uyghur-normalizer

Release 2.1.0 (2026-10-18).

Mirror. The canonical repository is github.com/UyghurReferenceTexts/uyghur-normalizer. Please open issues and pull requests there, and cite the Zenodo DOI below. The dataset viewer shows normalizer/provenance.tsv: every rule of the four maps with the dictionary row and scan image that attest it.

Normalizes Uyghur (ئۇيغۇر) text to the 2009 orthographic standard, and replaces terminology calqued from Chinese with Uyghur equivalents.

Two files, no dependencies, no network. Python 3.9+.

python uyghur_normalizer.py --in text.txt            # scan: report only
python uyghur_normalizer.py --in text.txt --apply    # write text_fixed.txt

It never edits in place, and by default it writes nothing at all.

Companion dataset: uyghur-headword-lists (Hugging Face), the audited headword lists of the dictionaries below. Every map in this repository is checked against them.

---

What it does

Three passes, in a fixed order:

pass	map	fixes
1	calque_map.tsv	terminology calqued from Chinese
2	era_map.tsv	pre-2009 spellings
3	imla_loan_map.tsv	loanword shape under the 2009 reform

The order is load-bearing. Pass 1 runs first so its replacements are themselves normalized by passes 2–3; pass 3 consumes pass 2's output.

Each pass can be switched off (--no-calque, --no-era, --no-loan). That matters more than it sounds: pass 1 makes editorial changes, and someone correcting OCR on a 1980s manuscript may want the orthography fixed and the vocabulary of the period left exactly as printed.

Two more passes are opt-in. Pass 0 corrects common misuses of words, from the 6-volume dictionary [2]; see Pass 0. Pass 4 handles forms produced by a specific machine-translation model; see Pass 4.

Orthographic authority

The standard followed throughout is the 2009 orthography dictionary, Hazirqi zaman Uyghur edebiy tilining imla lughiti (imla we teleppuz qa'idisi) — reference [4] below. Its rule prose and its word list are two parts of one book, and where they disagree the word list wins: the example lists inside the rule prose carry transcription noise and were not used as evidence.

The explanatory dictionaries [1] and [2] are used to arbitrate and to confirm that a form is actually attested. The 2006 Uyghur–Chinese dictionary [3] is the opposite kind of source: its table of countries, capitals and currencies records the older spellings of those names, so it supplies the left-hand side of a rule, never the right. Its headwords have not been used.

The target is the codified 2009 standard, not contemporary web or diaspora usage. The clearest case is the ـىيە class: this tool produces تېررىتورىيەسى, not the widely-seen تېررىتورىيىسى. If you want to match how Uyghur is written online today, this is the wrong tool.

---

Suffixes

A rule matches at a word boundary and may be followed by suffixes. Concatenating the replacement and the suffix is wrong whenever the two sides differ in vowel harmony, final-consonant voicing, or vowel-versus-consonant ending:

زۇڭتۇڭلار  →  پىرېزىدېنتلار    wrong, needs -لەر
گۇڭشېدە    →  كوممۇنادە        wrong, needs -دا
داشۆگە     →  ئالىي مەكتەپگە   wrong, needs -كە

So the suffix is re-derived morpheme by morpheme against the replacement. There are three outcomes:

• safe — both sides agree, so concatenation cannot change any allomorph.
• rewrite — they disagree and every morpheme is recognized; correct allomorphs are emitted.
• decline — a morpheme is unrecognized, or the replacement carries no harmony-bearing vowel. The text is left untouched and the case is reported.

Declining is deliberate. A normalizer that guesses is worse than one that says it doesn't know, because a plausible wrong form is harder to find later than an unchanged one. Everything declined is printed, and --review writes it to CSV so the cases can be added to a map explicitly.

Rule flags

An optional third column carries per-rule flags:

variant<TAB>standard<TAB>flags

nosfx — fire only on the bare form. Use when the replacement is already inflected:

سوغۇق ئۇرۇش	سوغۇقچىلىق ئۇرۇشى	nosfx

Without it, سوغۇق ئۇرۇشى matches this rule with ى captured as a suffix, and the possessive is applied a second time.

raise — the replacement's final ا/ە raises to ى before a suffix (ئامېرىكا + دا → ئامېرىكىدا).

Raising is off by default, which is a deliberate and slightly surprising choice. The rule is real and productive, but the exceptions are not marginal: the ـىيە class does not raise, and in a 2009-imla map it dominates. In the era map shipped here, 362 of the 364 ە-final replacements are in classes that must not raise. Turning raising on globally would mean:

پارتىيى + لەر  →  پارتىيىلەر

— the rule returning its own input, silently undoing the reform it exists to enforce while the report still claims a substitution. So it is opt-in, per rule.

Raising between suffixes (تىبەت + تا + مۇ → تىبەتتىمۇ) is always on and is not affected by the flag: only the locative and dative end in ا/ە, and only مۇ can follow them, so that case is narrow and well attested.

