English-German Machine Translation on WMT newstest2014
35BLEUBack-Translation with sampling
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Back-Translation with samplingTraining Data=back-translation with 226M newscrawl sentences2018.08 | 35 | 33.8 | |
| DeepLTraining Data=proprietary high-quality bitext2018.08 | 33.3 | — | |
| AdvAugLoss Config.=L_vrt + L_adv2020.06 | 29.57 | — | |
| Shaw et al. (2018)Training Data=WMT'16 bitext2018.08 | 29.2 | — | |
| Ahmed et al. (2017)Training Data=WMT'16 bitext2018.08 | 28.9 | — | |
| AdvAugLoss Config.=L_vrt2020.06 | 28.58 | — | |
| Vaswani et al. (2017)Training Data=WMT'16 bitext2018.08 | 28.4 | — | |
| Cheng et al. (2019)2020.06 | 28.34 | — | |
| AdvAugLoss Config.=L_mixup2020.06 | 28.08 | — | |
| Sato et al. (2019)2020.06 | 27.46 | — | |
| Vaswani et al. (2017)Loss Config.=L_clean2020.06 | 27.3 | — | |
| Gehring et al. (2017)Training Data=WMT'14 bitext2018.08 | 25.2 | — | |
| +synthetictraining_instances=44m (parallel) / 36m (synthetic), ensemble=ens-42015.11 | 23.8 | — | |
| +synthetictraining_instances=44m (parallel) / 36m (synthetic), ensemble=single2015.11 | 22.7 | — | |
| syntax-basedensemble=single2015.11 | 22.6 | — | |
| +monolingualtraining_instances=49m (parallel) / 49m (monolingual), ensemble=ens-42015.11 | 21.4 | — | |
| paralleltraining_instances=37m (parallel), ensemble=ens-42015.11 | 20.4 | — | |
| +monolingualtraining_instances=49m (parallel) / 49m (monolingual), ensemble=single2015.11 | 20.4 | — | |
| paralleltraining_instances=37m (parallel), ensemble=single2015.11 | 19.9 | — |