Obtaining Faithful Interpretations from Compositional Neural Networks