Morpheme Segmentation Using Bilingual Features

This paper presents an optimizing morphological segmentation metric for statistical machine translation performance. Unlike previous morpheme segmentation work for getting greater linguistic accuracy we focus on factors such as consistency, coverage and granularity, which directly affect MT performance. We propose a novel combination of dictionary information and statistical model, taking advantage of source-target bilingual features. Our method effectively integrates morpheme information while avoiding the complex calculations generated with the traditional usage of the morphemes. Experiments show that the approach outperforms previously proposed ones and provides an improvement of 1.03 and 0.89 BLEU results in both phrased-based and factored-based MT model on the Chinese-Mongolian translation task.