Personalizing Web Publishing via Information Extraction

because Web search and navigation are still underdeveloped. Although Web publishing is increasingly successful, it still requires too much time and effort to precisely locate specific information. This process is often tied to traditional solutions developed outside the Web scenario—for example, information retrieval (IR) models over hypertext rather than simple text documents. Moreover, even during navigation, users have limited freedom because the links Web sites provide express the information providers’ perspective rather than the users’ information needs. This means Web content is still underused. For example, searching portals managed according to humancentered policies still relies on traditional (that is, non-Web-inspired) technologies. We cannot easily personalize navigation paths according to the users’ information needs as long as specialized professionals are charged with the design, content, and authoring News agencies offer a meaningful example of this IR problem. Most newspaper portals deliver news streams by enrichment and topical categorization. News Web sites also associate previously released news items with new items as related texts. Related news links aim to increase the browsing capability and to speed up the search process. These news items are usually found by a combination of search methods that involve keyword matching and text clustering. The related news section generally includes news items that deal with the same people (such as George W. Bush, Stanley Kubrick, or Ronaldo), events (such as the Venezia Cinema Festival), or locations (such as Pakistan, Seattle, or Salt Lake City) as the initial article. However, this bag-of-words model does not provide any explicit (logically meaningful) representation for the people or events justifying the links. Related news lists are usually accepted by editorial boards, as in the case of Financial Times, where the related stories are called Editor’s Choice. The cost of this phase is strictly dependent on the accuracy of the applied retrieval methods. However, this is a critical issue because there is no significant way to judge a link without an explanation. Consequently, it is difficult for end users to know if a related link will interest them. This method is subject to obvious bias in the delivery process (editors might emphasize material according to strategic, economic, or political agendas), and users are left no real freedom. Although user profiles can explicitly describe user interests, there is no way to trace inferences over the information the technology offers. We suggest a new scenario. Imagine a Webpublishing portal where news is categorized and enriched by extracting explicit representations of its content (hereafter, objective representation [OR]), links among news items exist according to their ORs, and users can explicitly describe their information needs in conceptual profiles that constrain news categories and links. The main idea is that different classes of users (for example, company managers versus journalists) can select similar categories via different linking policies. When link types differ, a portal could make different hypertext documents available for the related user classes. The Namic system,1 a multilingual text classifiThe authors propose a