nguyenthanhasia/vsec-vietnamese-spell-correction

VSEC: Vietnamese Spell Correction Dataset

Dataset Description

VSEC (Vietnamese Spell Correction) is a comprehensive dataset for Vietnamese spelling error detection and correction, containing 9,341 sentences with 11,202 human-made misspellings across 5,211 unique error types. This dataset represents the largest publicly available collection of Vietnamese spelling errors with syllable-level annotations, making it an invaluable resource for developing and evaluating Vietnamese spell correction systems.

The dataset addresses both mistyped errors (occurring during typing) and misspelled errors (caused by regional pronunciation differences or word difficulty), providing a realistic foundation for training robust Vietnamese spell correction models. Each sentence includes detailed syllable-level annotations with error detection flags and correction suggestions.

Dataset Summary

• Total sentences: 9,341
• Total errors: 11,202
• Unique error types: 5,211
• Average errors per sentence: 1.20
• Language: Vietnamese
• Format: JSONL with syllable-level annotations
• Error coverage: 100% (all sentences contain at least one error)

Dataset Structure

Data Fields

Each example in the dataset contains the following fields:

• text (string): The original sentence containing spelling errors
• corrected_text (string): The corrected version of the sentence
• syllable_annotations (list): Detailed syllable-level annotations including:
  • syllable (string): The syllable text
  • is_correct (boolean): Whether the syllable is correctly spelled
  • corrections (list): List of correction suggestions for incorrect syllables
  • position (int): Position of the syllable in the sentence (0-indexed)
• error_count (int): Number of errors in the sentence
• error_positions (list): List of positions where errors occur
• correction_pairs (list): List of error-correction pairs with position information
• has_errors (boolean): Whether the sentence contains any errors

Data Example

{
  "text": "Thông qua công tác tuyên truyền, vận động này phụ huynh sẽ hiểu rõ hơn tầm quan trọng của việc gìn giữ môi trường sanh , sạch , đẹp.",
  "corrected_text": "Thông qua công tác tuyên truyền, vận động này phụ huynh sẽ hiểu rõ hơn tầm quan trọng của việc gìn giữ môi trường xanh , sạch , đẹp.",
  "error_count": 1,
  "error_positions": [51],
  "correction_pairs": [
    {
      "error": "sanh",
      "correction": "xanh", 
      "position": 51
    }
  ],
  "has_errors": true
}

Error Types

The dataset covers two main categories of Vietnamese spelling errors:

1. Mistyped Errors

Errors occurring during the typing process, typically caused by:

• Non-word errors: Resulting in syllables that don't exist in Vietnamese dictionaries
• Real-word errors: Valid Vietnamese syllables used in incorrect contexts

2. Misspelled Errors

Errors caused by:

• Regional pronunciation differences
• Difficulty with certain Vietnamese words
• Confusion between similar-sounding syllables

Dataset Creation

Source Data

The VSEC dataset was constructed from a large corpus of Vietnamese text sources, including news articles and educational materials. The original corpus contained approximately 14.8 million news articles and content from TaiLieu.VN.

Annotation Process

• Human annotators manually identified and corrected spelling errors
• Syllable-level annotations were created with error detection flags
• Correction suggestions were provided for each identified error
• Quality assurance processes ensured annotation consistency

Data Processing

The dataset underwent comprehensive preprocessing including:

• Noise character removal (emojis, line breaks)
• Case normalization (uppercase to lowercase conversion)
• Mark standardization using telex typing form mapping
• Merged syllable splitting using word segmentation algorithms

Uses

Intended Use

This dataset is primarily intended for:

• Research and development of Vietnamese spell correction systems
• Educational purposes in natural language processing courses
• Benchmarking spell correction algorithms for Vietnamese
• Training transformer-based models for Vietnamese error correction

Research Applications

• Developing sequence-to-sequence models for spell correction
• Studying Vietnamese linguistic error patterns
• Comparing different neural architectures for error correction
• Analyzing the effectiveness of subword tokenization for Vietnamese

Limitations

Technical Limitations

• Dataset focuses specifically on Vietnamese language
• Limited to syllable-level errors (not word-level semantic errors)
• May not cover all possible Vietnamese dialects and regional variations
• Annotation quality depends on human annotator expertise

Scope Limitations

• Primarily covers written text errors, not speech recognition errors
• May not represent all domains of Vietnamese text equally
• Error distribution may not match real-world typing error frequencies

Licensing and Legal Information

License

This dataset is released under a custom "other" license with specific terms for research and educational use.

Usage Terms

• Primary intended use: Research and educational purposes
• Commercial use: Users must ensure compliance with all applicable copyright laws and proceed at their own legal risk
• Attribution: Required when using this dataset (see citation below)

Important Legal Disclaimer

⚠️ COPYRIGHT NOTICE: This dataset is built from copyrighted source materials including news articles and educational content. While the dataset is made available for research purposes, users must be aware that:

1. Commercial Use Responsibility: For any commercial applications, users are solely responsible for ensuring compliance with copyright laws of all underlying source materials
2. Risk Assumption: Commercial users proceed entirely at their own legal risk
3. No Commercial License Granted: This release does not grant any commercial rights to the underlying copyrighted content
4. Research Focus: This dataset is optimized for academic research and educational applications

The dataset creators make no warranties regarding the copyright status of underlying materials and assume no liability for any copyright infringement claims arising from dataset usage.

Recommended Use Cases

✅ Recommended: Academic research, educational projects, non-commercial NLP development, algorithm benchmarking

⚠️ Use with Caution: Commercial applications (ensure proper legal review and copyright compliance)

Citation

If you use this dataset in your research, please cite the original paper:

@inproceedings{do2021vsec,
  title={Vsec: Transformer-based model for vietnamese spelling correction},
  author={Do, Dinh-Truong and Nguyen, Ha Thanh and Bui, Thang Ngoc and Vo, Hieu Dinh},
  booktitle={Pacific Rim International Conference on Artificial Intelligence},
  pages={259--272},
  year={2021},
  organization={Springer}
}

Dataset Card Authors

This dataset card was prepared to facilitate responsible use of the VSEC dataset while respecting the intellectual property rights of original content creators.

Additional Information

Additional Resources

For technical details and methodology, please refer to the original research paper.

Version History

• v1.0: Initial release with 9,341 sentences and 11,202 errors