Latent Semantic Analysis, or LSA, was one of the earliest methods for compressing distributional semantic models and learning implicit semantic information in the process (see http://lsa.colorado.edu/).

LSA usually refers specifically to the use of Singular Value Decomposition (SVD) to compress a term-document matrix, though the term has sometimes been used more generally to cover more types of distributional semantic methods.

Traditional LSA is implemented in SemanticVectors: instead of using java pitt.search.semanticvectors.BuildIndex, use java pitt.search.semanticvectors.LSA in just the same way.

For large corpora, LSA takes much more time and memory than RandomProjection.