Manual Review Guide for OCR Outputs

Comprehensive guide for manually verifying and correcting OCR-generated CSV files.

---

When is Manual Review Needed?

The OCR pipeline automatically flags outputs for manual review when:

1. Low OCR confidence (< 85%)
2. Validation errors detected:
   • Missing expected columns
   • Incorrect or missing sector names
   • Non-numeric values in data columns
   • Unrealistic values (> 500% or < -100%)
3. High proportion of missing values (> 10%)

---

Review Workflow Overview

1. Check needs_review.txt
      ↓
2. Open review package for each flagged file
      ↓
3. Compare PDF with CSV
      ↓
4. Correct errors in CSV
      ↓
5. Save corrected CSV to output folder
      ↓
6. Delete review folder
      ↓
7. Verify corrections

---

Step-by-Step Instructions

Step 1: Identify Files Needing Review

# Check the needs-review list
cat OCR/needs_review.txt

# Or view in editor
notepad OCR/needs_review.txt

Example output:

Files Needing Manual Review
================================================================================

Total files: 3

ns_11_1994.pdf
  Confidence: 78%
  Issues:
    - Missing columns
    - Unclear sector names in rows 3-4

ns_15_1994.pdf
  Confidence: 81%
  Issues:
    - Digit misrecognition in mining sector

...

Step 2: Open Review Package

For each flagged file, navigate to its review folder:

OCR/review/YYYY/ns_XX_YYYY/
├── original.pdf           # Scanned source document
├── preprocessed.png       # After image preprocessing
├── ocr_output.csv         # OCR result needing correction
└── issues.txt             # Detailed issues report

Example: For ns_11_1994.pdf, open:

• OCR/review/1994/ns_11_1994/

Step 3: Compare PDF with CSV

1. Open original.pdf in a PDF viewer
2. Open ocr_output.csv in Excel or text editor
3. Compare side-by-side:
   • Check sector names in first two columns
   • Verify numeric values column by column
   • Look for missing or transposed digits

Pro tip: Use dual monitors or split-screen for easier comparison

Step 4: Common OCR Errors to Look For

A. Digit Misrecognition

OCR Error	Should Be	Context
O (letter)	0 (zero)	In numeric values: 1O.5 → 10.5
l (lowercase L)	1 (one)	In numeric values: l2.3 → 12.3
S (letter)	5 (five)	At end of number: 9.S → 9.5
B (letter)	8 (eight)	In numeric values: B.4 → 8.4
Z (letter)	2 (two)	Less common: Z.5 → 2.5

Example from actual table:

# OCR output (WRONG):
pesca;fishing;-29.B;-26.4;-lO.5;12.S

# Corrected:
pesca;fishing;-29.8;-26.4;-10.5;12.5

B. Missing Decimal Points

Problem: OCR sometimes misses decimal points, especially in faded scans

# WRONG:
agropecuario;agriculture;37;36;42

# CORRECT:
agropecuario;agriculture;3.7;3.6;4.2

How to detect: Values > 100 are unusual for monthly/quarterly growth rates

C. Transposed or Missing Digits

Problem: Digits can be misread or completely missed

# WRONG (transposed 96 → 69):
manufactura;manufacturing;-3.5;-2.3;69.8

# CORRECT:
manufactura;manufacturing;-3.5;-2.3;96.8

D. Sector Name Errors

Problem: Spanish sector names with accents or special characters

# WRONG:
minena;mining;...           # Should be "mineria"
construcciOn;construction;  # Should be "construccion"
electncidad;electricity;    # Should be "electricidad"

# CORRECT:
mineria e hidrocarburos;mining and fuel;...
construccion;construction;...
electricidad y agua;electricity and water;...

