Tuning statistical machine translation parameters using perplexity

Statistical machine translation (SMT) involves many tasks including modeling, training, decoding, and evaluation. In this work, we present a methodology for optimizing the training process to get better translation quality using the well known GIZA ++ SMT toolkit. The methodology is based on adjusting the parameters of GIZA ++ that affect the generation of the translation model. When applying the methodology, an average improvement of 7% has been achieved in the translation quality.