Paper To Table

Extract structured data from academic papers (PDF/DOCX/TXT) into literature review tables (XLSX/CSV) with fidelity, batch support, and multi-domain handling.

Extract structured information from academic papers and populate literature review tables.

Quality Principles

1. Extraction Fidelity (质量约束)

• NEVER hallucinate information not present in the paper
• NEVER infer beyond what is explicitly stated
• Use "N/A" only when a field truly cannot be found after thorough search
• Distinguish between:
  • Explicitly stated facts → Extract directly
  • Implied but not stated → Mark as "N/A" or note as implied
  • Absent information → "N/A"
• Confidence scoring: Rate each extraction as High/Medium/Low confidence

2. Field Understanding (字段理解)

• Understand the semantic meaning of each table header
• Map paper content to headers based on meaning, not just keyword matching
• Handle synonyms and domain-specific terminology
• Recognize implicit information in context

3. Structured Depth (结构化深度)

• Extract at appropriate granularity
• Distinguish between study-level, experiment-level, and result-level information
• Preserve relationships between related fields
• Handle multi-experiment papers correctly

4. Batch Stability (批处理稳定性)

• Process papers independently (failure of one doesn't affect others)
• Log all operations for audit trail
• Support resume from interruption
• Validate outputs before writing to table

Workflow

Step 1: Identify Inputs

• Papers: Single file, multiple files, or folder path → 支持 PDF/DOCX/TXT
• Table template: XLSX 或 CSV，含表头定义结构
• Language: 自动检测或用户指定
• Domain: Psychology / Cognitive Neuroscience / Computer Science / Brain Science / General

Step 2: Read Table Headers

bash

python scripts/read_table.py <table_path>

输出：列名、数据类型约束、领域推断。

Step 3: Extract Paper Content

bash

python scripts/extract_paper.py <paper_path> --structured

自动处理格式：PDF→pdfplumber/PyMuPDF/OCR fallback；DOCX→python-docx；TXT→直接读取。

输出结构化 JSON：包含 full_text 和 sections（abstract/introduction/methods/results/discussion/conclusion）。

Step 4: LLM Extraction (Critical)

原则：只提取论文明确陈述的信息，绝不臆造。

输入：表头 + 论文全文/章节

输出格式（每个字段）：

json

{
  "FieldName": {
    "value": "extracted value 或 N/A",
    "confidence": "HIGH/MEDIUM/LOW",
    "source": "paper location"
  }
}

CRITICAL RULES：

1. JSON keys 必须与表头完全匹配（大小写敏感）
2. 缺失信息→"N/A"，不做推断
3. 多值用分号分隔
4. 保留原文语言
5. LOW confidence 字段需说明原因

提取优先级：Abstract→Methods→Results→Discussion→补充材料

Step 5: Validate & Write

bash

python scripts/write_table.py <table_path> '<json_data>' --validate

验证内容：JSON格式、键名匹配、无重复条目、数据类型合理。

重复检测：标题相似度>85%视为重复，跳过并警告。

Step 6: Report

报告：处理论文数、新增行数、跳过数（重复/错误）、LOW confidence 字段、输出路径。

Batch Processing

bash

python scripts/batch_process.py <papers_folder> <table_path> [output_folder]

• 独立处理每篇论文（单篇失败不影响其他）
• 自动生成日志 batch_log_YYYYMMDD_HHMMSS.json
• 支持断点续传（从日志恢复进度）

详细字段定义、提取策略 → references/extraction-patterns.md 质量检查清单 → references/quality-checklist.md

Domain Specializations

Psychology / Cognitive Neuroscience / Computer Science / Brain Science