English Question Formation (Hierarchical generalization)
99Partial AccuracyTree-GRU
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Tree-GRUArchitecture=Tree-structured GRU, Source=McCoy et al. (2020)2025.11 | 99 | — | |
| TfSource=Murty et al. (2023), Overtrained=true2025.11 | 92 | — | |
| Tf+Nd+NdArchitecture=Multi-Stack Attention Transformer2025.11 | 86.2 | 19.1 | |
| TfSource=Qin et al. (2025), Uses ground-truth parse trees=true2025.11 | 84 | — | |
| GRU+loc. attn.Architecture=GRU + local attention, Source=McCoy et al. (2020)2025.11 | 78 | — | |
| TfSource=Ahuja et al. (2025), Training objective=language modeling objective2025.11 | 75 | — | |
| Tf+NdArchitecture=Stack Attention Transformer2025.11 | 73.2 | 31.8 | |
| Tf+Sup+SupArchitecture=Transformer + Supervision + Supervision2025.11 | 69.7 | 16.1 | |
| TfArchitecture=Transformer2025.11 | 64.5 | 0.5 | |
| TfSource=Murty et al. (2023)2025.11 | 34 | — | |
| Tf+SupArchitecture=Transformer + Supervision2025.11 | 32.5 | 3.9 | |
| ON-LSTMArchitecture=Ordered Neurons LSTM, Source=McCoy et al. (2020)2025.11 | 5 | — |