next up previous
次へ: Results of our Machine 上へ: Experiments with our Machine 戻る: -gram Language Model

Decoder

We used ``Moses[7]'' as a decoder. In Japanese to English translation, the position of the verb is sometimes significantly changed from its original position. Thus, we set the ``distortion weight (weight-d)'' to ``0.2'' and ``distortion-limit'' to ``-1'' for standard statistical machine translation. However, our system has 2 stage machine translation and the output of first stage is "ENGLISH". In this case, the position of word did not move so widely. So, we set the ``distortion-limit'' to ``-6'' for second stage statistical machine translation for our system.

Table 2 indicates the other parameters. We did not optimize these parameters nor use a reordering model.


表 2: Parameters of moses.ini
ttable-limit 40 0      
weight-d 0.1        
weight-l 1.0        
weight-t 0.5 0.0 0.5 0.1 0.0
weight-w -1        
distortion-limit (-1 or 6)        


next up previous
次へ: Results of our Machine 上へ: Experiments with our Machine 戻る: -gram Language Model
Jin'ichi Murakami 平成22年7月5日