Predicting Machine Translation Performance on Low-Resource Languages: The Role of Domain Similarity