%BLEU on tune and test sets for DE→EN translation, comparing the baselines to our QPD model with target syntactic features (TgtTree) and then also with source syntax (+ TreeToTree). Here, merely using the additional round of tuning with the SSVM reranker improves the BLEU score to 19.9, which is statistically indistinguishable from the two QPD feature sets. Differences between Hiero and the three 19.9 numbers are at the border of statistical significance; the first two are statistically indistinguishable from Hiero but the third is different at p = 0.04.
German→English . | |||
---|---|---|---|
model . | notes . | tune . | test . |
Moses | Rampion, S = 200 | 16.2 | 19.0 |
Rampion, S = 500 | 16.2 | 19.2 | |
SSVM reranking | 16.9 | 19.9 | |
QPD | TgtTree | 17.2 | 19.9 |
TgtTree + TreeToTree | 17.1 | 19.9 | |
Hiero | Rampion | 17.1 | 20.1 |
German→English . | |||
---|---|---|---|
model . | notes . | tune . | test . |
Moses | Rampion, S = 200 | 16.2 | 19.0 |
Rampion, S = 500 | 16.2 | 19.2 | |
SSVM reranking | 16.9 | 19.9 | |
QPD | TgtTree | 17.2 | 19.9 |
TgtTree + TreeToTree | 17.1 | 19.9 | |
Hiero | Rampion | 17.1 | 20.1 |