Sparsemax and Relaxed Wasserstein for Topic Sparsity