satyask / skipgram-model Goto Github PK

Actually, the sliding window used by the skip-gram model is to collect word co-occurrence statistics. Yes, there is an invisible word co-occurrence matrix and the skip-gram model is doing singular value decomposition to it. The two vector representations correspond to left and right singular vectors. Is it possible to represent each word with only one eigenvector? Not possible if the invisible word co-occurrence matrix is not positive definite. It is highly likely that this matrix is not positive definite because the property of positive definite is very rare and abstract. However, I have discovered a way to modify any indefinite matrix into a positive definite matrix without changing the size relationship between any two matrix elements, i.e., if A_ij < A_mn then it is ensured that B_ij < B_mn for any ijmn, and yet A is indefinite and B is positive definite. Contact me if you are interested in representing each word with only one vector representation based on reasonable principles.