RoyirinRosmontis
1722c00249
- scripts/expand_real_dataset.py：AutoWeakS 全量 2215 条 + 500 条分层评测集 + 简历扩量（seed=2026 可复现） - pipeline 被动语料机制（passive_sources）：真实无日期 JD 参与接入/清洗/抽取验证（2915 条命中率 99.8%），不进图谱/演化/发现 - 新岗位发现第四指标'差异度'与双年窗口增长率口径 - 评测新增 real_jd_500 维度（99.91%）；测试会话隔离 jd_extra_dirs 防止本机配置拖慢套件
267 lines
10 KiB
Python
267 lines
10 KiB
Python
"""数据扩产脚本：在 build_real_dataset.py 的基础上扩充真实数据集（T1-T4）。
|
||
|
||
产物：
|
||
1. data/real/autoweaks_jd/jd_full.jsonl
|
||
—— AutoWeakS 全量中文 JD（去分词 + 技术岗过滤 + 全文去重），
|
||
保留原 200 条 jd_sample 记录（jid/原文不变），其余新增条目
|
||
publish_date 统一 2019-06-01（该子集不参与演化分析，仅作语料/评测）；
|
||
2. data/real/autoweaks_jd/jd_eval_500.jsonl + gold_answers_500.json
|
||
—— 真实 JD 评测集从 200 扩到 500：包含原 200 条，另按岗位类别分层补 300 条，
|
||
gold 沿用"SkillMatcher 初标"口径（人工抽检记录见 data/real/README.md）；
|
||
3. data/real/aistudio_resumes/：50 份扩到 200 份（保留 PII，沿用既有目录约定）；
|
||
4. data/real/batch_demo_resumes/：12 份扩到 32 份（落盘前必须经 mask_text 脱敏）。
|
||
|
||
英文数据集（T5）默认不接入，原因记录在 data/real/README.md。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import csv
|
||
import json
|
||
import random
|
||
import re
|
||
import sys
|
||
from collections import Counter
|
||
from pathlib import Path
|
||
|
||
ROOT = Path(__file__).resolve().parents[1]
|
||
sys.path.insert(0, str(ROOT))
|
||
sys.path.insert(0, str(ROOT / "scripts"))
|
||
|
||
from app.extract.base import SkillMatcher, load_lexicon # noqa: E402
|
||
from app.match.resume import mask_text # noqa: E402
|
||
from build_real_dataset import ( # noqa: E402
|
||
DST,
|
||
SRC,
|
||
SEED,
|
||
TECH_RE,
|
||
_PHONE_RE,
|
||
detokenize,
|
||
guess_title,
|
||
render_aistudio_resume,
|
||
render_chinese_resume,
|
||
)
|
||
|
||
AW_DIR = DST / "autoweaks_jd"
|
||
RESUME_DIR = DST / "aistudio_resumes"
|
||
DEMO_DIR = DST / "batch_demo_resumes"
|
||
|
||
|
||
# ---------- T1：AutoWeakS 全量中文 JD ----------
|
||
def build_autoweaks_full(matcher: SkillMatcher) -> list[dict]:
|
||
"""去分词 → 技术岗过滤 → 全文去重；保留原 200 条样本，其余追加。"""
|
||
pools: dict[str, list[str]] = {}
|
||
for fn, tag in [("Tecent_job_phrase.txt", "腾讯"), ("JD_job_phrase.txt", "京东")]:
|
||
lines = (SRC / "autoweaks" / fn).read_text(encoding="utf-8").splitlines()
|
||
seen, pool = set(), []
|
||
for ln in lines:
|
||
text = detokenize(ln)
|
||
if len(text) < 60 or not TECH_RE.search(text):
|
||
continue
|
||
if not matcher.find_names(text):
|
||
continue
|
||
if text in seen:
|
||
continue
|
||
seen.add(text)
|
||
pool.append(text)
|
||
pools[tag] = pool
|
||
|
||
sample_path = AW_DIR / "jd_sample.jsonl"
|
||
sample = [json.loads(ln) for ln in
|
||
sample_path.read_text(encoding="utf-8").splitlines() if ln.strip()]
|
||
sample_texts = {r["text"] for r in sample}
|
||
|
||
records = list(sample)
|
||
new_pairs: list[tuple[str, str]] = []
|
||
for tag in ("腾讯", "京东"):
|
||
for text in sorted(pools[tag]):
|
||
if text in sample_texts:
|
||
continue
|
||
new_pairs.append((tag, text))
|
||
|
||
for i, (company, text) in enumerate(new_pairs, start=len(sample) + 1):
|
||
records.append({
|
||
"jid": f"AW{i:04d}",
|
||
"job_title": guess_title(text),
|
||
"company": company,
|
||
"publish_date": "2019-06-01", # 无源日期，统一标记；该子集不参与演化分析
|
||
