Sentiment Classification on MR
90.8AccuracyRoBERTa (Gao et al., 2021)
Evaluation Results
| Method | Links | |
|---|---|---|
| RoBERTa (Gao et al., 2021)# Params=1.0x, Evaluation Protocol=Full fine-tuning2022.12 | 90.8 | |
| BERT+SCMModel type=Latent semantic tree models, Backbone=BERT2023.08 | 88.16 | |
| BERT (2019)Model type=Sequential models2023.08 | 87.65 | |
| SBERT2025.12 | 84.9 | |
| Kim et al., 2019Model type=Untagged tree (by external parser) models2023.08 | 83.8 | |
| NPM# Params=1.0x, Evaluation Protocol=Zero-shot, nonparametric=true2022.12 | 83.7 | |
| PICLShot=4-shot, Model=GPT-Neo (2.7B)2023.05 | 83.6 | |
| BILSTM+SCMModel type=Latent semantic tree models, Backbone=BiLSTM2023.08 | 83.41 | |
| BILSTM (1997)Model type=Sequential models2023.08 | 83.27 | |
| CNN+MCFAAdditional Context=Translation, N=ensemble2018.06 | 83.2 | |
| AdaSent2018.06 | 83.1 | |
| AdaSent2015.04 | 83.1 | |
| AdaSentNN category=Other2016.11 | 83.1 | |
| TopCNNAdditional Context=Topic, Granularity=ensemble2018.06 | 83 | |
| CNN+MCFAAdditional Context=Translation, N=102018.06 | 82.7 | |
| CNN+B1Additional Context=Translation, N=ensemble2018.06 | 82.6 | |
| CNN+MCFAAdditional Context=Translation, N=12018.06 | 82.3 | |
| BRNN2015.04 | 82.3 | |
| BLSTM-2DCNNNN category=ours2016.11 | 82.3 | |
| DSCNN2018.06 | 82.2 | |
| CNN+B2Additional Context=Translation, N=ensemble2018.06 | 82.2 | |
| SCD-RoBERTa_baseBackbone=RoBERTa-base2022.03 | 82.17 | |
| CNN+B2Additional Context=Translation, N=12018.06 | 82.1 | |
| CNN+B2Additional Context=Translation, N=102018.06 | 82.1 | |
| Dep-CNN2018.06 | 81.9 | |
| CNN+B1Additional Context=Translation, N=12018.06 | 81.9 | |
| Liu et al., 2017aModel type=Untagged tree (by external parser) models2023.08 | 81.9 | |
| BERT [CLS]-embeddingBackbone=BERT-base, Pooling=CLS2022.03 | 81.83 | |
| TopCNNAdditional Context=Topic, Granularity=word2018.06 | 81.7 | |
| CNN-Rule-qNetwork role=teacher network, Rule type=but-rule2016.03 | 81.7 | |
| Liu et al., 2017bModel type=Untagged tree (by external parser) models2023.08 | 81.7 | |
| RoBERTa# Params=1.0x, Evaluation Protocol=Zero-shot2022.12 | 81.7 | |
| CNN-Rule-pNetwork role=student network, Rule type=but-rule2016.03 | 81.6 | |
| USE_T+CNN (lrn w.e.)Universal Encoder=Transformer (USE_T), Transfer Level=Sentence, Embedding Initialization=learned on transfer task, Transfer Task Model=CNN2018.03 | 81.59 | |
| CNN-non-staticword vectors=fine-tuned pre-trained (word2vec)2014.08 | 81.5 | |
| CNN2018.06 | 81.5 | |
| CNN2015.04 | 81.5 | |
| DSCNNNN category=Other2016.11 | 81.5 | |
| BLSTM-2DPoolingNN category=ours2016.11 | 81.5 | |
| USE TUniversal Encoder=Transformer (USE_T), Transfer Level=Sentence, Embedding Initialization=None, Transfer Task Model=None2018.03 | 81.44 | |
| CNN+B1Additional Context=Translation, N=102018.06 | 81.4 | |
| USE_T+DAN (lrn w.e.)Universal Encoder=Transformer (USE_T), Transfer Level=Sentence, Embedding Initialization=learned on transfer task, Transfer Task Model=DAN2018.03 | 81.36 | |
| USE_T+DAN (w2v w.e.)Universal Encoder=Transformer (USE_T), Transfer Level=Sentence & Word, Embedding Initialization=word2vec skip-gram, Transfer Task Model=DAN2018.03 | 81.32 | |
| TopCNNAdditional Context=Topic, Granularity=sentence2018.06 | 81.3 | |
| CNNModel variant=non-static2016.03 | 81.3 | |
