Stability vs. Effectiveness: Improved Sentence-Level Combination of Machine Translation Based on Weighted MBR

摘要

We describe an improved strategy to combine the outputs of machine translation on sentence-level balancing the stability and the effectiveness of the combination. The new method alternates the classical MBR-based sentence-level combination with weighted Minimum Bayes Risk (wMBR). During the calculation of the risk, we weight the hypotheses with the performance of the MT system, which is measured by the automatic evaluation metrics on the development data. In experiments, the wMBR-based method stably achieve better results than other sentence-level methods and get the best position in CWMT08 evaluation track outperforming the other word-level and sentence-level combination systems.