"source": "autoweaks",
|
||
"url": "",
|
||
"text": text,
|
||
})
|
||
|
||
with (AW_DIR / "jd_full.jsonl").open("w", encoding="utf-8") as f:
|
||
for r in records:
|
||
f.write(json.dumps(r, ensure_ascii=False) + "\n")
|
||
print(f"[T1] AutoWeakS 全量：腾讯池 {len(pools['腾讯'])} 京东池 {len(pools['京东'])}"
|
||
f" → jd_full {len(records)} 条（含原样本 {len(sample)} 条）")
|
||
return records
|
||
|
||
|
||
# ---------- T2：真实 JD 评测集 500 条（分层） ----------
|
||
def build_eval_500(matcher: SkillMatcher, full: list[dict]) -> tuple[list[dict], dict]:
|
||
rng = random.Random(SEED)
|
||
sample = [json.loads(ln) for ln in
|
||
(AW_DIR / "jd_sample.jsonl").read_text(encoding="utf-8").splitlines() if ln.strip()]
|
||
sample_ids = {r["jid"] for r in sample}
|
||
|
||
by_title: dict[str, list[dict]] = {}
|
||
for r in full:
|
||
by_title.setdefault(r["job_title"], []).append(r)
|
||
sample_counts = Counter(r["job_title"] for r in sample)
|
||
|
||
remaining = {
|
||
title: [r for r in group if r["jid"] not in sample_ids]
|
||
for title, group in by_title.items()
|
||
}
|
||
for v in remaining.values():
|
||
rng.shuffle(v)
|
||
# 优先补齐样本覆盖少的岗位类别（样本数升序，剩余池大者优先）
|
||
order = sorted(remaining, key=lambda t: (sample_counts.get(t, 0), -len(remaining[t])))
|
||
|
||
picked = list(sample)
|
||
picked_ids = set(sample_ids)
|
||
while len(picked) < 500:
|
||
advanced = False
|
||
for title in order:
|
||
if not remaining[title]:
|
||
continue
|
||
rec = remaining[title].pop(0)
|
||
if rec["jid"] in picked_ids:
|
||
continue
|
||
picked.append(rec)
|
||
picked_ids.add(rec["jid"])
|
||
advanced = True
|
||
if len(picked) >= 500:
|
||
break
|
||
if not advanced:
|
||
break
|
||
|
||
gold: dict[str, dict] = {}
|
||
for r in picked:
|
||
gold[r["jid"]] = {
|
||
"job_title": r["job_title"],
|
||
"responsibilities": [],
|
||
"required_skills": matcher.find_names(r["text"]),
|
||
"preferred_skills": [],
|
||
"scenarios": matcher.find_scenarios(r["text"]),
|
||
}
|
||
|
||
with (AW_DIR / "jd_eval_500.jsonl").open("w", encoding="utf-8") as f:
|
||
for r in picked:
|
||
f.write(json.dumps(r, ensure_ascii=False) + "\n")
|
||
(AW_DIR / "gold_answers_500.json").write_text(
|
||
json.dumps(gold, ensure_ascii=False, indent=1), encoding="utf-8")
|
||
cats = sorted(Counter(r["job_title"] for r in picked).items())
|
||
print(f"[T2] 真实 JD 评测集：{len(picked)} 条（含原样本 {len(sample)} 条），"
|
||
f"岗位类别 {len(cats)} 类")
|
||
return picked, gold
|
||
|
||
|
||
# ---------- T3：AIStudio 简历 50 → 200 ----------
|
||
def build_aistudio_extra(matcher: SkillMatcher) -> None:
|
||
target_total = 200 # 50（既有）→ 200，幂等：够了就跳过
|
||
rng = random.Random(SEED + 4)
|
||
gold_path = RESUME_DIR / "gold.json"
|
||
gold = json.loads(gold_path.read_text(encoding="utf-8"))
|
||
need = target_total - len(gold)
|
||
if need <= 0:
|
||
print(f"[T3] AIStudio 简历：已有 {len(gold)} 份，目标 {target_total}，跳过")
|
||
return
|
||
used = {g["source_id"] for g in gold.values()}
|
||
|
||
data = json.loads((SRC / "aistudio" / "train.json").read_text(encoding="utf-8"))
|
||
cands = []
|
||
for key, rec in data.items():
|
||
if key in used:
|
||
continue
|
||
blob = json.dumps([rec.get("工作经历"), rec.get("项目经历")], ensure_ascii=False)
