Machine Translation on WMT
3.2PPLTeacher
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TeacherModel=Transf. (8-layer)2020.09 | 3.2 | 28.7 | 29.2 | |
| Vanilladata=Full2020.09 | 5.2 | 23.8 | 24.5 | |
| Vanilla2020.09 | 5.5 | 22 | 23.1 | |
| SeqKDdata=Full2020.09 | 5.6 | 24.7 | 25.3 | |
| ImitKDdata=Full2020.09 | 5.6 | 25.3 | 25.9 | |
| ImitKD*data=Full, training=student/teacher dataset mixture2020.09 | 5.6 | 25 | 25.8 | |
| ImitKD2020.09 | 7.4 | 24.6 | 25.5 | |
| ImitKD*training=student/teacher dataset mixture2020.09 | 8.4 | 25.3 | 25.8 | |
| SeqKD2020.09 | 9 | 24.9 | 25.5 | |
| ImitKDdata=25x less data2020.09 | 15 | 17.8 | 19 | |
| ImitKD*data=25x less data, training=student/teacher dataset mixture2020.09 | 17.8 | 18.6 | 19.5 | |
| Vanilladata=25x less data2020.09 | 18.7 | 13.8 | 15.1 | |
| SeqKDdata=25x less data2020.09 | 42.3 | 17.1 | 17.9 |