ANU/ACSys TREC-5 Experiments.

Abstract

A number of experiments conducted within the framework of the TREC-6 conference and using a completely re-engineered version of the PArallel Document Retrieval Engine (PADRE97) are reported. Passage-based pseudo relevance feedback combined with a variant of City University's Okapi BM25 scoring function achieved best average precision, best recall and best precision@20 in the Long-topic Automatic Adhoc category. The same basic method was used as the basis for successful submissions in the Manual Adhoc, Filtering and VLC tasks. A new BM25-based method of scoring concept intersections was shown to produce a small but significant gain in precision on the Manual Adhoc task while the relevance feedback scheme produced a significant improvement in recall for all of the Adhoc query sets to which it was applied. 1 Introduction The work reported here comprises a number of text retrieval experiments conducted within the framework of TREC-6 and addressing questions of interest in the following re...