Machine Translation on German→Romansh human evaluation set (test)
77Win RateHuman reference
Evaluation Results
| Method | Links | |
|---|---|---|
| Human referenceEvaluation protocol=document-level accuracy ranking, Reference system=Gemini 3 Pro (preview)2026.03 | 77 | |
| + dictionary promptingEvaluation protocol=document-level accuracy ranking, Reference system=Gemini 3 Pro (preview)2026.03 | 74 | |
| LR→HR augmentationEvaluation protocol=document-level accuracy ranking, Reference system=Gemini 3 Pro (preview)2026.03 | 62 | |
| Human referenceEvaluation protocol=document-level accuracy ranking, Reference system=LR→HR augmentation2026.03 | 61 | |
| Human referenceEvaluation protocol=document-level accuracy ranking, Reference system=+ dictionary prompting2026.03 | 48 | |
| LR→HR augmentationEvaluation protocol=document-level accuracy ranking, Reference system=+ dictionary prompting2026.03 | 29 | |
| + dictionary promptingEvaluation protocol=document-level accuracy ranking, Reference system=LR→HR augmentation2026.03 | 22 | |
| + dictionary promptingEvaluation protocol=document-level accuracy ranking, Reference system=Human reference2026.03 | 15 | |
| LR→HR augmentationEvaluation protocol=document-level accuracy ranking, Reference system=Human reference2026.03 | 13 | |
| Gemini 3 Pro (preview)Evaluation protocol=document-level accuracy ranking, Reference system=+ dictionary prompting2026.03 | 10 | |
| Gemini 3 Pro (preview)Evaluation protocol=document-level accuracy ranking, Reference system=Human reference2026.03 | 6 | |
| Gemini 3 Pro (preview)Evaluation protocol=document-level accuracy ranking, Reference system=LR→HR augmentation2026.03 | 4 |