Well, I've started with a collection of documents. The aim is to extract keywords for each document.
I've made a document-term matrix, to which I applied an singular value decomposition. I've made a new matrix (an approximation of the Original document-term matrix) with the truncated SVD, so I've reduced the rank.
Now I'm wondering if you can use this matrix approximation for extracting new keywords. Is there a theory about this? Or maybe with term similarities?
