Latent semantic indexing (LSI) : TREC-3 report
Susan Dumais · 1994
This paper reports on recent developments of the Latent Semantic Indexing (LSI) retrieval method for TREC-3. LSI uses a reduced-dimension vector space to represent words and documents. An important aspect of this representation is that the association between terms is automatically captured, explicitly represented, and used to improve retrieval. We used LSI for both TREC-3 routing and adhoc tasks. For the routing tasks an LSI space was constructed using the training documents. We compared profiles constructed using just the topic words (no training) with profiles constructed using the average of relevant documents (no use of the topic words). Not surprisingly, the centroid of the relevant documents was 30% better than the topic words. This simple feedback method was quite good compared to the routing performance of other systems. Various combinations of information from the topic words and relevant documents provide small additional improvements in performance. For the adhoc task we c...