Comparison of classification accuracy (%) on various domains in Amazon review dataset by using fine-tuning (FT), soft prompt (SP) [27] and the proposed meta soft prompt (MSP) where the length of soft prompt is set to be 2 (denoted by SP† and MSP†) and 5 (denoted by SP‡ and MSP‡). The highest number among different methods is shown by bold.
| Domain | FT | SP† | MSP† | SP‡ | MSP‡ |
|---|---|---|---|---|---|
| Books | 80.8 | 85.2 | 83.6 | 86.8 | 88.0 |
| DVD | 79.3 | 83.2 | 83.4 | 84.4 | 85.6 |
| Electronics | 79.4 | 82.1 | 82.6 | 84.8 | 85.1 |
| Kitchen | 79.5 | 84.5 | 86.2 | 86.1 | 87.6 |
| Domain | FT | SP† | MSP† | SP‡ | MSP‡ |
|---|---|---|---|---|---|
| Books | 80.8 | 83.6 | 86.8 | ||
| DVD | 79.3 | 83.2 | 84.4 | ||
| Electronics | 79.4 | 82.1 | 84.8 | ||
| Kitchen | 79.5 | 84.5 | 86.1 |