File size: 10,619 Bytes
6342c6d | 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 | """Extract speaker-segmented Chinese debate transcripts from .doc / .docx files
into a JSONL dataset.
Source layout in this repo:
docs/ — 107 .doc files (legacy OLE Word97-2003)
机器-订单转写结果-*/ — 43 folders × ~45 .docx files (modern OOXML)
Each file is an ASR transcription of a single debate session. Word's text-export
delimits the speaker-tagged segments with `\r` between segments and `\v` (VT,
0x0B) between the header (`说话人N HH:MM`) and the body text.
We extract via Word COM (works for both .doc and .docx) because the alternative
Python-only paths (python-docx, mammoth) only handle .docx and pip cannot reach
PyPI from this network.
Output: data/_train_raw.jsonl with one record per file.
"""
from __future__ import annotations
import json
import re
import sys
import time
import traceback
from pathlib import Path
import pythoncom
import win32com.client
REPO = Path(__file__).resolve().parent.parent
OUT_DIR = REPO / "data"
OUT_JSONL = OUT_DIR / "_train_raw.jsonl"
ERR_LOG = OUT_DIR / "_extract_errors.log"
# ---------- filename metadata parsers ----------
# Three filename patterns observed across the two source generations.
#
# Format A (docs/ named files): "<topic>丨<round> <team_a>vs<team_b>_音频.<ext>..."
# 丨 is U+FF5C — the full-width vertical bar separator used in the original titles.
FORMAT_A_RE = re.compile(
r"^(?P<topic>.+?)丨(?P<round>.+?) (?P<team_a>.+?)vs(?P<team_b>.+?)_音频",
re.IGNORECASE,
)
# Format B (early order folders): "<id>-<round> <team_a> VS <team_b> <topic>-<NNN>P ..."
# Uppercase `VS` surrounded by whitespace; topic follows the teams.
FORMAT_B_RE = re.compile(
r"^(?P<id>\d+)-(?P<round>\S+?)\s+(?P<team_a>\S+?)\s+VS\s+(?P<team_b>\S+?)\s+(?P<topic>.+?)-\d{3,4}P"
)
# Format C (later order folders): "<id>-<date>日 <round-tag> <topic> <team_a>vs<team_b>-<NNN>P"
# Lowercase `vs` glued to team names; an optional `<N月N日>` date prefix on
# the round-tag (e.g. `7月27日 中学团队赛`); topic precedes the teams.
FORMAT_C_RE = re.compile(
r"^(?P<id>\d+)-(?P<head>.+?)\s+(?P<team_a>\S+?)vs(?P<team_b>\S+?)-\d{3,4}P"
)
FORMAT_C_DATE_RE = re.compile(r"^(?P<date>\d+月\d+\s*日)\s+(?P<rest>.+)$")
def parse_filename(stem: str) -> dict:
"""Return metadata dict from a filename stem. Always has the four keys;
values are None when no pattern matches."""
out = {"topic": None, "round": None, "team_a": None, "team_b": None}
m = FORMAT_A_RE.search(stem)
if m:
out.update({
"topic": m["topic"].strip(),
"round": m["round"].strip(),
"team_a": m["team_a"].strip(),
"team_b": m["team_b"].strip(),
})
return out
m = FORMAT_B_RE.search(stem)
if m:
out.update({
"topic": m["topic"].strip(),
"round": m["round"].strip(),
"team_a": m["team_a"].strip(),
"team_b": m["team_b"].strip(),
})
return out
m = FORMAT_C_RE.search(stem)
if m:
head = m["head"].strip()
out["team_a"] = m["team_a"].strip()
out["team_b"] = m["team_b"].strip()
date_m = FORMAT_C_DATE_RE.match(head)
if date_m:
rest = date_m["rest"].strip()
parts = rest.split(maxsplit=1)
if len(parts) == 2:
out["round"] = f"{date_m['date'].strip()} {parts[0].strip()}"
out["topic"] = parts[1].strip()
else:
out["round"] = f"{date_m['date'].strip()} {rest}"
else:
# No date prefix — take the first whitespace-delimited token as
# round, remainder as topic.
parts = head.split(maxsplit=1)
if len(parts) == 2:
out["round"] = parts[0].strip()
out["topic"] = parts[1].strip()
else:
out["round"] = head
return out
return out
# ---------- transcript segmentation ----------
# Header inside one segment: "说话人N HH:MM" then VT then body
SEGMENT_HEADER_RE = re.compile(r"^\s*说话人(\d+)\s+(\d{1,2}:\d{2})\v(.*)$", re.DOTALL)
def parse_segments(text: str) -> list[dict]:
"""Split Word-exported text into a list of {speaker_id, timestamp, text}."""
