What Is Named Entity Recognition (NER)?
Written byFlavio Amiel Founder, Roborank
Named entity recognition (NER) is the natural-language-processing task of locating entities in text and tagging each one with its type — such as person, organization, or location. It converts unstructured writing into labeled spans a machine can act on, and it is the first step in turning a document into a structured set of real-world things.
- NER identifies spans of text and classifies each as an entity type; the classic CoNLL-2003 benchmark uses four types — person (PER), location (LOC), organization (ORG), and miscellaneous (MISC).
- Most systems use BIO tagging, marking the Beginning and Inside of each entity span and Outside for everything else.
- CoNLL-2003, built from Reuters news text and introduced by Tjong Kim Sang and De Meulder in 2003, remains the standard English NER benchmark; top models score span-based F1 around 94.
- The open-source spaCy library ships production NER out of the box; its English models label around 18 entity types, well beyond the four CoNLL categories.
How Named Entity Recognition Works
Named entity recognition is the task of scanning text and tagging each entity with its type. Given a sentence, an NER system marks which spans are people, which are places, which are organizations, and which are none of the above — turning a wall of words into labeled, machine-usable units. It is the doorway to entity-based understanding: nothing downstream, from an entity salience score to a knowledge graph link, can happen until an entity has first been found and named.
Most systems represent the output with BIO tagging. Each token gets one of three prefixes: B- for the beginning of an entity, I- for a token inside (continuing) the same entity, and O for a token outside any entity. This lets a model treat a multi-word name as a single span. In the standard illustration, the sentence “Mark Watney visited Mars” is tagged B-PER I-PER O B-LOC — “Mark Watney” is one person entity, “Mars” is one location, and “visited” belongs to neither.
Entity types vary by model. Research benchmarks tend to use a small, fixed set, while production libraries recognize many more. The open-source spaCy library ships ready-to-use NER whose English models label roughly 18 types — including dates, money, products, and events — far beyond the handful used to score academic systems.
Example of Named Entity Recognition
The canonical worked example is the CoNLL-2003 shared task, introduced by Erik F. Tjong Kim Sang and Fien De Meulder in 2003 and still the reference benchmark for English NER. It was built from Reuters newswire text and annotates four entity types: person (PER), location (LOC), organization (ORG), and miscellaneous (MISC) — the last covering things like nationalities and events that don’t fit the first three.
CoNLL-2003 defined how the field measures success: models are scored with span-based F1, meaning a prediction only counts as correct if it gets both the exact boundaries and the type of an entity right. That strictness is what makes the benchmark meaningful. Two decades on, leading models reach an F1 of roughly 94 on the English test set — for example, the ACE document-context model reported by Wang and colleagues in 2021 reached 94.6. Researchers now note that near-human performance has effectively been reached on this particular dataset, which says as much about the benchmark’s age as about the models.
The practical takeaway for anyone writing for search: recognition is imperfect, and it is the step your content most directly influences. An entity written plainly and named consistently is easy to tag correctly; one hidden behind pronouns, abbreviations, or unusual phrasing is easy to miss or mislabel. Clean recognition is the price of admission to every entity feature that follows.
The thing people underestimate is that NER is upstream of almost everything in entity SEO. Before a system can decide which entity a page is about, or hand it a salience score, or link it to a knowledge graph, it has to first find the entity in the raw text and label it correctly — and that step is imperfect. Ambiguous names, novel brands, and entities that share a spelling with a common word all trip recognizers up. When I see a page whose topic a search engine seems to have misidentified, the root cause is often that the intended entity was never cleanly recognized in the first place: it was written in a way that let the parser mistag it or miss it. Naming entities fully and consistently, rather than leaning on pronouns and clever paraphrase, is a quiet but real way to make sure the machine tags what you meant.
Frequently Asked Questions
What is named entity recognition?
What entity types does NER use?
What is BIO tagging in NER?
How accurate is named entity recognition?
The Bottom Line
Named entity recognition is the parsing step that finds the real-world things inside a block of text and stamps each with a label. It is the foundation the rest of entity understanding is built on: before a system can score, link, or reason about an entity, NER has to spot it and name its type. Get recognized cleanly, and everything downstream — salience, linking, knowledge features — has something reliable to work from.
Sources
- Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition — Tjong Kim Sang & De Meulder, ACL Anthology (2003)
- Named Entity Recognition (task, benchmarks, BIO tagging) — NLP-progress
- spaCy — Industrial-strength Natural Language Processing — Explosion AI
Rank & Cash — the weekly SEO breakdown
One practical teardown a week on ranking in search and getting cited by AI. No fluff.