Focused Concept Detection and Polarity Measurement of Web Video Using Social Context

Introduction Large amount of news, documentary videos and videos of real world events are uploaded and shared online. Content creation is not restricted to the media professional and news channels alone, more and more videos of similar interest and events are also created and republished by users to express their opinion about the event. As a result a single query fetches thousands of videos with multiple viewpoints but unfiltered. It would be ideal if the results were grouped or presented in a way to reflect the focused content and the emotional polarity. For example a query of "Copenhagen summit on climate change" should reflect which videos have pro climate change views, which videos express the other side of the divide, and some videos are neutral but focus on related issues concerning the main topic so the user will be able to browse efficiently with minimal effort. The above task requires two subtasks 1) identifying the focused concept, 2) measuring the opinion polarity. Related Studies Bermingham et al.[1] analyzed Youtube data using sentiment analysis and social network analysis in the context of online radicalization. W.-H.Lin and A.Hauptmann,’s[2] study focused on the visual variation and tag usage to describe the perspective differences among web videos. Our work is more in line with the above study whereas we consider not only the tag space but also the user views and their profiles in order to detect the perspective and their polarity more robustly. Our Approach Identification of focused concept has been performed with semantic analysis of rich contextual cues available around the video, whereas identification of opinion polarity is achieved with a combined approach of user profile and sentiment analysis techniques. Focused Concept Detection (FCD) We treated FCD as document topic detection where the video is represented by a bag of words feature including its title, tags, descriptions and other related contextual metadata. For topic identification, a focused sub graph is extracted from the feature graph where the nodes are the tags and the weighted link between two nodes is the semantic distance. Polarity Measurement Since state of the art existing sentiment analysis techniques are all based on text documents or blog posts where textual content is more or less explicit and descriptive the same cannot be assumed for texts around videos. In web videos, the tags are mere freeform keywords without any explicit grammar rules and the user comments are mostly subjective and short. A direct adaptation of sentiment analysis to these texts will be sub optimal so we additionally used user profile and his/her network structure in measuring the polarity score of the video Analysis Experimental result of 10 different recent issues (climate change, Google-china dispute etc). show that most issue based queries comes with multiple user perspectives and the opinion is both pro and against the issues.One interesting observation is that user profile and their network structure is a great indicator of opinion polarity which can be exploited for a better categorizations and presentation of video search results. Application of present approach can primarily be in two main areas, Recommendation and personalization of videos can be fine grained as per the user interest and faceted presentation of the search result where the user can clearly see subtopics under the main query as well as different viewpoints of a subtopic.