NER - Information Extraction

What is Named Entity Recognition?

Named Entity Recognition (NER) is the NLP task of identifying and classifying named entities in text into predefined categories such as persons, organizations, locations, dates, quantities, and other domain-specific types. NER transforms unstructured text into structured information by finding the "who, what, where, when" of text.

NER is a foundational component in information extraction pipelines, powering applications like search engines (understanding queries), question answering (finding answer entities), content recommendation (extracting topics), knowledge graph construction (populating nodes), and financial analysis (identifying companies and events in news).

Entity Categories

Standard Entity Types (CoNLL, OntoNotes)

PERSON	"Albert Einstein", "Dr. Smith", "Mary"
ORGANIZATION	"Google", "United Nations", "MIT"
LOCATION	"Paris", "Mount Everest", "the Nile"
DATE	"January 2024", "last Tuesday", "the 90s"
TIME	"3:00 PM", "noon", "midnight"
MONEY	"$50 million", "€100", "500 rupees"
PERCENT	"25%", "a third", "half"

Domain-Specific Entities

Biomedical NER	Gene, Protein, Disease, Drug, Chemical
Legal NER	Court, Judge, Statute, Case
Financial NER	Ticker, Index, Currency, Metric

NER Approaches

Rule-Based NER

- Capitalized word sequences	potential entity
- "Mr./Mrs./Dr." + Name	PERSON
- Lookup in city/country dictionary	LOCATION
- Regex: \$[0-9,]+	MONEY
Pros	High precision for known patterns
Cons	Low recall (misses variations), maintenance burden

Sequence Labeling with BIO Tags

Most NER systems frame it as sequence labeling using BIO encoding:

B-TAG	Beginning of an entity
I-TAG	Inside (continuation) of an entity
O	Outside any entity
Barack	B-PER
Obama	I-PER
visited	O
New	B-LOC
York	I-LOC
City	I-LOC
yesterday	O

CRF-Based NER (Traditional ML)

P(B-PER	I-PER) = high
P(B-PER	I-LOC) = 0 (illegal transition!)
P(O	I-PER) = 0 (must start with B)

Deep Learning NER (BiLSTM-CRF)

Example

Architecture: Input tokens → Character CNN + Word Embeddings ↓ Bidirectional LSTM (contextual features) ↓ CRF Layer (enforces valid tag sequences) ↓ BIO tag sequence Why CRF on top of LSTM? LSTM independently scores each position CRF captures tag transition constraints Prevents illegal sequences like O → I-PER

Transformer-Based NER (BERT + Token Classification)

Modern state-of-the-art

Input: "[CLS] Barack Obama visited New York [SEP]"

BERT: Contextual embeddings for each token

Linear: Each token embedding → entity logits

Output: B-PER I-PER O B-LOC I-LOC

Fine-tuning BERT for NER

- Add linear classification head

- Train on labeled NER data (CoNLL-2003)

- 2-4 epochs, lr=2e-5

- F1 score: ~93% (English)

Worked Example: NER Processing

Input	"Apple Inc. CEO Tim Cook announced new products in Cupertino on March 15, 2024"
Step 1	Tokenization
Step 2	Model prediction (BIO tags)
Apple	B-ORG
Inc	I-ORG
.	I-ORG
CEO	O
Tim	B-PER
Cook	I-PER
announced	O
new	O
products	O
in	O
Cupertino	B-LOC
on	O
March	B-DATE
15	I-DATE
,	I-DATE
2024	I-DATE
Step 3	Entity extraction
ORG	"Apple Inc."
PER	"Tim Cook"
LOC	"Cupertino"
DATE	"March 15, 2024"

Challenges in NER

Entity Boundary Detection

"New York City"	one entity or "New York" + "City"?
"Bank of America"	one ORG or "Bank" + LOC("America")?
"The Lord of the Rings"	where does the title start/end?

Nested Entities

ORG	"University of California, Berkeley"
LOC	"California"
LOC	"Berkeley"
Solutions	Flat NER (longest span), layered NER, span-based models

Ambiguity

"Apple"	Company (ORG) or fruit (not an entity)?
"Jordan"	Person (Michael Jordan) or Country?
"Washington"	Person, State, or City?
Context resolves: "Apple stock rose"	ORG
"Apple pie recipe"	not an entity

Evaluation Metrics

Predicted	[Tim Cook]_PER [Cupertino CA]_LOC
Gold	[Tim Cook]_PER [Cupertino]_LOC
Tim Cook	Exact match ✓ (correct entity + correct type)
Cupertino CA	Wrong boundary (predicted too much) ✗
State-of-the-art (CoNLL-2003 English)	F1 ≈ 93-94%

Applications

Application	Entity Types	Example
Search engines	All	Query understanding, knowledge panels
Healthcare	Disease, Drug, Gene	Clinical note processing
Finance	Company, Amount, Date	News analysis, compliance
Legal	Person, Org, Statute	Document review, case analysis
Social media	Person, Org, Product	Brand monitoring, trending

Interview Questions

Q: Why use BIO encoding instead of just labeling entities? A: BIO (or BIOES) encoding solves the boundary problem. Without B/I distinction, adjacent entities of the same type merge: "[John][Smith]" both labeled PER would become one entity "John Smith." B-PER, B-PER correctly identifies two separate person entities.

Q: How does NER handle unseen entities? A: Modern NER models generalize through contextual features (not memorization). BERT-based NER recognizes that "Zyloth Corp announced..." likely contains an ORG at "Zyloth Corp" even if never seen before—because the context pattern (X + announced) matches learned ORG patterns.

