|
Hi, I'd appreciate some clarifications about the use / guarantees of the CCA method in the context of NLP. In particular, CCA is generally described as a method operating on two random vectors, and its guarantees are also expressed in terms of random vectors. It is not clear to me how this translates to the prominent use in the NLP literature, in which CCA is described as operating on two matrices (supposedly samples from the random vectors) with an equal number of rows (alternatively, on pairs of concrete column vectors). How does one translate between these two views of CCA, and what are the guarantees of CCA in the two matrices case? |
