The characteristic of the pattern-based machine translation method is that high-quality translation results can be obtained if the input sentence matches the translation pattern. However, translation patterns are made by hand in pattern-based machine translation. In comparison, phrase-based SMT is very popular and low in cost. However, the rule-based machine translation system is better than the phrase-based statistical machine for Japanese-English translation. In our opinion, this is caused by the reordering model, and we believe it is not a statistical phenomenon.
To overcome these problems, we proposed pattern-based statistical machine translations. In such translation, the reordering problem is no longer a problem because translation patterns are used. By using phrase-based SMT tools, we can implement this translation method automatically. In the NTCIR-10 patent task, we obtained high-quality translation sentences under certain conditions. The proposed method was especially effective in the human evaluation in the A-rank and B-rank classification.
In the future, we will make translation patterns with phrase-based Japanese-English word pairs in order to increase the number of translations in the A rank.