Latent Semantic Analysis (LSA) is text mining dimension reduction technique akin to Principal Component Analysis. It assigns each document different "loadings" on the topics (the reduced dimensions). The input to LSA is a term document matrix (often modified using tf-idf). So each document has a bag-of-words count (I'm going to call this a vector) of the different terms that appeared in it. You could also cluster the different term vectors using a clustering algorithm such as k-means. The difference between clustering and LSA is that clustering algorithms assign each document to a specific "cluster" while LSA assigns a set of topic loadings to each document. Using an example, a clustering algorithm might be able to cluster documents about "medicine" vs. "sports" using the term vector for each document but would do a bad job assigning documents about "sports medicine" or "sport injuries" to the correct cluster since the documents are about multiple topics. LSA (or LDA or another topic modeling algorithm) would in theory show that certain documents (such as the "sports injuries" documents) have loadings in both the "sports" and "medicine" topics.