Recognize merged PDF table headers before extraction

Where the extra work starts

Check how the columns are represented before relying on the spreadsheet so the review starts with the correct meaning for each field.

For supplier reports and technical tables, start with ParseMill’s synthetic material-register example. Compare the source references and output with the register you need, then discuss a document pilot with an agreed scope and fee. The self-service trial covers invoice PDFs.

ParseMill's invoice workflow recognizes supported single-word and multiword invoice headings, but does not promise general interpretation of merged or multi-level headers. For a complex report table, use the hierarchy map to assess a separate reviewed method. An invoice trial establishes fit only for the invoice layouts you actually check.

A practical way to get it right

Inspect the heading region separately from the data. Record each top-level group, subordinate label and unit. Trace one data cell upward to every heading needed to describe it. If the meaning depends on two levels, preserve both in a manual schema map.

Create distinct field roles for review, such as measured result and permitted limit, without pretending those combined labels were printed as one source heading. Keep the original label and position beside each proposed field. Ask the reviewer to confirm the mapping before any transformation.

Choose a manual or otherwise suitable reviewed method when the parser cannot represent the hierarchy. If testing another tool, verify the relationship for every subcolumn and retain the transformation rules. Do not score a rectangular output as correct merely because it has the expected number of columns.

W3C: PDF table markup distinguishes header and data cells and supports column spans. That structural relationship explains why a merged heading must be mapped to its actual subordinate columns.

What this looks like on a real task

Synthetic example. The people, references and figures below are invented to demonstrate the method.

The repeated Unit labels belong to different groups. A flat list of Measurement and Specification would omit the subcolumn meanings, while a flat list of Unit twice remains ambiguous. Copy the hierarchy, value and field-role columns. The schema labels are your reviewed mapping, not automatically extracted semantic knowledge.

Catch the cases that create rework

Do not fill a merged heading across every nearby column without verifying its span. Borders, spacing and notes may indicate a different group boundary. A table can also change hierarchy on a continuation page, so inspect each relevant heading region.

If the mapping cannot be resolved from the source, keep the fields unresolved and request clarification. Do not infer a specification limit merely from the larger of two numbers.

Microsoft: Promote or demote column headers notes that automatic header identification can be wrong. A promoted row of labels still needs semantic review against the complete source hierarchy.

Decide whether the change is worth keeping

The accepted map should let another reviewer identify every value's complete heading and unit context. Record unsupported automatic-mapping cases explicitly and preserve the source page. Use the map to select the next method, and measure any manual reconstruction effort before assuming that extraction offers a practical benefit for this layout.