Conventional pattern-based machine translation costs a lot because its translation patterns are made manually. In return, the output of such translation is grammatical and tends to be a good translation. In comparison, statistical machine translation is low in cost because it uses only source and target sentence pairs. However, such translation often outputs ungrammatical translations. In our opinion, this is caused by the reordering model. We believe it is not a statistical phenomenon.
To overcome the above mentioned problems, we proposed pattern-based statistical machine translation. We focused on the corresponding word pairs between the source language and the target language that can be automatically found with GIZA++. GIZA++ gets the source and target word pairs automatically from source and target sentence pairs. The Japanese-English translation patterns can then be made from these Japanese-English word pairs. The steps of the proposed method are described below.