Turkish Text Normalization Resources

Part of a Turkish text preprocessing toolkit built for CMPE 561 (Natural Language Processing), Application Project 1, Boğaziçi University, 2026. All code is our own; no external NLP library is used in the systems. Status: progress report, 20 October 2026.

These are the resources used by our rule- and corpus-based normalizer, which converts noisy or non-standard Turkish (missing Turkish characters, chat abbreviations, repeated letters, colloquial suffixes, attached clitics, wrong casing) into standard Turkish orthography. Companion repositories: stopwords, sentence_splitter.

How the normalizer uses the data

1. Text-level cleanup (Unicode NFC, invisible characters, dotted i, quote/apostrophe variants, whitespace, repeated punctuation): fixed rules, no data.
2. Sentence state: is the sentence ASCII-typed, noisy (at least 10% of lowercase words unknown to wiki_vocab.tsv) or all caps?
3. Word cascade, first stage that applies wins; every proposal must be a word attested in Wikipedia: (1) normalization.tsv lexicon → (2) repeated letters → (3) diacritic restoration by Wikipedia frequency → (4) colloquial suffix rules → (5) proper noun + case suffix → (6) clitic splitting (mi/de/ki) → (7) character 6-gram language model for unknown ASCII words. Stages 4–7 run only on noisy text.
4. Casing from the most frequent Wikipedia spelling; context step for ambiguous pairs (su/şu) with wiki_context.tsv; question-particle vowel harmony.

See figures/flow_overview.png and figures/flow_cascade.png (labels in Turkish).

Files

File	Format	Contents	Used in
lexicon/normalization.tsv	surface ⇥ normalized ⇥ source	550 entries: 492 word pairs derived from MultiLexNorm Turkish train + 58 hand-written	cascade stage 1 (slm → selam)
lexicon/normalization_manual.tsv	same	58 hand-written entries (chat abbreviations, spoken forms, spelling variants)	merged into the lexicon; wins on conflict
wikipedia/wiki_vocab.tsv	word ⇥ freq	654,804 case-sensitive word forms seen ≥ 3 times in 267,494 Turkish Wikipedia articles (59.6M tokens)	validity check, diacritic restoration, casing, rule validation, char-LM training
wikipedia/wiki_context.tsv	word ⇥ L/R ⇥ neighbour ⇥ count	left/right neighbour counts for 5,240 diacritic-ambiguous words (neighbours ASCII-folded)	context step (su an → şu an)
results/	JSON, Markdown	all dev/test scores, ablation, error categories
figures/	PNG	test accuracy vs. leave-as-is baseline, ablation, social-media error categories, flow diagrams

Lines starting with # in the TSV files are provenance comments.

Examples from wiki_vocab.tsv: çok 120822, cok 113, Ankara 28265, ANKARA 81, ankara 16, Ankara'da 3705. Example from wiki_context.tsv: şu R an 2338 vs. su R an 7.

Results (test, scored once)

Test case	Accuracy	Leave-as-is baseline	ERR
BOUN test, clean (does it damage good text?)	99.56%	100%	–
BOUN test, diacritics removed (synthetic)	98.27%	62.0%	0.954
BOUN test, all caps (synthetic)	95.65%	18.6%	0.947
MultiLexNorm Turkish test (real tweets)	80.48%	64.0%	0.458

Thresholds were tuned on BOUN dev and on MultiLexNorm train (5-fold cross-validation, lexicon rebuilt from the training folds in every fold). No test sentence was used to build any resource.

Sources

Source	Licence	Use
Turkish Wikipedia, dump 20231101 (wikimedia/wikipedia, 20231101.tr, shard 1 of 2)	CC BY-SA 3.0 / GFDL	wiki_vocab.tsv, wiki_context.tsv, char LM
MultiLexNorm Turkish (bitbucket.org/robvanderg/multilexnorm, data/tur, commit 31695c6); Çolakoğlu, Sulubacak & Tantuğ 2019, ACL SRW, doi:10.18653/v1/P19-2037; van der Goot et al. 2021, W-NUT	not stated; raw files not redistributed here	train: lexicon pairs, tuning; test: evaluation only
UD Turkish-BOUN (UniversalDependencies/UD_Turkish-BOUN); Türk et al. 2022, LRE	CC BY-SA 4.0	evaluation text only (dev: tuning, test: scored once)

If you use the lexicon, please also cite the MultiLexNorm and Çolakoğlu et al. papers.

Main literature

Torunoğlu & Eryiğit 2014 (cascaded normalization of Turkish social media, LASM); Eryiğit & Torunoğlu-Selamet 2017 (NLE); Adalı & Eryiğit 2014 (vowel and diacritic restoration, LASM); Arslan 2016 (deASCIIfication for IR, IP&M); Çolakoğlu et al. 2019; van der Goot et al. 2021.