Home/Building Blocks/Named Entity Recognition
Text→Structured Data

Named Entity Recognition

Extract named entities (people, organizations, locations, dates) from text. Key for information extraction and knowledge graphs.

How Named Entity Recognition Works

A technical deep-dive into Named Entity Recognition. From token classification to zero-shot approaches with GLiNER and LLMs.

1. What is NER2. BIO Tagging3. Architecture4. Models5. Benchmarks6. Code
1

What is Named Entity Recognition?

NER identifies and classifies named entities in text into predefined categories. Think of it as structured information extraction from unstructured text.

NER in Action

Barack Obama was the 444th Presidentof the UUnited Statesfrom 22009to 22017

PERSONORDINALGPEDATE

Common Entity Types

PERSON
Names of people
"Albert Einstein"
ORG
Organizations, companies, agencies
"OpenAI"
LOC
Locations, countries, cities
"New York"
DATE
Dates and time expressions
"January 2024"
MONEY
Monetary values
"$1.5 billion"
GPE
Geopolitical entities
"United States"
PRODUCT
Products, objects, vehicles
"iPhone 15"
EVENT
Named events
"World War II"
Information Extraction

Extract structured data from documents, emails, and news articles. Build knowledge graphs automatically.

Document Understanding

Identify key entities in contracts, legal documents, and medical records for downstream processing.

Search Enhancement

Power entity-based search, auto-linking, and semantic navigation in content management systems.

2

BIO/IOB Tagging Scheme

NER is formulated as token classification. Each token gets a tag. BIO tagging handles multi-word entities by marking the Beginning,Inside, and Outside of entities.

BIO Tags Explained

B
Beginning
First token of an entity
I
Inside
Continuation of entity
O
Outside
Not an entity

Example: "Elon Musk founded SpaceX in California"

TokenElonMuskfoundedSpaceXinCalifornia
BIO TagB-PERSONI-PERSONOB-ORGOB-LOC

Why Not Just Label Tokens?

Problem: Adjacent Entities

"John Smith met Jane Doe" - without BIO, we cannot tell where one PERSON ends and another begins.

John(PER) Smith(PER) met(O) Jane(PER) Doe(PER) // ambiguous!
Solution: BIO Tagging

B- marks entity boundaries, I- continues them. Now we know there are two distinct people.

John(B-PER) Smith(I-PER) met(O) Jane(B-PER) Doe(I-PER) // clear!
Variants: BIOES (adds Single and End tags), BILOU (similar with Last/Unit). BIO is most common due to simplicity. More tags = slightly better accuracy but harder to learn.
3

Token Classification Architecture

Modern NER uses a BERT-style encoder + linear classification head. Each token gets its own prediction based on contextual embeddings.

BERT-NER Architecture

Input Text
Raw text to analyze
"Elon Musk founded SpaceX
->
Tokenization
Split into subwords
["Elon", "Musk", "founded
->
BERT Encoder
Contextual embeddings
[768-dim] x 5 tokens
->
Linear Layer
Per-token classification
[num_labels] x 5 tokens
->
Softmax
Tag probabilities
B-PER, I-PER, O, B-ORG, I
Text Classification (CLS token)

Uses only [CLS] embedding for whole-document label.

[CLS] token1 token2 ... [SEP]
  |
  v
class_label
Token Classification (NER)

Uses ALL token embeddings, one label per token.

[CLS] token1 token2 token3 [SEP]
        |     |     |
        v     v     v
      B-PER  I-PER   O

Subword Alignment Challenge

BERT tokenizes words into subwords. "SpaceX" becomes ["Space", "##X"]. But we only have one label for "SpaceX". Solutions:

First Subword Only
Label first subword, ignore rest (-100 loss mask)
Copy to All
Copy label to all subwords of the word
Post-Process
Aggregate predictions during inference
4

NER Models Comparison

From production-ready spaCy to zero-shot GLiNER. Choose based on your constraints.

ModelTypeSpeedAccuracyNotes
spaCy (en_core_web_trf)Pre-trained~10K tokens/secF1: 89%Production-ready, 18 entity types, GPU optional
BERT-NERFine-tuned~5K tokens/secF1: 92%Fine-tune on CoNLL-2003, requires training
RoBERTa-NERFine-tuned~4K tokens/secF1: 93%Better than BERT, larger model
GLiNERZero-Shot~2K tokens/secF1: 85-90%Any entity type without training
GPT-4 / ClaudeLLM~100 tokens/secF1: 80-95%Most flexible, expensive at scale
Use spaCy when:
  • - You need production-ready, tested solution
  • - Standard entity types (PER, ORG, LOC, etc.)
  • - CPU inference is acceptable
Use Fine-tuned BERT when:
  • - You have labeled training data
  • - Domain-specific entity types
  • - Maximum accuracy is critical
Use GLiNER when:
  • - No labeled data available
  • - Entity types change frequently
  • - Quick prototyping / exploration
Use LLMs when:
  • - Complex, nuanced entity definitions
  • - Need explanations with extractions
  • - Low volume, high-value documents
*
GLiNER: Zero-Shot NER

