DSIR: the First TREC-7 Attempt

This paper describes our first large-scale retrieval attempt in TREC-7 using DSIR. DSIR is a vector space based retrieval system in which semantic similarity between words, documents and queries, is interpreted in terms of geometric proximity of vectors in a multi-dimensional space. A co-occurrence matrix computed directly from the collection is used to build the underlying semantic space. We have implemented DSIR on a cluster of lowcost PC Pentium-class machines, and chosen the PVM message-passing library to manage our distributed DSIR version. Although our first adhoc retrieval results are quite poor in terms of recall-precision measure, we believe that more work and experiments have to be explored in order to obtain more promising retrieval performance.