Recent Experiments with INQUERY

Past TREC experiments by the University of Massachusetts have focused primarily on ad-hoc query creation. Substantial eeort was directed towards automatically translating TREC topics into queries, using a set of simple heuristics and query expansion. Less emphasis was placed on the routing task, although results were generally good. The Spanish experiments in TREC-3 concentrated on simple indexing, sophisticated stemming, and simple methods of creating queries. The TREC-4 experiments were a departure from the past. The ad-hoc experiments involved \\ne tuning" existing approaches, and modiications to the INQUERY term weighting algorithm. However, much of the research focus in TREC-4 was on the routing, Spanish, and collection merging experiments. These tracks more closely match our broader research interests in document routing, document ltering, distributed IR, and multilingual retrieval. The University of Massachusetts' experiments were conducted with version 3.0 of the INQUERY information retrieval system. INQUERY is based on the Bayesian inference network retrieval model. It is described elsewhere 7, 5, 12, 11], so this paper focuses on relevant diierences to the previously published algorithms. For the ad-hoc retrieval experiments, the major change to the system was a new estimation technique for term weighting. We also continued to reene our analysis techniques for the TREC topics, our use of passage retrieval, and query expansion using InFinder 1. 1.1 Query Processing TREC topics 201{250 diier from earlier TREC topics in that the elds were removed. This change makes the TREC topics even more dissimilar from user queries in an online system than in the past. The TREC topics observe the niceties of grammar, punctuation and, especially, polite periphrasis. In an online system, users typically discard grammar, punctuation and any non-functional verbiage in an eeort to get the information they want. The removal of the eld created a set of topics that resembled essay questions. Much of our TREC processing this year focussed on creating queries that more closely resemble \real" online queries, by stripping oo the polite circumlocution and its accompanying grammar. As a result, in addition to the standard \stop-phrase" program distributed with INQUERY, which removes the occasional polite circumlocution, we resurrected an old program for removing additional verbiage that is likely to be content-free, especially in questions. For example, in topic 201 1 Formerly called PhraseFinder.