The effect of optimization by RIBES in statistical machine translation

Abstract

Parameter optimization is indispensable for improving the performance of statistical machine translation. MERT is a standard optimization tool that uses BLEU as the typical objective function. However, for translation between very different language such as English - from/to - Japanese translation, BLEU has very low correlation with human judgement. Therefore, it is better to use more human-like evaluation function like RIBES instead of BLEU as the objective function in the optimization process. In this paper, we investigate effects of RIBES used as the objective function of MERT and found that RIBES achieves better than BLEU when SMT is forced to determine the word order which is very different from the source sentence.