A quality-inspection table reports inspected and accepted counts for batches A and B. If all four numbers are recognized but assigned to the wrong batch, character accuracy can look excellent while the conclusions are wrong. PDF table extraction must establish which value belongs to which row, column and header group—not merely export a spreadsheet. The method below uses constructed data throughout.
Separate four different recognition problems
First locate the table region without treating titles, footers or surrounding prose as data. Next recover rows, columns and cells spanning multiple rows or columns. Then identify header roles: for example, a parent batch label above inspected and accepted counts. Finally map recognized text back to the right cells while preserving source-page locations. An early error can affect many downstream fields.
Microsoft's Table Transformer project separates table detection from structure recognition, and its annotations cover rows, columns, blank cells and headers.[1] Tables are therefore two-dimensional structured objects, not strings concatenated in reading order. Benchmark performance does not guarantee correctness on enterprise scans, especially with different layouts, languages or imaging conditions.
Record region, structure, recognition and business-mapping errors separately. That distinction helps determine whether to improve scanning, change parsing, adjust a structure model or correct field definitions. More prompting cannot reliably recover header relationships already discarded by upstream parsing.
Preserve a structural intermediate representation
Keep at least the file and version, page, table identifier, cell identifier, row and column spans, raw text and source coordinates. Specify coordinate units, origin and scaling so rotation or cropping remains reversible to the source. For multi-page tables, a logical field may have multiple evidence locations rather than one forced page number.
Represent merged cells through their explicit spans before duplicating anything into flat records. Expand hierarchical headers into ordered paths such as batch A / accepted count, rather than retaining only the final repeated label. Preserve row-group labels and their scope as well. Business fields should be generated from these relationships through confirmed mapping rules.
A blank cell can mean intentionally empty, continuation of a group above, not applicable or recognition failure. Inherit labels only when both layout and business convention support it; never forward-fill numeric regions indiscriminately. A dash is not inherently zero. Preserve the original symbol and unresolved state when its meaning is uncertain.
Validation before ingesting LLM structured outputs remains necessary. This article adds an earlier obligation: establish that each field comes from the correct two-dimensional position and header path. A schema-valid object can faithfully preserve an incorrect column assignment; format validation cannot replace source checks.
Constructed example: every character is right, but batches are swapped
Suppose a table contains one inspection item, dimensions. Its columns are batch A inspected 100, batch A accepted 98, batch B inspected 80 and batch B accepted 72. Each batch header spans two columns, with inspected and accepted repeated underneath. The correct accepted proportions are 98% for A and 90% for B. These numbers illustrate an error mechanism only.
If the parser loses the parent batch headers, it retains two groups of identically named fields and may assign the second group to batch A. All four numbers remain correctly recognized, and the count relationships still look plausible, while the batch conclusions are reversed. Even overall totals of 180 inspected and 170 accepted remain unchanged. Total reconciliation alone cannot expose this mistake.
Acceptance should verify each number's composite identity: inspection item, batch, measure and unit. Trace it to the supporting cell. Reference annotations for critical fields should include both value and attribution, not merely an unordered answer set. A review interface should place the source region beside candidate fields so a reviewer can see column shifts directly.
If the upper header is cropped out, the system may recognize the four numbers but cannot invent the batch labels. Correct handling is to report missing header evidence and request the complete page or human review, not infer batch order from numerical magnitude. Withholding unsupported mappings is itself an acceptance behavior.
Establish page continuity before removing repeated headers
Stitching should consider table identifiers, continuation markers, column counts and relative positions, header paths, units, neighboring content and row identifiers. No single clue proves continuity: a document may contain several tables with the same layout, or a new page may begin a different batch or unit convention. Record the evidence for each join and stop automatic merging when it conflicts.
Only after confirming that pages belong to one logical table should the system identify and remove repeated continuation headers from the data stream. Preserve their source records in the evidence trail. When a row continues onto the next page, determine whether the text actually extends the same item. An empty first column alone is insufficient; the source may genuinely omit an item label.
Page subtotals, table totals and detail rows need distinct roles before deciding which enter downstream calculations. Treating a subtotal as detail double-counts it. Footnotes defining units, exclusions or special meanings should attach to the affected region rather than become a numeric field in the last row.
In a constructed counterexample, page one uses millimeters and page two has the same column names but uses centimeters. Joining on labels alone creates inconsistent values. Even where explicit conversion is permitted, retain the original unit, conversion rule and source. Silent normalization that discards evidence is not acceptable extraction.
Choose parsing routes according to observed error types
For digitally generated PDFs, embedded text and positions are a useful starting point. Scans usually require optical character recognition, or OCR, to obtain text candidates. Both may need layout and table-structure recognition. Embedded text need not follow visible reading order, and OCR text does not establish rows and columns. Compare complete outputs against one annotated set rather than checking only whether text is selectable.
Docling documents table-structure and cell-matching controls, including a matching adjustment that can help when multiple columns are incorrectly merged.[3] This is a diagnostic option to test, not a universal recommendation to disable a setting. Pin the actual version and configuration, vary the factor under investigation and check other layouts for regressions.
Image-based models can assist with complex headers and exceptional regions, but should return verifiable positions and relationships. A second model reading the same incorrectly linearized text does not provide independent validation merely because it agrees. Original images, visible rules, coordinates and alternative parses can reveal shared-source errors.
For cost control, process clear, regular tables first and route structural conflicts, poor images and missing critical fields to more expensive recognition or human review. Derive routing conditions from annotated error analysis rather than interpreting a model's reported confidence as a probability of safe business release.
Combine structural metrics with business-field acceptance
GriTS compares predicted and reference tables in matrix form, providing a method for evaluating table structure.[2] Such metrics help diagnose overall quality, but critical fields still need jointly correct values, header paths, units and sources. A high average can conceal one header error affecting an entire column, so it is insufficient as the sole automatic-ingestion gate.
Report at least four measures: detection of expected tables; correct attribution of critical cells; the share of tables with every agreed critical field correct; and discovered errors among automatically released outputs. State each denominator. Missed tables must not disappear from end-to-end success counts, and unreviewed outputs must not be treated as verified correct.
Stratify materials by scanned versus digital, single-level versus hierarchical headers, one-page versus multi-page, and regular versus irregular merged cells. Separate near-duplicate templates used in training or tuning from independent acceptance files. Reference annotations may permit defined structural equivalences, but specify them before evaluation rather than changing the standard afterward to improve scores.
Build a small counterexample suite. Swapping parent headers should change batch attribution. Removing a parent header should trigger review. Inserting repeated continuation headers should not create detail records. Removing a page should prevent claims of completeness. Adding a subtotal should not double-count values. These are proposed tests; no model experiment or measured score is reported here.
Deliver results that remain traceable to the original
A maintainable delivery package includes source versions, parsing configuration, the structural intermediate representation, business-mapping rules, validation states and an exception list. Human corrections should record the changed cell, supporting evidence and affected downstream fields. A model upgrade or rerun should not silently overwrite confirmed corrections.
If the target system accepts only flat records, export the final fields while retaining structure and evidence separately, linked through stable identifiers. Before scaling, establish review effort and backlog capacity. If complex tables largely require manual handling, narrow automation scope or improve inputs instead of concealing pending items inside an apparently successful spreadsheet.
The objective is not a visually convincing reconstruction. It is correct attribution and traceable support for every business fact. Joint acceptance of two-dimensional structure, field meaning and automatic-release conditions makes documents more dependable inputs to data workflows.