Extracting bilingual multi-word expressions for low-resource statistical machine translation

Abstract

Improving the performance of statistical machine translation is often a significant problem, especially in low language resource scenarios such as Chinese-Mongolian SMT. In this paper, we propose a method to improve the performance of Chinese-Mongolian SMT system using multi-word expressions, which is also a pilot study for this language pair. We extract MWEs from the phrase-table then integrate the MWEs into SMT system by various strategies. Experimental results indicate our method outperforms a baseline model by 0.81 BLEU points on Test-All and 1.54 BLEU points on Test-MWE.