Proper

Abstract

Pivot-based Statistical Machine Translation uses a pivot language as a “bridge” to translate from source language to target language. However, one weakness of pivot-based SMT is that the noises in source-pivot translation and pivot-target translation often be transferred and amplified in the source-pivot-target translation. In this paper, we apply several popular pruning methods in traditional SMT to pivot-based SMT, and compare the performance and application scenarios of these pruning methods. Finally, we try to combine these pruning methods. Experimental results on European Parliament data show that our combined method leads to significant improvements over the baseline system.