AMR Generation on LDC2015E86 (test)
35.3BLEUDCGCN
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DCGCNextra silver data=0.3M, ensemble=true2021.02 | 35.3 | — | |
| GIGA-20MPre-training Corpus=Gigaword, Training Size=20M sentences2017.04 | 33.8 | — | |
| MSeq2seq+Anonadditional training data=20M Gigaword sentences2018.05 | 33.8 | — | |
| LSTMextra silver data=20M2021.02 | 33.8 | — | |
| Graph2seq+charLSTM+copyadditional training data=2M Gigaword sentences2018.05 | 33.6 | — | |
| GRNextra silver data=2M2021.02 | 33.6 | — | |
| DCGCNextra silver data=0.3M2021.02 | 33.2 | — | |
| GIGA-2MPre-training Corpus=Gigaword, Training Size=2M sentences2017.04 | 32.3 | — | |
| MSeq2seq+Anonadditional training data=2M Gigaword sentences2018.05 | 32.3 | — | |
| Seq2seq+charLSTM+copyadditional training data=2M Gigaword sentences2018.05 | 31.7 | — | |
| RA-Trans-F-oursTraining Time (s/step)=0.61, Loss Configuration=Loss 1 + Loss 2 (Both)2021.02 | 31.41 | 36.2 | |
| RA-Trans-F-oursTraining Time (s/step)=0.52, Loss Configuration=Loss 2 (linearized graph)2021.02 | 31.13 | 36.1 | |
| RA-Trans-F-oursTraining Time (s/step)=0.38, Loss Configuration=Loss 1 (triple relations)2021.02 | 30.47 | 35.5 | |
| RA-Trans-SADescription=self attention to model relation path2021.02 | 29.66 | 35.45 | |
| RA-Trans-F-oursTraining Time (s/step)=0.25, Loss Configuration=Baseline2021.02 | 29.11 | 35 | |
| Graph2seq+charLSTM+copyadditional training data=200K Gigaword sentences2018.05 | 28.2 | — | |
| GIGA-200kPre-training Corpus=Gigaword, Training Size=200k sentences2017.04 | 27.4 | — | |
| MSeq2seq+Anonadditional training data=200K Gigaword sentences2018.05 | 27.4 | — | |
| Seq2seq+charLSTM+copyadditional training data=200K Gigaword sentences2018.05 | 27.4 | — | |
| PBMT*Training Corpus=LDC2014T122017.04 | 26.9 | — | |
| PBMTadditional training data=None2018.05 | 26.9 | — | |
| DCGCNDescription=graph neural network2021.02 | 25.7 | — | |
| SNRGadditional training data=None2018.05 | 25.6 | — | |
| GRAPHEncoder=Graph-based GCNSEQ, Modeling reentrancies=true2019.03 | 24.4 | 23.6 | |
| TREEEncoder=Tree-based GCNSEQ, Modeling reentrancies=false2019.03 | 23.93 | 23.32 | |
| Graph2seq+charLSTM+copyadditional training data=None2018.05 | 23.3 | — | |
| Song et al. (2018)Encoder=Graph-based2019.03 | 23.3 | — | |
| GRNDescription=graph neural network2021.02 | 23.28 | — | |
| TREETOSTR2017.04 | 23 | — | |
| Tree2Stradditional training data=None2018.05 | 23 | — | |
| Graph2seq+copyadditional training data=None2018.05 | 22.7 | — | |
| TSP2017.04 | 22.4 | — | |
| AMR-ONLYTraining Data=AMR corpus only2017.04 | 22 | — | |
| MSeq2seq+Anonadditional training data=None2018.05 | 22 | — | |
| Konstas et al. (2017)Encoder=Sequential2019.03 | 22 | — | |
| LSTMDescription=multi-layer LSTM on linearized AMRs2021.02 | 22 | — | |
| SEQEncoder=Sequential2019.03 | 21.43 | 21.53 |