Natural Language Inference on SNLI 1.0 (test)
90.67AccuracyBERT-base w/ Z-Aug
Evaluation results
| Method | Links | |
|---|---|---|
| BERT-base w/ Z-AugBackbone=BERT-base, Debiasing Dataset=Z-Aug Z(DG + DSNLI)2022.03 | 90.67 | |
| BERT-base baselineBackbone=BERT-base, Training Data=DSNLI2022.03 | 90.45 | |
| BERT-base + PoE w/ DSNLIBackbone=BERT-base, Debiasing Strategy=PoE, Training Data=DSNLI2022.03 | 90.25 | |
| PoETraining Data=SNLI, Debiasing Strategy=Product of Experts2022.03 | 90.11 | |
| DFLTraining Data=SNLI, Debiasing Strategy=Debiased Focal Loss2022.03 | 89.57 | |
| BERT-base w/ Par-ZBackbone=BERT-base, Debiasing Dataset=Par-Z Z(DSNLI) U Z(DG+)2022.03 | 88.11 | |
| BERT-base w/ Seq-ZBackbone=BERT-base, Debiasing Dataset=Seq-Z Z(DG -> Z(DSNLI))2022.03 | 88.08 | |
| BERT-base + PoE w/ Seq-ZBackbone=BERT-base, Debiasing Strategy=PoE, Training Data=Seq-Z Z(DG+ -> Z(DSNLI))2022.03 | 87.65 | |
| Multiple DSADimension=2400D, Parameters (m)=7.0, T(s)/epoch=1982018.08 | 87.4 | |
| Single DSADimension=600D, Parameters (m)=2.1, T(s)/epoch=1352018.08 | 86.8 | |
| Reinforced self-attention networkDimension=300D, Parameters (m)=3.1, T(s)/epoch=6222018.08 | 86.3 | |
| Distance-based self-attention networkDimension=1200D, Parameters (m)=4.7, T(s)/epoch=6932018.08 | 86.3 | |
| Gumbel TreeLSTMDimension=600D, Parameters (m)=10.02018.08 | 86 | |
| Residual stacked encodersDimension=600D, Parameters (m)=29.02018.08 | 86 | |
| Directional self-attention networkDimension=300D, Parameters (m)=2.4, T(s)/epoch=5872018.08 | 85.6 | |
| CNN (Dense) with self-attentionDimension=600D, Parameters (m)=2.4, T(s)/epoch=1212018.08 | 84.6 | |
| BiLSTM with self-attentionDimension=600D, Parameters (m)=2.82018.08 | 84.2 | |
| Ens. AdvClsTraining Data=SNLI, Debiasing Strategy=Ensemble Adversarial Classifier2022.03 | 84.09 | |
| AdvClsTraining Data=SNLI, Debiasing Strategy=Adversarial Classifier2022.03 | 83.56 |