What is table extraction?
Table extraction is the recovery of tabular structure from a document: the rows and the columns as well as the association between them. It is substantially harder than extracting isolated fields, because a correct cell value placed in the wrong row produces a plausible and wrong record. Accuracy therefore has to be measured at the level of the row rather than the cell.
Also called: Table recognition
Why a correctly read cell in the wrong row is the worst outcome
An isolated field that extracts wrongly is usually obvious, since a date in a name field announces itself. A table that extracts wrongly produces numbers that are individually correct and collectively meaningless, with nothing on the face of the record to indicate it. In an accreditation annexure or a schedule of equipment, that record then supports a decision nobody will question until an inspection. Table extraction is therefore the step where cell-level accuracy figures mislead most and row-level measurement is mandatory.
- A cell read correctly into the wrong row yields a record that looks valid.
- Merged and spanning cells have no single correct flat representation.
- Tables continuing across a page break are routinely resolved as two tables.
- Cell-level accuracy figures flatter a pipeline that cannot preserve rows.
How table extraction works
- Table regions are located during layout analysis before any cell is read.
- Row and column boundaries are inferred from ruling lines or from alignment.
- Header rows are identified and mapped to the fields of the destination schema.
- Each cell is recognised and associated with its row and its column.
- Tables spanning pages are joined by matching headers and continuing row sequence.
- Control totals printed in the document are used to verify the extracted result.
How table extraction is evaluated
Table extraction is judged at the level of the row, since that is the unit a downstream record consumes. Measure the following on real annexures:
- exact row match rate against a hand-marked reference
- cell accuracy reported separately from row accuracy
- header mapping accuracy across form variants
- handling of merged cells and of tables continuing across pages
- agreement between extracted totals and totals printed in the document
Limitations, risks and common failure modes
- Cell accuracy is quoted while row integrity is never measured at all.
- Tables without ruling lines are segmented on alignment that varies by printer.
- An empty cell shifts every subsequent value one column to the left.
- A continuation table acquires its header row as a data row.
- Footnotes inside a table are extracted as additional data rows.