Interpretable and Steerable Sequence Learning via Prototypes