|
||
if not TECH_RE.search(blob):
|
||
continue
|
||
if not _PHONE_RE.match(str(rec.get("电话", ""))):
|
||
continue
|
||
name = str(rec.get("姓名", "")).strip()
|
||
if not re.fullmatch(r"[一-龥]{2,4}", name):
|
||
continue
|
||
cleaned = dict(rec)
|
||
cleaned["姓名"] = name
|
||
cands.append((key, cleaned))
|
||
rng.shuffle(cands)
|
||
picked = cands[:need]
|
||
|
||
start = len(gold) # 编号接续
|
||
for i, (key, rec) in enumerate(picked, start=start + 1):
|
||
fname = f"aistudio_{i:03d}.txt"
|
||
text = render_aistudio_resume(rec)
|
||
(RESUME_DIR / fname).write_text(text, encoding="utf-8")
|
||
gold[fname] = {
|
||
"name": rec["姓名"].strip(),
|
||
"phone": rec["电话"],
|
||
"skills": matcher.find_names(text),
|
||
"source_id": key,
|
||
}
|
||
gold = dict(sorted(gold.items()))
|
||
gold_path.write_text(json.dumps(gold, ensure_ascii=False, indent=1), encoding="utf-8")
|
||
print(f"[T3] AIStudio 简历：候选 {len(cands)} → 新增 {len(picked)} 份"
|
||
f"（目录合计 {len(gold)} 份）")
|
||
|
||
|
||
# ---------- T4：chinese_resume 批量演示 12 → 32（必须脱敏） ----------
|
||
def build_demo_extra() -> None:
|
||
target_total = 32 # 12（既有）→ 32，幂等：够了就跳过
|
||
rng = random.Random(SEED + 3)
|
||
with (SRC / "chinese_resume" / "Chinese_resume_data.csv").open(encoding="utf-8") as f:
|
||
rows = [r for r in csv.DictReader(f) if r.get("筛选结果") == "通过"]
|
||
|
||
existing = sorted(p.name for p in DEMO_DIR.glob("demo_*.txt"))
|
||
need = target_total - len(existing)
|
||
if need <= 0:
|
||
print(f"[T4] chinese_resume 批量演示：已有 {len(existing)} 份，目标 {target_total}，跳过")
|
||
return
|
||
used_slugs = {re.sub(r"\.txt$", "", name).split("_", 2)[-1] for name in existing}
|
||
by_intent: dict[str, list[dict]] = {}
|
||
for r in rows:
|
||
by_intent.setdefault(r["意向岗位"], []).append(r)
|
||
intents = sorted(by_intent, key=lambda k: -len(by_intent[k]))
|
||
|
||
picked: list[dict] = []
|
||
# 第一轮：优先覆盖已有文件未覆盖的意向岗位
|
||
for intent in intents:
|
||
if len(picked) >= need:
|
||
break
|
||
slug = re.sub(r"\W+", "", intent)
|
||
if slug in used_slugs:
|
||
continue
|
||
picked.append(rng.choice(by_intent[intent]))
|
||
used_slugs.add(slug)
|
||
# 第二轮：从其余通过样本补齐（避免覆盖已有文件）
|
||
rest = [r for intent in intents for r in by_intent[intent] if r not in picked]
|
||
rng.shuffle(rest)
|
||
for row in rest:
|
||
if len(picked) >= need:
|
||
break
|
||
picked.append(row)
|
||
|
||
start = len(existing) # 既有 12 份，编号接续
|
||
for i, row in enumerate(picked, start=start + 1):
|
||
raw = render_chinese_resume(row)
|
||
masked = mask_text(raw) # 脱敏后才落盘，原始 PII 不写入 data/
|
||
slug = re.sub(r"\W+", "", row["意向岗位"])
|
||
(DEMO_DIR / f"demo_{i:02d}_{slug}.txt").write_text(masked, encoding="utf-8")
|
||
print(f"[T4] chinese_resume 批量演示：新增 {len(picked)} 份"
|
||
f"（目录合计 {start + len(picked)} 份，全部已脱敏）")
|
||
|
||
|
||
def main() -> None:
|
||
if not SRC.exists():
|
||
raise SystemExit(f"源数据目录不存在: {SRC}")
|
||
matcher = SkillMatcher(load_lexicon())
|
||
full = build_autoweaks_full(matcher)
|
||
build_eval_500(matcher, full)
|
||
build_aistudio_extra(matcher)
|
||
build_demo_extra()
|
||
print(f"数据扩产完成，产物写入 {DST}")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|