File size: 17,679 Bytes
2f9e43a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 | """
将标注转移到 OCR 后的 JSON 文本 chunk 中
------------------------------------------
支持两种 chunk 模式 × 两种文档模式:
chunk: "length" (SentenceSplitter) / "structure" (DFS-based)
doc: "single" (V1.csv, 单文档问题) / "cross" (cross.xlsx, 跨文档问题)
用法:
python transfer_annotations.py [chunk_mode] [doc_mode]
例:
python transfer_annotations.py length single # 默认
python transfer_annotations.py structure cross # structure chunk + cross 标注
"""
import os
import re
import json
import sys
import pandas as pd
from llama_index.core.node_parser import SentenceSplitter
# ======================== 配置 ========================
import os
CSV_PATH = os.path.join(
"Report-Level Dataset",
"ClimRetrieve_ReportLevel_V1.csv"
)
CROSS_XLSX_PATH = os.path.join(
"Expert-Annotated Relevant Sources Dataset",
"ClimRetrieve_cross.xlsx"
)
MINERU_DIR = "MinerU_Reports"
# ---- Chunk 模式 ----
CHUNK_MODE = "structure" # "length" 或 "structure"
# ---- 文档模式 ----
# "single" : 使用 V1.csv 的 relevant_text (单文档问题)
# "cross" : 使用 cross.xlsx 的 Relevant (跨文档问题)
DOC_MODE = "cross"
# length 模式参数
CHUNK_SIZE = 350 # 目标 chunk 大小 (词数)
CHUNK_OVERLAP = 50 # overlap 大小 (词数)
# structure 模式参数
STRUCTURE_MAX_TOKENS = 550
RELEVANCE_THRESHOLD = 1 # 只使用 relevance >= 此阈值的 relevant_text (single 模式)
MIN_MATCH_LEN = 30 # relevant_text 最短长度, 太短的跳过(避免误匹配)
# ======================== 工具函数 ========================
def word_count_tokenizer(text: str) -> list:
"""按空格分词, 返回 token 列表, 用于 SentenceSplitter 按词数切分."""
return text.split()
def normalize_text(text: str) -> str:
"""标准化文本: 去除多余空白, 小写化."""
text = re.sub(r'\s+', ' ', text).strip().lower()
return text
def load_content_list_json(json_path: str) -> str:
"""加载 content_list.json, 提取所有 text 类型的文本并拼接."""
with open(json_path, 'r', encoding='utf-8') as f:
content_list = json.load(f)
texts = []
for item in content_list:
if item.get('type') == 'text' and item.get('text'):
texts.append(item['text'].strip())
# 用换行连接各段文字
full_text = "\n".join(texts)
return full_text
def build_report_name_mapping(csv_reports, mineru_folders):
"""建立 CSV report 名称 -> MinerU 文件夹名称 的映射."""
mapping = {}
for csv_name in csv_reports:
# 去掉 .pdf 后缀
stripped = csv_name.replace(".pdf", "").strip()
# 精确匹配 (忽略大小写和尾部空格)
found = False
for folder in mineru_folders:
if stripped.lower() == folder.strip().lower():
mapping[csv_name] = folder
found = True
break
if not found:
# 尝试模糊匹配: 基于词集重叠
csv_words = set(stripped.lower().split())
best_match = None
best_score = 0
for folder in mineru_folders:
folder_words = set(folder.lower().split())
overlap = len(csv_words & folder_words)
total = max(len(csv_words), len(folder_words))
score = overlap / total if total > 0 else 0
if score > best_score:
best_score = score
best_match = folder
if best_match and best_score >= 0.8:
mapping[csv_name] = best_match
print(f" [模糊匹配] CSV '{csv_name}' -> MinerU '{best_match}' (score={best_score:.2f})")
else:
print(f" [未匹配] CSV '{csv_name}' (best: '{best_match}', score={best_score:.2f})")
return mapping
def find_content_list_json(mineru_dir, folder_name):
"""在 MinerU 文件夹中找到 _content_list.json 文件."""
folder_path = os.path.join(mineru_dir, folder_name)
for f in os.listdir(folder_path):
if f.endswith('_content_list.json'):
return os.path.join(folder_path, f)
return None
def check_chunk_contains_relevant_text(chunk_normalized: str, relevant_text_normalized: str) -> bool:
"""检查 chunk 是否包含 relevant_text.
