table-extractor

Table Extractor Skill

概要

このスキルでは、PDF から表を高精度で抽出するための標準ツールである camelot を使用します。結合セルのある複雑な表、罫線のない表、複数ページにわたるレイアウトにも高精度で対応します。

使用方法

1. 表を含む PDF を指定します
2. 必要に応じてページや表の検出方法を指定します
3. 表を pandas DataFrame として抽出します

プロンプトの例：

• 「この PDF からすべての表を抽出して」
• 「このレポートの 5 ページ目にある表を取得して」
• 「このドキュメントから罫線のない表を抽出して」
• 「PDF の表を Excel 形式に変換して」

ドメイン知識

camelot の基本

import camelot

# Extract tables from PDF
tables = camelot.read_pdf('document.pdf')

# Access results
print(f"Found {len(tables)} tables")

# Get first table as DataFrame
df = tables[0].df
print(df)

抽出方法

方法	用途	説明
lattice	罫線のある表	線や罫線を使って表を検出します
stream	罫線のない表	テキストの位置を使います

# Lattice method (default) - for tables with visible borders
tables = camelot.read_pdf('document.pdf', flavor='lattice')

# Stream method - for borderless tables
tables = camelot.read_pdf('document.pdf', flavor='stream')

ページの選択

# Single page
tables = camelot.read_pdf('document.pdf', pages='1')

# Multiple pages
tables = camelot.read_pdf('document.pdf', pages='1,3,5')

# Page range
tables = camelot.read_pdf('document.pdf', pages='1-5')

# All pages
tables = camelot.read_pdf('document.pdf', pages='all')

高度なオプション

Lattice オプション

tables = camelot.read_pdf(
    'document.pdf',
    flavor='lattice',
    line_scale=40,              # Line detection sensitivity
    copy_text=['h', 'v'],       # Copy text across merged cells
    shift_text=['l', 't'],      # Shift text alignment
    split_text=True,            # Split text at newlines
    flag_size=True,             # Flag super/subscripts
    strip_text='\n',            # Characters to strip
    process_background=False,   # Process background lines
)

Stream オプション

tables = camelot.read_pdf(
    'document.pdf',
    flavor='stream',
    edge_tol=500,               # Edge tolerance
    row_tol=10,                 # Row tolerance
    column_tol=0,               # Column tolerance
    strip_text='\n',            # Characters to strip
)

表の範囲の指定

# Extract from specific area (x1, y1, x2, y2)
# Coordinates from bottom-left, in PDF points (72 points = 1 inch)
tables = camelot.read_pdf(
    'document.pdf',
    table_areas=['72,720,540,400'],  # One area
)

# Multiple areas
tables = camelot.read_pdf(
    'document.pdf',
    table_areas=['72,720,540,400', '72,380,540,200'],
)

列の指定

# Manually specify column positions (for stream method)
tables = camelot.read_pdf(
    'document.pdf',
    flavor='stream',
    columns=['100,200,300,400'],  # X positions of column separators
)

抽出結果の操作

import camelot

tables = camelot.read_pdf('document.pdf')

for i, table in enumerate(tables):
    # Access DataFrame
    df = table.df
    
    # Table metadata
    print(f"Table {i+1}:")
    print(f"  Page: {table.page}")
    print(f"  Accuracy: {table.accuracy}")
    print(f"  Whitespace: {table.whitespace}")
    print(f"  Order: {table.order}")
    print(f"  Shape: {df.shape}")
    
    # Parsing report
    report = table.parsing_report
    print(f"  Report: {report}")

エクスポートオプション

import camelot

tables = camelot.read_pdf('document.pdf')

# Export to CSV
tables[0].to_csv('table.csv')

# Export to Excel
tables[0].to_excel('table.xlsx')

# Export to JSON
tables[0].to_json('table.json')

# Export to HTML
tables[0].to_html('table.html')

# Export all tables
for i, table in enumerate(tables):
    table.to_excel(f'table_{i+1}.xlsx')

視覚的なデバッグ

import camelot

# Enable visual debugging
tables = camelot.read_pdf('document.pdf')

# Plot detected table areas
camelot.plot(tables[0], kind='contour').show()

# Plot text on table
camelot.plot(tables[0], kind='text').show()

# Plot detected lines (lattice only)
camelot.plot(tables[0], kind='joint').show()
camelot.plot(tables[0], kind='line').show()

