Interpreting Neural Networks for and with Natural Language