Faster Parallel Training of Word Embeddings