Q: What is the difference between NER and entity linking? A: NER identifies entity mentions and their types ("Obama" = PER). Entity linking disambiguates and connects to a knowledge base ("Obama" → Barack_Obama in Wikipedia, not Michelle_Obama). Entity linking requires NER first, then adds disambiguation.

What is Named Entity Recognition?

Entity Categories

Standard Entity Types (CoNLL, OntoNotes)

PERSON	"Albert Einstein", "Dr. Smith", "Mary"
ORGANIZATION	"Google", "United Nations", "MIT"
LOCATION	"Paris", "Mount Everest", "the Nile"
DATE	"January 2024", "last Tuesday", "the 90s"
TIME	"3:00 PM", "noon", "midnight"
MONEY	"$50 million", "€100", "500 rupees"
PERCENT	"25%", "a third", "half"

Domain-Specific Entities

Biomedical NER	Gene, Protein, Disease, Drug, Chemical
Legal NER	Court, Judge, Statute, Case
Financial NER	Ticker, Index, Currency, Metric

NER Approaches

Rule-Based NER

- Capitalized word sequences	potential entity
- "Mr./Mrs./Dr." + Name	PERSON
- Lookup in city/country dictionary	LOCATION
- Regex: \$[0-9,]+	MONEY
Pros	High precision for known patterns
Cons	Low recall (misses variations), maintenance burden

Sequence Labeling with BIO Tags

Most NER systems frame it as sequence labeling using BIO encoding:

B-TAG	Beginning of an entity
I-TAG	Inside (continuation) of an entity
O	Outside any entity
Barack	B-PER
Obama	I-PER
visited	O
New	B-LOC
York	I-LOC
City	I-LOC
yesterday	O

CRF-Based NER (Traditional ML)

P(B-PER	I-PER) = high
P(B-PER	I-LOC) = 0 (illegal transition!)
P(O	I-PER) = 0 (must start with B)

Deep Learning NER (BiLSTM-CRF)

Example

Transformer-Based NER (BERT + Token Classification)

Modern state-of-the-art

Input: "[CLS] Barack Obama visited New York [SEP]"

BERT: Contextual embeddings for each token

Linear: Each token embedding → entity logits

Output: B-PER I-PER O B-LOC I-LOC

Fine-tuning BERT for NER

- Add linear classification head

- Train on labeled NER data (CoNLL-2003)

- 2-4 epochs, lr=2e-5

- F1 score: ~93% (English)

Worked Example: NER Processing

Input	"Apple Inc. CEO Tim Cook announced new products in Cupertino on March 15, 2024"
Step 1	Tokenization
Step 2	Model prediction (BIO tags)
Apple	B-ORG
Inc	I-ORG
.	I-ORG
CEO	O
Tim	B-PER
Cook	I-PER
announced	O
new	O
products	O
in	O
Cupertino	B-LOC
on	O
March	B-DATE
15	I-DATE
,	I-DATE
2024	I-DATE
Step 3	Entity extraction
ORG	"Apple Inc."
PER	"Tim Cook"
LOC	"Cupertino"
DATE	"March 15, 2024"

Challenges in NER

Entity Boundary Detection

"New York City"	one entity or "New York" + "City"?
"Bank of America"	one ORG or "Bank" + LOC("America")?
"The Lord of the Rings"	where does the title start/end?

Nested Entities

ORG	"University of California, Berkeley"
LOC	"California"
LOC	"Berkeley"
Solutions	Flat NER (longest span), layered NER, span-based models

Ambiguity

"Apple"	Company (ORG) or fruit (not an entity)?
"Jordan"	Person (Michael Jordan) or Country?
"Washington"	Person, State, or City?
Context resolves: "Apple stock rose"	ORG
"Apple pie recipe"	not an entity

Evaluation Metrics

Predicted	[Tim Cook]_PER [Cupertino CA]_LOC
Gold	[Tim Cook]_PER [Cupertino]_LOC
Tim Cook	Exact match ✓ (correct entity + correct type)
Cupertino CA	Wrong boundary (predicted too much) ✗
State-of-the-art (CoNLL-2003 English)	F1 ≈ 93-94%

Applications

Application	Entity Types	Example
Search engines	All	Query understanding, knowledge panels
Healthcare	Disease, Drug, Gene	Clinical note processing
Finance	Company, Amount, Date	News analysis, compliance
Legal	Person, Org, Statute	Document review, case analysis
Social media	Person, Org, Product	Brand monitoring, trending

What is Named Entity Recognition?

Entity Categories

Standard Entity Types (CoNLL, OntoNotes)

Domain-Specific Entities

NER Approaches

Rule-Based NER

Sequence Labeling with BIO Tags

CRF-Based NER (Traditional ML)

Deep Learning NER (BiLSTM-CRF)

Transformer-Based NER (BERT + Token Classification)

Modern state-of-the-art

Fine-tuning BERT for NER

Worked Example: NER Processing

Challenges in NER

Entity Boundary Detection

Nested Entities

Ambiguity

Evaluation Metrics

Applications

Interview Questions

Continue learning this concept

What is Named Entity Recognition?

Entity Categories

Standard Entity Types (CoNLL, OntoNotes)

Domain-Specific Entities

NER Approaches

Rule-Based NER

Sequence Labeling with BIO Tags

CRF-Based NER (Traditional ML)

Deep Learning NER (BiLSTM-CRF)

Transformer-Based NER (BERT + Token Classification)

Modern state-of-the-art

Fine-tuning BERT for NER

Worked Example: NER Processing

Challenges in NER

Entity Boundary Detection

Nested Entities

Ambiguity

Evaluation Metrics

Applications

Interview Questions

Continue learning this concept