LSI does not require any sort of “semantically related” keyword list. LSI is based on an non-parametric statistical technique called principal component analysis, which decomposes functions into their axes of maximum variance. Basically, documents are converted into extremely sparse vectors of keyword counts (one dimension for each keyword). This is mildly problematic because doing any sort of calculation on a million+ dimension vector isn’t really feasible on a large scale. In order to solve this problem, LSI/PCA is used to transform the vector space into a low rank approximation (typically anywhere from 4-15 dimensions). Once the documents have been transformed into points in a low dimensional vector space, they can be quickly compared using the standard inner product for similarity.