Controllable Text Generation (Sentiment + Tense + Formality) on Yelp (test)
97Accuracy (Sentiment)GPT2-FT
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GPT2-FTBackbone=GPT2-large, Supervision=Fully-supervised2022.08 | 97 | 92 | 87 | 91.9 | 10.3 | 36.8 | |
| LATENTOPSBackbone=GPT2-large, Approach=Latent space / VAE2022.08 | 97 | 92 | 93 | 93.7 | 25.8 | 21.1 | |
| PPLMBackbone=GPT2-large, Approach=Plug-and-play2022.08 | 82 | 57 | 56 | 59.8 | 17.5 | 30.5 | |
| FUDGEBackbone=GPT2-large, Approach=Plug-and-play2022.08 | 67 | 64 | 62 | 55.6 | 11.5 | 35.9 |