Improving Word Embedding Factorization for Compression Using Distilled Nonlinear Neural Decomposition