Removing Biases from Trainable MT Metrics by Using Self-Training

Miloš Stanojević

Most trainable machine translation (MT) metrics train their weights on human judgments of state-of-the-art MT systems outputs. This makes trainable metrics biases in many ways. One…