Top five predicted words based on the probability for the masked token where text representations using hard prompt and meta soft prompt are compared. The examples of product reviews of the domains (a) Music and (b) Magazine in FDU-MTL dataset are shown. Red shows that the predicted word is seen as an irrelevant or wrong sentiment for the given product review. Blue shows that the predicted word is viewed as a correct sentiment.
| input text: This album contains only rap and no rock songs. This was very disappointing to say the least. → negative review | |
|---|---|
| hard prompt | meta soft prompt |
| 1. positive | 1. bad |
| 2. worth | 2. unacceptable |
| 3. negative | 3. disappointing |
| 4. lacking | 4. terrible |
| 5. good | 5. wrong |
| input text: This album contains only rap and no rock songs. This was very disappointing to say the least. → negative review | |
|---|---|
| hard prompt | meta soft prompt |
| 1. positive | 1. bad |
| 2. worth | 2. unacceptable |
| 3. negative | 3. disappointing |
| 4. lacking | 4. terrible |
| 5. good | 5. wrong |