Improving phrase-based SMT model with Flattened Bilingual Parse Tree

Abstract

Phrase orders influence much on translation quality. However, general phrase based methods take only the source side information for phrase orderings. We instead propose a bilingual parse structure, Flattened Bilingual Parse Tree (FBPT), for better describing the inner structure of bilingual sentences and then for better translations. The main idea is to extract phrase pairs with orientation features under the help of FBPT structure. Such features can help maintain better sentence generations during translation. Furthermore, the FBPT structure can be learned automatically from parallel corpus with lower costs without the need of complex linguistic parsing. Evaluations on MT08 translation task indicate that 7% relative improvement on BLEU can be achieved compared to distortion based method (like Pharaoh).