Towards Compositional Interpretability for XAI