Expected Error Minimization with Ultraconservative Update for SMT
Lemao Liu, Tiejun Zhao, Taro Watanabe, Hailong Cao, Conghui Zhu · 2012
Minimum error rate training is a popular method for parameter tuning in statistical machine translation (SMT). However, the optimization objective function may change drastically at each optimization step, which may induce MERT instability. We propose an alternative tuning method based on an ultraconservative update, in which the combination of an expected task loss and the distance from the parameters in the previous round are minimized with a variant of gradient descent. Experiments on test datasets of both Chinese-to-English and Spanish-to-English translation show that our method can achieve improvements over MERT under the Moses system.