Ablating the dual-path roll-in policy hurts EDITOR on soft-constrained MT, but still outperforms LevT, confirming that reposition and dual-path imitation learning both benefit EDITOR.
| . | BLEU↑ . | RIBES↑ . | CPR↑ . | Lat. ↓ . |
|---|---|---|---|---|
| Ro-En | ||||
| EDITOR | 33.1 | 85.0 | 86.8 | 108.98 |
| -dual-path | 32.2 | 84.4 | 74.8 | 119.61 |
| LevT | 31.6 | 83.4 | 80.3 | 121.80 |
| En-De | ||||
| EDITOR | 28.2 | 81.6 | 88.4 | 121.65 |
| -dual-path | 27.2 | 80.4 | 78.7 | 130.85 |
| LevT | 27.1 | 80.0 | 75.6 | 127.00 |
| En-Ja | ||||
| EDITOR | 45.3 | 85.7 | 91.3 | 109.50 |
| -dual-path | 44.0 | 83.9 | 80.0 | 154.10 |
| LevT | 42.8 | 84.0 | 74.3 | 161.17 |
| . | BLEU↑ . | RIBES↑ . | CPR↑ . | Lat. ↓ . |
|---|---|---|---|---|
| Ro-En | ||||
| EDITOR | 33.1 | 85.0 | 86.8 | 108.98 |
| -dual-path | 32.2 | 84.4 | 74.8 | 119.61 |
| LevT | 31.6 | 83.4 | 80.3 | 121.80 |
| En-De | ||||
| EDITOR | 28.2 | 81.6 | 88.4 | 121.65 |
| -dual-path | 27.2 | 80.4 | 78.7 | 130.85 |
| LevT | 27.1 | 80.0 | 75.6 | 127.00 |
| En-Ja | ||||
| EDITOR | 45.3 | 85.7 | 91.3 | 109.50 |
| -dual-path | 44.0 | 83.9 | 80.0 | 154.10 |
| LevT | 42.8 | 84.0 | 74.3 | 161.17 |