Abstract
This paper presents a novel approach for web video categorization by leveraging Wikipedia categories (WikiCs) and content-duplicated open resources (CDORs). Unlike previous methods that limit themselves to specific media forms, this framework utilizes commercial search engines and Wikipedia's structured knowledge to identify discriminative concepts and collect diverse external resources. By categorizing videos in the space spanned by WikiCs rather than raw tags, the proposed model achieves significant performance gains. Experimental results show that combining WikiCs and CDORs outperforms models relying on only one source or state-of-the-art textual and social approaches.
Addressing the Semantic Gap in Video Categorization
Categorizing web videos is a persistent challenge in information retrieval. User-provided tags are often sparse, noisy, or irrelevant, while raw visual analysis is computationally expensive and frequently fails to capture high-level semantic meaning.
This research addresses the "semantic gap" by proposing that a single event—such as a volcanic eruption—is documented across the web in multiple forms, including news articles, blogs, images, and other videos. The authors define these as Content-Duplicated Open Resources (CDORs). By identifying the core concepts of a video, the system can tap into this web-scale redundancy to find more reliable descriptive data.
The Categorization Pipeline
The authors introduce a three-stage pipeline to move from noisy tags to structured categories:
- Discriminative Concept Identification: The system extracts entities from video metadata using a "longest match" principle against Wikipedia titles (e.g., identifying "Expo 2010" as a single entity). These are then classified into buckets such as Person, Location, or Other Proper Nouns based on their Wikipedia category keywords.
- CDOR Collection: Using the identified concepts, the system generates a search query to fetch titles and snippets from a commercial search engine. These snippets provide an enriched textual context for the video.
- WikiC Voting Categorization: The video and its associated CDORs are mapped to Enriched Wikipedia Categories (EWikiCs). A voting algorithm then assigns a video category. To ensure accuracy, votes are weighted by the category's entropy; highly specific categories (e.g., "Volcanic landforms") carry more weight than broad ones (e.g., "Living people").
Experimental Evaluation
The framework was tested on the MCG-WEB dataset, consisting of approximately 15,000 YouTube videos across 15 standard categories. The study compared four models:
- RC-SVM: SVM trained on raw concepts.
- EC-SVM: SVM trained on enriched concepts (including CDORs).
- RC-WikiC: WikiC voting on raw concepts.
- EC-WikiC: WikiC voting on enriched concepts.
The EC-WikiC model achieved a 23.3% improvement in Mean Average Precision (MAP) over the baseline RC-SVM. The inclusion of CDORs alone improved performance by roughly 10%, highlighting the power of external web resources in overcoming metadata scarcity. The system was particularly effective in well-defined categories like "Gaming" and "Autos," where Wikipedia concepts are highly specific.
Conclusion and Future Directions
By using Wikipedia as a "cognitive bridge" and leveraging the redundancy of the open web, this method significantly improves the accuracy of video categorization. While the current implementation relies on search snippets, future work could explore the use of full-text CDOR content and more advanced disambiguation techniques to further refine the semantic mapping.
Summary updated