Indri at TREC 2004: Terabyte Track

This paper provides an overview of experiments carried out at the TREC 2004 Terabyte Track using the Indri search engine. Indri is an efficient, effective distributed search engine. Like INQUERY, it is based on the inference network framework and supports structured queries, but unlike INQUERY, it uses language modeling probabilities within the network which allows for added flexibility. We describe our approaches to the Terabyte Track, all of which involved automatically constructing structured queries from the title portions of the TREC topics. Our methods use term proximity information and HTML document structure. In addition, a number of optimization procedures for efficient query processing are explained.