A learned evaluation metric for text generation that uses a pre-trained model to assess translation quality beyond surface-level n-gram matching.