Formality Style Transfer on GYAFC (test)
1.55Formality (E&M)Ours (best)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Ours (best)Backbone=T5-large, Training=SSL framework2022.03 | 1.55 | 1.41 | 1.88 | 1.85 | 1.87 | 1.88 | |
| Chawla'sdescription=uses language model discriminators and maximizing mutual information to improve a pretrained BART-Large model, along with a cycle-reconstruction loss2022.03 | 1.46 | 1.22 | 1.85 | 1.8 | 1.87 | 1.88 | |
| Ours (base)Backbone=T5-large2022.03 | 1.42 | 1.28 | 1.84 | 1.82 | 1.86 | 1.95 |