Skip to Main Content
Table 7 

%BLEU on tune and test sets for ZH→EN translation, showing the contribution of feature sets in our QPD model. Both QPD models are significantly better than the best Moses numbers on test sets 1 and 2, but not on test set 3. The full QPD model is significantly better than the version with only TgtTree features on test set 1 but statistically indistinguishable on the other two test sets. Hiero is significantly better than the full QPD model on test set 2 but not on the other two.

Chinese→English
model
notes
tune
test 1
test 2
test 3
test avg.
Moses Rampion, S = 200 36.0 35.5 34.3 31.3 33.7 
Rampion, S = 500 36.2 36.1 34.6 31.8 34.2 
SSVM reranking 36.3 36.1 34.6 31.8 34.2 
 
QPD TgtTree 37.1 36.8 35.3 32.0 34.7 
TgtTree + TreeToTree 37.3 37.2 35.5 31.9 34.9 
 
Hiero Rampion 37.3 37.4 36.1 32.1 35.2 
Chinese→English
model
notes
tune
test 1
test 2
test 3
test avg.
Moses Rampion, S = 200 36.0 35.5 34.3 31.3 33.7 
Rampion, S = 500 36.2 36.1 34.6 31.8 34.2 
SSVM reranking 36.3 36.1 34.6 31.8 34.2 
 
QPD TgtTree 37.1 36.8 35.3 32.0 34.7 
TgtTree + TreeToTree 37.3 37.2 35.5 31.9 34.9 
 
Hiero Rampion 37.3 37.4 36.1 32.1 35.2 
Close Modal

or Create an Account

Close Modal
Close Modal