SmartVideoRanking, which enablesusers to search for videos on the basis of the detected emotions. We estimate viewer emotions from time-synchronized commentson videos and estimate the usefulness of each emotion by usingsupport vector machine regression. Experimental results showthat: (1) Spearman's rank correlation between the estimatedusefulness scores and gold standard data was 0.7547, (2) emotions associated with videos vary from one query to anotherand it is therefore meaningful to detect emotions according to aninput query, and (3) rankings based on viewer emotions enableusers to browse videos that do not appear at the top ofconventional search results. We also conduct a user study anddemonstrate SmartVideoRanking's capability to search for videos.