For the second stage, we constructed an
``English''-English phrase table using Giza++
[Och and Ney2003] and ``train-model.perl''[Koehn
2007]. We set
default values for the
parameters. Also, 60,000 of the 100,000 training
sentences matched the Japanese-English patterns. Thus, we used these
60,000``English'' sentences to make an
``English''-English phrase table.