Compositional Attention Networks for Interpretability in Natural Language Question Answering