OCR evidence checks: protect signs, decimals, footnotes
Treat OCR output as a proposed transcription, especially when a sign, decimal, footnote, or column position can change a research conclusion. Verify consequential cells against the original page image before using them in calculations. Correct prose around a table does not establish that its numbers were extracted correctly.
Tesseract’s quality guidance discusses image processing, segmentation, and table-recognition limitations. Those considerations justify checking document quality and layout. They do not provide an accuracy guarantee for a particular report, language, model, or scan.
Preserve the original visual evidence
Save the source document identity and page number. Keep the rendered page available while inspecting extracted text. If the page contains merged headers, rotated labels, shaded footnotes, or narrow columns, record those layout features before deciding how the output should be interpreted.
Separate transcription from normalization. First record what the page shows, including parentheses, dashes, symbols, and units. Then create a normalized numeric field under an explicit rule. A dash might mean zero, not applicable, or not disclosed depending on the source; do not choose one meaning just to make a spreadsheet calculate.
| Cell record | Example of required detail |
|---|---|
| Visual location | Page, table, row, column |
| Source appearance | Characters as printed |
| Proposed extraction | OCR text before correction |
| Interpretation | Sign, unit, period, and missing status |
| Verified value | Corrected result with reviewer note |
| Downstream use | Formula or claim using the cell |
Retain corrections as a log. Replacing the extracted file without recording the change makes later discrepancies harder to investigate.
Hypothetical sign error
Assume a fictional report prints operating cash flow as (12.5) in a table denominated in millions. The extraction returns 12.5. If the analyst treats it as positive, a cash-use discussion can become a cash-generation discussion. The original page and table convention are necessary to resolve the sign.
Now suppose a neighboring column contains 1.25, but the extraction assigns that value to the wrong year because a merged header was flattened. Every digit may be correctly recognized while the meaning is still wrong. This is a layout error rather than a character error.
Both examples are hypothetical. Their lesson is to verify the complete cell context, not merely compare digit strings. A valid value needs the correct sign, unit, row, column, and relevant footnote.
Choose checks by consequence
Inspect all cells supporting the main conclusion, all unusual signs, and all inputs used in a material ratio. Add reconciliation checks such as subtotals where the source provides them, while remembering that a matching total can conceal offsetting errors. Do not rely on one aggregate equation as proof of every input.
For a long table, review a distributed sample of ordinary cells and every high-impact cell. State which portions received direct visual checks. A sampled review should not be described as a full cell-by-cell verification. If errors cluster around one layout feature, reopen similar cells throughout the document.
When image quality is poor, try obtaining an official text-based version or a better scan. Improving contrast or segmentation can help extraction, but the corrected image remains a transcription aid rather than a new source. Preserve the original document so the evidence chain stays clear.
Recalculate after corrections
Once a cell is corrected, identify every dependent formula and sentence. Rerun the calculation and revise the narrative if needed. A research memo can remain wrong after the table is fixed if its summary still uses the old sign or period.
The workflow is unsuitable for a precise conclusion when the original image itself is illegible. Mark the cell unresolved and seek another source; do not let the model invent the missing digit. Common mistakes include deleting parentheses, assuming a dash is zero, losing units, trusting a clean-looking CSV, and correcting numbers without updating dependent claims.
FAQ
Is OCR confidence enough to accept a financial number?
No. A confidence value does not establish the correct row, period, sign, or footnote. Consequential cells still require appropriate source checks.
Can a table total replace visual review?
It is a useful consistency check, but not a complete substitute. Offset errors and incorrect column assignments can survive a matching total.