NoisyWikiTr

A Turkish spelling-correction benchmark built from Turkish Wikipedia sentences with synthetic spelling errors. It pairs noisy sentences with their clean versions, for training and evaluating Turkish spell checkers. To our knowledge it is the first large-scale spelling-correction benchmark for Turkish.

Released with the paper Leveraging Large Language Models for Spelling Correction in Turkish (PeerJ Computer Science, 2025).

• Code: github.com/cgturhan/trspell
• Paper: PeerJ Computer Science 11:e2889
• Zenodo (original release, v1, 25 April 2025): 10.5281/zenodo.15281473
• Models: cguzel on Hugging Face

Dataset structure

Files on Hugging Face (Parquet, one shared schema for all splits):

File	Split	Rows
data/train.parquet	train	687,008
data/validation.parquet	validation	85,871
data/test.parquet	test	85,824

The original CSV files (NoisyWikiTr_train.csv, NoisyWikiTr_val.csv, NoisyWikiTr_test_final.csv, 512.8 MB) are on Zenodo. They are identical except that the CSV train and validation files do not have the avg_edit_distance and noise_type columns.

Statistics computed from the released files:

	Train	Validation	Test
Sentences	687,008	85,871	85,824
Tokens (clean)	10,413,448	1,301,968	1,290,313
Noisy tokens	995,595	124,827	130,193
Mean noise ratio	10.9%	10.9%	12.0%

Noise counts by type:

Noise type	Train	Validation	Test
Common misspelling (replacement)	216,203	26,950	28,083
Substitution	189,746	23,762	25,569
Transposition	189,093	23,884	25,452
Vowel deletion	185,561	23,125	24,910
Insertion	119,879	15,038	15,963
Vowel insertion	69,837	8,859	9,613
Separation	20,765	2,626	0
Deletion	4,511	583	603

Columns

Column	Description
Unnamed: 0	index of the sentence in the source corpus
Sentences	original Wikipedia sentence, before preprocessing
cleaned	preprocessed, correct sentence (the target)
corrupted	the same sentence with synthetic spelling errors (the input)
noise_label	per-token list of the noise applied (None = unchanged)
n_noises	number of corrupted tokens in the sentence
n_deletion, n_insertion, n_transposition, n_substitution, n_vowel_insertion, n_vowel_deletion, n_seperation, n_common	number of tokens per noise type (n_common = replaced by a common misspelling)
n_asciification	count of Turkish-to-ASCII character errors (0 in this release)
noise_ratio	corrupted tokens / tokens in the sentence
edit_distances	edit distance of each corrupted token
avg_edit_distance	mean of edit_distances per sentence
noise_type	per-token noise type list (test split only; empty in train and validation)

For spelling correction, use corrupted as input and cleaned as target.

How it was built

Source. Raw Turkish Wikipedia sentences from erenfazlioglu/turkishwikipedia2023 on Hugging Face (about 1.6 million clean sentences after preprocessing).

Preprocessing.

• HTML tags, numbers and special characters removed
• Unicode NFKC normalization
• Lowercasing with Turkish-aware rules (standard case-conversion libraries mishandle Turkish characters such as I/ı and İ/i)

Synthetic noise. About 10% of the tokens in each sentence are corrupted, following the word-noising strategy of Sakaguchi et al. (2017) and Jayanthi et al. (2020, NeuSpell). Noise types:

Noise type	What it does
Deletion	removes one character at a random position
Vowel deletion	removes a vowel
Insertion	inserts a keyboard-adjacent letter (Turkish QWERTY layout)
Vowel insertion	inserts a vowel from {a, e, i, o, u, ü, ö}
Permutation	swaps two adjacent characters
Substitution	replaces a character with a keyboard-adjacent letter
Replacement	replaces a word with a common Turkish misspelling (from a Turkish Wikipedia list of frequent misspellings)
Separation	splits a word in two (training set only, not evaluated)

Insertion and substitution are limited to edit distance 1; replacement can exceed it.

Test settings. Models are evaluated on two test variants:

• Uniform noise: noise types applied uniformly
• Realistic noise: noise based on typical user typing behaviour (still synthetic, not human-written errors)

Limitations

• All errors are synthetic; there is no human-written error test set yet.
• Separation errors appear only in training data and are not evaluated.
• Wikipedia text is encyclopedic, so informal language (social media, chat) is under-represented.

How to use

from datasets import load_dataset

ds = load_dataset("cguzel/NoisyWikiTr")
example = ds["test"][0]
print(example["corrupted"])  # noisy input
print(example["cleaned"])    # correct target

Citation

If you use this dataset, please cite the paper and the dataset:

@dataset{guzelturhan2025noisywikitr,
  author    = {G{\"u}zel Turhan, Ceren},
  title     = {NoisyWikiTr},
  year      = {2025},
  publisher = {Zenodo},
  version   = {v1},
  doi       = {10.5281/zenodo.15281473},
  url       = {https://doi.org/10.5281/zenodo.15281473}
}

@article{guzelturhan2025trspell,
  title   = {Leveraging large language models for spelling correction in Turkish},
  author  = {G{\"u}zel Turhan, Ceren},
  journal = {PeerJ Computer Science},
  volume  = {11},
  pages   = {e2889},
  year    = {2025}
}