| RoBERTa [CLS]-embeddingBackbone=RoBERTa-base, Pooling=CLS2022.03 | 81.27 | |
| USE_T+CNN (w2v w.e.)Universal Encoder=Transformer (USE_T), Transfer Level=Sentence & Word, Embedding Initialization=word2vec skip-gram, Transfer Task Model=CNN2018.03 | 81.18 | |
| CNN-multichannelword vectors=pre-trained (word2vec), channels=22014.08 | 81.1 | |
| CNN-multichannel2016.03 | 81.1 | |
| IS-BERT_baseBackbone=BERT-base2022.03 | 81.09 | |
| CNN-AnaNN category=CNN2016.11 | 81.02 | |
| CNN-staticword vectors=static pre-trained (word2vec)2014.08 | 81 | |
| BLSTM-AttNN category=ours2016.11 | 81 | |
| RoBERTa (Gao et al., 2021)# Params=1.0x, Evaluation Protocol=Zero-shot2022.12 | 80.8 | |
| SimCSE-BERT_baseBackbone=BERT-base2022.03 | 80.74 | |
| LM+TSED+PT+2LEmb.Size=50, Teacher Model Architecture=LSTM2019.05 | 80.33 | |
| STM+TSED+PT+2LEmb.Size=50, Teacher Model Architecture=LSTM2019.05 | 80.09 | |
| BLSTMNN category=ours2016.11 | 80 | |
| SCLie2025.12 | 79.8 | |
| SCNN2025.12 | 79.7 | |
| Sent-Parser2014.08 | 79.5 | |
| NBSVM2014.08 | 79.4 | |
| NB-SVM2015.04 | 79.4 | |
| NLM+TSED+PT+2LEmb.Size=50, Teacher Model Architecture=LSTM2019.05 | 79.33 | |
| LSTM-400Emb.Size=4002019.05 | 79.28 | |
| F-Dropout2014.08 | 79.1 | |
| CNN-400Emb.Size=4002019.05 | 79.07 | |
| LM+TSED+PT+2LEmb.Size=50, Teacher Model Architecture=CNN2019.05 | 79.06 | |
| MV-RNN2014.08 | 79 | |
| MNB2014.08 | 79 | |
| G-Dropout2014.08 | 79 | |
| MNB2015.04 | 79 | |
| MV-RecNN2015.04 | 79 | |
| G-Dropout2016.03 | 79 | |
| MVRNN (2012)Model type=Untagged tree (by external parser) models2023.08 | 79 | |
| NLM+TSEDEmb.Size=50, Teacher Model Architecture=LSTM2019.05 | 78.89 | |
| STM+TSEDEmb.Size=50, Teacher Model Architecture=LSTM2019.05 | 78.85 | |
| STM+TSED+PT+2LEmb.Size=50, Teacher Model Architecture=CNN2019.05 | 78.77 | |
| TreeLSTM (2015)Model type=Untagged tree (by external parser) models2023.08 | 78.7 | |
| Avg. BERT embeddingsBackbone=BERT-base, Pooling=Avg2022.03 | 78.66 | |
| LM+TSEDEmb.Size=50, Teacher Model Architecture=LSTM2019.05 | 78.61 | |
| NLM+TSED+PT+2LEmb.Size=50, Teacher Model Architecture=CNN2019.05 | 78.6 | |
| USE_D+CNN (lrn w.e.)Universal Encoder=Deep Averaging Network (USE_D), Transfer Level=Sentence, Embedding Initialization=learned on transfer task, Transfer Task Model=CNN2018.03 | 78.49 | |
| USE_D+CNN (w2v w.e.)Universal Encoder=Deep Averaging Network (USE_D), Transfer Level=Sentence & Word, Embedding Initialization=word2vec skip-gram, Transfer Task Model=CNN2018.03 | 78.2 | |
| GPT-2 kNN-LM# Params=2.2x, Evaluation Protocol=Zero-shot2022.12 | 78.2 | |
| NLM+TSEDEmb.Size=50, Teacher Model Architecture=CNN2019.05 | 78.1 | |
| CNN-50Emb.Size=502019.05 | 78.07 | |
| STM+TSEDEmb.Size=50, Teacher Model Architecture=CNN2019.05 | 77.81 | |
| CCAE2014.08 | 77.8 | |
| RAE2014.08 | 77.7 | |
| RAE2015.04 | 77.7 | |
| LM+TSEDEmb.Size=50, Teacher Model Architecture=CNN2019.05 | 77.63 | |
| USE_D+DAN (lrn w.e.)Universal Encoder=Deep Averaging Network (USE_D), Transfer Level=Sentence, Embedding Initialization=learned on transfer task, Transfer Task Model=DAN2018.03 | 77.57 | |
| DPCLie2025.12 | 77.4 | |
| Tree-CRF2014.08 | 77.3 | |
| GloVe embeddings (avg.)Model=GloVe (avg.)2022.03 | 77.25 | |
| CBOW2015.04 | 77.2 | |
| RNN2015.04 | 77.2 | |
| DPCNN2025.12 | 77.2 | |
| LSTM-50Emb.Size=502019.05 | 77.16 |