Dependency-based Chinese-English Statistical Machine Translation

We present a Chinese-English Statistical Machine Translation (SMT) system based on dependency tree mappings. We use a state-of-the-art dependency parser to parse the English translation of the Penn Chinese Treebank to make it bilingual and then learn a tree-to-tree dependency mapping model. We also train a phrase-based translation model and collect a bilingual phrase lexicon to bootstrap a treelet translation model. For decoding, we use the same dependency parser on Chinese, using a log-linear framework to integrate the teamed translation model with a variety of dependency tree based probability models, and then find the best English dependency tree by dynamic programming. Finally the English tree is flattened to produce the translation. We evaluate our system on the 863 and NIST 2005 Chinese-English MT test data and find that the dependency-based model significantly outperforms Caravan, our phrase-based SMT system which participated in NIST 2006 and IWSLT 2006.