Chinese Spelling Correction on CSCD-NS
79.71Sentence Correction F1 ScoreQwen3-32B-RL
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-32B-RLparameters=32B2025.12 | 79.71 | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-14B-RLparameters=14B2025.12 | 76.34 | — | — | — | — | — | — | — | — | — | — | — | |
| PLOME + LCSTS-IME-2MArchitecture=Token-level classification, Pseudo Dataset=LCSTS-IME-2M2022.11 | 74.42 | 81.2 | 72.21 | 76.44 | 79.05 | 70.3 | 84.21 | 73.81 | 78.67 | 82 | 71.88 | 76.6 | |
| ReLM-D2C2025.12 | 74 | — | — | — | — | — | — | — | — | — | — | — | |
| C-LLM2025.12 | 73.8 | — | — | — | — | — | — | — | — | — | — | — | |
| SM BERT + LCSTS-IME-2MArchitecture=Token-level classification, Pseudo Dataset=LCSTS-IME-2M2022.11 | 73.62 | 79.19 | 74.86 | 76.97 | 75.75 | 71.6 | 82.39 | 77.93 | 80.1 | 78.63 | 74.37 | 76.44 | |
| BERT + LCSTS-IME-2MArchitecture=Token-level classification, Pseudo Dataset=LCSTS-IME-2M2022.11 | 72.96 | 78.98 | 73.6 | 76.2 | 75.63 | 70.47 | 82.19 | 75.75 | 78.84 | 78.84 | 72.67 | 75.63 | |
| SCOPE2025.12 | 71.7 | — | — | — | — | — | — | — | — | — | — | — | |
| ReLM2025.12 | 69.5 | — | — | — | — | — | — | — | — | — | — | — | |
| Doubao2025.12 | 69.45 | — | — | — | — | — | — | — | — | — | — | — | |
| DeepSeek-32Bparameters=32B2025.12 | 67.32 | — | — | — | — | — | — | — | — | — | — | — | |
| SMBERT2025.12 | 67.22 | — | — | — | — | — | — | — | — | — | — | — | |
| Gmini 2.52025.12 | 66.29 | — | — | — | — | — | — | — | — | — | — | — | |
| SM BERTArchitecture=Token-level classification, Pseudo Dataset=None2022.11 | 66.2 | 80.87 | 64.78 | 71.94 | 74.42 | 59.62 | 84.46 | 65.35 | 73.68 | 77.5 | 59.97 | 67.62 | |
| Baichuan2-13B + LCSTS-IME-2MArchitecture=Generative/LoRA, Pseudo Dataset=LCSTS-IME-2M2022.11 | 66.1 | 67.82 | 67.35 | 67.58 | 66.33 | 65.87 | 61.67 | 73.91 | 67.24 | 60.06 | 71.98 | 65.48 | |
| PLOMEArchitecture=Token-level classification, Pseudo Dataset=None2022.11 | 65.23 | 79.78 | 57.23 | 66.65 | 78.09 | 56.01 | 83.48 | 57.99 | 68.44 | 81.49 | 56.61 | 66.81 | |
| Baichuan2-7B + LCSTS-IME-2MArchitecture=Generative/LoRA, Pseudo Dataset=LCSTS-IME-2M2022.11 | 64.44 | 66.94 | 66.13 | 66.54 | 64.84 | 64.05 | 60.63 | 72.57 | 66.07 | 58.55 | 70.08 | 63.8 | |
| BERTArchitecture=Token-level classification, Pseudo Dataset=None2022.11 | 64.06 | 79.16 | 65.83 | 71.88 | 70.55 | 58.66 | 83 | 67.01 | 74.15 | 73.59 | 59.41 | 65.75 | |
| Baichuan2-13BArchitecture=Generative/LoRA, Pseudo Dataset=None2022.11 | 61.42 | 67.53 | 60.23 | 63.67 | 65.14 | 58.11 | 60.07 | 64.62 | 62.26 | 57.49 | 61.86 | 59.6 | |
| DeepSeek-14Bparameters=14B2025.12 | 60.18 | — | — | — | — | — | — | — | — | — | — | — | |
| Claude 3.72025.12 | 59.07 | — | — | — | — | — | — | — | — | — | — | — | |
| Baichuan2-7BArchitecture=Generative/LoRA, Pseudo Dataset=None2022.11 | 56.35 | 64.98 | 53.04 | 58.41 | 62.7 | 51.17 | 57.1 | 56.92 | 57.01 | 54.72 | 54.55 | 54.63 | |
| GPT-42025.12 | 54.41 | — | — | — | — | — | — | — | — | — | — | — | |
| Qwen3-32Bparameters=32B2025.12 | 54.41 | — | — | — | — | — | — | — | — | — | — | — | |
| GPT4Architecture=Generative, Evaluation Protocol=10-shot In-context Learning (ICL)2022.11 | 54.28 | 58.37 | 59.71 | 59.03 | 53.67 | 54.9 | 58.4 | 63.6 | 60.89 | 52.34 | 57 | 54.57 | |
| Qwen3-14Bparameters=14B2025.12 | 53.78 | — | — | — | — | — | — | — | — | — | — | — | |
| ChatGPT2025.12 | 52.5 | — | — | — | — | — | — | — | — | — | — | — | |
| ChatGPTArchitecture=Generative, Evaluation Protocol=10-shot In-context Learning (ICL)2022.11 | 51.14 | 59.74 | 51.6 | 55.38 | 55.17 | 47.66 | 60.41 | 55.73 | 57.98 | 54.84 | 50.59 | 52.63 | |
| MDCSpell+ARM2025.12 | 48.93 | — | — | — | — | — | — | — | — | — | — | — | |
| PGT (BERT)2025.12 | 48.57 | — | — | — | — | — | — | — | — | — | — | — | |
| BART + LCSTS-IME-2MArchitecture=Sequence-to-sequence, Pseudo Dataset=LCSTS-IME-2M2022.11 | 46.22 | 42.06 | 54.29 | 47.4 | 41.01 | 52.95 | 40.87 | 75.97 | 53.15 | 39.68 | 73.75 | 51.6 | |
| SoftMask2025.12 | 44.48 | — | — | — | — | — | — | — | — | — | — | — | |
| MDCSpell2025.12 | 42.08 | — | — | — | — | — | — | — | — | — | — | — | |
| BARTArchitecture=Sequence-to-sequence, Pseudo Dataset=None2022.11 | 38.47 | 38.73 | 46.05 | 42.08 | 35.41 | 42.11 | 36.97 | 63.32 | 46.69 | 33.3 | 57.04 | 42.05 | |
| BERT2025.12 | 25.49 | — | — | — | — | — | — | — | — | — | — | — |