Language Models for Searching in Web Corpora

We describe our participation in the TREC 2004 Web and Terabyte tracks. For the web track, we employ mixture language models based on document full-text, incoming anchortext, and documents titles, with a range of webcentric priors. We provide a detailed analysis of the effect on relevance of document length, URL structure, and link topology. The resulting web-centric priors are applied to three types of topics?distillation, home page, and named page?and improve effectiveness for all topic types, as well as for the mixed query set. For the terabyte track, we experimented with building an index just based on the document titles, or on the incoming anchor texts. Very selective indexing leads to a compact index that is effective in terms of early precision, catering for the typical web searcher behavior.