Latent semantic indexing (LSI) fails for TREC collections

Abstract

The aim of latent semantic indexing (LSI) is to uncover the relationships between terms, hidden concepts, and documents. LSI uses the matrix factorization technique known as singular value decomposition (SVD). In this paper, we apply LSI to standard benchmark collections. We find that LSI yields poor retrieval accuracy on the TREC 2, 7, 8, and 2004 collections. We believe that the negative result is robust, because we try more LSI variants than any previous work. First, we show that using Okapi BM25 weights for terms in documents improves the performance of LSI. Second, we derive novel scoring methods that implement the ideas of query expansion and score regularization in the LSI framework. Third, we show how to combine the BM25 method with LSI methods. All proposed methods are evaluated experimentally on the four TREC collections mentioned above. The experiments show that the new variants of LSI improve upon previous LSI methods. Nevertheless, no way of using LSI achieves a worthwhile improvement in retrieval accuracy over BM25.