Field extraction is the identification of specific, named values inside a document and their assignment to defined fields. In commercial real estate, it pulls a lease's base rent, commencement date, renewal option, and square footage into labeled data points rather than free text. It is the targeted layer of document extraction, where individual values are found and scored.
How Does Field Extraction Work?
Field extraction works by pairing each value in a document with the field it belongs to, a task known as key-value extraction. The system locates a label or contextual cue, then binds the nearby value to it, so "Base Rent: $32.50 PSF" becomes the field base_rent with value 32.50. Each pairing gets a 0 to 100% confidence score, per Nanonets.
Underneath, this draws on named entity recognition (NER), which tags spans of text as entities of known types such as dates, amounts, or parties. Template-based OCR and rule-based NER can be slow, brittle, or inaccurate on varied layouts, per Nanonets, which is why model-based extraction generalizes better across the nonstandard leases CRE routinely handles. Each extracted field carries its confidence score forward so low-confidence values can be routed to a reviewer.
Term | Meaning |
Field | A defined slot in the output, such as base_rent or expiration_date |
Value | The specific text bound to that field, such as $32.50 or 2029-12-31 |
Key-value pair | The bound field and value together |
Confidence score | A 0 to 100% certainty estimate the model assigns per field |
Coverage | Fields successfully extracted versus fields expected |
Why Field Extraction Matters
Field extraction matters because underwriting depends on specific values, not on a document being readable. A lease can be perfectly digitized and still useless until base rent, escalations, and expiration are pulled into structured fields. Field-level accuracy, coverage, confidence, and human touch rate are the core metrics of document extraction quality, per Astera, and each is measured field by field.
Accuracy is uneven across fields, which is the operator's real problem. A clearly labeled square-footage value extracts near-perfectly; a percentage-rent breakpoint buried in a narrative clause does not. Because each field carries a confidence score, a pipeline can flag exactly the values it is unsure about rather than forcing a full re-read. That per-field signal is what lets a reviewer spend minutes, not hours, on a lease.
Example
Field extraction is easiest to see when you count fields, not documents. An analyst extracts 50 defined fields from a single office lease. The model returns a value and a confidence score for each, and the pipeline auto-accepts anything above a 90% threshold.
Metric | Value |
Fields defined | 50 |
Fields extracted (coverage) | 48 |
Coverage rate | 96% |
Fields above 90% confidence | 44 |
Fields flagged for review | 4 |
Review time at 2 min per field | 8 minutes |
The model covers 48 of 50 defined fields, a 96% coverage rate, and leaves 2 fields blank where no value was found. Of the 48 extracted, 44 clear the 90% confidence threshold and pass automatically. The remaining 4 low-confidence fields, plus the 2 uncovered ones, go to a reviewer, who resolves all 6 in roughly 12 minutes. Coverage and confidence are the two dials here: coverage measures how much was found, confidence measures how much can be trusted without a human. The threshold and per-field review time are illustrative inputs; the calculation follows from them.
Variations and Edge Cases
Field extraction changes with the field's structure and where it sits on the page. A value in a labeled table cell is straightforward; a value implied across two clauses is not. The variants below shift accuracy and the rate at which fields get flagged.
Variant | Treatment |
Structured field | Labeled table cell or form slot; high confidence, low review |
Semi-structured field | Consistent label but variable position; moderate confidence |
Unstructured field | Value implied in narrative prose; low confidence, high review |
Derived field | Computed from other fields, such as annual rent from PSF and area |
Multi-value field | One field with several entries, such as staged rent escalations |
Field Extraction vs Document Extraction
Field extraction is often used interchangeably with document extraction, but one is a subset of the other. Document extraction is the whole pipeline that turns a file into structured data, including OCR, layout analysis, and validation. Field extraction is the specific step inside it that identifies individual named values and binds them to defined fields.
Document extraction is the process; field extraction is the unit of work and the unit of measurement. When a vendor quotes accuracy, it is almost always field-level accuracy, because that is the number that decides whether the base rent in your model is right.
Frequently Asked Questions
What is field extraction in document processing?Field extraction is the step in document processing that identifies specific named values and assigns them to defined fields. In a lease it pulls values like base rent, square footage, and expiration date into labeled, structured data rather than leaving them as free text.
What is the difference between field extraction and document extraction?Document extraction is the full pipeline that converts a document into structured data, including OCR, layout analysis, and validation. Field extraction is the step within it that identifies individual named values and binds each to a defined field. Field extraction is a subset of document extraction.
Why does each extracted field get a confidence score?Each field gets a 0 to 100% confidence score because accuracy is uneven across a document. The score is an uncertainty signal that lets a pipeline auto-accept high-confidence values and route low-confidence ones to a human reviewer, per Nanonets and Iteration Layer.
Related Terms
Document Extraction
Confidence Score
Source Citation
Human-in-the-Loop
Rent Roll