E. Column Misalignment

Problem: OCR sometimes merges or splits columns incorrectly

# WRONG (values shifted):
sectores_economicos;economic_sectors;1992_ene;1992_feb
agropecuario agriculture;and livestock;3.7;3.6

# CORRECT:
sectores_economicos;economic_sectors;1992_ene;1992_feb
agropecuario;agriculture and livestock;3.7;3.6

F. Missing Negative Signs

Problem: Negative signs can be lost, especially for small fonts

# WRONG (should be negative):
pesca;fishing;29.8;26.4;55.9

# CORRECT:
pesca;fishing;-29.8;-26.4;-55.9

How to detect: Check if values seem unrealistic (fishing sector with consistent huge growth)

---

Step 5: Correction Workflow

Using Excel (Recommended)

1. Open CSV in Excel:

   File → Open → Select "ocr_output.csv"
   

2. Check delimiter:

   • Should be semicolon (;)
   • Excel may auto-detect or ask

3. Correction checklist:

   • First column: Spanish sector names correct
   • Second column: English sector names correct
   • Remaining columns: All numeric values
   • No missing decimal points
   • No digit misrecognitions (O→0, l→1, etc.)
   • Negative signs present where needed
   • No column misalignments

4. Save corrected version:

   File → Save As → CSV (Semicolon delimited) (*.csv)
   

Using Text Editor (Advanced)

For quick fixes, use VS Code, Notepad++, or similar:

1. Open ocr_output.csv

2. Use Find & Replace for systematic errors:

   Find:    ;O\.     (semicolon + O + decimal)
   Replace: ;0.      (semicolon + 0 + decimal)
   
   Find:    ;l(\d)   (semicolon + l + digit)
   Replace: ;1$1     (semicolon + 1 + digit)
   

3. Manually fix unique errors

4. Save with UTF-8 encoding

---

Step 6: Save Corrected CSV

Important: Save to the correct output location!

# Correct location format:
OCR/output/table_1/YYYY/ns-XX-YYYY.csv
# OR
OCR/output/table_2/YYYY/ns-XX-YYYY.csv

Example for ns_11_1994.pdf (Table 1):

Source:      OCR/review/1994/ns_11_1994/ocr_output.csv
Destination: OCR/output/table_1/1994/ns-11-1994.csv

Note the filename change: ns_11_1994 → ns-11-1994 (underscore → hyphen)

---

Step 7: Delete Review Folder

After saving corrected CSV:

# Delete review folder for this file
rm -rf OCR/review/1994/ns_11_1994/

# Or manually delete in file explorer

Why delete? Signals that review is complete and prevents re-flagging

---

Step 8: Verify Corrections

Run verification to ensure corrections are valid:

# Option 1: Verify specific file
python OCR/ocr_processors/validator.py OCR/output/table_1/1994/ns-11-1994.csv

# Option 2: Verify all outputs
python scripts/run_ocr_pipeline.py --verify

Expected output:

✓ Validation passed

Confidence Scores:
  OCR: 78.0%
  Combined: 88.0%  # Improved after manual correction
  Status: PASS

---

Quality Control Checklist

Before marking a file as reviewed, verify:

• All sector names present:

  • agropecuario / agriculture and livestock
  • pesca / fishing
  • mineria e hidrocarburos / mining and fuel
  • manufactura / manufacturing
  • electricidad y agua / electricity and water
  • construccion / construction
  • comercio / commerce
  • otros servicios / other services
  • pbi / gdp

• Numeric columns:

  • All values are numbers (no letters)
  • Decimal points present where needed
  • Negative signs correct
  • No unrealistic outliers (check: -100% to +500% range)

• Column structure:

  • First column: Spanish sector names (semicolon)
  • Second column: English sector names (semicolon)
  • Remaining columns: Numeric data (semicolon-separated)

• File naming:

  • Saved to correct output folder (table_1 or table_2)
  • Filename uses hyphens: ns-XX-YYYY.csv

---

Tips for Efficient Review

1. Prioritize by Confidence

Review lowest-confidence files first:

# Sort needs_review.txt by confidence
sort -t: -k2 -n OCR/needs_review.txt

2. Batch Similar Years

Process files from the same year together to spot patterns:

# Review all 1994 files at once
ls OCR/review/1994/

3. Use Excel Formulas for Validation

Create a validation column in Excel:

