TREC10 Web and interactive tracks at CSIRO
Nick Craswell, David Hawking, Ross G. Wilkinson, Mingfang Wu · 2002
s time increases. 2. The web track 2.1. Topic distillation In topic distillation we used the following forms of evidence: . BM25 on content. Pages returned should be relevant. We indexed the .GOV corpus and applied BM25, sometimes with stemming sometimes without. . BM25 on content and referring anchor text. An alternative to content-only BM25 is to include referring anchor text words in the BM25 calculation (content and anchors). . In-link counting and filtering. We expected pages with more in-links to be potentially better answers, and we differentiated between on-host and off-host links. We also eliminated many results on the grounds that they had insufficient in-links. . URL length. We expected short URLs to be better answers than long URLs . BM25 score aggregation. We expected sites with many BM25-matching pages to be better than those with few. Table 1 reports the results for our topic distillation runs. Our (non-submitted) content-only achieved better performance than an