Table extraction is the process of recovering the rows, columns, and cells of a table from a document into a structured grid a system can query. In commercial real estate it turns a rent roll or an offering memorandum cash flow into aligned records, preserving not just the text but the position of each value in its row and column.
How Does Table Extraction Work?
Table extraction works in two stages: table detection, which locates the table on the page, and table structure recognition, which maps every cell to its row and column. Detection draws a bounding box. Structure recognition rebuilds the grid so that a base rent value stays tied to its unit and its lease dates.
The hard part is structure, not text. On the PubTables-1M benchmark, Microsoft's Table Transformer (TATR) model scores are reported with GriTS, a grid table similarity metric that measures cell correctness by topology, location, and content. A 2603 PDF-parser study found parsers reaching 74% text accuracy but only 35% structure preservation, a gap that explains why extracted rent rolls often read correctly yet misalign a column. Merged cells, spanning headers, and borderless tables cause most of that structural loss.
Stage | Function |
Table detection | Locates the table region on the page |
Structure recognition | Assigns each cell to a row and column |
Cell text recovery | Reads the value inside each cell |
Header association | Links data cells to their column headers |
Why Table Extraction Matters
Table extraction matters because the numbers that drive underwriting live in tables, and a misplaced cell corrupts a model silently. A rent roll with 120 units feeds occupancy, in-place rent, and lease expiration schedules. If extraction shifts one column, every derived total is wrong with no error thrown, which is why structure accuracy is the metric that matters.
The accuracy signal is measurable. PubTables-1M aggregates nearly one million tables to train and score models, and after correcting benchmark annotation errors, TATR's exact-match accuracy on the ICDAR-2013 benchmark rose to 81% when trained on combined data, per the Table Transformer research. General PDF-to-table tools cite 95 to 99% accuracy on clean, ruled tables, but that figure falls on scanned or borderless layouts common in older CRE documents.
Example
Table extraction is clearest on a short rent roll excerpt. An analyst runs a four-unit section of a rent roll through an extraction model, which returns each cell tied to its row and column header.
Unit | Tenant | SF | Annual Rent |
101 | Harbor Coffee | 1,200 | 42,000 |
102 | Vacant | 900 | 0 |
103 | Ridgeline PT | 2,400 | 78,000 |
104 | Delta Nails | 1,100 | 39,600 |
From the extracted grid a system totals occupied square footage as 1,200 plus 2,400 plus 1,100, which equals 4,700 SF, against 5,600 SF total, for 83.9% physical occupancy. In-place annual rent sums to 42,000 plus 78,000 plus 39,600, which equals 159,600. Blended rent on occupied space is 159,600 divided by 4,700, which equals $33.96 per square foot. A single shifted column would have folded Unit 102's zero into an occupied row and broken all three figures.
Variations and Edge Cases
Table extraction behaves differently as table structure grows more irregular. The variants below trade simple layouts against the merged and nested structures common in financial schedules.
Variant | Treatment |
Ruled tables | Visible gridlines guide detection; highest accuracy |
Borderless tables | Structure inferred from whitespace; error-prone |
Spanning headers | One header covers several columns; needs association logic |
Merged cells | A value covers multiple rows or columns; a common failure point |
Multi-page tables | A rent roll continues across pages; needs header carry-forward |
Table Extraction vs Optical Character Recognition
Table extraction is often confused with OCR, but they answer different questions. OCR, optical character recognition, converts an image of a page into raw text with no sense of rows or columns. Table extraction takes that text, or a digital table, and rebuilds the grid so each value knows its row and column position.
OCR answers "what characters are on this page." Table extraction answers "which row and column does this number belong to." OCR run alone on a rent roll returns a stream of numbers with no way to tell a unit number from its rent, which is why table extraction is the step that makes a table queryable.
Frequently Asked Questions
What is table extraction?Table extraction is the process of recovering the rows, columns, and cells of a table from a document into a structured grid a system can query. In commercial real estate it turns a rent roll or offering memorandum cash flow into aligned records where each value keeps its row and column position.
How accurate is table extraction?General PDF-to-table tools cite 95 to 99% accuracy on clean, ruled tables. Accuracy falls on borderless, merged, or scanned layouts, and a 2603 study found parsers preserving only 35% of table structure while reading 74% of text correctly, so structure is the weaker metric.
Is table extraction the same as OCR?No. OCR converts a page image into raw text with no sense of rows or columns. Table extraction rebuilds the grid so each value keeps its row and column position. OCR is often an input step to table extraction, not a replacement for it.
Related Terms
Optical Character Recognition
Structured Data Extraction
Field Extraction
Data Normalization
Rent Roll