polish-dynaword / src /fetch_njkp.py
1am's picture
Add source: nkjp1m
3358140 verified
Raw
History Blame
13.4 kB
#!/usr/bin/env python3
"""Fetch the NKJP 1-million-word subcorpus (Podkorpus Milionowy NKJP 1.2) and
build a DynaWord-style parquet shard from the TEI source.
Downloads + extracts the IPI PAN tarball under /tmp, then walks every sample
folder and emits one parquet row per <div> in text.xml. Each <div> is a single
contiguous excerpt from one source document (its <ab> paragraphs joined by
newlines, ellipses stripped); across <div>s the excerpts are unrelated, so they
are kept as separate rows rather than merged.
The same minimal gates as src/build_dynaword.py are applied per passage (drop
< 200 chars, drop non-Polish by diacritic ratio, exact sha1 dedup; the OCR gate
is inapplicable - NKJP1M is not OCR), so the shard matches what the DynaWord
build would keep. Per-row `created` is the sample's TEI publication date
(<date type="published">) when the header records one. Token counts use
tiktoken cl100k (encode_ordinary), same as build_dynaword.py, and the emitted
stats are recomputed from the written parquet.
Text extraction + the stats report are adapted from tmp/count_tokens.py.
Usage:
python3 src/fetch_njkp.py --out . --workers 8
"""
from __future__ import annotations
import argparse
import hashlib
import json
import re
import ssl
import sys
import tarfile
import time
from concurrent.futures import ProcessPoolExecutor
from datetime import date
from pathlib import Path
from urllib.error import URLError
from urllib.request import urlopen, Request
from xml.etree import ElementTree as ET
import pyarrow as pa
import pyarrow.parquet as pq
import tiktoken
URL = ("https://clip.ipipan.waw.pl/NationalCorpusOfPolish"
"?action=AttachFile&do=get&target=NKJP-PodkorpusMilionowy-1.2.tar.gz")
ROOT_NAME = "NKJP-PodkorpusMilionowy-1.2"
UA = {"User-Agent": "polish-dynaword/0.1 (+research; openly-licensed corpus)"}
KEY = "nkjp1m" # source name / parquet stem, consumed by build_dynaword
LICENSE = "CC-BY" # stated on the NKJP download page (clip.ipipan.waw.pl)
AUTHOR = "NKJP" # compiled/distributed by the NKJP Consortium (IPI PAN)
# Gates - identical thresholds to src/build_dynaword.py so the shard matches
# what the DynaWord build itself would keep.
MIN_CHARS = 200
MIN_POLISH_RATIO = 0.005
POLISH_RE = re.compile(r"[ąćęłńóśźżĄĆĘŁŃÓŚŹŻ]")
ALPHA_RE = re.compile(r"[^\W\d_]", re.UNICODE)
NS = "{http://www.tei-c.org/ns/1.0}"
DIV_TAG, AB_TAG, DATE_TAG = f"{NS}div", f"{NS}ab", f"{NS}date"
_ELLIPSIS = re.compile(r"…|\.{3,}") # ellipsis markers stripped from text
# Same schema as src/build_dynaword.py so the shard drops straight into DynaWord.
SCHEMA = pa.schema([
("id", pa.string()), ("text", pa.string()), ("source", pa.string()),
("added", pa.string()), ("created", pa.string()), ("token_count", pa.int64()),
("license", pa.string()), ("author", pa.string()),
])
_ENC = None # per-worker tiktoken encoder
def download(url: str, dst: Path) -> Path:
if dst.exists() and dst.stat().st_size:
print(f" cached {dst}", flush=True)
return dst
print(f" downloading {url}", flush=True)
dst.parent.mkdir(parents=True, exist_ok=True)
def stream(ctx):
with urlopen(Request(url, headers=UA), timeout=120, context=ctx) as r, \
dst.open("wb") as f:
while chunk := r.read(1 << 20):
f.write(chunk)
try:
stream(None)
except URLError as e:
# clip.ipipan.waw.pl ships an incomplete cert chain Python rejects (curl
# accepts it); retry unverified for this known source only.
if not isinstance(e.reason, ssl.SSLError):
raise
print(" ! TLS verify failed; retrying unverified", file=sys.stderr, flush=True)
stream(ssl._create_unverified_context())
return dst
def extract_archive(archive: Path, dest: Path) -> Path:
# The tarball has no top folder, so extract into a dedicated dir to scope
# the text.xml scan.
root = dest / ROOT_NAME
if root.is_dir():
print(f" already extracted {root}", flush=True)
return root
print(f" extracting {archive} -> {root}", flush=True)
root.mkdir(parents=True)
with tarfile.open(archive, "r:gz") as tar:
tar.extractall(root, filter="data") # py3.12+ safe extraction
return root
def _clean(text: str) -> str:
"""Strip ellipsis markers (… and "...") and tidy the whitespace they leave."""
