This paper describes the retrieval approach that UC Berkeley used in the 2003 INEX evaluation, and the subsequent analysis and correction of search failures in the “official runs”.
As in last year’s INEX, our primary approach is a combination of probabilistic methods using a Logistic Regression (LR) algorithm for estimation of document (article) relevance and/or element relevance, along with Boolean constraints. This year we also used data fusion techniques to combine results from multiple probabilistic retrieval algorithms, specifically the Okapi BM-25 algorithm, and multiple search elements for any given query.