Natural Language Inference on MultiNLI matched (test)
85.38AccuracyBERT-base + PoE
Evaluation Results
| Method | Links | |
|---|---|---|
| BERT-base + PoEDebiasing Strategy=Z-Aug2022.03 | 85.38 | |
| BERT-baseDebiasing Strategy=Z-Aug2022.03 | 85.12 | |
| BERT-base + PoETraining Data=D_MNLI2022.03 | 84.69 | |
| PoETraining Data=MNLI2022.03 | 84.11 | |
| BERT-baseTraining Data=D_MNLI baseline2022.03 | 84.11 | |
| Regularized-confTraining Data=MNLI2022.03 | 84.1 | |
| BERT-baseDebiasing Strategy=Seq-Z2022.03 | 83.41 | |
| BERT-baseDebiasing Strategy=Par-Z2022.03 | 83.27 | |
| DRCN+ELMoensemble=true, embedding=ELMo2018.05 | 82.3 | |
| LM-Transformerensemble=false2018.05 | 82.1 | |
| DRCNensemble=true2018.05 | 80.6 | |
| CAFEensemble=true2018.05 | 80.2 | |
| DIINEnsemble=true2017.09 | 80 | |
| DIINensemble=true2018.05 | 80 | |
| Learned-MixinTraining Data=MNLI2022.03 | 79.5 | |
| DRCNensemble=false2018.05 | 79.1 | |
| DIINEnsemble=false2017.09 | 78.8 | |
| DIINensemble=false2018.05 | 78.8 | |
| CAFEensemble=false2018.05 | 78.7 | |
| CAFE2017.11 | 78.7 | |
| DIINExtra SNLI training set=true2017.11 | 77.8 | |
| KIM2017.11 | 77.2 | |
| ESIM2017.11 | 76.8 | |
| Gated-Att BiLSTMEnsemble=true2017.09 | 74.9 | |
| Shortcut-Stacked encoderEnsemble=false2017.09 | 74.6 | |
| Shortcut stacked BiLSTM# Params=140M, Additional training data=SNLI dataset2018.09 | 74.6 | |
| 2-layer Bi-CAS-LSTM# Params=6.8M2018.09 | 74.6 | |
| SS BILSTM2017.11 | 74.6 | |
| ESIM GloVeAuxiliary Information=GloVe vectors2017.06 | 74.58 | |
| Shortcut-Stacked BiLSTM encodersSNLI Mix=true, Parameter count=140.2m2017.12 | 74.5 | |
| Shortcut-Stacked BiLSTM2018.08 | 74.5 | |
| 3-layer Bi-CAS-LSTM# Params=8.6M2018.09 | 74.2 | |
| Distance-based Self-Attention NetworkSNLI Mix=false, Parameter count=4.7m2017.12 | 74.1 | |
| 2-layer CAS-LSTM# Params=2.9M2018.09 | 74 | |
| BiLSTM with generalized pooling# Params=18M2018.09 | 73.8 | |
| 3-layer CAS-LSTM# Params=4.8M2018.09 | 73.8 | |
| HBMPdimensions=600D2018.08 | 73.7 | |
| Deep Gated Attn. BiLSTM encodersSNLI Mix=false, Parameter count=11.6m2017.12 | 73.5 | |
| Deep Gated Attn. BiLSTM encoders2018.08 | 73.5 | |
| BiLSTM with gated pooling# Params=12M2018.09 | 73.5 | |
| Gated BiLSTM2017.11 | 73.5 | |
| Gated-Att BiLSTMEnsemble=false2017.09 | 73.2 | |
| ESIMEnsemble=false2017.09 | 72.3 | |
| ESIMensemble=false2018.05 | 72.3 | |
| InnerAttEnsemble=true2017.09 | 72.2 | |
| BiLSTM + Inner-attentionSNLI Mix=true2017.12 | 72.1 | |
| InnerAttEnsemble=false2017.09 | 72.1 | |
| BiLSTM + Inner-attention2018.08 | 72.1 | |
| ESIM dictionaryAuxiliary Information=dictionary definitions2017.06 | 71.45 | |
| Directional Self-Attention NetworkSNLI Mix=false, Parameter count=2.4m2017.12 | 71 | |
| DiSAN2017.11 | 71 | |
| BiLSTM + enhanced embedding + max poolingSNLI Mix=false2017.12 | 70.7 | |
| BiLSTM + enh embed + max pooling2018.08 | 70.7 | |
| ESIM spellingAuxiliary Information=spelling2017.06 | 68.89 | |
| ESIM baselineAuxiliary Information=none2017.06 | 68.55 | |
| Cha-level Intra-attention BiLSTM encodersSNLI Mix=true2017.12 | 67.9 | |
| BiLSTMSNLI Mix=true2017.12 | 67.5 | |
| BiLSTM2018.08 | 67.5 | |
| BiLSTMEnsemble=false2017.09 | 67 | |
| BiLSTM# Params=null2018.09 | 66.9 | |
| BILSTM2017.11 | 66.9 | |
| CBOWSNLI Mix=true2017.12 | 66.2 | |
| CBOW2018.08 | 66.2 | |
| CBOW# Params=null2018.09 | 64.8 | |
| CBOW2017.11 | 64.8 |