Feasibility of Minimum Error Rate Training with a Human- Based Automatic Evaluation Metric

Minimum error rate training (MERT) involves choosing parameter values for a machine translation (MT) system that maximize performance on a tuning set as measured by an automatic evaluation metric, such as BLEU. The method is best when the system will eventually be evaluated using the same metric, but in reality, most MT evaluations have a human-based component. Although performing MERT with a human-based metric seems like a daunting task, we describe a new metric, RYPT, which takes human judgments into account, but only requires human input to build a database that can be reused over and over again, hence eliminating the need for human input at tuning time. In this investigative study, we analyze the diversity (or lack thereof) of the candidates produced during MERT, we describe how this redundancy can be used to our advantage, and show that RYPT is a better predictor of translation quality than BLEU. Feasibility of Minimum Error Rate Training with a HumanBased Automatic Evaluation Metric Omar F. Zaidan and Chris Callison-Burch (ozaidan|[email protected]) Dept. of Computer Science, Johns Hopkins University Abstract. Minimum error rate training (MERT) involves choosing parameter values for a machine translation (MT) system that maximize performance on a tuning set as measured by an automatic evaluation metric, such as BLEU. The method is best when the system will eventually be evaluated using the same metric, but in reality, most MT evaluations have a human-based component. Although performing MERT with a human-based metric seems like a daunting task, we describe a new metric, RYPT, which takes human judgments into account, but only requires human input to build a database that can be reused over and over again, hence eliminating the need for human input at tuning time. In this investigative study, we analyze the diversity (or lack thereof) of the candidates produced during MERT, we describe how this redundancy can be used to our advantage, and show that RYPT is a better predictor of translation quality than BLEU. Minimum error rate training (MERT) involves choosing parameter values for a machine translation (MT) system that maximize performance on a tuning set as measured by an automatic evaluation metric, such as BLEU. The method is best when the system will eventually be evaluated using the same metric, but in reality, most MT evaluations have a human-based component. Although performing MERT with a human-based metric seems like a daunting task, we describe a new metric, RYPT, which takes human judgments into account, but only requires human input to build a database that can be reused over and over again, hence eliminating the need for human input at tuning time. In this investigative study, we analyze the diversity (or lack thereof) of the candidates produced during MERT, we describe how this redundancy can be used to our advantage, and show that RYPT is a better predictor of translation quality than BLEU.