Statistical Translation Model Based On Source Syntax Structure

Syntax-based statistical translation model is proved to be better than phrasebased model, especially for language pairs with very different syntax structures, such as Chinese and English. In this talk I will introduce a serial of statistical translation models based on source syntax structure. The tree-based model uses the one best syntax tree for translation. The forest-based model uses a compact forest which encodes exponential number of syntax trees in a polynomial spaces and lead to better performance. The joint parsing and translation model produces source parse trees, using the source side of the translation rules instead of separate parsing rules, and generate translations on the target side simultaneously, which outperforms the forest-based model. Some extensions of these models are introduced also.