Like Greg's PCA approach and unlike VSM, these term representations capture relationships between words: two documents with no common words can still have high similarity.
Main difference from PCA: SVD on cooccurrences, not correlations (preserves sparsity).
