A new model for statistical translation ispresented. A novel feature of this modelis that the alignments it produces are hierarchicallyarranged. The generative processbegins by splitting the input sentencein two parts. Each of the parts istranslated by a recursive application ofthe model and the resulting translationare then concatenated. If the sentenceis small enough, a simpler model (in ourcase IBM’s model 1) is applied.The training of the model is explained. Finally,the model is evaluated using the corporafrom a large vocabulary shared task.
展开▼