return re.sub(r"[ \t]{2,}", " ", _ELLIPSIS.sub(" ", text)).strip()
def _polish_ratio(text: str) -> float:
"""Fraction of letters that are Polish-specific diacritics (build_dynaword)."""
letters = ALPHA_RE.findall(text)
return len(POLISH_RE.findall(text)) / len(letters) if letters else 0.0
def _norm_date(raw: str) -> str:
"""Normalize a TEI @when value to an ISO date or bare year, else "".
Keeps full YYYY-MM-DD, keeps bare YYYY, and salvages a leading 4-digit year
from anything else (e.g. "YYYY-MM", a stray trailing space). Values whose
year is implausible as a publication year (NKJP1M has a few malformed ones,
such as "200") are treated as missing.
"""
raw = (raw or "").strip()
if re.fullmatch(r"\d{4}-\d{2}-\d{2}", raw):
return raw if 1500 <= int(raw[:4]) <= 2014 else ""
m = re.match(r"(\d{4})", raw)
return m.group(1) if m and 1500 <= int(m.group(1)) <= 2014 else ""
def _published_date(header_path: Path) -> str:
"""Publication date from a sample's TEI header (<date type='published'>)."""
try:
tree = ET.parse(header_path)
except Exception: # noqa: BLE001 - missing/unreadable header -> no date
return ""
for el in tree.iter(DATE_TAG):
if el.get("type") == "published":
d = _norm_date(el.get("when") or (el.text or ""))
if d:
return d
return ""
def extract_passages(xml_path: str) -> list[str]:
"""One passage per TEI <div> (its <ab> paragraphs joined by newlines).
<ab> blocks inside a <div> are adjacent source paragraphs (real coherence);
different <div>s are unrelated sampled excerpts, so each becomes its own row.
Adapted from tmp/count_tokens.py, which instead merged every <ab> per file.
"""
passages, current = [], []
for _, elem in ET.iterparse(xml_path, events=("end",)):
if elem.tag == AB_TAG:
# itertext() also captures any nested inline markup text.
text = _clean("".join(elem.itertext()))
if text:
current.append(text)
elem.clear()
elif elem.tag == DIV_TAG:
if current:
passages.append("\n".join(current))
current = []
elem.clear()
if current: # <ab> outside any <div> (not expected) - keep as one passage
passages.append("\n".join(current))
return passages
def process_file(xml_path: str) -> tuple[list[dict], list[int]]:
"""Sample file -> (kept row dicts per surviving <div>, [read, short, lang]).
Gates run per passage; dedup is deferred to main() so it can span files.
"""
global _ENC
if _ENC is None:
_ENC = tiktoken.get_encoding("cl100k_base")
folder_dir = Path(xml_path).parent
folder = folder_dir.name
created = _published_date(folder_dir / "header.xml")
try:
passages = extract_passages(xml_path)
except Exception: # noqa: BLE001 - skip unreadable samples, don't crash
return [], [0, 0, 0]
read = short = lang = 0
kept = []
for i, text in enumerate(passages):
read += 1
text = text.strip()
if len(text) < MIN_CHARS:
short += 1
continue
if _polish_ratio(text) < MIN_POLISH_RATIO:
lang += 1
continue
kept.append({
"id": f"{KEY}_{folder}_{i}", # passage index kept stable across gating
"text": text,
"created": created,
"tokens": len(_ENC.encode_ordinary(text)),
"chars": len(text),
"sha1": hashlib.sha1(text.encode("utf-8")).digest(),
})
return kept, [read, short, lang]
def build_stats(parquet_path: Path, gate: dict) -> dict:
"""Recompute the sidecar stats directly from the written parquet.
Counts (kept/chars/tokens/licenses/authors/dates) come from the bytes on
disk; the drop_* tallies are build-time artifacts carried over from `gate`.
A cross-check asserts read == kept + drops so the two can't silently drift.