---

Pass 0: common misuses

The 6-volume izahliq lughet [2] lists common misuses as their own headwords and marks the correct form with توغرىسى, e.g. ئاپارات‎ → ئاپپارات‎. normalizer/usage_map.tsv turns those entries into rules:

python uyghur_normalizer.py --in text.txt --usage --apply

It is off by default because it corrects word choice rather than spelling, and would also "correct" a regional or literary form used on purpose.

Pass 0 runs before pass 1, so a pre-2009 spelling in [2]'s correct form is normalized by passes 2–3: گىراپا‎ → گرافىك‎ → گىرافىك‎.

The map is generated by tools/build_usage_map.py, not edited by hand. A pair is left out when its wrong form is a headword of [4] or [1] or a word inside one (ئال‎ "take", گۈندە‎ in ئارا گۈندە‎), when it has three letters or fewer, or when its correct form is not attested in [4] or [1]. Every rule's correct form is attested in a row of [4] or [1] checked against the scan; normalizer/provenance.tsv gives the row and the [2] entry it came from. Where [2]'s form is correct but older than [4] and [1], the newer form is set in tools/usage_target_overrides.tsv.

---

Pass 4: model-specific fixes

Machine translation models produce forms that are not in their fine-tuning data and cannot be fixed by correcting that data. The finding this pass exists for: a rule was added to calque_map.tsv, the training corpus was verified at 100% compliance — zero occurrences of the wrong form — and every checkpoint still emitted it. Fine-tuning did not overwrite what the base model had already learned.

Correcting the corpus cannot fix a form that was never in the corpus. That requires an output-side pass:

python uyghur_normalizer.py --in output.txt --fix-map my_fixes.tsv --apply

There is no default path, and no map of this kind ships enabled. These rules describe one model, not the language. examples/output_fix_map.examples.tsv shows the shape and the level of evidence expected — every rule cites the checkpoint and the measured cost. Yours will be different, because your base model is different.

Do not add these rules to calque_map.tsv. A corpus-compliance check counts occurrences in the corpus; adding forms that were never there makes the check report violations that do not exist.

---

Vetting new rules

check_new_rules.py audits a batch of proposed rules before any of them reach a map. It writes nothing to the maps: a clean TSV of what is safe to append, and a CSV of everything a human has to decide.

python check_new_rules.py --xlsx new_terms.xlsx \
    --maps normalizer/*_map.tsv \
    --target-map normalizer/imla_loan_map.tsv \
    --out-dir review

Codes: REVERSE (an existing map already rewrites the other way — one of the two is wrong), CONFLICT, CHAIN, DUPLICATE, NAMING (word counts differ, so it is a terminology change rather than an orthographic one), SUBSTRING, SHORT, HYGIENE.

CHAIN is scoped to a single map on purpose. Passes run calque → era → imla_loan, so era producing a form that imla_loan then rewrites is the design. A chain within one map is a real defect, and it is often subtler than it looks. A worked example from this repo's own history: a dictionary batch proposed both

كىنگستون → كىڭستون      (Kingston, Jamaica)
كىڭستون  → كىڭستوۋن     (Kingstown, St Vincent)

The 2011 spelling of Kingston is identical to the 2006 spelling of Kingstown. No rule ordering can separate them, and the second rule would corrupt every correctly-spelled Kingston. It was dropped. String rewriting has limits, and the useful thing a tool can do is find them before a human does.

---

Input formats

Plain text, .jsonl (rewrites translation.ug), and EN: / UG: benchmark blocks (only UG: lines are touched) are detected automatically.

Arabic presentation forms are folded and text is NFC-normalized on load, so badly-encoded OCR output still matches.

---

Tests

python -m unittest discover -s tests -v

tests/test_suffix.py covers the morphology. tests/test_maps.py enforces the invariants on the map data itself: no duplicate left-hand sides, no rule mapping a form to itself, no map rewriting into its own left-hand side, NFC throughout, no padded fields, recognized flags only, every rule fires on its own input, and the pipeline is idempotent — normalizing twice equals normalizing once.

The idempotence test is the one that catches real damage. It is how the Kingston collision above was found.

tests/test_lexicon.py runs the pipeline over every headword and sub-entry of [4] and [1] (tests/data/lexicon_snapshot.tsv.gz, built from uyghur-headword-lists). Those are already in the target orthography, so nothing may change. Only rows checked by hand against the scan can fail the test: the remaining rows were extracted by a model trained on web Uyghur, which may have written a web form the book does not print. Hits on those are reported instead, and tools/lexicon_review.py writes them out with their row ids and scan images. Genuine exceptions are listed, with their scans, in tests/data/lexicon_exceptions.tsv.

After the headword lists change, rebuild the derived files:

python tools/build_lexicon_snapshot.py --imla <2009 xlsx> --izahliq <2011 vol1> <2011 vol2>
python tools/build_usage_map.py --vol1990 <vol1> ... <vol6>
python tools/build_provenance.py

---

Limitations

• Rules are literal strings, so a form the map has never seen is not corrected.
• Vowel raising alters the stem, so ئامېرىكىدا does not match a rule whose left-hand side is ئامېرىكا. Inflected forms may need their own entries.
• Raised suffix allomorphs are not recognized on the input side: شىزاڭدىمۇ is declined rather than derived, and wants an explicit entry.
• The tool has no idea what a word means. ئۈزۈك سەپ and مىللىي دۆلەت are terminological preferences, not corrections, and reasonable translators disagree. Read calque_map.tsv before trusting it on your text.

---

Maps as data

The maps are the substance here; the script is a few hundred lines of regex around them. They were built word by word against the sources below, and the 2009 word list was treated as final throughout. Frequency-zero rules are kept: a rule that has never fired is dormant, not wrong.

---

Sources

The spelling rules (passes 0, 2 and 3) are traceable to these four books. The terminology rules of pass 1 come from the corpora listed after them.

[1] ھازىرقى زامان ئۇيغۇر تىلىنىڭ ئىزاھلىق لۇغىتى (قىسقارتىلمىسى) Hazirqi zaman Uyghur tilining izahliq lughiti (qisqartilmisi). Sh.U.A.R. Milletler Til-Yéziq Xizmiti Komitéti (comp.); chief compilers Hemdulla Abdurehman Imam, Perhat Nur, Esqer Abduqadir. Shinjang Xelq Neshriyati, Ürümchi, 2011. ISBN 978-7-228-13933-0. (1st ed. 1999, chief editors Abliz Yaqup and Ghenizat Gheyurani.) → Attestation and arbitration.

[2] ئۇيغۇر ئەدەبىي تىلىنىڭ ئىزاھلىق لۇغىتى، 1–6 توم Uyghur edebiy tilining izahliq lughiti, vols. 1–6. Compilers Abliz Yaqup, Ghenizat Gheyurani, Ismail Qadir, Hemdulla Abduraxman, Perhat Nur, Abliz Emet, Esqer Abduqadir, Abduzahir Tahir, Ablikim Réhimjan, Zayit Héwil. Milletler Neshriyati, Beijing, 1990–1999. ISBN 7-105-00928-4 (vol. 1, 1990); 7-105-01431-8 (vol. 2, 1991); 7-105-01692-2 (vol. 3, 1992); 7-105-02248-5 (vol. 4, 1994); 7-105-02578-6 (vol. 5, 1996); 7-105-03227-8 (vol. 6, 1999). → Attestation for words absent from [1]. Source of the usage map (pass 0).

[3] ئۇيغۇرچە - خەنزۇچە لۇغەت Uyghurche–Xenzuche lughet. Sh.U.A.R. Milletler Til-Yéziq Xizmiti Komitéti (comp.); responsible editor Yaqup Muhemmetrozi. Milletler Neshriyati, Beijing, 2006. ISBN 7-105-07803-0. → Its table of countries, capitals and currencies only: the pre-reform spellings on the left-hand side of the country and capital rules in imla_loan_map.tsv, aligned against [4]. The dictionary's headwords have not been extracted or used.

[4] ھازىرقى زامان ئۇيغۇر ئەدەبىي تىلىنىڭ ئىملا لۇغىتى (ئىملا ۋە تەلەپپۇز قائىدىسى) Hazirqi zaman Uyghur edebiy tilining imla lughiti (imla we teleppuz qa'idisi). Mirsultan Osmanof, Tahir Abduweli, Abdughappar Abduraxman, Memet'éli Abdurehim, Anargül Abdurehim, Enwer Exmet. Shinjang Xelq Neshriyati, Ürümchi, 2009. ISBN 978-7-228-12817-4. The orthographic standard followed throughout this repository. The rule prose cited in the map headers as the imla qa'idisi is the introductory section of this book, not a separate work.

None of these are distributed here. They are in print, and a correction is far easier to act on if it cites one of them by page.

Corpora. The terminology of calque_map.tsv (pass 1) does not come from [3]. Its first rules were mined from the parallel Chinese–Uyghur law texts of prc-laws-bilingual. A few were added during the audit of [1], [2] and [4], from prc-primary-documents and from the web. The rules added on 2026-10-18 rest on cross-references and [خەن] entries in [1], [2] and [4]; the evidence row of each is given in a comment above it.

---

Contributing and corrections

A correction is more valuable than a new rule, and both are welcome.

If you think a rule is wrong, open an issue. Include the form, the rule as it stands, what you think it should be, and a citation — book and page from the list above. "This looks wrong to me" is a fine issue too, but it will sit until someone can check it against a dictionary.

If you want to add rules, run them through check_new_rules.py first and say in the pull request what it reported. Rules that arrive already vetted get merged; rules that arrive as a raw list get vetted by whoever reviews them, which is slower.

Whether a form is correct is a discussion, not a patch. Open the issue before the pull request. The tests will tell you if a rule breaks the map; they cannot tell you if a rule is bad Uyghur.

---

License

Code (uyghur_normalizer.py, check_new_rules.py, tests/*.py, tools/*.py): MIT.

Data (normalizer/*.tsv, examples/*.tsv, tests/data/, tools/*.tsv): CC BY 4.0.