Web video categorization based on Wikipedia categories and content-duplicated open resources

Zhineng Chen, Juan Cao, Yicheng Song, Yongdong Zhang, Jintao Li

TL;DR

This paper introduces a robust framework for web video categorization that moves beyond noisy user tags by leveraging the structured knowledge of Wikipedia. By identifying core concepts in a video's metadata, the system fetches "Content-Duplicated Open Resources" (CDORs) from the wider web and uses a novel Wikipedia Category (WikiC) voting algorithm to achieve a 23.3% performance boost over standard methods.

What Problem Does This Paper Address?

Categorizing web videos (like those on YouTube) is notoriously difficult. User-provided tags are often sparse, noisy, or irrelevant, and analyzing raw video pixels is computationally expensive and frequently fails to capture high-level "semantic" meaning.

While some researchers have tried using external images or related videos to help, they usually limit themselves to one or two media types. This paper addresses the "semantic gap" by arguing that a single event (like a volcanic eruption) is recorded across the web in many forms—news articles, blogs, images, and other videos. The authors call these Content-Duplicated Open Resources (CDORs). The challenge lies in accurately identifying the "core" of a video to find these resources and then translating that diverse information into a reliable category label.

Methods

The authors propose a three-stage pipeline:

  1. Discriminative Concept Identification: Instead of treating all tags equally, the system uses Wikipedia to find "discriminative concepts." It employs a "longest match" principle (e.g., recognizing "Expo 2010" as one specific concept rather than "Expo" and "2010" separately). It then uses a heuristic to classify these concepts into three buckets: Person, Location, and Other Proper Nouns based on specific keywords in their Wikipedia categories (e.g., "births" indicates a person).
  2. CDOR Collection: The system constructs a search query using the most discriminative concepts and issues it to a commercial search engine (Google). It collects the titles and snippets of the top results that contain all the query terms. These snippets act as an "enriched" textual description of the video's content.
  3. WikiC Voting Categorization: Rather than using a standard classifier like an SVM on raw words, the authors map the concepts found in the video and its CDORs to Enriched Wikipedia Categories (EWikiCs). They then use a voting algorithm where each WikiC "votes" for a video category. Crucially, the weight of a vote is determined by the category's entropy—if a Wikipedia category (like "Living People") appears across many different types of videos, its vote counts for less than a highly specific category.

Datasets

The study utilizes the MCG-WEB benchmark dataset:

  • Training Set: 10,000 videos randomly selected from a pool of ~80,000 YouTube videos (crawled Dec 2008 – Feb 2009).
  • Testing Set: 5,887 "Most Viewed" YouTube videos (crawled Mar 2009 – Jul 2009).
  • Ground Truth: The 15 standard YouTube categories (e.g., Gaming, Science & Technology, Pets & Animals).

Key Insights and Ideas

  • Wikipedia as a Cognitive Bridge: Wikipedia categories are curated by humans and reflect how we naturally organize knowledge. Mapping a messy tag like "Mickey Mouse" to categories like "Fictional Mice" provides a semantic link that raw text analysis misses.
  • The Power of Redundancy: The "aha moment" here is that we don't need to perfectly understand the video itself if we can find five news articles and three blog posts describing the exact same event. This redundancy filters out the noise of individual user tags.
  • Longer is Better: The "longest match" principle is a simple but effective way to handle entity recognition, ensuring that specific entities aren't broken down into generic, meaningless terms.
  • Specificity Matters: By using an entropy-based voting system, the model automatically learns to ignore "stop-word" categories in Wikipedia that are too broad to be useful for classification.

Key Results

The authors compared four schemes: SVMs on raw concepts (RC-SVM), SVMs on enriched concepts (EC-SVM), and their WikiC voting on both.

  • Overall Boost: The best-performing model (EC-Wiki) achieved a 23.3% improvement in Mean Average Precision (MAP) over the baseline RC-SVM.
  • CDOR Effectiveness: Simply adding the external search results (CDORs) improved the SVM classifier by 9.4% and the WikiC voting by 11.2%.
  • State-of-the-Art Comparison: Their method outperformed then-current state-of-the-art approaches (which combined textual, social, and author data) by 3.4%, despite using a simpler feature set.
  • Category Wins: The system was particularly effective for well-defined categories like "Gaming" and "Autos," where the Wikipedia concepts are highly specific.

Limitations and Future Work

  • Snippet Dependency: Due to the difficulty of crawling full websites, the authors only used Google's generated titles and snippets. They acknowledge that using the full text of the CDORs would likely provide even richer knowledge.
  • Vague Categories: The system struggled with broad categories like "People & Blogs" or "Travel & Events." These categories are so diverse that they lack the "rules of thumb" or specific Wikipedia category clusters needed for the voting algorithm to be decisive.
  • Ambiguity: While they implemented a basic disambiguation step (checking for overlapping categories among tags), highly ambiguous terms (like "Apple") still pose a challenge if the surrounding tags are also sparse.

Summary updated

Abstract

This paper presents a novel approach for web video categorization by leveraging Wikipedia categories (WikiCs) and open resources describing the same content as the video, i.e., content-duplicated open resources (CDORs). Note that current approaches only collect CDORs within one or a few media forms and ignore CDORs of other forms. We explore all these resources by utilizing WikiCs and commercial search engines. Given a web video, its discriminative Wikipedia concepts are first identified and classified. Then a textual query is constructed and from which CDORs are collected. Based on these CDORs, we propose to categorize web videos in the space spanned by WikiCs rather than that spanned by raw tags. Experimental results demonstrate the effectiveness of both the proposed CDOR collection method and the WikiC voting categorization algorithm. In addition, the categorization model built based on both WikiCs and CDORs achieves better performance compared with the models built based on only one of them as well as state-of-the-art approach.
BibTeX
@inproceedings{chen2010web,
  author = {Zhineng Chen and Juan Cao and Yicheng Song and Yongdong Zhang and Jintao Li},
  title = {{Web video categorization based on Wikipedia categories and content-duplicated open resources}},
  booktitle = {ACM Multimedia},
  year = {2010},
  doi = {10.1145/1873951.1874162},
}

Loading related content...

Figure