Distilled Wasserstein Learning for Word Embedding and Topic Modeling