lm-spell/sinhala-spell-correction-dataset

Sinhala Spell Correction Dataset

A Sinhala spell correction dataset used for training and evaluating neural spell correction models as part of the LMSpell project.

Dataset Description

This dataset combines data from previously published Sinhala spell correction resources and applies additional cleaning to improve its suitability for training neural spell correction models.

The dataset originates from the benchmark introduced by Sonnadara et al. (2021) and was subsequently extended with additional real-world error data by Sudesh et al. (2022).

As part of the LMSpell project, we cleaned the data contributed by the latter work by removing duplicate entries before incorporating it into the dataset used for our experiments.

Previous authors have replaced all numbers and special characters from this dataset. The test set was additionally annotated and reviewed by two Sinhala language professionals.

Additional Test Datasets

Two additional test datasets were created from the original clean test set using the error-injection methodology introduced by Sonnadara et al. (2021).

The datasets contain different proportions of artificially introduced spelling errors:

• test_set_41p.csv — approximately 41% of the words contain spelling errors.
• test_set_65p.csv — approximately 65% of the words contain spelling errors.

These datasets were created to evaluate model performance under controlled error rates and provide additional test settings beyond the original test set.

Dataset Lineage

Original Dataset

The original Sinhala spell correction benchmark was introduced by:

Sonnadara, C., Ranathunga, S., & Jayasena, S. (2021). Sinhala spell correction: A novel benchmark with neural spell correction.

The implementation and related resources from the original work are available at:

https://github.com/chason94/SinNeuSpellCorrector

This work established a benchmark and dataset for Sinhala spell correction and provided an important foundation for subsequent research in the area.

Additional Data

Additional Sinhala spelling-error data was introduced by:

Sudesh, P., Dashintha, D., Lakshan, R., & Dias, G. (2022, July). Erroff: A tool to identify and correct real-word errors in Sinhala documents. In 2022 Moratuwa Engineering Research Conference (MERCon) (pp. 1–6). IEEE.

This work contributed additional real-world Sinhala spelling errors.

LMSpell Contribution

As part of the LMSpell: Spell Correction with Pre-Trained Language Models project, we processed the data obtained from the Erroff work and removed duplicate entries before combining it with the existing spell correction data.

This cleaning step was performed to reduce duplicated training examples and provide a cleaner dataset for fine-tuning pretrained language models. We additionally created two test datasets from the original clean test set using the error-injection methodology introduced by Sonnadara et al. (2021).

Data Processing

The dataset preparation performed for LMSpell included duplicate removal from the additional data obtained from the Erroff work and preparation of the resulting data for model fine-tuning and evaluation.

For the complete dataset preparation methodology, experimental setup, and evaluation procedure, please refer to the LMSpell paper.

Research

This dataset was used in our research on pretrained language models for Sinhala spell correction:

LMSpell: Spell Correction with Pre-Trained Language Models

A. Gunathilake, N. Karunarathna, T. Bandaranayake, S. Ranathunga, N. de Silva and N. Jayatilleke, "LMSpell: Spell Correction with Pre-Trained Language Models," 2026 Moratuwa Engineering Research Conference (MERCon), Moratuwa, Sri Lanka, 2026, pp. 503-508.

• Paper: https://doi.org/10.1109/MERCon71835.2026.11691371
• Extended Preprint: https://arxiv.org/html/2512.05414v3
• Project: https://github.com/Nadil-K/lm-spell

The LMSpell paper provides further details on dataset preparation, model fine-tuning, evaluation, and experimental results.

Citation

Original Dataset

@misc{sonnadara2021sinhala,
  title={Sinhala spell correction: A novel benchmark with neural spell correction},
  author={Sonnadara, Charana and Ranathunga, Surangika and Jayasena, Sanath},
  year={2021},
  publisher={Research-Gate}
}

Additional Data

@inproceedings{sudesh2022erroff,
  title={Erroff: a tool to identify and correct real-word errors in Sinhala documents},
  author={Sudesh, Pasindu and Dashintha, Dilan and Lakshan, Rashmika and Dias, Gihan},
  booktitle={2022 Moratuwa Engineering Research Conference (MERCon)},
  pages={1--6},
  year={2022},
  organization={IEEE}
}

LMSpell

@INPROCEEDINGS{11691371,
  author={Gunathilake, Akesh and Karunarathna, Nadil and Bandaranayake, Tharusha and Ranathunga, Surangika and de Silva, Nisansa and Jayatilleke, Nevidu},
  booktitle={2026 Moratuwa Engineering Research Conference (MERCon)},
  title={LMSpell: Spell Correction with Pre-Trained Language Models},
  year={2026},
  pages={503-508},
  doi={10.1109/MERCon71835.2026.11691371}
}