File size: 10,476 Bytes
11632a3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 | #!/usr/bin/env python3
"""
Balance the augmented corpus to include more multi-word examples.
Current distribution: 98.5% single, 0.0% two-word, 1.5% multi
Target distribution: ~40% single, ~30% two-word, ~30% multi
Strategy:
1. Extract single-word vocabulary with transliterations
2. Generate two-word compound pairs using Syriac patterns:
- Construct state: noun + d- + noun (e.g., beyt d-ʾabrāhām)
- Proclitic combos: b-/w-/l- + word + word
3. Downsample single-word examples
4. Output balanced corpus
"""
import json
import random
from collections import defaultdict
from pathlib import Path
# Proclitics for combining
PROCLITICS_WEST = {
"ܒ": "b-", # in/with
"ܕ": "d-", # of/that
"ܘ": "w-", # and
"ܠ": "l-", # to/for
}
PROCLITICS_EAST = {
"ܒ": "b-",
"ܕ": "d-",
"ܘ": "w-",
"ܠ": "l-",
}
def load_corpus(path: Path) -> list[dict]:
"""Load JSONL corpus."""
entries = []
with open(path) as f:
for line in f:
entries.append(json.loads(line))
return entries
def extract_vocabulary(entries: list[dict]) -> dict[str, str]:
"""Extract single-word vocabulary: syriac -> transliteration."""
vocab = {}
for entry in entries:
t = entry["transliteration"]
src = t["src"]
tgt = t["tgt"]
# Only single words, skip proclitics
if " " not in src and not tgt.startswith(("b-", "d-", "w-", "l-")):
vocab[src] = tgt
return vocab
def generate_two_word_pairs(
vocab: dict[str, str],
dialect: str,
count: int,
) -> list[dict]:
"""Generate two-word compound pairs."""
pairs = []
words = list(vocab.items())
proclitic_map = PROCLITICS_WEST if dialect == "west" else PROCLITICS_EAST
# Sample word pairs
random.shuffle(words)
used = set()
for i in range(0, len(words) - 1, 2):
if len(pairs) >= count:
break
syr1, lat1 = words[i]
syr2, lat2 = words[i + 1]
# Skip if either has vowel marks that might cause issues
key = (syr1, syr2)
if key in used:
continue
used.add(key)
# Pattern 1: Simple juxtaposition (word1 word2)
pairs.append(
{
"transliteration": {
"src": f"{syr1} {syr2}",
"tgt": f"{lat1} {lat2}",
"title": "compound",
"dialect": dialect,
"source": "synthetic-2word",
}
}
)
# Pattern 2: Construct state with d- (word1 d-word2)
pairs.append(
{
"transliteration": {
"src": f"{syr1} ܕ{syr2}",
"tgt": f"{lat1} d-{lat2}",
"title": "construct",
"dialect": dialect,
"source": "synthetic-construct",
}
}
)
# Pattern 3: Proclitic + word1 + word2
for syr_pro, lat_pro in proclitic_map.items():
if syr_pro == "ܕ": # Skip d- since we have construct
continue
pairs.append(
{
"transliteration": {
"src": f"{syr_pro}{syr1} {syr2}",
"tgt": f"{lat_pro}{lat1} {lat2}",
"title": "proclitic-pair",
"dialect": dialect,
"source": "synthetic-proclitic",
}
}
)
return pairs[:count]
def generate_multi_word_phrases(
vocab: dict[str, str],
dialect: str,
count: int,
) -> list[dict]:
"""Generate 3, 4, and 5-word phrases."""
phrases = []
words = list(vocab.items())
random.shuffle(words)
i = 0
while len(phrases) < count and i + 4 < len(words):
syr1, lat1 = words[i]
syr2, lat2 = words[i + 1]
syr3, lat3 = words[i + 2]
syr4, lat4 = words[i + 3]
syr5, lat5 = words[i + 4]
i += 5
# 3-word patterns
phrases.append(
{
"transliteration": {
"src": f"{syr1} ܕ{syr2} ܘ{syr3}",
"tgt": f"{lat1} d-{lat2} w-{lat3}",
"title": "phrase",
"dialect": dialect,
"source": "synthetic-3word",
}
}
)
phrases.append(
{
"transliteration": {
"src": f"{syr1} {syr2} {syr3}",
"tgt": f"{lat1} {lat2} {lat3}",
"title": "phrase",
"dialect": dialect,
"source": "synthetic-3word",
}
}
)
# 4-word patterns
phrases.append(
{
"transliteration": {
"src": f"{syr1} {syr2} ܕ{syr3} {syr4}",
"tgt": f"{lat1} {lat2} d-{lat3} {lat4}",
"title": "phrase",
"dialect": dialect,
"source": "synthetic-4word",
}
}
)
phrases.append(
{
"transliteration": {
"src": f"ܒ{syr1} {syr2} ܘ{syr3} {syr4}",
"tgt": f"b-{lat1} {lat2} w-{lat3} {lat4}",
"title": "phrase",
"dialect": dialect,
"source": "synthetic-4word",
}
}
)
# 5-word patterns
phrases.append(
{
"transliteration": {
"src": f"{syr1} ܕ{syr2} {syr3} ܘ{syr4} {syr5}",
"tgt": f"{lat1} d-{lat2} {lat3} w-{lat4} {lat5}",
"title": "phrase",
"dialect": dialect,
"source": "synthetic-5word",
}
}
)
phrases.append(
{
"transliteration": {
"src": f"{syr1} {syr2} {syr3} {syr4} {syr5}",
"tgt": f"{lat1} {lat2} {lat3} {lat4} {lat5}",
"title": "phrase",
"dialect": dialect,
"source": "synthetic-5word",
}
}
)
return phrases[:count]
def balance_corpus(
entries: list[dict],
dialect: str,
target_single_ratio: float = 0.40,
target_two_ratio: float = 0.30,
target_multi_ratio: float = 0.30,
) -> list[dict]:
"""Balance corpus with target distribution."""
# Categorize existing entries
single = []
two_word = []
multi = []
for entry in entries:
src = entry["transliteration"]["src"]
words = src.split()
if len(words) == 1:
single.append(entry)
elif len(words) == 2:
two_word.append(entry)
else:
multi.append(entry)
print(f"Original distribution:")
print(f" Single: {len(single):>8}")
print(f" Two: {len(two_word):>8}")
print(f" Multi: {len(multi):>8}")
# Extract vocabulary for synthetic generation
vocab = extract_vocabulary(entries)
print(f" Vocabulary size: {len(vocab)}")
# Calculate target counts
# Use multi-word count as anchor (keep all existing multi-word)
existing_multi = len(multi)
# Target: enough examples that each category is well-represented
# Use the multi count scaled up as reference
target_multi = max(existing_multi, 100_000)
target_two = int(target_multi * target_two_ratio / target_multi_ratio)
target_single = int(target_multi * target_single_ratio / target_multi_ratio)
print(f"\nTarget counts:")
print(f" Single: {target_single:>8}")
print(f" Two: {target_two:>8}")
print(f" Multi: {target_multi:>8}")
# Generate synthetic two-word pairs
needed_two = max(0, target_two - len(two_word))
if needed_two > 0:
print(f"\nGenerating {needed_two} synthetic two-word pairs...")
synthetic_two = generate_two_word_pairs(vocab, dialect, needed_two)
two_word.extend(synthetic_two)
print(f" Generated: {len(synthetic_two)}")
# Generate synthetic multi-word phrases (3, 4, 5 words)
needed_multi = max(0, target_multi - len(multi))
if needed_multi > 0:
print(
f"\nGenerating {needed_multi} synthetic multi-word phrases (3-5 words)..."
)
synthetic_multi = generate_multi_word_phrases(vocab, dialect, needed_multi)
multi.extend(synthetic_multi)
print(f" Generated: {len(synthetic_multi)}")
# Downsample single-word examples
if len(single) > target_single:
print(f"\nDownsampling single-word from {len(single)} to {target_single}...")
random.shuffle(single)
single = single[:target_single]
# Combine
balanced = single + two_word + multi
random.shuffle(balanced)
print(f"\nFinal distribution:")
final_single = sum(
1 for e in balanced if len(e["transliteration"]["src"].split()) == 1
)
final_two = sum(
1 for e in balanced if len(e["transliteration"]["src"].split()) == 2
)
final_multi = sum(
1 for e in balanced if len(e["transliteration"]["src"].split()) >= 3
)
total = len(balanced)
print(f" Single: {final_single:>8} ({100*final_single/total:.1f}%)")
print(f" Two: {final_two:>8} ({100*final_two/total:.1f}%)")
print(f" Multi: {final_multi:>8} ({100*final_multi/total:.1f}%)")
print(f" Total: {total:>8}")
return balanced
def main():
data_dir = Path(__file__).parent
for dialect in ["west", "east"]:
print(f"\n{'='*60}")
print(f"Processing {dialect.capitalize()} dialect")
print("=" * 60)
input_path = data_dir / f"syriac_{dialect}_augmented_corpus.jsonl"
output_path = data_dir / f"syriac_{dialect}_balanced_corpus.jsonl"
if not input_path.exists():
print(f" Skipping - {input_path} not found")
continue
entries = load_corpus(input_path)
balanced = balance_corpus(entries, dialect)
# Write output
with open(output_path, "w") as f:
for entry in balanced:
f.write(json.dumps(entry, ensure_ascii=False) + "\n")
print(f"\nWritten to: {output_path}")
if __name__ == "__main__":
random.seed(42)
main()
|