Grammatical Error Correction on JFLEG
63.3GLEUTransformer + Pre-training + Iterative refinement + Ensemble Decoding (8)
Evaluation results
| Method | Links | |
|---|---|---|
| Transformer + Pre-training + Iterative refinement + Ensemble Decoding (8)Backbone=Transformer, Pre-training=true, Iterative refinement=true, Ensemble size=8, Decoding Strategy=Sequential2019.10 | 63.3 | |
| Lichtarge et al.pseudo-data=true, ensemble=true2020.05 | 63.3 | |
| Base + FB learning and inferencetraining=fluency boost learning, inference=fluency boost inference2018.07 | 62.42 | |
| Human performance2018.07 | 62.37 | |
| ELECTRA-VERNetTop-K=5, Backbone=ELECTRA2021.05 | 62.07 | |
| ELECTRA-VERNetTop-K=4, Backbone=ELECTRA2021.05 | 62.05 | |
| BERT-fuse GED + R2Lpseudo-data=true, ensemble=true, R2L=true2020.05 | 62 | |
| Kaneko et al.Configuration=Ensemble2020.05 | 62 | |
| BERT-fuse (GED)R2L=true2021.05 | 62 | |
| ELECTRA-VERNetTop-K=3, Backbone=ELECTRA2021.05 | 61.8 | |
| Lichtarge et al.pseudo-data=true, ensemble=false2020.05 | 61.6 | |
| ELECTRA-VERNetTop-K=2, Backbone=ELECTRA2021.05 | 61.6 | |
| SMT-NMT hybrid2018.07 | 61.5 | |
| SMT + NMT (Bi-GRU) + LM + Ensemble Decoding (4) + SpellcheckBackbone=SMT + NMT (Bi-GRU), Language Model (LM) usage=true, Ensemble size=4, Spellcheck integration=true, Decoding Strategy=Sequential2019.10 | 61.5 | |
| Base + FB learningtraining=fluency boost learning2018.07 | 61.41 | |
| Kiyono et al.*pseudo-data=true, ensemble=true2020.05 | 61.4 | |
| Basic GECR2L=true2021.05 | 61.4 | |
| BERT-fuse GEDpseudo-data=true, ensemble=false2020.05 | 61.3 | |
| Kaneko et al.Configuration=Single2020.05 | 61.3 | |
| BERT-fuse (GED)R2L=false2021.05 | 61.3 | |
| BERT-VERNetTop-K=3, Backbone=BERT2021.05 | 61.24 | |
| BERT-fusepseudo-data=true, ensemble=false2020.05 | 61.2 | |
| BERT-fuse maskpseudo-data=true, ensemble=false2020.05 | 61.2 | |
| Grundkiewicz et al.pseudo-data=true, ensemble=true2020.05 | 61.2 | |
| Kiyono et al.Configuration=Ensemble2020.05 | 61.2 | |
| BERT-VERNetTop-K=2, Backbone=BERT2021.05 | 61.17 | |
| BERT-VERNetTop-K=4, Backbone=BERT2021.05 | 61.16 | |
| BERT-VERNetTop-K=5, Backbone=BERT2021.05 | 61.1 | |
| Transformer + Pre-training + LM + Spellcheck + Ensemble Decoding (4)Backbone=Transformer, Pre-training=true, Language Model (LM) usage=true, Spellcheck integration=true, Ensemble size=4, Decoding Strategy=Sequential2019.10 | 61 | |
| PIEBackbone=Transformer, Language Model (LM) usage=true, Pre-training=true, Spellcheck integration=true, Iterative refinement=true, Ensemble size=5, Decoding Strategy=Iterative2019.10 | 61 | |
| w/o BERTpseudo-data=true, ensemble=false2020.05 | 61 | |
| Basic GECR2L=false2021.05 | 61 | |
| Base convolutional seq2seq2018.07 | 60.87 | |
| Awasthi et al.pseudo-data=true, ensemble=false2020.05 | 60.3 | |
| Adapted-transformer2018.07 | 59.9 | |
| Kiyono et al.pseudo-data=true, ensemble=false2020.05 | 59.7 | |
| Kiyono et al.Configuration=Single2020.05 | 59.7 | |
| CNN + LM + Ensemble Decoding (4) + SpellcheckBackbone=CNN, Language Model (LM) usage=true, Ensemble size=4, Spellcheck integration=true, Decoding Strategy=Sequential2019.10 | 57.5 | |
| NUS182018.07 | 57.47 | |
| BART-basedConfiguration=Single2020.05 | 57.3 | |
| BART-basedConfiguration=Ensemble2020.05 | 57.2 | |
| NUS172018.07 | 56.78 | |
| BERT-fuse GEDpseudo-data=false, ensemble=false2020.05 | 54.4 | |
| BERT-fuse maskpseudo-data=false, ensemble=false2020.05 | 54.2 | |
| BERT-fusepseudo-data=false, ensemble=false2020.05 | 54.1 | |
| BERT-initpseudo-data=false, ensemble=false2020.05 | 53 | |
| w/o BERTpseudo-data=false, ensemble=false2020.05 | 52.7 |