In this work, Fused Gromov-Wasserstein distance is applied to improve the performance of different NLP tasks, such as Machine Translation, abstractive summarization, etc
Gromov-Wasserstein distance only or Wasserstein distance only also tested, the machine translation task on EN-VI dataset simply shows FGWD is better than both W or GW.
| Modle | EN-VI uncased | EN-VI cased | EN-DE uncased | EN-DE cased |
|---|---|---|---|---|
| transformer base | 29.25 | 28.46 | 25.60 | 25.12 |
| transformer + W | 29.49 | 28.68 | TBD | TBD |
| transformer + GW | 28.65 | 28.34 | TBD | TBD |
| transformer + FGW | 29.92 | 29.09 | 26.05 | 25.54 |
Wasserstein distance:
Gromov-Wasserstein distance:
Fused Gromov-Wasserstein distance:
