Minimum error training of log-linear translation models

Recent work on training of log-linear interpolation models for statistical machine translation reported performance improvements by optimizing parameters with respect to translation quality, rather than to likelihood oriented criteria. This work presents an alternative and more direct training procedure for log-linear interpolation models. In addition, we point out the subtle interaction between log-linear models and the beam search algorithm. Experimental results are reported on two Chinese-English evaluation sets, C-Star 2003 and Nist 2003, by using a statistical phrase-based model derived from Model 4. By optimizing parameters with respect to the BLUE score, performance relative improvements by 9.6% and 2.8% were achieved, respectively.