build_dataset.py

# -*- coding: utf-8 -*-
"""
Turkish Text Normalization (TN/ITN) dataset builder.
Deterministic, rule-based generation of written<->spoken Turkish pairs.
Reproducible: fixed seed. No external data sources.
"""
import csv, json, os, random

SEED = 42
random.seed(SEED)

ONES  = ["", "bir","iki","üç","dört","beş","altı","yedi","sekiz","dokuz"]
TENS  = ["", "on","yirmi","otuz","kırk","elli","altmış","yetmiş","seksen","doksan"]
SCALES = ["", "bin", "milyon", "milyar"]

def three_digit(n):           # 0..999 -> words
    out = []
    h = n // 100
    if h == 1: out.append("yüz")
    elif h > 1: out.append(ONES[h] + " yüz")
    t = (n % 100) // 10
    if t: out.append(TENS[t])
    o = n % 10
    if o: out.append(ONES[o])
    return " ".join(out)

def num_to_tr(n):             # 0..10^12-1 -> Turkish words (cardinal)
    if n == 0: return "sıfır"
    chunks = []
    x = n
    while x > 0:
        chunks.append(x % 1000); x //= 1000
    words = []
    for i in range(len(chunks)-1, -1, -1):
        c = chunks[i]
        if c == 0: continue
        if i == 1 and c == 1:
            words.append("bin")           # "bin", not "bir bin"
        else:
            words.append(three_digit(c))
            if i > 0: words.append(SCALES[i])
    return " ".join(words)

ORDINAL_LAST = {
 "bir":"birinci","iki":"ikinci","üç":"üçüncü","dört":"dördüncü","beş":"beşinci",
 "altı":"altıncı","yedi":"yedinci","sekiz":"sekizinci","dokuz":"dokuzuncu",
 "on":"onuncu","yirmi":"yirminci","otuz":"otuzuncu","kırk":"kırkıncı","elli":"ellinci",
 "altmış":"altmışıncı","yetmiş":"yetmişinci","seksen":"sekseninci","doksan":"doksanıncı",
 "yüz":"yüzüncü","bin":"bininci","milyon":"milyonuncu","milyar":"milyarıncı",
}
def ordinal_tr(n):
    w = num_to_tr(n).split()
    w[-1] = ORDINAL_LAST[w[-1]]
    return " ".join(w)

MONTHS = ["","ocak","şubat","mart","nisan","mayıs","haziran","temmuz","ağustos","eylül","ekim","kasım","aralık"]

def dec_to_tr(int_part, frac_str):
    return num_to_tr(int_part) + " virgül " + " ".join(num_to_tr(int(d)) if d!="0" else "sıfır" for d in frac_str) \
        if False else num_to_tr(int_part) + " virgül " + num_to_tr(int(frac_str)) if not frac_str.startswith("0") else \
        num_to_tr(int_part) + " virgül " + " ".join("sıfır" if d=="0" else num_to_tr(int(d)) for d in frac_str)

rows = []
def add(cat, written, spoken):
    spoken = " ".join(spoken.split())
    rows.append((cat, written, spoken))

# 1) Cardinals
seen=set()
for _ in range(9000):
    r = random.random()
    if r < 0.5: n = random.randint(0, 100)
    elif r < 0.8: n = random.randint(101, 9999)
    elif r < 0.95: n = random.randint(10000, 999999)
    else: n = random.randint(10**6, 10**9)
    if n in seen: continue
    seen.add(n)
    add("cardinal", str(n), num_to_tr(n))

# 2) Ordinals
seen=set()
for _ in range(3500):
    n = random.choice([random.randint(1,100), random.randint(1,9999)])
    if n in seen: continue
    seen.add(n)
    add("ordinal", f"{n}.", ordinal_tr(n))

# 3) Decimals
seen=set()
for _ in range(3500):
    ip = random.randint(0, 9999)
    fl = random.choice([1,1,2,2,2,3])
    fr = "".join(random.choice("0123456789") for _ in range(fl))
    key=(ip,fr)
    if key in seen: continue
    seen.add(key)
    frac_words = " ".join("sıfır" if d=="0" else num_to_tr(int(d)) for d in fr) if fr[0]=="0" else num_to_tr(int(fr))
    add("decimal", f"{ip},{fr}", f"{num_to_tr(ip)} virgül {frac_words}")

# 4) Percentages
seen=set()
for _ in range(2500):
    whole = random.random() < 0.7
    if whole:
        v = random.randint(0,100); w=f"%{v}"; sp=f"yüzde {num_to_tr(v)}"
    else:
        ip=random.randint(0,100); fr=random.choice(["5","25","5","75","50"]); w=f"%{ip},{fr}"
        frac_words = num_to_tr(int(fr))
        sp=f"yüzde {num_to_tr(ip)} virgül {frac_words}"
    if w in seen: continue
    seen.add(w); add("percentage", w, sp)

# 5) Currency
CUR = [("TL","lira"),("TL","türk lirası"),("$","dolar"),("€","avro"),("₺","lira")]
seen=set()
for _ in range(3200):
    n = random.choice([random.randint(1,999), random.randint(1,99999), random.randint(1,9999)])
    sym,word = random.choice(CUR)
    if sym in ("$","€","₺"): w=f"{sym}{n}"
    else: w=f"{n} {sym}"
    key=(n,sym,word)
    if key in seen: continue
    seen.add(key); add("currency", w, f"{num_to_tr(n)} {word}")

# 6) Dates  DD.MM.YYYY
seen=set()
for _ in range(3200):
    d=random.randint(1,28); m=random.randint(1,12); y=random.randint(1900,2035)
    w=f"{d:02d}.{m:02d}.{y}"
    if w in seen: continue
    seen.add(w)
    add("date", w, f"{num_to_tr(d)} {MONTHS[m]} {num_to_tr(y)}")

# 7) Times HH:MM
seen=set()
for _ in range(2200):
    h=random.randint(0,23); mn=random.randint(0,59)
    w=f"{h:02d}:{mn:02d}"
    if w in seen: continue
    seen.add(w)
    if mn==0: sp=f"saat {num_to_tr(h)}"
    elif mn==30 and random.random()<0.5: sp=f"saat {num_to_tr(h)} buçuk"
    else: sp=f"saat {num_to_tr(h)} {num_to_tr(mn)}"
    add("time", w, sp)

# dedup globally on (written, spoken)
uniq = {}
for cat,w,s in rows:
    uniq[(w,s)] = (cat,w,s)
data = list(uniq.values())
random.shuffle(data)

# split 90/10
cut = int(len(data)*0.9)
train, test = data[:cut], data[cut:]

os.makedirs("data", exist_ok=True)
def dump(path, part):
    with open(path,"w",encoding="utf-8",newline="") as f:
        wr=csv.writer(f); wr.writerow(["id","category","written","spoken"])
        for i,(cat,w,s) in enumerate(part): wr.writerow([i,cat,w,s])
dump("data/train.csv", train)
dump("data/test.csv", test)

from collections import Counter
cats = Counter(c for c,_,_ in data)
stats = {"total": len(data), "train": len(train), "test": len(test),
         "by_category": dict(sorted(cats.items())), "seed": SEED}
json.dump(stats, open("stats.json","w",encoding="utf-8"), ensure_ascii=False, indent=2)
print(json.dumps(stats, ensure_ascii=False, indent=2))
print("\n--- ornekler ---")
for cat,w,s in random.sample(data, 14):
    print(f"[{cat:11}] {w:16} -> {s}")