Classifying the Hungarian Web

Classifying the Hungarian Web Andras Kornai Metacarta Inc. 875 Massachusetts Ave. Cambridge, MA 02139 andras@kornai.com David Twomey CEHQ, Inc. 145 Rosemary Street Ste H Needham, MA 02494 dtwomey@theworld.com Marc Krellenstein Reed-Elsevier Inc. 200 Wheeler Rd. Burlington, MA 01803 m.krellenstein@elsevier.com Fruvdmiliffess TeragraraCorp. 236 Huntington Ave. Boston, MA 02115 veress@cs.bu.edu Michael Mulligan divine Inc. 1 Wayside Road Burlington, MA 01803 mulligan@alum.mit.edu Alec Wysoker deNovis Inc. One Cranberry Hill, Suite 203 Lexington, MA 02421 alecw@pobox.com Abstract In this paper we present some lessons learned from building viz s la, the keyword search and topic classification system used on the largest Hungarian portal, [ origo .hu]. Based on a sim- ple statistical language model, and the large-scale supporting evidence from vizsla, we argue that in topic classi- fication only positive evidence matters. 0 Introduction Novices are often attracted to menu-based por- tals because these are easy to navigate. As they get more familiar with the web, users soon re- alize that their portal covers only a tiny fraction of the web, and move to keyword search engines. But as their information needs and sophistication grow, so does their frustration with simple key- word search. As a result seemingly obscure fea- tures, such as boolean searches, wildcards, and topic classification become increasingly relevant to them. To most users, the ideal system would be one that combines the ease of navigation provided e.g. by Yahoo with the near-exhaustive coverage provided e.g. by Google. But topic classification the Yahoo way, by professional editors, is expen- sive, and the results of using amateur editors, as in dmoz, are often highly questionable. One way to address the problem of low edito- rial bandwidth is to automate the topic classifi- cation process. Section 1 of this paper describes [origo.hu], a Hungarian portal that uses both man- ual and automatic topic classification, and gives a brief overview of the keyword search and au- toclassification technology developed by Northern Light Technology (NLT, now part of divine Inc) that is deployed on the Hungarian web, which cur- rently has about 20 million unique pages. As we shall see, this is a very successful system, both in terms of standard performance measures and in terms of end-user satisfaction. In Section 2 we turn to the main question of the paper: why is this algorithm, which is in many ways closer to classic TF-IDF than modern TREC- style topic detection systems, performing so well? We present a formal analysis of what we take to be the essential part of the topic classification prob- lem, and argue that the characteristics revealed by this analysis justify the use of methods that are simpler than generally thought acceptable. We of- fer our conclusions in Section 3. 1 [origo.hu] [origo .hu] (the square brackets are part of the branding) is owned and operated by Axelero Inc, the largest Hungarian ISP. It is by far the most popular web portal in Hungary: according to the visitor number statistics published by Median Inc. (see www . webaudit .hu for current numbers), …

Content truncated to 500 words