Word Embeddings as Tuples of Feature Probabilities