# Save plot
fig = camelot.plot(tables[0])
fig.savefig('debug.png')

複数ページにわたる表の処理

import camelot
import pandas as pd

def extract_multipage_table(pdf_path, pages='all'):
    """Extract and combine tables that span multiple pages."""
    
    tables = camelot.read_pdf(pdf_path, pages=pages)
    
    # Group tables by similar structure (columns)
    table_groups = {}
    
    for table in tables:
        cols = tuple(table.df.columns)
        if cols not in table_groups:
            table_groups[cols] = []
        table_groups[cols].append(table.df)
    
    # Combine similar tables
    combined = []
    for cols, dfs in table_groups.items():
        if len(dfs) > 1:
            # Combine and deduplicate header rows
            combined_df = pd.concat(dfs, ignore_index=True)
            combined.append(combined_df)
        else:
            combined.append(dfs[0])
    
    return combined

ベストプラクティス

1. 両方の方法を試す：罫線のある表には Lattice、罫線のない表には Stream を使用します
2. 精度スコアを確認する：通常、90% 以上であれば良好です
3. 視覚的なデバッグを行う：抽出結果を確認します
4. 範囲を指定する：複数の種類の表を含む PDF で使用します
5. ヘッダーを処理する：多くの場合、最初の行は特別な処理が必要です

よくあるパターン

表の一括抽出

import camelot
from pathlib import Path
import pandas as pd

def batch_extract_tables(input_dir, output_dir):
    """Extract tables from all PDFs in directory."""
    
    input_path = Path(input_dir)
    output_path = Path(output_dir)
    output_path.mkdir(exist_ok=True)
    
    results = []
    
    for pdf_file in input_path.glob('*.pdf'):
        try:
            tables = camelot.read_pdf(str(pdf_file), pages='all')
            
            for i, table in enumerate(tables):
                # Skip low accuracy tables
                if table.accuracy < 80:
                    continue
                
                output_file = output_path / f"{pdf_file.stem}_table_{i+1}.xlsx"
                table.to_excel(str(output_file))
                
                results.append({
                    'source': str(pdf_file),
                    'table': i + 1,
                    'page': table.page,
                    'accuracy': table.accuracy,
                    'output': str(output_file)
                })
        
        except Exception as e:
            results.append({
                'source': str(pdf_file),
                'error': str(e)
            })
    
    return results

表の抽出方法を自動検出

import camelot

def smart_extract_tables(pdf_path, pages='1'):
    """Try both methods and return best results."""
    
    # Try lattice first
    lattice_tables = camelot.read_pdf(pdf_path, pages=pages, flavor='lattice')
    
    # Try stream
    stream_tables = camelot.read_pdf(pdf_path, pages=pages, flavor='stream')
    
    # Compare and return best
    results = []
    
    if lattice_tables and lattice_tables[0].accuracy > 70:
        results.extend(lattice_tables)
    elif stream_tables:
        results.extend(stream_tables)
    
    return results

例

例 1：財務諸表の抽出

import camelot
import pandas as pd

def extract_financial_tables(pdf_path):
    """Extract financial tables from annual report."""
    
    # Extract all tables
    tables = camelot.read_pdf(pdf_path, pages='all', flavor='lattice')
    
    financial_data = {
        'income_statement': None,
        'balance_sheet': None,
        'cash_flow': None,
        'other_tables': []
    }
    
    for table in tables:
        df = table.df
        text = df.to_string().lower()
        
        # Identify table type
        if 'revenue' in text or 'sales' in text:
            if 'operating income' in text or 'net income' in text:
                financial_data['income_statement'] = df
        elif 'asset' in text and 'liabilities' in text:
            financial_data['balance_sheet'] = df
        elif 'cash flow' in text or 'operating activities' in text:
            financial_data['cash_flow'] = df
        else:
            financial_data['other_tables'].append({
                'page': table.page,
                'data': df,
                'accuracy': table.accuracy
            })
    
    return financial_data

financials = extract_financial_tables('annual_report.pdf')
if financials['income_statement'] is not None:
    print("Income Statement found:")
    print(financials['income_statement'])

例 2：科学データの抽出

import camelot
import pandas as pd

def extract_research_data(pdf_path, pages='all'):
    """Extract data tables from research paper."""
    
