File size: 8,378 Bytes
5952424 | 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 | #!/usr/bin/env python3
"""Stage 5b (documentary-clean variant): per-sentence documentary-hit masks.
Matching layers against the stage-4b documentary index, for the three trainable
tiers (pristine, repaired, bronze):
a) exact skeleton hash
b) bag (sorted-words) hash
c) word-8-gram windows, marking every sentence they overlap
d) document-level MinHash-LSH match (est J >= DOC_J) against the stage-4c
documentary signature index -- catches whole-RECORD near-duplicates that
(a)-(c) can miss: scanned SOURCEBOOK volumes (Dittenberger's Sylloge,
Schwyzer's Dialectorum Graecarum Exempla Epigraphica Potiora, Cagnat's
Inscriptiones Graecae ad Res Romanas Pertinentes, and similar epigraphic/
papyrological corpora catalogued as ordinary "literary" books) reproduce
documentary text with enough OCR noise / editorial apparatus / formatting
drift that individual 8-grams can slip through while the document is still
substantially the same content. A doc-level match flags EVERY sentence of
the record, so 06b_doc_clean.py drops the whole record (not just the
overlapping span) -- the same "wholesale over partial" policy already used
for Greek-origin bronze.
Output: work/doc_clean/masks/<tier>/shard_NNNNN.parquet, row-aligned with
work/sentences/<tier>/shard_NNNNN.parquet: (rid, masks list<uint16>).
A nonzero mask entry means "this sentence textually collides with some
documentary text (any PHI/TM digit)".
"""
import glob
import json
import os
import sys
from concurrent.futures import ProcessPoolExecutor
import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from common import h64, NGRAM
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
OUT = os.path.join(ROOT, "work", "doc_clean")
G = {}
MASK_DOC = np.uint16(1 << 13) # same bit stage 4b/4c index entries carry
N_PERM = 128
DOC_BANDS, DOC_ROWS = 64, 2
DOC_J = 0.50 # doc-level match threshold (same as stage 5's edition-level net)
DOC_MIN_WORDS = 35 # both sides must be substantial; short records are
DOC_RATIO = 3.0 # protected by exact/bag sentence matching instead
MERSENNE61 = np.uint64((1 << 61) - 1)
rng = np.random.RandomState(20260709) # same perms as stage 2 / 4c / 5
PERM_A = rng.randint(1, 1 << 28, size=N_PERM).astype(np.uint64)
PERM_B = rng.randint(0, 1 << 32, size=N_PERM).astype(np.uint64)
BAND_MIX = np.uint64(0x9E3779B97F4A7C15)
def setup_globals():
G["keys"] = np.load(os.path.join(OUT, "index_keys.npy"), mmap_mode="r")
G["masks"] = np.load(os.path.join(OUT, "index_masks.npy"), mmap_mode="r")
print("doc index: %d keys" % len(G["keys"]), flush=True)
z = np.load(os.path.join(OUT, "minhash_documentary.npz"))
G["d_nwords"] = z["nwords"]
sigs = z["sigs"]
bk = band_keys(sigs) # (m, 64)
order = np.argsort(bk, axis=0, kind="stable")
G["lsh_sorted"] = np.take_along_axis(bk, order, axis=0)
G["lsh_idx"] = order # original documentary row per sorted slot
G["lsh_sigs"] = sigs
print("doc minhash: %d documentary records indexed for LSH" % len(sigs), flush=True)
def lookup(q):
keys, masks = G["keys"], G["masks"]
pos = np.searchsorted(keys, q)
pos[pos >= len(keys)] = len(keys) - 1
hit = keys[pos] == q
out = np.zeros(len(q), dtype=np.uint16)
out[hit] = masks[pos[hit]]
return out
def minhash_from_ngram_hashes(hs):
if len(hs) == 0:
return np.zeros(N_PERM, dtype=np.uint64)
h = np.asarray(hs, dtype=np.uint64) & np.uint64(0xFFFFFFFF)
v = (PERM_A[:, None] * h[None, :] + PERM_B[:, None]) % MERSENNE61
return v.min(axis=1)
def band_keys(sig):
"""sig (..,128) -> (..,64) uint64 band keys."""
s = sig.reshape(sig.shape[:-1] + (DOC_BANDS, DOC_ROWS))
return (s[..., 0] * BAND_MIX + s[..., 1])
def doc_lsh_hit(sig, qwords):
"""True if some documentary record has est Jaccard >= DOC_J to sig.
Guarded against degenerate matches: both sides must have >= DOC_MIN_WORDS
and sizes within DOC_RATIO of each other (short/boilerplate records are
protected by the exact/bag/8-gram sentence layers instead).
