BARTSCORE: Evaluating Generated Text as Text Generation