策略:
1. 直接子串匹配 (标准化后)
2. 如果 relevant_text 较长, 尝试匹配其中连续子句 (取前/后 50% 词)
"""
# 直接子串匹配
if relevant_text_normalized in chunk_normalized:
return True
# 尝试匹配 relevant_text 的前半部分和后半部分
# (因为 OCR 可能在句子边界处有差异)
words = relevant_text_normalized.split()
if len(words) >= 8:
# 取前 50% 的词
front_part = ' '.join(words[:int(len(words) * 0.5)])
if len(front_part) >= MIN_MATCH_LEN and front_part in chunk_normalized:
return True
# 取后 50% 的词
back_part = ' '.join(words[int(len(words) * 0.5):])
if len(back_part) >= MIN_MATCH_LEN and back_part in chunk_normalized:
return True
return False
def normalize_chunk_for_matching_raw(chunk_text: str) -> str:
"""标准化 chunk 原始文本 (保留标题行) 用于匹配."""
return normalize_text(chunk_text)
def build_chunk_match_views(chunk_text: str) -> dict:
"""
为同一 chunk 构建双通道匹配视图:
- raw: 保留标题行
- stripped: 去除 markdown 标题行
若两者文本相同, 仅保留 raw.
"""
raw_norm = normalize_chunk_for_matching_raw(chunk_text)
stripped_norm = normalize_chunk_for_matching(chunk_text)
if raw_norm == stripped_norm:
return {"raw": raw_norm}
return {"raw": raw_norm, "stripped": stripped_norm}
def match_relevant_text_multi_view(chunk_views: dict, relevant_text_normalized: str) -> tuple:
"""
在多个文本视图上匹配 relevant_text.
返回:
(is_matched: bool, hit_channels: list[str])
"""
hit_channels = []
for channel, chunk_norm in chunk_views.items():
if check_chunk_contains_relevant_text(chunk_norm, relevant_text_normalized):
hit_channels.append(channel)
return (len(hit_channels) > 0), hit_channels
# ======================== 主流程 ========================
def chunk_report_length(json_path, splitter):
"""[length 模式] 加载 OCR 文本并用 SentenceSplitter 切分."""
full_text = load_content_list_json(json_path)
if len(full_text.strip()) == 0:
return [], []
chunks = splitter.split_text(full_text)
# 返回 (chunk_texts, chunk_extras) — extras 为空 dict 列表
return chunks, [{} for _ in chunks]
def chunk_report_structure(json_path, max_tokens):
"""[structure 模式] 用 Structure-based DFS Chunking 切分."""
from Experiments.structure_chunker import structure_chunk_document
chunks_data = structure_chunk_document(json_path, max_tokens=max_tokens)
if not chunks_data:
return [], []
texts = [c["text"] for c in chunks_data]
extras = [{"section_path": " > ".join(c["metadata"]["section_path"]),
"document_id": c["metadata"]["document_id"]} for c in chunks_data]
return texts, extras
def normalize_chunk_for_matching(chunk_text: str) -> str:
"""
标准化 chunk 文本用于 relevant_text 匹配.
structure chunk 带有 Markdown 标题行, 匹配时需要去除.
"""
lines = chunk_text.split('\n')
content_lines = [l for l in lines if not l.strip().startswith('#')]
return normalize_text(' '.join(content_lines))
def _build_output_dir(chunk_mode, doc_mode):
"""构建输出目录名."""
