yagmurtuncer/turkish-text-normalization

🇹🇷 Turkish Text Normalization (TN / ITN)

A deterministic, rule-based dataset of Turkish written ↔ spoken pairs for Text Normalization (TN) and Inverse Text Normalization (ITN) — mapping digit/symbol forms (1.500 TL, %25, 15.07.2026) to their fully spoken Turkish words (bin beş yüz lira, yüzde yirmi beş, on beş temmuz iki bin yirmi altı) and back.

This is a common, high-value preprocessing step for Turkish ASR post-processing and TTS front-ends, where numbers, dates, currencies and percentages must be verbalized.

Provenance & honesty: every pair is generated programmatically with transparent linguistic rules (Turkish cardinal/ordinal number grammar, month names, decimal reading conventions). No text is scraped and no private data is used. The full generator is included (build_dataset.py) so the dataset is 100% reproducible from a fixed seed.

---

📦 Contents

Split	Rows
Train	15,398
Test	1,711
Total	17,109

Fields

Column	Description
id	Row index within the split
category	One of: cardinal, ordinal, decimal, percentage, currency, date, time
written	Digit / symbol form (the "written" surface form)
spoken	Fully verbalized Turkish words (lowercase)

Category distribution (unique pairs)

Category	Count	Example (written → spoken)
cardinal	4,259	2026 → iki bin yirmi altı
decimal	3,497	530,5 → beş yüz otuz virgül beş
date	3,082	15.07.2026 → on beş temmuz iki bin yirmi altı
currency	3,051	1500 TL → bin beş yüz lira
ordinal	1,683	523. → beş yüz yirmi üçüncü
time	1,101	14:30 → saat on dört buçuk
percentage	436	%25 → yüzde yirmi beş

---

🚀 Usage

from datasets import load_dataset

ds = load_dataset("yagmurtuncer/turkish-text-normalization")

# ITN (spoken -> written) or TN (written -> spoken)
ex = ds["train"][0]
print(ex["written"], "→", ex["spoken"])

Typical uses:

• Train/evaluate a seq2seq normalizer (written → spoken for TTS, or spoken → written for ASR)
• Rule-engine regression tests for Turkish verbalization
• Data augmentation for Turkish ASR/TTS pipelines

---

🏗️ How it was built (reproducible)

python build_dataset.py   # regenerates data/ deterministically (seed = 42)
python validate.py        # 6 data-quality checks, all must pass

The generator implements Turkish number grammar directly:

• Cardinals up to the billions (bin not bir bin, bir milyon, correct yüz/bin scaling)
• Ordinals via a vetted last-word suffix map (dört → dördüncü, on → onuncu)
• Dates with Turkish month names, times (incl. buçuk), currency (TL / $ / € / ₺), percentages (yüzde …), and decimals (virgül …)

Conventions

• spoken is lowercased Turkish (letters + spaces only — enforced by validate.py)
• Decimal fractions with a leading zero are read digit-by-digit (0,05 → sıfır virgül sıfır beş); otherwise the fraction is read as a whole number (3,14 → üç virgül on dört)
• Pairs are globally de-duplicated on (written, spoken) — counts reflect unique examples, not inflated repetitions

---

✅ Data Quality

validate.py enforces: non-empty fields · valid category set · spoken charset · global (written, spoken) uniqueness · all categories present · every written contains a digit. All checks pass on the released data.

---

⚠️ Limitations

• Rule-generated: covers standard verbalizations, not every colloquial reading (e.g. clock time also has çeyrek geçe styles not exhaustively included)
• Small percentage count is intentional — whole percentages only span 0–100, and duplicates are removed rather than padded
• Not a substitute for a full production normalizer; intended for training, testing and prototyping

---

📄 License & Citation

Released under CC-BY-4.0.

@misc{tuncer_turkish_text_normalization,
  title  = {Turkish Text Normalization (TN/ITN)},
  author = {Nur Yağmur Tuncer},
  year   = {2026},
  url    = {https://huggingface.co/datasets/yagmurtuncer/turkish-text-normalization}
}