Research‎ > ‎

TV News Archive - Multimodal Topic Detection


TV has been the most influential news channel for decades. Understanding its behaviors and impacts is a core problem in communication and journalism, but it has been challenging to manage the massive amount of multimodal data. Our goals in this project are to develop a fully-automated computational pipeline to process news videos by large-scale computer vision, machine learning, and NLP techniques, to detect news events, actors, sentiments, topics and their associations, and to systematically quantify the characteristics and effects of the mass media such as media bias or agenda-setting. 

  • Joint Image-Text News Topic Detection and Tracking by Multimodal Topic And-Or Graph
    Weixin Li, Jungseock Joo, Hang Qi, and Song-Chun Zhu 
    IEEE Transactions on Multimedia. Accepted

UCLA NewsScape TV News Archive

  • Visual Parsing of News Videos: Scene, On screen texts, Face, ..

  • Our multimodal news topic representation jointly captures key actors and actions (who, when, where) within each news topic, using both news texts (captions) and visuals.  

  • Topic Tracking

  • Multimodal Topic Detection Dataset will be released soon.