"""
if qwords < DOC_MIN_WORDS:
return False
bk = band_keys(sig[None, :])[0] # (64,)
srt, sidx = G["lsh_sorted"], G["lsh_idx"]
cands = []
for b in range(DOC_BANDS):
col = srt[:, b]
lo = np.searchsorted(col, bk[b], side="left")
hi = np.searchsorted(col, bk[b], side="right")
if hi > lo:
cands.append(sidx[lo:hi, b])
if not cands:
return False
cand = np.unique(np.concatenate(cands))
cw = G["d_nwords"][cand]
ok = (cw >= DOC_MIN_WORDS) & (cw <= qwords * DOC_RATIO) & \
(cw * DOC_RATIO >= qwords)
cand = cand[ok]
if len(cand) == 0:
return False
est = (G["lsh_sigs"][cand] == sig[None, :]).mean(axis=1)
return bool((est >= DOC_J).any())
def process_shard(args):
tier, path, out_path = args
t = pq.read_table(path, columns=["rid", "skels"])
rids = t["rid"].to_pylist()
skels_col = t["skels"].to_pylist()
n = len(rids)
exact_q, bag_q, gram_q = [], [], []
exact_loc, gram_loc = [], []
gram_range = [] # per record: (start, end) into gram_q
for i in range(n):
stream, wsent = [], []
for si, sk in enumerate(skels_col[i]):
w = sk.split()
exact_q.append(h64(sk))
bag_q.append(h64(" ".join(sorted(w))))
exact_loc.append((i, si))
stream.extend(w)
wsent.extend([si] * len(w))
g0 = len(gram_q)
if len(stream) >= NGRAM:
for j in range(len(stream) - NGRAM + 1):
gram_q.append(h64(" ".join(stream[j:j + NGRAM])))
gram_loc.append((i, wsent[j], wsent[j + NGRAM - 1]))
gram_range.append((g0, len(gram_q)))
gram_arr = np.array(gram_q, dtype=np.uint64)
em = lookup(np.array(exact_q, dtype=np.uint64)) if exact_q else np.zeros(0, np.uint16)
bm = lookup(np.array(bag_q, dtype=np.uint64)) if bag_q else np.zeros(0, np.uint16)
gm = lookup(gram_arr) if gram_q else np.zeros(0, np.uint16)
sent_masks = [np.zeros(len(skels_col[i]), dtype=np.uint16) for i in range(n)]
for (i, si), m1, m2 in zip(exact_loc, em, bm):
if m1 or m2:
sent_masks[i][si] |= m1 | m2
for (i, s0, s1), m1 in zip(gram_loc, gm):
if m1:
sent_masks[i][s0:s1 + 1] |= m1
n_doc_hits = 0
for i in range(n):
g0, g1 = gram_range[i]
if g1 <= g0:
continue
qwords = (g1 - g0) + NGRAM - 1
if doc_lsh_hit(minhash_from_ngram_hashes(gram_arr[g0:g1]), qwords):
sent_masks[i][:] |= MASK_DOC # whole-record echo: flag every sentence
n_doc_hits += 1
out = pa.table({
"rid": rids,
"masks": pa.array([m.tolist() for m in sent_masks],
type=pa.list_(pa.uint16())),
})
pq.write_table(out, out_path, compression="zstd")
n_contaminated = sum(1 for m in sent_masks if m.any())
return n, n_contaminated, n_doc_hits
def main():
setup_globals()
tasks = []
for tier in ("pristine", "repaired", "bronze"):
os.makedirs(os.path.join(OUT, "masks", tier), exist_ok=True)
for p in sorted(glob.glob(os.path.join(ROOT, "work", "sentences",
tier, "shard_*.parquet"))):
out = os.path.join(OUT, "masks", tier, os.path.basename(p))
tasks.append((tier, p, out))
print("%d shards" % len(tasks), flush=True)
workers = max(4, min(16, (os.cpu_count() or 12) - 8))
stats = {}
with ProcessPoolExecutor(max_workers=workers) as ex:
for (tier, p, _), (nr, nc, nd) in zip(
tasks, ex.map(process_shard, tasks, chunksize=1)):
s = stats.setdefault(tier, {"records": 0, "contaminated": 0,
"doc_lsh_hits": 0})
s["records"] += nr
s["contaminated"] += nc
s["doc_lsh_hits"] += nd
with open(os.path.join(OUT, "stage5b_stats.json"), "w") as f:
json.dump(stats, f, indent=2)
print(json.dumps(stats, indent=2), flush=True)
if __name__ == "__main__":
main()
|