Towards Explainable Evaluation Metrics for Machine Translation