TiSpell: A Semi-Masked Methodology for Tibetan Spelling Correction

📄 Paper: TiSpell: A Semi-Masked Methodology for Tibetan Spelling Correction covering Multi-Level Error with Data Augmentation

🧑‍💻 Author: Yutong Liu, Xiao Feng, Ziyue Zhang, Yongbin Yu*, Cheng Huang, Fan Gao, Xiangxiang Wang*, Ban Ma-bao, Manping Fan, Thupten Tsering, Gadeng Luosang, Renzeng Duojie, Nyima Tashi

📦 Repository: Official implementation of TiSpell.

---

🧠 Overview

TiSpell is a Tibetan spelling correction algorithm specifically designed for multi-level orthographic errors. It proposes a semi-masked methodology that jointly models character-level, syllable-level, and word-level errors. With integrated data augmentation strategies, TiSpell improves robustness and accuracy in real-world spelling correction tasks. It leverages pre-trained language models and introduces an end-to-end correction architecture.

---

✨ Features

• ✅ Handles character-, syllable-, and word-level spelling errors
• ✅ Combines semi-masked modeling with structural reconstruction
• ✅ Includes multiple data augmentation techniques (perturbation, phonetic substitution, etc.)
• ✅ Fully compatible with Huggingface Transformers for easy integration and customization

---

🗂️ Project Structure

TiSpell/
├── dataloader/ # Data loading utilities
├── dataset/ # Preprocessed and raw datasets
├── images/ # Visualizations
├── model/ # Model architecture
├── pretrained_models/ # Checkpoints and pre-trained weights
├── scripts/ # Training and evaluation scripts
├── LICENSE
├── README.md
├── compute_parameter.py # Parameter counting utility
├── data_analysis.py # Exploratory data analysis
├── infer.py # Inference script
├── metrics.py # Evaluation metrics
├── option.py # Argument parsing
├── plot.py # Visualization utilities
├── train.py # Training script
└── requirements.txt # Python dependencies

---

🚀 Quick Start

🔧 1. Install Dependencies

pip install -r requirements.txt

📁 2. Prepare Dataset

Download the Tibetan News Classification dataset from Huggingface and place it under the dataset/ directory. Ensure that the dataset is formatted in the following structure:

TiSpell/
└── dataset/
    └── tibetan_news_classification/
        ├── 政务类
        ├── 教育类
        ├── 文化类
        ├── 旅游类
        ├── 时政类
        ├── 民生类
        ├── 法律类
        ├── 科技类
        ├── 经济类
        └── 艺术类
            ├── 0.txt
            ├── 1.txt
            ├── 2.txt
            └── ...

🏋️‍♂️ 3. Train the Model

python main.py

⚙️ Configuration

You can customize training and evaluation parameters in option.py, including:

• Learning rate / Batch size
• Training epochs
• Weight decay

📌 Citation

If you find TiSpell helpful in your research, please cite our work:

@misc{liu2025tispellsemimaskedmethodologytibetan,
      title={TiSpell: A Semi-Masked Methodology for Tibetan Spelling Correction covering Multi-Level Error with Data Augmentation}, 
      author={Yutong Liu and Feng Xiao and Ziyue Zhang and Yongbin Yu and Cheng Huang and Fan Gao and Xiangxiang Wang and Ma-bao Ban and Manping Fan and Thupten Tsering and Cheng Huang and Gadeng Luosang and Renzeng Duojie and Nyima Tashi},
      year={2025},
      eprint={2505.08037},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.08037}, 
}

📝 License

This project is licensed under the MIT License. See the LICENSE file for more details.