Compositionality Decomposed: How do Neural Networks Generalise? (Extended Abstract)