Constructing morpheme-based translation model for Mongolian-Chinese SMT

The data sparsity and the morphological difference between Chinese and Mongolian are the main problems in Mongolian-Chinese statistical machine translation (SMT). In this paper, we propose a novel method to construct morpheme-based translation model by using Mongolian morpheme as the pivot language. First, we train Mongolian-Morpheme SMT and Morpheme-Chinese SMT system, achieving a balance in the morphology of the language pair. Then we construct a new phrase table via these two systems to enrich translation knowledge without any additional bilingual resources, which is suitable for low-resource language pairs. Finally we incorporate this phrase table to our Mongolian-Chinese SMT system in two ways: by using multiple decoding paths and by the combination of two phrase tables. Experimental results demonstrate the effectiveness of our method with a maximum improvement of 1.37 BLEU points increment over baseline.