CLARIT TREC-8 Experiments in Searching Web Data.
Jeffrey Bennett, Tong Xiang, David Andreoff Evans · 1999
CLARITECH submitted two baseline content-only runs and completed two additional content+link runs in the TREC-8 Web Track. These represent our first serious attempt to deal with Web data, and our first automatic runs in several years. The first question was whether CLARIT would perform as well on Web data as on more traditional text. We found that, with extensive pre-processing of the raw data prior to indexing, the automatic retrieval system actually performed better on Web data than on Ad Hoc data. For the link runs, we implemented a version of the HITS algorithm [Kleinberg 1997], originally developed at IBM. Our version optimized HITS for the CLARIT environment, but also reflected some constraints imposed by limited resources. Unable to develop and sufficiently test our own matrix-processing library in time, we used a commercial product for the number crunching. Performance on the link runs was poor, but failure analysis suggests many ways to improve it. 1 Introduction Even casual...