    # Try lattice for bordered tables
    tables = camelot.read_pdf(pdf_path, pages=pages, flavor='lattice')
    
    if not tables or all(t.accuracy < 70 for t in tables):
        # Fall back to stream for borderless
        tables = camelot.read_pdf(pdf_path, pages=pages, flavor='stream')
    
    extracted_data = []
    
    for table in tables:
        df = table.df
        
        # Clean up the DataFrame
        # Set first row as header if it looks like one
        if not df.iloc[0].str.contains(r'\d').any():
            df.columns = df.iloc[0]
            df = df[1:]
            df = df.reset_index(drop=True)
        
        extracted_data.append({
            'page': table.page,
            'accuracy': table.accuracy,
            'data': df
        })
    
    return extracted_data

data = extract_research_data('research_paper.pdf')
for i, item in enumerate(data):
    print(f"Table {i+1} (Page {item['page']}, Accuracy: {item['accuracy']}%):")
    print(item['data'].head())

例 3：請求書の明細項目

import camelot

def extract_invoice_items(pdf_path):
    """Extract line items from invoice."""
    
    # Usually invoices have bordered tables
    tables = camelot.read_pdf(pdf_path, flavor='lattice')
    
    line_items = []
    
    for table in tables:
        df = table.df
        
        # Look for table with typical invoice columns
        header_text = ' '.join(df.iloc[0].astype(str)).lower()
        
        if any(term in header_text for term in ['quantity', 'qty', 'amount', 'price', 'description']):
            # This looks like a line items table
            df.columns = df.iloc[0]
            df = df[1:]
            
            for _, row in df.iterrows():
                item = {}
                for col in df.columns:
                    col_lower = str(col).lower()
                    value = row[col]
                    
                    if 'desc' in col_lower or 'item' in col_lower:
                        item['description'] = value
                    elif 'qty' in col_lower or 'quantity' in col_lower:
                        item['quantity'] = value
                    elif 'price' in col_lower or 'rate' in col_lower:
                        item['unit_price'] = value
                    elif 'amount' in col_lower or 'total' in col_lower:
                        item['amount'] = value
                
                if item:
                    line_items.append(item)
    
    return line_items

items = extract_invoice_items('invoice.pdf')
for item in items:
    print(item)

例 4：表の比較

import camelot
import pandas as pd

def compare_pdf_tables(pdf1_path, pdf2_path):
    """Compare tables between two PDF versions."""
    
    tables1 = camelot.read_pdf(pdf1_path)
    tables2 = camelot.read_pdf(pdf2_path)
    
    comparisons = []
    
    # Match tables by shape and position
    for t1 in tables1:
        best_match = None
        best_score = 0
        
        for t2 in tables2:
            if t1.df.shape == t2.df.shape:
                # Calculate similarity
                try:
                    similarity = (t1.df == t2.df).mean().mean()
                    if similarity > best_score:
                        best_score = similarity
                        best_match = t2
                except:
                    pass
        
        if best_match:
            comparisons.append({
                'page1': t1.page,
                'page2': best_match.page,
                'similarity': best_score,
                'identical': best_score == 1.0,
                'diff': pd.DataFrame(t1.df != best_match.df)
            })
    
    return comparisons

comparison = compare_pdf_tables('report_v1.pdf', 'report_v2.pdf')

制限事項

• 暗号化された PDF には対応していません
• 画像ベースの PDF には、事前の OCR 処理が必要です
• 非常に複雑な結合セルは調整が必要になる場合があります
• 回転した表には事前処理が必要です
• 大きな PDF はページごとに処理する必要がある場合があります

インストール

pip install camelot-py[cv]

# Additional dependencies
# macOS
brew install ghostscript tcl-tk

# Ubuntu
apt-get install ghostscript python3-tk

リソース

• camelot のドキュメント
• GitHub リポジトリ
• 他のツールとの比較