Comparison of RA (%) and SA (%) using BERT, SimCSE, two-view CL (our SimCSE) and two-view AT (our AT) under adversarial perturbations of TextFooler and BERT-Attack. Yelp and IMDB datasets are evaluated
| Perturbations | Methods | Yelp | IMDB | ||
|---|---|---|---|---|---|
| RA | SA | RA | SA | ||
| TextFooler (Jin et al., 2020) | BERT (Devlin et al., 2019; Li et al., 2020b) | 6.6 | 95.6 | 13.6 | 90.9 |
| SimCSE (Gao et al., 2021) | 7.9 | 96.4 | 15.7 | 92.0 | |
| Ours (SimCSE) | 20.3 | 95.1 | 23.8 | 88.3 | |
| Ours (AT) | 35.3 | 94.5 | 39.3 | 87.9 | |
| BERT-Attack (Li et al., 2020b) | BERT (Devlin et al., 2019; Li et al., 2020b) | 5.1 | – | 11.4 | – |
| SimCSE (Gao et al., 2021) | 6.7 | – | 13.7 | – | |
| Ours (SimCSE) | 18.4 | – | 20.4 | – | |
| Ours (AT) | 31.9 | – | 35.3 | – | |
| Perturbations | Methods | Yelp | |||
|---|---|---|---|---|---|
| TextFooler ( | 6.6 | 95.6 | 13.6 | 90.9 | |
| SimCSE ( | 7.9 | 15.7 | |||
| Ours (SimCSE) | 20.3 | 95.1 | 23.8 | 88.3 | |
| Ours ( | 94.5 | 87.9 | |||
| BERT-Attack ( | 5.1 | – | 11.4 | – | |
| SimCSE ( | 6.7 | – | 13.7 | – | |
| Ours (SimCSE) | 18.4 | – | 20.4 | – | |
| Ours ( | – | – | |||
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.