We collected large Japanese-English parallel corpus from electronic medias of 698,973 parallel sentences. And we made long phrase table. Our statistical machine translation system is similar to a statistic example based machine translation system. We utilized standard statistical machine translation tools like "moses"[2] and "GIZA++"[1].
We used this data and these tools, and obtained good results for simple sentences and acceptable results for complex and compound sentences for the contest of IWSLT07. We obtain 0.4321 BLEU score. In analyzing the error results, we found that unknown words decrease the BLEU score. Also we do not optimize parameters of "moses"[2] and other programs.
We will optimize these parameters and will add a procedure of unknown words, that will enable our system to achieve better performance.