Named entity recognition is a natural language processing task that locates spans of text and classifies each into a predefined category such as a party, a date, or a money amount. In commercial real estate, NER pulls tenant names, commencement dates, and rent figures out of lease and offering memorandum text into labeled fields.
How Named Entity Recognition Works
Named entity recognition works by scanning text token by token, deciding where each entity begins and ends, and assigning each span a category label. Classic NER covers persons, organizations, and locations; a CRE model is trained on domain labels like tenant, landlord, commencement date, base rent, CAM charge, and renewal option so the output maps to abstraction fields.
The pipeline usually runs after optical character recognition converts a scanned lease to text. NER then reads that text and emits structured spans. Modern systems use transformer models fine-tuned on annotated documents, or large language models prompted to extract and label entities, replacing older rule-based patterns that broke on lease-language variation.
Output is scored against a human-annotated answer key using precision, recall, and F1. On the standard CoNLL-2003 benchmark, deep-learning NER passed 91 F1 by 2016, and recent models report roughly 95.6% F1, a level Primer AI describes as on par with human labeling. Domain documents like leases are harder than clean benchmark text.
Why Named Entity Recognition Matters
Named entity recognition matters because it is the step that turns a lease from an unsearchable PDF into queryable data. Without reliable entity extraction, a portfolio's rent steps, renewal windows, and co-tenancy triggers stay locked in prose, and the transcription errors that create liability, a missed renewal window or a misapplied rent step, go undetected until they cost money.
The efficiency case is documented. Industry reports on AI lease abstraction describe per-lease review time falling 85%, from about 2 hours to 17 minutes, while accuracy stays above 95%. In a document context the classification detail is decisive: a model that reads "$4,500" but cannot tell base rent from a CAM charge produces a field that still needs human review, so entity type accuracy, not just span detection, is what governs usable output.
Example
Consider one clause from a commercial lease and the labeled spans an NER model returns. The clause: "Tenant Acme Logistics LLC shall pay Base Rent of $28,750 per month commencing March 1, 2026." The model detects four spans and assigns a type to each.
Text span | Entity type | Abstraction field |
Acme Logistics LLC | Organization / Tenant | Tenant name |
$28,750 per month | Money / Base rent | Monthly base rent |
March 1, 2026 | Date / Commencement | Commencement date |
per month | Duration / Frequency | Payment frequency |
Suppose a test set contains 100 true entities, the model returns 96 spans, 92 of them correct. Precision is 92 / 96 = 0.958. Recall is 92 / 100 = 0.92. F1 is the harmonic mean, 2 x (0.958 x 0.92) / (0.958 + 0.92) = 0.939, or about 93.9%.
Variations and Edge Cases
Variation | Description |
Rule-based NER | Regular-expression and dictionary patterns; brittle on lease-language variation |
Transformer NER | Fine-tuned models like BERT trained on labeled documents; current standard |
LLM extraction | Large language models prompted to extract and label; flexible, needs validation |
Nested NER | Handles an entity inside another, such as a date range inside a rent-escalation clause |
Edge cases in CRE documents break weaker models. A single "$4,500" may be base rent, additional rent, or a security deposit depending on surrounding clause context. Defined terms like "the Premises" carry meaning set pages earlier, and amendments override the base lease, so span detection without document-level context yields fields that still require human review.
Named Entity Recognition vs Intelligent Document Processing
Named entity recognition is often confused with intelligent document processing. NER is one component: it finds and classifies entity spans within already-extracted text. Intelligent document processing is the full pipeline, ingestion, OCR, classification, NER, validation, and export to a system of record.
The distinction matters when scoping accuracy. NER accuracy is measured per entity span with precision, recall, and F1. IDP accuracy is measured at the document or field level after validation and human review, so a strong NER model is necessary but not sufficient for a trustworthy abstraction output.
Frequently Asked Questions
What entities does NER extract from a commercial lease?CRE named entity recognition extracts tenant and landlord names, commencement and expiration dates, base rent and escalation amounts, CAM and other charges, renewal options, and premises identifiers, mapping each span to an abstraction field so the lease becomes queryable data.
How accurate is named entity recognition?On the clean CoNLL-2003 benchmark, recent NER models report about 95.6% F1, which Primer AI describes as near human-level. Real lease documents are harder because of defined terms, amendments, and context-dependent amounts, so production accuracy is validated per field.
Is named entity recognition the same as OCR?No. Optical character recognition converts a scanned image to machine-readable text. Named entity recognition reads that text and identifies which spans are entities and what type each is. NER runs after OCR in a document extraction pipeline.
Related Terms
Intelligent Document Processing
Optical Character Recognition
Lease Abstraction