# Flag non-numeric values in column C
=IF(ISNUMBER(C2), "OK", "ERROR")

# Flag unrealistic values
=IF(OR(C2>500, C2<-100), "OUTLIER", "OK")

4. Keep Original PDF Open

Don't close the PDF - you may need to check multiple times

5. Document Patterns

If you notice systematic OCR errors (e.g., all 'O' should be '0'), note them for potential pipeline improvements

---

Troubleshooting

Problem: Excel Changes Formatting

Symptom: Dates become dates (1994_ene → 1/1994), decimals change

Solution:

1. Open Excel
2. File → Open (don't double-click CSV)
3. Select "Text Import Wizard"
4. Choose "Delimited" → Semicolon
5. Set all columns as "Text" format
6. Complete import
7. Save as CSV (Semicolon delimited)

Problem: Can't Read PDF (Too Blurry)

Solution:

1. Check preprocessed.png in review folder
2. If preprocessing improved quality, reference that
3. If still unclear, note in issues and flag for re-scan
4. Make best-effort correction based on context

Problem: Entire Row Missing

Solution:

1. Check if sector is in PDF
2. If yes, add manually:

   pesca;fishing;-29.8;-26.4;...
   

3. Match column structure to other rows

Problem: Extra Columns or Rows

Solution:

1. Check against PDF to determine correct structure
2. Remove spurious columns/rows
3. Ensure row count matches expected sectors (8 + GDP = 9 rows)

---

Examples

Example 1: Complete Review of ns_11_1994.pdf

Initial issues.txt:

Issues Detected (3):
  1. Column 'sectores_economicos' has wrong name: 'sectores economicos'
  2. Missing Spanish sector: construccion
  3. Unrealistic values > 500% detected

Review process:

1. Open files:

   • original.pdf shows clear table
   • ocr_output.csv has errors

2. Identified errors:

   • Header: sectores economicos (space) → sectores_economicos (underscore)
   • Row 6 missing: construccion / construction
   • manufactura row: 5O2.3 → 50.2 (O→0, extra digit?)

3. Corrections made:

   # Fixed header
   sectores_economicos;economic_sectors;...
   
   # Added missing row
   construccion;construction;27.7;12.8;28.8;...
   
   # Fixed unrealistic value
   manufactura;manufacturing;-3.5;-2.3;50.2;...  # was 5O2.3
   

4. Saved to: OCR/output/table_1/1994/ns-11-1994.csv

5. Verified: ✓ All checks passed

6. Deleted: OCR/review/1994/ns_11_1994/

---

Example 2: Systematic Digit Error

Problem: All 'O' (letter O) should be '0' (zero) in one column

Solution using Find & Replace in Excel:

1. Select the problematic column
2. Ctrl+H (Find & Replace)
3. Find: O
4. Replace: 0
5. Options: Match case ✓
6. Replace All

Result: All 47 instances corrected instantly

---

Recording Your Work

Create a Review Log (Optional)

Keep track of corrections for documentation:

review_log.txt:

2026-01-07:
- ns_11_1994.pdf: Fixed O→0 errors (5 instances), added missing construccion row
- ns_15_1994.pdf: Corrected transposed digits in mining sector (96.8 not 69.8)
- ns_22_1994.pdf: Fixed missing negative signs in pesca sector

Total reviewed: 3 files
Time: ~45 minutes

---

Summary

Efficient manual review workflow:

1. ✓ Check needs_review.txt for flagged files
2. ✓ Open review package (PDF + CSV + issues)
3. ✓ Compare and identify errors systematically
4. ✓ Correct using Excel or text editor
5. ✓ Save to correct output location with proper filename
6. ✓ Delete review folder
7. ✓ Verify corrections
8. ✓ Move to next file

Expected time: 15-30 minutes per file (depending on error severity)

Target: ~5-10% of 228 PDFs = 12-23 files needing review = 3-6 hours total

---

Need Help?

• Technical issues: Check OCR/README.md troubleshooting section
• Validation errors: Review error messages in issues.txt
• Unclear scans: Reference preprocessed.png for enhanced version
• Questions: Contact jj.cruza@up.edu.pe

---

Remember: Manual review is a normal and expected part of OCR workflows. Your corrections ensure the highest quality dataset for publication!