On Explaining with Attention Matrices