"""
t = pq.read_table(parquet_path)
d = t.to_pydict()
n = t.num_rows
dated = [c for c in d["created"] if c]
years = sorted({int(c[:4]) for c in dated})
read = gate["read"]
drops = gate["drop_short"] + gate["drop_lang"] + gate["drop_dup"] + gate["drop_ocr"]
assert n == read - drops, f"gate arithmetic: kept {n} != read {read} - drops {drops}"
return {
"read": read,
"kept": n,
"drop_short": gate["drop_short"],
"drop_lang": gate["drop_lang"],
"drop_dup": gate["drop_dup"],
"drop_ocr": gate["drop_ocr"],
"chars": sum(len(x) for x in d["text"]),
"tokens": sum(d["token_count"]),
"licenses": {LICENSE: n},
"authors_with_value": sum(1 for a in d["author"] if a),
"documents_with_created": len(dated),
"created_range": f"{years[0]}-{years[-1]}" if years else "",
"license": LICENSE,
"stats_recomputed_from_parquet": True,
}
def report(rows: list[dict], stats: dict) -> None:
"""Aggregate cl100k token stats. (adapted from count_tokens.py)"""
toks = sorted(r["tokens"] for r in rows)
total_tok, total_chars = sum(toks), sum(r["chars"] for r in rows)
def pct(p: float) -> int:
return toks[max(0, min(len(toks) - 1, round(p / 100 * (len(toks) - 1))))]
print("\n" + "=" * 60)
print("NKJP cl100k token statistics")
print("=" * 60)
print(f"Passages kept: {len(rows):,}")
print(f"Total tokens: {total_tok:,}")
print(f"Total characters: {total_chars:,}")
if toks:
print(f"Mean tokens/pass.: {total_tok / len(toks):,.1f}")
print(f"Median / p90 / max: {pct(50):,} / {pct(90):,} / {max(toks):,}")
if total_tok:
print(f"Chars per token: {total_chars / total_tok:.2f}")
if stats.get("created_range"):
print(f"Created range: {stats['created_range']} "
f"({stats['documents_with_created']:,}/{len(rows):,} dated)")
def main() -> int:
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("--out", default=".",
help="DynaWord root; shard -> <out>/data/nkjp1m/nkjp1m.parquet")
ap.add_argument("--tmp", default="/tmp", help="Download + extraction dir")
ap.add_argument("--workers", type=int, default=None, help="Process pool size")
ap.add_argument("--added", default=date.today().isoformat(),
help="Value for the 'added' column (default: today)")
args = ap.parse_args()
t0 = time.time()
tmp = Path(args.tmp).expanduser()
root = extract_archive(download(URL, tmp / f"{ROOT_NAME}.tar.gz"), tmp)
files = sorted(str(p) for p in root.rglob("text.xml"))
if not files:
print(f"error: no text.xml under {root}", file=sys.stderr)
return 1
print(f"Found {len(files)} text.xml files under {root}", flush=True)
read = short = lang = 0
collected = [] # kept row dicts, in sorted file + passage order (deterministic)
with ProcessPoolExecutor(max_workers=args.workers) as pool:
for i, (kept, st3) in enumerate(pool.map(process_file, files, chunksize=16), 1):
read += st3[0]; short += st3[1]; lang += st3[2]
collected.extend(kept)
if i % 1000 == 0 or i == len(files):
print(f" processed {i}/{len(files)} files, {len(collected):,} kept passages",
file=sys.stderr)
# Exact dedup (sha1), first-wins over the deterministic order above.
seen, rows, drop_dup = set(), [], 0
for r in collected:
if r["sha1"] in seen:
drop_dup += 1
continue
seen.add(r["sha1"])
rows.append(r)
n = len(rows)
out = Path(args.out).expanduser().resolve() / "data" / KEY / f"{KEY}.parquet"
out.parent.mkdir(parents=True, exist_ok=True)
pq.write_table(pa.table({
"id": [r["id"] for r in rows],
"text": [r["text"] for r in rows],
"source": [KEY] * n,
"added": [args.added] * n,
"created": [r["created"] for r in rows],
"token_count": [r["tokens"] for r in rows],
"license": [LICENSE] * n,
"author": [AUTHOR] * n,
}, schema=SCHEMA), out, compression="zstd")
# DynaWord-style sidecar stats, recomputed from the parquet just written.
gate = {"read": read, "drop_short": short, "drop_lang": lang,
"drop_dup": drop_dup, "drop_ocr": 0}
stats = build_stats(out, gate)
out.with_name(f"{KEY}.stats.json").write_text(
json.dumps(stats, ensure_ascii=False, indent=2) + "\n")
report(rows, stats)
print(f"\nWrote {n:,} passages from {len(files):,} files "
f"(read {read:,}, -short {short:,} -lang {lang:,} -dup {drop_dup:,}) "
f"-> {out} in {round(time.time() - t0)}s", flush=True)
return 0
if __name__ == "__main__":
raise SystemExit(main())