base = "OCR_Chunked_Annotated"
parts = [base]
if chunk_mode == "structure":
parts.append("structure")
if doc_mode == "cross":
parts.append("cross")
return "_".join(parts)
def load_annotation_source(doc_mode):
"""
根据 doc_mode 加载标注数据, 返回统一格式:
question_doc_relevant: dict[(report, question)] -> list[str] (标准化后的 relevant texts)
all_reports: list[str]
all_question_docs: list[(question, report)]
"""
if doc_mode == "single":
df = pd.read_csv(CSV_PATH, index_col=0)
print(f" 加载 V1.csv: {len(df)} 行, {df['report'].nunique()} 报告, {df['question'].nunique()} 问题")
print(f" relevance 分布:\n{df['relevance'].value_counts().sort_index().to_string()}")
high_rel = df[df['relevance'] >= RELEVANCE_THRESHOLD].copy()
print(f" relevance >= {RELEVANCE_THRESHOLD}: {len(high_rel)} 行, {high_rel['relevant_text'].nunique()} 唯一 relevant_text")
all_reports = list(df['report'].unique())
question_doc_relevant = {}
for report in all_reports:
report_qs = df[df['report'] == report]['question'].unique()
for q in report_qs:
q_rel = high_rel[(high_rel['report'] == report) & (high_rel['question'] == q)]
rel_texts = q_rel['relevant_text'].dropna().unique()
normalized = [normalize_text(rt) for rt in rel_texts if len(str(rt).strip()) >= MIN_MATCH_LEN]
question_doc_relevant[(report, q)] = normalized
return question_doc_relevant, all_reports
elif doc_mode == "cross":
df = pd.read_excel(CROSS_XLSX_PATH)
if 'Unnamed: 0' in df.columns:
df = df.drop(columns=['Unnamed: 0'])
print(f" 加载 cross.xlsx: {len(df)} 行, {df['Document'].nunique()} 报告, {df['Question'].nunique()} 问题")
print(f" Source Relevance Score 分布:\n{df['Source Relevance Score'].value_counts().sort_index().to_string()}")
all_reports = list(df['Document'].unique())
question_doc_relevant = {}
for _, row in df.iterrows():
report = row['Document']
question = row['Question']
relevant = row.get('Relevant', '')
if pd.isna(relevant) or len(str(relevant).strip()) < MIN_MATCH_LEN:
continue
key = (report, question)
if key not in question_doc_relevant:
question_doc_relevant[key] = []
norm = normalize_text(str(relevant))
if norm not in question_doc_relevant[key]:
question_doc_relevant[key].append(norm)
print(f" (report, question) 对数: {len(question_doc_relevant)}")
return question_doc_relevant, all_reports
else:
raise ValueError(f"未知 DOC_MODE: {doc_mode}")
def main():
global CHUNK_MODE, DOC_MODE
# 命令行参数: python transfer_annotations.py [chunk_mode] [doc_mode]
if len(sys.argv) > 1 and sys.argv[1] in ("length", "structure"):
CHUNK_MODE = sys.argv[1]
if len(sys.argv) > 2 and sys.argv[2] in ("single", "cross"):
DOC_MODE = sys.argv[2]
output_dir = _build_output_dir(CHUNK_MODE, DOC_MODE)
print("=" * 60)
print(f"标注转移 (CHUNK={CHUNK_MODE}, DOC={DOC_MODE})")
print(f"输出目录: {output_dir}")
print("=" * 60)
# ---- Step 1: 加载标注数据 ----
print("\nStep 1: 加载标注数据")
print("=" * 60)
question_doc_relevant, all_reports = load_annotation_source(DOC_MODE)
# ---- Step 2: 建立 report 名称映射 ----
print("\n" + "=" * 60)
print("Step 2: 建立 report 名称映射")
print("=" * 60)
mineru_folders = [f for f in os.listdir(MINERU_DIR)
if os.path.isdir(os.path.join(MINERU_DIR, f))]
report_mapping = build_report_name_mapping(all_reports, mineru_folders)
print(f"\n 成功映射: {len(report_mapping)} / {len(all_reports)}")
# ---- Step 3: 切分 chunk + 转移标注 ----
print("\n" + "=" * 60)
print(f"Step 3: 切分 chunk ({CHUNK_MODE}), 转移标注 ({DOC_MODE})")
print("=" * 60)
splitter = None
if CHUNK_MODE == "length":
splitter = SentenceSplitter(
chunk_size=CHUNK_SIZE,
chunk_overlap=CHUNK_OVERLAP,
tokenizer=word_count_tokenizer,
)
os.makedirs(output_dir, exist_ok=True)
# 收集每个 report 关联的 questions
report_questions_map = {}
for (report, question) in question_doc_relevant.keys():
if report not in report_questions_map:
report_questions_map[report] = set()
report_questions_map[report].add(question)
all_results = []
for report_name in all_reports:
