Lexicalized reordering model is adopted in state-of-the-art phrase-based machine translation systems to help formulate a better word reordering of translation results. The most widely-used MSD (Monotone, Swap, Discontinuous) reordering model is designed generically and has been used in every language pair without customization. However, in the scenarios of translation between Chinese and English, the word reordering distance tends to be long due to the syntax difference between English and Chinese, in which case MSD model is likely to deliver unappropriate results. Based on intensive investigation on large English-Chinese bilingual corpus, we redesign the orientation set of the reordering model and propose a new lexicalized reordering model MLR (Monotone, LeftDiscontinuous, RightDiscontinuous), which is tailored for C2E and E2C MT. MLR can handel long-distance word reordering well. The superiority of MLR is verified in our empirical studies and has already been applied to Youdao online translation system (http://fanyi.youdao.com)~1.
展开▼
