The Ultimate Guide to Chinese Spelling & Grammar Correction: Champion Models in Action

Do you struggle with confusing “的,” “得,” and “地” in Chinese writing? Or worry about typos in important documents? This guide reveals award-winning AI tools that have dominated NLP competitions for three consecutive years – complete with practical implementation tutorials.

1. Core Technology Breakdown

1.1 Evolution of Champion Models

This project has won three consecutive championships in authoritative competitions:

• 🏆 2024 CCL Champion (Research Paper)
• 🏆 2023 NLPCC-NaCGEC Champion
• 🏆 2022 FCGEC Champion

1.2 Model Capability Matrix

2. Performance Benchmarks

2.1 Spelling Correction (F1 Scores)

2.2 Grammar Correction Dominance (NaCGEC Dataset)

# Champion model evaluation:
Precision = 0.743
Recall = 0.7294
F0.5 = 0.7402  # 10 points above Huawei's solution

3. 4-Step Implementation Guide

3.1 Environment Setup

conda create -n zh_correct python=3.10
conda activate zh_correct
pip install transformers vllm==0.8.5

3.2 Single-Sentence Correction (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("twnlp/ChineseErrorCorrector3-4B")
tokenizer = AutoTokenizer.from_pretrained("twnlp/ChineseErrorCorrector3-4B")

# Professional correction prompt
prompt = "As a text correction expert, fix grammatical errors in: "
text_input = "对待每一项工作都要一丝不够。"  # Contains error

messages = [{"role": "user", "content": prompt + text_input}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Execute correction
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))
# Output: 对待每一项工作都要一丝不苟。 (Corrected)

3.3 Batch Processing (VLLM – Production Recommended)

from vllm import LLM, SamplingParams

llm = LLM(model="twnlp/ChineseErrorCorrector3-4B")
sampling_params = SamplingParams(max_tokens=512)

# Batch input example
batch_texts = [
    "这个苹果非常好吃。",  # Correct
    "他昨天去图书馆读书。",  # Correct
    "会议将在明天下午三点钟举行。"  # Correct
]

results = llm.generate(batch_texts, sampling_params)
for output in results:
    print(f"Corrected: {output.outputs[0].text}")

3.4 Engineering Deployment

1. Clone repository:

   git clone https://github.com/TW-NLP/ChineseErrorCorrector
   cd ChineseErrorCorrector
   

2. Configure settings:

   # config.py key settings:
   DEFAULT_CKPT_PATH = "twnlp/ChineseErrorCorrector3-4B"  # Model selection
   USE_VLLM = True  # Enable high-performance inference
   

3. Run batch processing:

   python main.py
   

4. Training Datasets Explained

4.1 Dataset Composition

5. Frequently Asked Questions

5.1 Which model should I choose?

• Professional Editing: ChineseErrorCorrector3-4B
• Real-time Systems: ChineseErrorCorrector-1.5B
• Domain-Specific Texts: Use custom training tools

5.2 What error types are covered?

1. Spelling Errors: Character confusion (e.g., “帐号” → “账号”)
2. Grammar Errors: Sentence structure issues (e.g., “原因是…造成”)
3. Collocation Errors: Measure word misuse (e.g., “一个书”)
4. Redundancy: Duplicate words (e.g., “大约半小时左右”)

5.3 How to improve domain-specific results?

# Legal document enhancement example:
prompt = "As a legal document proofreader, correct errors in this clause: "
text_input = "甲方应于签约后三十日个工作日内付款。"  # Contains error

6. Technical Deep Dive

6.1 Multi-Stage Training

1. Data Augmentation: 14 grammatical error patterns
2. Iterative Training: Multi-round fine-tuning with 2M samples
3. Domain Adaptation: Legal/medical/official text optimization

6.2 Champion Model Architecture

graph LR
A[Input Text] --> B(Error Detection)
B --> C{Error Type}
C -->|Spelling| D[Character Similarity Analysis]
C -->|Grammar| E[Dependency Parsing]
D & E --> F[Correction Generation]
F --> G[Corrected Output]

7. Academic Reference

@inproceedings{wei2024automated,
  title={Automated Detection, Correction and Fluency Assessment of Grammatical Errors in Chinese Composition},
  author={Wei, Tian},
  booktitle={Proceedings of the 23rd Chinese National Conference on Computational Linguistics},
  pages={278--284},
  year={2024}
}

Project evolution: