File size: 8,472 Bytes
dd89885 | 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 | #!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
build_parallel_split.py -- assemble a leak-free seen/unseen Arabic->English split.
Why this is needed
------------------
The project's own bilingual material is not a benchmark:
* its biggest "parallel" sources are LEXICONS, not sentences
(lisan_farahidian 57,886 / semantic_anchor 7,583 / Basran_Heritage_Matrix 137,961);
* the same work appears in several files, so pairs must be de-duplicated ACROSS files;
* the Grand-100 evaluation set is 100% contained in training data.
What this does
--------------
1. Collects sentence-level pairs from every parallel file.
2. Drops lexicon-style sources by name AND by a form heuristic (no-space Arabic side, or an
English side that is a bare gloss).
3. De-duplicates on (arabic, english) so a pair cannot sit in two splits via two files.
4. Holds out WHOLE WORKS for test -- never sentences from a work that also contributes training
pairs -- so the split is leak-free at source level by construction.
5. VERIFIES the split: counts 13-gram overlap between test and train on both the Arabic and the
English side. Anything above ~0 means the split is not clean and must be reported.
Outputs: /workspace/mt_split/{train,dev,test}.jsonl + split_report.json
"""
import argparse
import collections
import json
import pathlib
import random
import re
import sys
D = pathlib.Path('/workspace/rootformer_v12/v18_next_root_morph/data')
FILES = [
'pure_gold_bilingual_train.jsonl', 'pure_gold_bilingual_val.jsonl',
'unified_basran_andalusian_train.jsonl', 'unified_basran_andalusian_val.jsonl',
'sovereign_classical_transmute_corpus.jsonl',
'grand_neural_transmute_corpus.jsonl', 'lisan_3pillar_master_pairs.jsonl',
]
# sources that are lexical resources, not sentence translation
# Only genuinely lexical resources: root->gloss tables. NOTE: 'Heritage_Matrix' is NOT here --
# it contains real sentence translations as well as lexeme rows, and excluding it by name threw
# away ~138k legitimate pairs. The form heuristic below is what separates a gloss from a sentence.
LEXICON_PAT = re.compile(r'lisan_farahidian|semantic_anchor|master_pairs|3pillar|'
r'root_invariant|wazn_derivation', re.I)
AR = re.compile(r'[\u0600-\u06FF]')
ARABIC_ONLY = re.compile(r'^[\u0600-\u06FF\s]+$')
def is_sentence_pair(a, e):
"""Reject lexicon entries masquerading as sentence pairs."""
if not a or not e:
return False
aw, ew = a.split(), e.split()
if not (4 <= len(aw) <= 60 and 3 <= len(ew) <= 70):
return False
# a lexical gloss is a single Arabic token (or bare root) with a short English gloss
if len(aw) == 1 and len(ew) <= 8:
return False
if len(aw) == 1 and len(a.strip()) <= 4:
return False
if ARABIC_ONLY.match(e.strip()): # untranslated Arabic on the English side
return False
return True
def ngrams(tokens, n):
return {tuple(tokens[i:i + n]) for i in range(max(0, len(tokens) - n + 1))}
def main():
ap = argparse.ArgumentParser()
ap.add_argument('--test-works', nargs='*', default=None,
help='work names to hold out entirely (default: auto-pick)')
ap.add_argument('--overlap-ngram', type=int, default=13)
ap.add_argument('--out-dir', default='/workspace/mt_split')
args = ap.parse_args()
pairs = {} # (ar,en) -> work
skipped_lex = 0
per_work = collections.Counter()
for fn in FILES:
p = D / fn
if not p.exists():
print(f' [skip missing] {fn}')
continue
n = 0
for line in open(p, encoding='utf-8', errors='ignore'):
line = line.strip()
if not line:
continue
try:
d = json.loads(line)
except Exception:
continue
a = (d.get('arabic') or '').strip()
e = (d.get('english') or '').strip()
src = str(d.get('source') or d.get('book') or 'unknown')
if LEXICON_PAT.search(src):
skipped_lex += 1
continue
if not is_sentence_pair(a, e):
continue
pairs.setdefault((a, e), src)
per_work[src] += 1
n += 1
print(f' {fn:<46} kept {n}')
print(f'\n[*] distinct sentence pairs: {len(pairs)} (lexicon rows skipped: {skipped_lex})')
print(f'[*] distinct works: {len(per_work)}')
# ---- pick held-out works -------------------------------------------------------------
if args.test_works:
test_works = set(args.test_works)
else:
# choose works with >=300 pairs so the test set is meaningful
# hold out whole works until ~15% of pairs are in test, preferring small/medium works
# so the split stays usable for training
total = sum(per_work.values())
target = 0.15 * total
cand = sorted([(w, c) for w, c in per_work.items() if 300 <= c <= 20000],
key=lambda kv: kv[1])
test_works, acc = set(), 0
for w, c in cand:
if acc + c > target * 1.25:
continue
test_works.add(w); acc += c
if acc >= target:
break
if not test_works:
test_works = {max(per_work.items(), key=lambda kv: kv[1])[0]}
print(f'[*] held-out works: {sorted(test_works)}')
train, test = [], []
for (a, e), w in pairs.items():
rec = {'arabic': a, 'english': e, 'work': w}
(test if w in test_works else train).append(rec)
random.Random(0).shuffle(train)
random.Random(1).shuffle(test)
dev = train[:len(train) // 20]
train = train[len(train) // 20:]
print(f'[*] train {len(train)} | dev {len(dev)} | test {len(test)}')
# ---- VERIFY: 13-gram overlap, both sides --------------------------------------------
N = args.overlap_ngram
report = {'n_train': len(train), 'n_dev': len(dev), 'n_test': len(test),
'test_works': sorted(test_works), 'overlap_ngram': N}
for side in ('arabic', 'english'):
tr = set()
for r in train:
tr |= ngrams(r[side].split(), N)
hit = 0
tot = 0
for r in test:
g = ngrams(r[side].split(), N)
if not g:
continue
tot += len(g)
hit += len(g & tr)
cov = sum(1 for r in test if ngrams(r[side].split(), N) & tr)
report[f'{side}_test_ngrams'] = tot
report[f'{side}_test_ngrams_seen_in_train'] = hit
report[f'{side}_test_rows_with_any_overlap'] = cov
report[f'{side}_overlap_pct'] = 100.0 * hit / max(tot, 1)
print(f' {side:<8} test {N}-grams={tot} seen-in-train={hit} '
f'({100.0*hit/max(tot,1):.3f}%) rows affected={cov}/{len(test)}')
# ---- CLEAN test: drop any test row that shares a 13-gram with train -------------------
tr_ar, tr_en = set(), set()
for r in train:
tr_ar |= ngrams(r['arabic'].split(), N)
tr_en |= ngrams(r['english'].split(), N)
clean = [r for r in test
if not (ngrams(r['arabic'].split(), N) & tr_ar)
and not (ngrams(r['english'].split(), N) & tr_en)]
report['n_test_clean'] = len(clean)
report['n_test_dropped_for_overlap'] = len(test) - len(clean)
print(f'[*] CLEAN test (zero {N}-gram overlap with train): {len(clean)} rows '
f'({len(test)-len(clean)} dropped)')
out = pathlib.Path(args.out_dir)
out.mkdir(parents=True, exist_ok=True)
for name, rows in (('train', train), ('dev', dev), ('test', test), ('test_clean', clean)):
with open(out / f'{name}.jsonl', 'w', encoding='utf-8') as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + '\n')
# also a plain text version for sacrebleu-style tooling
with open(out / 'test_clean.ar', 'w', encoding='utf-8') as fa, \
open(out / 'test_clean.en', 'w', encoding='utf-8') as fe:
for r in clean:
fa.write(r['arabic'].replace('\n', ' ') + '\n')
fe.write(r['english'].replace('\n', ' ') + '\n')
json.dump(report, open(out / 'split_report.json', 'w'), indent=2)
print(f'\nwrote {out}/ (train/dev/test.jsonl, test.ar, test.en, split_report.json)')
print('A test-ngram overlap of 0.000% means the held-out works genuinely never occur in train.')
if __name__ == '__main__':
main()
|