Skip to Main Content
Table 11 

Feature ablation experiments for UR→EN translation with string-to-tree features, showing the drop in BLEU when separately removing word (Word), cluster (Clust), and configuration (Cfg) feature sets. * = significantly worse than TgtTree. Removing word features causes no significant difference. Removing cluster features results in a significant difference on both test sets, and removing configuration features results in a significant difference on test 2 only.

Urdu→English
model
notes
tune
test 1
test 2
test avg. (Δ)
Moses SSVM reranking 24.9 24.4 24.7 24.6 
 
QPD TgtTree = Word + Clust + Cfg 25.8 25.4 25.5 25.4 
TgtTree − Word 25.6 25.0 25.5 25.2 (− 0.2) 
TgtTree − Clust 25.4 24.8* 24.9* 24.9 (− 0.5) 
TgtTree − Cfg 25.1 25.1 25.0* 25.0 (− 0.4) 
Urdu→English
model
notes
tune
test 1
test 2
test avg. (Δ)
Moses SSVM reranking 24.9 24.4 24.7 24.6 
 
QPD TgtTree = Word + Clust + Cfg 25.8 25.4 25.5 25.4 
TgtTree − Word 25.6 25.0 25.5 25.2 (− 0.2) 
TgtTree − Clust 25.4 24.8* 24.9* 24.9 (− 0.5) 
TgtTree − Cfg 25.1 25.1 25.0* 25.0 (− 0.4) 
Close Modal

or Create an Account

Close Modal
Close Modal