segs: list[dict] = []
for chunk in text.split("\r"):
m = SEGMENT_HEADER_RE.match(chunk)
if not m:
continue
body = m.group(3).strip()
body = body.replace("\v", " ").replace("\x07", "") # strip stray VT/BEL
segs.append({
"speaker_id": int(m.group(1)),
"timestamp": m.group(2),
"text": body,
})
return segs
def normalize_full_text(text: str) -> str:
"""Flatten Word's `\r` / `\v` line separators to regular newlines so the raw
transcript column is human-readable."""
return text.replace("\v", "\n").replace("\r", "\n").strip()
# ---------- iteration ----------
def iter_source_files() -> list[tuple[Path, str]]:
"""Return (path, source_collection) tuples for every transcript file."""
items: list[tuple[Path, str]] = []
docs = REPO / "docs"
if docs.is_dir():
for f in sorted(docs.iterdir()):
if f.name.startswith("~$") or not f.is_file():
continue
if f.suffix.lower() in {".doc", ".docx"}:
items.append((f, "docs"))
for d in sorted(REPO.iterdir()):
if not (d.is_dir() and d.name.startswith("机器-订单")):
continue
for f in sorted(d.iterdir()):
if f.name.startswith("~$") or not f.is_file():
continue
if f.suffix.lower() in {".doc", ".docx"}:
items.append((f, d.name))
return items
# ---------- main extract loop ----------
def _record_id(path: Path, source: str) -> str:
return f"{source}/{path.stem}" if source != "docs" else path.stem
def _load_done_ids() -> set[str]:
"""Resume support: any id already in OUT_JSONL is considered done."""
if not OUT_JSONL.exists():
return set()
ids: set[str] = set()
with open(OUT_JSONL, "r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
try:
ids.add(json.loads(line)["id"])
except Exception:
pass
return ids
def _spawn_word():
word = win32com.client.gencache.EnsureDispatch("Word.Application")
word.Visible = False
word.DisplayAlerts = False
return word
def _quit_word(word) -> None:
try:
word.Quit(SaveChanges=False)
except Exception:
pass
def main() -> int:
OUT_DIR.mkdir(exist_ok=True)
sources = iter_source_files()
done = _load_done_ids()
todo = [(p, s) for (p, s) in sources if _record_id(p, s) not in done]
print(
f"Found {len(sources)} source files; {len(done)} already extracted; "
f"{len(todo)} to do.",
flush=True,
)
pythoncom.CoInitialize()
word = _spawn_word()
written = len(done)
failed: list[tuple[str, str]] = []
consecutive_word_restarts = 0
t0 = time.time()
# Append mode so resumes don't truncate prior progress.
with open(OUT_JSONL, "a", encoding="utf-8") as out:
for i, (path, source) in enumerate(todo, 1):
attempt = 0
text: str | None = None
while attempt < 3:
attempt += 1
try:
doc = word.Documents.Open(
str(path), ReadOnly=True, AddToRecentFiles=False
)
try:
text = doc.Range().Text
finally:
doc.Close(SaveChanges=False)
break
except Exception as e:
msg = repr(e)
print(f" retry {attempt} on {path.name}: {msg}", flush=True)
# Likely Word died — kill and respawn.
_quit_word(word)
try:
word = _spawn_word()
consecutive_word_restarts += 1
except Exception as e2:
failed.append((str(path), f"respawn failed: {e2!r}"))
text = None
break
# If Word has died too many times in a row, bail.
if consecutive_word_restarts > 5:
failed.append((str(path), "too many consecutive Word restarts"))
text = None
break
if text is None:
failed.append((str(path), "exhausted retries"))
continue
else:
consecutive_word_restarts = 0
meta = parse_filename(path.stem)
segs = parse_segments(text)
rec = {
"id": _record_id(path, source),
"source_collection": source,
"source_file": path.name,
"source_format": path.suffix.lower().lstrip("."),
"topic": meta["topic"],
"round": meta["round"],
"team_a": meta["team_a"],
"team_b": meta["team_b"],
"num_segments": len(segs),
"num_chars": len(normalize_full_text(text)),
"transcript": normalize_full_text(text),
"segments": segs,
}
out.write(json.dumps(rec, ensure_ascii=False) + "\n")
out.flush()
written += 1
if i % 25 == 0:
rate = i / max(time.time() - t0, 0.001)
print(
f" {i}/{len(todo)} (+{written - len(done)} this run, "
f"{rate:.1f}/s, last: {path.name[:60]})",
flush=True,
)
_quit_word(word)
if failed:
with open(ERR_LOG, "a", encoding="utf-8") as ef:
for p, err in failed:
ef.write(f"{p}\t{err}\n")
elapsed = time.time() - t0
print(
f"\nDone. {written} total records in JSONL, "
f"{written - len(done)} new this run, {len(failed)} failed. "
f"Elapsed {elapsed:.1f}s. → {OUT_JSONL}",
flush=True,
)
if failed:
print(f" failures logged at {ERR_LOG}", flush=True)
return 0 if written else 1
if __name__ == "__main__":
try:
sys.exit(main())
except Exception:
traceback.print_exc()
sys.exit(2)
|