Word Embedding based on Low-Rank Doubly Stochastic Matrix Decomposition