Natural Language Inference on QNLI (test) (SEAT Bias)
0.01SEAT e-size (Names: Career/Family)Gender-tuning-random
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Gender-tuning-randomBackbone=BERT, Masking=Random2023.07 | 0.01 | 0.13 | 0.3 | 0.23 | 0.25 | 0.15 | 0.178 | 91.61 | |
| Gender-tuningBackbone=BERT2023.07 | 0.02 | 0.04 | 0.08 | 0.16 | 0.21 | 0.04 | 0.091 | 91.32 | |
| OriginBackbone=RoBERTa2023.07 | 0.04 | 0.22 | 0.53 | 0.48 | 0.47 | 0.36 | 0.35 | 92.03 | |
| Sent-DBackbone=BERT2023.07 | 0.05 | 0.004 | 0.08 | 0.62 | 0.71 | 0.44 | 0.321 | 90.6 | |
| FairFBackbone=BERT2023.07 | 0.1 | 0.2 | 0.32 | 0.28 | 0.24 | 0.16 | 0.222 | 90.8 | |
| Gender-tuningBackbone=RoBERTa2023.07 | 0.17 | 0.04 | 0.09 | 0.03 | 0.53 | 0.52 | 0.23 | 92.09 | |
| OriginBackbone=BERT2023.07 | 0.26 | 0.15 | 0.58 | 0.58 | 0.08 | 0.52 | 0.365 | 91.3 | |
| Gender-tuning-randomBackbone=RoBERTa, Masking=Random2023.07 | 0.38 | 0.1 | 0.16 | 0.06 | 0.57 | 0.47 | 0.29 | 92.51 |