print(f"\n 处理报告: {report_name}")
if report_name not in report_mapping:
print(f" [跳过] 未找到对应的 MinerU 文件夹")
continue
folder_name = report_mapping[report_name]
json_path = find_content_list_json(MINERU_DIR, folder_name)
if json_path is None:
print(f" [跳过] 未找到 content_list.json")
continue
# 切分 chunk
if CHUNK_MODE == "length":
chunks, extras = chunk_report_length(json_path, splitter)
else:
chunks, extras = chunk_report_structure(json_path, STRUCTURE_MAX_TOKENS)
if not chunks:
print(f" [跳过] 无 chunk 产生")
continue
print(f" 切分为 {len(chunks)} 个 {CHUNK_MODE} chunk")
report_questions = report_questions_map.get(report_name, set())
if not report_questions:
print(f" [跳过] 该报告无关联问题")
continue
# 对每个 chunk × question 匹配
for chunk_idx, chunk_text in enumerate(chunks):
chunk_views = build_chunk_match_views(chunk_text)
chunk_word_count = len(chunk_text.split())
for q in report_questions:
rel_norms = question_doc_relevant.get((report_name, q), [])
is_relevant = False
matched_logs = {}
matched_channels = set()
for rt_norm in rel_norms:
hit, channels = match_relevant_text_multi_view(chunk_views, rt_norm)
if hit:
is_relevant = True
if rt_norm not in matched_logs:
matched_logs[rt_norm] = set()
matched_logs[rt_norm].update(channels)
matched_channels.update(channels)
matched_texts = []
for rt_norm in sorted(matched_logs.keys()):
channel_tag = "+".join(sorted(matched_logs[rt_norm]))
matched_texts.append(f"[{channel_tag}] {rt_norm[:80]}...")
result = {
'report': report_name,
'chunk_idx': chunk_idx,
'chunk_text': chunk_text,
'chunk_word_count': chunk_word_count,
'question': q,
'is_relevant': 1 if is_relevant else 0,
'matched_relevant_texts': "; ".join(matched_texts) if matched_texts else "",
'num_matched': len(matched_logs),
'matched_channels': ",".join(sorted(matched_channels)) if matched_channels else "",
}
if extras and chunk_idx < len(extras) and extras[chunk_idx]:
result.update(extras[chunk_idx])
all_results.append(result)
# 统计
report_results = [r for r in all_results if r['report'] == report_name]
total_pairs = len(report_results)
relevant_pairs = sum(1 for r in report_results if r['is_relevant'])
print(f" (chunk, question) 对总数: {total_pairs}, 标为 relevant: {relevant_pairs}")
# ---- Step 4: 保存 ----
print("\n" + "=" * 60)
print("Step 4: 保存结果")
print("=" * 60)
result_df = pd.DataFrame(all_results)
tag_parts = []
if CHUNK_MODE != "length":
tag_parts.append(CHUNK_MODE)
if DOC_MODE != "single":
tag_parts.append(DOC_MODE)
tag = ("_" + "_".join(tag_parts)) if tag_parts else ""
output_csv = os.path.join(output_dir, f"ocr_chunks_annotated{tag}.csv")
result_df.to_csv(output_csv, index=False, encoding='utf-8-sig')
print(f" 完整结果: {output_csv}")
print(f" 总行数: {len(result_df)}")
print(f" 标为 relevant: {result_df['is_relevant'].sum()}")
unique_chunks = result_df.drop_duplicates(subset=['report', 'chunk_idx']).shape[0]
print(f" 唯一 chunk 数: {unique_chunks}")
# chunk 列表
dedup_cols = ['report', 'chunk_idx', 'chunk_text', 'chunk_word_count']
if 'section_path' in result_df.columns:
dedup_cols += ['section_path', 'document_id']
chunks_only = result_df.drop_duplicates(subset=['report', 'chunk_idx'])[dedup_cols].reset_index(drop=True)
chunks_json = os.path.join(output_dir, f"ocr_chunks_all{tag}.json")
chunks_only.to_json(chunks_json, orient='records', force_ascii=False, indent=2)
print(f" chunk 列表: {chunks_json} ({len(chunks_only)} chunks)")
# 汇总
summary = result_df.groupby('report').agg(
total_chunks=('chunk_idx', 'nunique'),
total_pairs=('is_relevant', 'count'),
relevant_pairs=('is_relevant', 'sum'),
).reset_index()
summary['relevant_ratio'] = summary['relevant_pairs'] / summary['total_pairs']
summary_path = os.path.join(output_dir, f"annotation_summary{tag}.csv")
summary.to_csv(summary_path, index=False)
print(f" 汇总: {summary_path}")
print(summary.to_string())
if __name__ == "__main__":
main()
|