A Regularized Framework for Sparse and Structured Neural Attention