Who am I? And if so, why?

Named Entity Recognition is how machines pick the names out of a sentence — and decide what each one is. Every name in a piece of text is quietly asking a question: who am I, and if so, why? Is "Fisher" a person or a profession? Is "Ford" a family, a company, or a place? Is "Orange" a fruit, a color, or a town in France?

Humans answer these questions without noticing. Teaching a machine to do the same is the job of Named Entity Recognition (NER) — one of the foundational tasks in natural language processing.

What a "named entity" actually is

"Named entity" is a slightly odd term invented by computational linguists. It refers, more or less, to proper names: words that point to a single, unique thing in the real world. There is only one Eiffel Tower, one Marie Curie, one Beethoven's Ninth Symphony. Ordinary words live in dictionaries; named entities live in encyclopedias. That distinction matters, because there is no finite list you can memorize — new people, products, and places appear every day.

A NER system has two jobs at once: decide that a span of text is a name, and then assign it a type. The classic set, designed for news-style text, covers:

• People — "Ada Lovelace," "Dr. Skiggs"
• Locations — "Los Angeles," "the North Sea"
• Organizations — "European Commission," "British Telecom"
• Dates and times — "March 2016," "next Tuesday"
• Quantities and money — "€250M," "12 percent"

Specialized domains extend this with their own classes: chemical compounds, genes, legal citations, ICD codes, ticker symbols, even "all offshore wind parks in Europe" if that is what you happen to be tracking.

A multilingual headache

Some names travel unchanged across languages — "Los Angeles" looks the same in English, Spanish, French, and German. Others mutate: Nürnberg / Nuremberg / Norimberga / Núremberg are all the same city. Scripts change too. A NER system that works in one language rarely works in another out of the box, which is why multilingual NER is a field of its own.

How machines learn to spot names

NER has gone through four broad eras, and modern systems often blend them.

1. Lists and rules (gazetteers). The earliest approach: keep big lists of known names — countries, major cities, large companies, football teams — and match against them. Lists are precise but never complete, so they were paired with hand-written rules: if an unknown word follows an academic title, it is probably a person ("Dr. ___"); "Institute of ___" usually marks an organization; "west of ___" usually marks a place.

2. Statistical models. Instead of hand-coding rules, researchers trained models such as Conditional Random Fields (CRFs) on annotated text. The model learns, from examples, which contextual clues predict which entity type. The work shifts from writing rules to labeling data.

3. Neural networks. Architectures like BiLSTM-CRF learned to represent words and their context as vectors, capturing patterns that are hard to write down by hand. Word embeddings let the model generalize from "London" to other city names it had rarely seen.

4. Transformers and large language models. Today, transformer models like BERT and its many descendants set the standard, reading a whole sentence in both directions before deciding what each word is. Large language models push this further: with a good prompt and a few examples, an LLM can extract entities zero-shot or few-shot, without task-specific training — useful when you have no labeled data for a new domain. Lightweight open models such as GLiNER now offer flexible, custom entity types on modest hardware.

Why it's still hard

If NER were solved, this article would be shorter. The stubborn problems:

• Ambiguity in context. "He lived in Napoleonic times" matches the pattern "lived in ___," but "Napoleonic" is not a place. Every rule is a heuristic that fails somewhere.
• Coreference. A single entity is referred to many ways across a passage — "Ford Motor Company," then "Ford Motors," then "Ford," then "it." Recognizing that all of these point to the same thing (and resolving pronouns) is one of the hardest parts of understanding a document.
• Nested and overlapping entities. "Bank of England" is an organization that contains a location. Flat taggers struggle with names inside names.
• Domain shift. A model trained on news collapses on clinical notes, legal contracts, or social media slang. Adapting to a new domain usually means new labeled data.

How we know if it's any good

NER quality is measured with two numbers that pull against each other:

• Recall — of all the real entities in the text, how many did the system find?
• Precision — of all the spans the system flagged, how many were actually correct?

Loosen the system and recall goes up but precision falls (more guesses, more mistakes). Tighten it and the reverse happens. The combined score, the F1, rewards getting both right. On clean research benchmarks like CoNLL-2003 or OntoNotes, top systems exceed 0.90 F1; in messy production settings, 0.75–0.85 is more realistic. The number depends on language, domain, text type, and how much training data you have — so treat any single headline figure with suspicion.

Where NER actually gets used

• Better search and retrieval. Search for "chess player Fisher" and you want documents about a player named Fisher, not stories of fishermen who play chess. NER tells the two apart.
• Information extraction and media monitoring. Tracking every article that mentions a company, product, or person depends on reliably catching those names — "Electrolux in China," "offshore wind development," "dairy exports to Thailand."
• Opinion mining. Once you know which entity a sentence is about, you can ask what attitude it expresses toward that entity: "Electrolux fails to save jobs in Orange."
• Knowledge graphs and document understanding. Entities are the nodes; the relations between them are the edges. NER is the first step in turning unstructured text into a structured, queryable picture of the world.

Try it yourself

If you want to see NER in action, open-source tools make it easy to start: spaCy and Stanza for production pipelines, Hugging Face Transformers for state-of-the-art models, and GLiNER for custom entity types without retraining. A few lines of Python will tag a paragraph in seconds.

So, who am I?

Back to the question every name is asking. The honest answer is that a machine never knows in the human sense — it estimates, from context and patterns, the most likely answer and a confidence to go with it. That is enough to power search engines, monitor the news, and build the knowledge graphs behind modern AI. Not bad for a system whose entire job is helping each word in a sentence figure out: who am I, and if so, why?

For more on how machines make sense of language, see our guide on modeling human natural language understanding, or browse all our coverage of multimodal AI.

Frequently asked questions about NER

What is the difference between NER and entity linking?

NER finds a name and labels its type ("Paris" is a location). Entity linking goes one step further and connects that mention to a specific record in a knowledge base — disambiguating "Paris" the city in France from "Paris" the person.

Do I need labeled data to do NER?

Not always. Classic and neural models are trained on labeled examples, but modern large language models can extract entities zero-shot or few-shot from a prompt, which is useful when you have little or no annotated data for your domain.