YouCat: Weakly Supervised Youtube Video Categorization System from Meta Data & User Comments using WordNet & Wikipedia

Proceedings of COLING 2012: Technical Papers, pages 1865–1882, COLING 2012, Mumbai, December 2012. YouCat : Weakly Supervised Youtube Video Categorization System from Meta Data & User Comments using WordNet & Wikipedia Subhabrata Mukherjee†, Pushpak Bhattacharyya‡ †IBM India Research Lab ‡Dept. of Computer Science and Engineering, IIT Bombay subhabmu@in.ibm.com, pb@cse.iitb.ac.in ABSTRACT In this paper, we propose a weakly supervised system, YouCat, for categorizing Youtube videos into different genres like Comedy, Horror, Romance, Sports and Technology The system takes a Youtube video url as input and gives it a belongingness score for each genre. The key aspects of this work can be summarized as: (1) Unlike other genre identification works, which are mostly supervised, this system is mostly unsupervised, requiring no labeled data for training. (2) The system can easily incorporate new genres without requiring labeled data for the genres. (3) YouCat extracts information from the video title, meta description and user comments (which together form the video descriptor). (4) It uses Wikipedia and WordNet for concept expansion. (5) The proposed algorithm with a time complexity of O(|W|) (where (|W|) is the number of words in the video descriptor) is efficient to be deployed in web for real-time video categorization. Experimentations have been performed on real world Youtube videos where YouCat achieves an F-score of 80.9%, without using any labeled training set, compared to the supervised, multiclass SVM F-score of 84.36% for single genre prediction. YouCat performs better for multi-genre prediction with an F-Score of 90.48%. Weak supervision in the system arises out of the usage of manually constructed WordNet and genre description by a few root words. KEYWORDS : Youtube, Genre Prediction, Comments, Metadata, Wikipedia, WordNet 1865 1 INTRODUCTION In recent times there has been an explosion in the number of online videos. With the gradually increasing multimedia content, the task of efficient query-based video retrieval has become important. The proper genre or category identification of the video is essential for this purpose. The automatic genre identification of videos has been traditionally posed as a supervised classification task of the features derived from the audio, visual content and textual features. Whereas some works focus on classifying the video based on the meta data (text) provided by the uploader (Cui et al., 2010; Borth et al., 2009, Filippova et al., 2011), other works attempt to extract low-level features by analyzing the frames, signals, audio etc. along with textual features (Ekenel et al., 2010; Yang et al., 2007). There have been some recent advances in incorporating new features for classification like the social content comprising of the user connectivity (Zhang et al., 2011; Yew et al., 2011), comments (Filippova et al., 2011), interest etc. All the above approaches pose the genre prediction task as supervised classification requiring a large amount of training data. It has been argued that a serious challenge for supervised classification is the availability and requirement of manually labeled data (Filippova et. al, 2010; Wu et. al, 2010; Zanetti et. al, 2008). For example, consider a video with the descriptor …

Content truncated to 500 words