Introduced on day 41
Latent semantic analysis (LSA)
A dense representation obtained by factorising a TF-IDF matrix with truncated SVD, so that words and messages that co-occur share a small number of latent dimensions.
In detail
LSA uses no labels and can be fit on a large unlabelled pool of messages. It lets a classifier transfer weight between words that appear together, which helps with vocabulary mismatch. It also merges words that merely co-occur, has no notion of word order, and depends on the corpus it was fit on. Neural sentence encoders pursue the same goal with learned, contextual representations.