Cross-language Projection of Dependency Trees for Tree-to-tree Machine Translation

Syntax-based machine translation (MT) is an attractive approach for introducing additional linguistic knowledge in corpus-based MT. Previous studies have shown that treeto-string and string-to-tree translation models perform better than tree-to-tree translation models since tree-to-tree models require two high quality parsers on the source as well as the target language side. In practice, high quality parsers for both languages are difficult to obtain and thus limit the translation quality. In this paper, we explore a method to transfer parse trees from the language side which has a high quality parser to the side which has a low quality parser to obtain transferred parse trees. We then combine the transferred parse trees with the original low quality parse trees. In our tree-to-tree MT experiments we have observed that the new combined trees lead to better performance in terms of BLEU score compared to when the original low quality trees and the transferred trees are used separately.