Named entity recognition (NER), explained simply

In plain words

Read a news story and your eye catches the important bits: who, where, when, how much. Named entity recognition, or NER, teaches software to do the same.

A "named entity" is anything with a proper name or a precise value: a person, a city, a company, a date, a sum of money. NER finds each one and puts a label on it: this is a person, this is a place, this is money.

That turns messy text into tidy facts a computer can use. It can help search engines understand what a page is about, help banks pull names and amounts from forms, and help hospitals or courts hide personal details before sharing documents.

The hard part is context. "Jordan" can be a person or a country, and "Apple" a fruit or a company. Good NER reads the words around a name to decide.

Everyday example

Highlighter pens on a ticket booking

Give a friend a booking message and four highlighters: pink for people, blue for places, green for dates, yellow for money. "Riya booked Mumbai to Delhi on 14 March for ₹6,200" comes back with every key detail colour-coded. NER does that, for millions of messages, without getting tired. The colours are its labels.

Where the example breaks: Your friend knows Riya is a person because they know people. The software only learned patterns from labelled examples, so unusual names, spellings or languages it rarely saw can fool it.

How it actually works

NER is one of the classic jobs in natural language processing. It is usually treated as a labelling task on each word.

Split the text into tokens

The sentence is broken into words or tokens. Each one will get a tag.

Tag each token with a "B", "I" or "O"

A common scheme marks the beginning of a name ("B"), a word inside it ("I"), or a word outside any name ("O"). So "New" gets "B-place" and "Delhi" gets "I-place". This lets multi-word names like "State Bank of India" stay together.

Learn from labelled examples

Older systems used hand-written rules and lists of names. Modern ones are supervised models trained on text that people tagged by hand. Since 2018, transformer models like BERT have pushed accuracy up, because they read the words on both sides of a name.

Or simply ask a language model

Today a large language model can often do NER from a plain instruction, returning names as structured output. Dedicated small models are still common when speed, cost or privacy matter.

Task first set

The Sixth Message Understanding Conference (MUC-6), evaluated in September 1995

Classic labels

Person, organisation, location, plus dates, times, money and percentages

Classic benchmark

CoNLL-2003: Reuters news in English and German, with four types: person, location, organisation, miscellaneous

Score jump

Best English score in 2003: 88.8 F1. BERT in 2018: 92.8

F1 is a single score that balances finding every name (recall) against not tagging wrong ones (precision). 100 is perfect.

Questions people ask

What is an example of named entity recognition?

In "Meera flew to London on Monday", NER tags Meera as a person, London as a place and Monday as a date. Each name gets a type, so a computer can file it, search for it or hide it.

What are the types of named entities?

The classic set is person, organisation and location, often with dates, times, money and percentages. Many tools add more, like products, events, laws or languages, and specialist models add things like medicines or genes.

Is NER part of NLP?

Yes. It is one of the oldest and most common natural language processing tasks, sitting alongside sentiment analysis and summarising. It is often the first step in pulling facts out of documents.

Can ChatGPT do named entity recognition?

Yes. You can paste text and ask a chatbot to list the people, places and organisations. It often does well. For millions of documents, a small dedicated NER model is usually faster and cheaper, and can run on your own servers.

Is NER the same as entity linking?

No. NER finds a name and says what type it is. Entity linking goes one step further: it decides which real thing the name points to, for example whether "Hyderabad" means the city in India or the one in Pakistan.