MIT Press

Table 2:

Results for cascaded and multi-task models under full training data conditions.

Model	BLEU
Cascade	32.45
Direct	35.30

Basic two-stage	34.36
APM	35.31
APM + cross connections	36.51
APM + cross conn. + additional loss	36.70

Best APM w/o block dropout	36.04

Model	BLEU
Cascade	32.45
Direct	35.30

Basic two-stage	34.36
APM	35.31
APM + cross connections	36.51
APM + cross conn. + additional loss	36.70

Best APM w/o block dropout	36.04