GLiNER can extract any entity type you define without training. Just describe what you want: "pharmaceutical_company", "medical_condition", "dosage". It uses a generative approach to match entity spans to your labels. Great for prototyping before investing in fine-tuning.

5

Benchmarks and Evaluation

Standard datasets for evaluating NER systems. F1 score is the primary metric.

DatasetLanguageEntitiesSizeSOTA
CoNLL-2003English4 types22K sent.94.6% (LUKE)
OntoNotes 5.0English18 types77K sent.92.4% (LUKE)
WNUT-17English6 types5K sent.60.4% (emerging entities)
MultiNERD10 languages15 types164K sent.varies by lang

Understanding F1 Score for NER

Precision
Of entities predicted, how many are correct?
TP / (TP + FP)
Recall
Of actual entities, how many did we find?
TP / (TP + FN)
F1 Score
Harmonic mean of precision and recall
2 * P * R / (P + R)

NER evaluation is strict: entity must match both the span (exact boundaries) AND the type. Partial matches are typically counted as wrong.

CoNLL-2003: The Standard Benchmark

Reuters news articles with 4 entity types: PER, ORG, LOC, MISC. Most papers report results here. Current SOTA is ~94.6% F1.

PER: Person namesORG: OrganizationsLOC: LocationsMISC: Miscellaneous
6

Code Examples

Get started with NER in Python. From simple spaCy to zero-shot GLiNER.

spaCypip install spacy && python -m spacy download en_core_web_trf
Recommended Start
import spacy

# Load transformer-based model
nlp = spacy.load("en_core_web_trf")

text = "Apple Inc. was founded by Steve Jobs in Cupertino, California in 1976."

# Process text
doc = nlp(text)

# Extract entities
for ent in doc.ents:
    print(f"{ent.text:20} {ent.label_:10} {ent.start_char}-{ent.end_char}")

# Output:
# Apple Inc.           ORG        0-10
# Steve Jobs           PERSON     27-37
# Cupertino            GPE        41-50
# California           GPE        52-62
# 1976                 DATE       66-70

Quick Reference

For Production
  • - spaCy (en_core_web_trf)
  • - Fine-tuned BERT/RoBERTa
  • - F1 > 90% on standard types
For Prototyping
  • - GLiNER (zero-shot)
  • - LLM with structured output
  • - Define any entity type
Key Metrics
  • - F1 Score (precision + recall)
  • - Exact span matching
  • - Per-entity-type breakdown

Use Cases

  • ✓Information extraction
  • ✓Knowledge graph building
  • ✓Resume parsing
  • ✓News analysis
  • ✓Legal document processing

Architectural Patterns

Sequence Labeling

Tag each token with entity type (BIO scheme).

Pros:
  • +Fast
  • +Well-understood
  • +Good for standard entities
Cons:
  • -Fixed entity types
  • -Needs labeled data

Span Extraction

Predict start/end positions of entity spans.

Pros:
  • +Handles nested entities
  • +More flexible
Cons:
  • -Slower
  • -More complex training

LLM-Based NER

Use LLMs to extract entities via prompting.

Pros:
  • +Zero-shot for new entity types
  • +Context-aware
Cons:
  • -Expensive
  • -Inconsistent formats

Implementations

API Services

AWS Comprehend

AWS
API

Managed NER. Custom entity training available.

Open Source

spaCy

MIT
Open Source

Production NLP. Fast, accurate, many languages.

GitHub

GLiNER

Apache 2.0
Open Source

Zero-shot NER. Generalize to any entity type.

HuggingFace

BERT-NER

Apache 2.0
Open Source

Classic BERT-based NER. Good accuracy.

HuggingFace

Flair

MIT
Open Source

State-of-the-art NER. Stacked embeddings.

GitHub

Benchmarks

CoNLL-2003 →OntoNotes 5.0 →

Quick Facts

Input
Text
Output
Structured Data
Implementations
4 open source, 1 API
Patterns
3 approaches

Related Blocks

Text Classification
Text → Structured Data
Document Extraction
Document → Structured Data

Have benchmark data?

Help us track the state of the art for named entity recognition.

Submit Results