Experiments on probabilistic word embeddings: word similarity tasks interpretability of the components sparsity structure downstream tasks