Table 4:
BLEU scores for the MTNT test sets with some of the objectives deactivated for training the zero-shot NMT system that synthesizes $PR1-R2S$. The configurations using #1 synthetic data were trained exclusively on this data. “*” denotes systems significantly worse than using all the objectives with a p-value < 0.05.
Lossesfr→enen→fr
Zero-Shot NMT
AE+BT+MT 21.4 22.4
BT+MT 20.4* 22.3
AE+BT 19.2* 21.2*
BT 0.1* 0.4*

#1: only with $PR1-R2S$ synthesized from PL1-L2
AE+BT+MT 19.0 18.0
BT+MT 17.7* 17.0*
AE+BT 6.6* 0.7*
BT 0.0* 0.2*

#2: PL1-L2 + $PR1-R2S$ synthesized from MR2
AE+BT+MT 29.3 26.8
BT+MT 28.5* 25.9*
AE+BT 28.3* 25.0*
BT 13.2* 12.1*
Lossesfr→enen→fr
Zero-Shot NMT
AE+BT+MT 21.4 22.4
BT+MT 20.4* 22.3
AE+BT 19.2* 21.2*
BT 0.1* 0.4*

#1: only with $PR1-R2S$ synthesized from PL1-L2
AE+BT+MT 19.0 18.0
BT+MT 17.7* 17.0*
AE+BT 6.6* 0.7*
BT 0.0* 0.2*

#2: PL1-L2 + $PR1-R2S$ synthesized from MR2
AE+BT+MT 29.3 26.8
BT+MT 28.5* 25.9*
AE+BT 28.3* 25.0*
BT 13.2* 12.1*
Close Modal