Machine translation as tree labeling

Abstract

We present the main ideas behind a new syntax-based machine translation system, based on reducing the machine translation task to a tree-labeling task. This tree labeling is further reduced to a sequence of decisions (of four varieties), which can be discriminatively trained. The optimal tree labeling (i.e. translation) is then found through a simple depth-first branch-and-bound search. An early system founded on these ideas has been shown to be competitive with Pharaoh when both are trained on a small subsection of the Europarl corpus.