Training Set Similarity Based Parameter Selection for Statistical Machine Translation

Log-linear model based statistical machine translation systems (SMT) are usually composed of multiple feature functions. Each feature function is assigned a weight as a model parameter. In this paper, we consider that different input source sentences may have discrepant needs for model parameters. To adapt the model to different inputs, we propose a model parameters selection method for log-linear model based SMT systems. The method is mainly based on the characteristics of different feature functions themselves without any assumption on unseen test sets. Experimental results on two language pairs (Zh-En and Ug-Zh) show that our method leads to the improvements up to 2.4 and 2.2 BLEU score respectively, and it also shows the good interpretability of our proposed method.