Why Does Latent Semantic Analysis Work?

Warning to regular readers: this post is a bit more theoretical than average for this blog.

Peter Turney has a nice post today about “SVD, Variance, and Sparsity“. It’s actually a follow-up to a post last year entitled “Why Does SVD Improve Similarity Measurement?” that apparently has remained popular despite its old age in blog years.

For readers unfamiliar with singular value decomposition (SVD), I suggest a brief detour to the Wikipedia entry on latent semantic analysis (also known as latent semantic indexing). In a nutshell, latent semantic analysis is an information retrieval techinque that applies SVD to the term-document matrix of a corpus in order to reduce this sparse, high-dimensional matrix to a denser, lower-dimensional matrix whose dimensions correspond to the “latent” topics in the corpus.

Back to Peter’s thesis. He’s observed that document similarity is more accurate in the lower-dimensional vector space produced by SVD than in the space defined by the original term-document matrix. This isn’t immediately obvious; after all, SVD is a lossy approximation of the term-document matrix, so you might expect accuracy to decrease.

In his 2007 post, Peter offers three hypotheses for why SVD improves the similarity measure:

1. High-order co-occurrence: Dimension reduction with SVD is sensitive to high-order co-occurrence information (indirect association) that is ignored by PMI-IR and cosine similarity measures without SVD. This high-order co-occurrence information improves the similarity measure.
2. Latent meaning: Dimension reduction with SVD creates a (relatively) low-dimensional linear mapping between row space (words) and column space (chunks of text, such as sentences, paragraphs, or documents). This low-dimensional mapping captures the latent (hidden) meaning in the words and the chunks. Limiting the number of latent dimensions forces a greater correspondence between words and chunks. This forced correspondence between words and chunks (the simultaneous equation constraint) improves the similarity measurement.
3. Noise reduction: Dimension reduction with SVD removes random noise from the matrix (it smooths the matrix). The raw matrix contains a mixture of signal and noise. A low-dimensional linear model captures the signal and removes the noise. (This is like fitting a messy scatterplot with a clean linear regression equation.)

In today’s follow-up post, he drills down on this third hypothesis, noting that noise can come from either variance and sparsity. He then proposes independently adjusting the sparsity-smoothing and variance-smoothing effects of SVD to split this third hypothesis into two sub-hypotheses.

I haven’t done much work with latent semantic analysis. But work that I’ve done with other statistical information retrieval techinques, such as using Kullback-Leibler divergence to measure the signal of a document set, suggest a similar benefit from preprocessing steps that reduce noise. Now I’m curious about the relative benefits of variance vs. sparsity reduction.