Handle a repeated table caption without treating it as data

Give each non-data row a reason

Identify it consistently so removing the noise does not also remove a real record with similar wording.

For multipage invoices, use ParseMill to review extracted invoice data beside the source and export separate summary and line-item sheets. Include a page break in your trial and check that every required line made it through. Other report types need their own fit assessment.

Keep an exclusion record containing the raw text, physical page, source position and reviewed role. This makes removal auditable and gives you a way to restore an incorrectly classified row.

Annotate an invented example

The final row is intentionally awkward. Its first field contains Results, but its aligned fields describe a real observation in the fixture. A single-keyword filter would remove useful data. The continuation caption also differs slightly from the original, so exact text equality alone would miss it.

Microsoft: Table.SelectRows retains rows satisfying a supplied condition. Make the condition depend on a reviewed role or complete signature rather than an unexamined keyword.

Apply a conservative removal process

1. Retain raw extracted lines with page and position information. Classify candidate captions by inspecting the source above and below each occurrence.
2. Record the full caption family, including continuation variants. Distinguish the caption from the column headings beneath it.
3. Remove only records explicitly classified as captions or headings from the data output. Keep the excluded records in a separate audit table.
4. Compare the retained rows with the source at each removed boundary. Check that the first data row after a caption was not swallowed into the exclusion.

Microsoft: Table.RowCount counts table rows. Reconcile raw, excluded and retained counts using the same extraction unit, so a multiline caption is not accidentally counted as one row in one place and two elsewhere.

Watch for captions that carry context

A caption may specify a test condition, date range or material group. Excluding it from flat records does not mean discarding its meaning. Store applicable context separately with the source range it governs. If that range is unclear, keep the context unassigned pending review.

A line that resembles a caption can also be a section summary or a substantive remark. Classify it from the source structure rather than appearance alone. Explain unusual exclusions so another reviewer can reproduce the decision.

Verify the expected exclusions

For the five listed lines, the two captions and one heading are excluded; both data rows remain. Review the complete fixture before calculating a real table count, because the illustration lists only selected lines. The outcome should be fewer non-data rows without a loss of observations.

Choose an invoice with a repeated heading and a genuine line description that uses the same wording. Check both in the ParseMill workbook. This makes false deletion a practical buying test instead of accepting a tidy-looking export at face value.