Using Network Structure to Learn Category Classification in Wikipedia

Wikipedia, founded in 2001, is a massive, collaboratively-edited online encyclopedia. As of November 2011, the site has more than 20 million articles available in 282 different languages. The site is maintained by a vibrant community of editors, who supervise the editing process to ensure quality and consistency. The body of a Wikipedia article consists primarily of text, images, and hyperlinks to other articles. In addition to article content, each page is also placed into various relevant categories of articles. The editors of Wikipedia maintain this category hierarchy, manually labeling articles with the most appropriate categories. These categories are often used as the gold standard for semantic NLP problems, such as finding document topics [9]. Categories can also be useful in navigation of Wikipedia, whether simply finding related articles or attempting to find longer paths through the network. However, as Wikipedia continues to grow, manually labeling categories becomes more and more difficult. With over 80,000 categories, the monumental task requires its own Uncategorized Task Force of Wikipedia editors, whose goal is to categorize uncategorized articles. Even so, many articles remain uncategorized or undercategorized [13]. Fortunately, many articles in Wikipedia have a wealth of information, from the content of the text to the hyperlink network. These features suggest that an automatic classifier of Wikipedia may be a tractable problem. A classifier with reasonable accuracy could greatly lighten the load of the task force, improve the coverage of articles, and possibly identify and correct miscategorized articles. In this paper, we focus on network features to achieve this goal.