MIT Press

Skip to Main Content

Table 5:

Ablation study for machine translation on WMT16 Ro-En. Results of Stern et al. (2019) are based on greedy decoding with the EOS penalty.

Model Variants	dev	test
Baseline L2R	32.53	31.82
SAO default	33.60	32.47
no bootstrap	32.86	31.88
no bootstrap, no noise	32.64	31.72
bootstrap from R2L order	33.12	32.02
bootstrap from SYN order	33.09	31.93
Stern et al. (2019) - Uniform	29.99	28.52
Stern et al. (2019) - Binary	32.27	30.66

Model Variants	dev	test
Baseline L2R	32.53	31.82
SAO default	33.60	32.47
no bootstrap	32.86	31.88
no bootstrap, no noise	32.64	31.72
bootstrap from R2L order	33.12	32.02
bootstrap from SYN order	33.09	31.93
Stern et al. (2019) - Uniform	29.99	28.52
Stern et al. (2019) - Binary	32.27	30.66