The Inside Story: Towards Better Understanding of Machine Translation Neural Evaluation Metrics