Named-entity recognition

A support ticket says “Acme Corp was double-billed $4,200 on 30 June”. A person reads the facts instantly: a customer, an amount, a date. Named-entity recognition (NER) teaches software to do the same — to find the spans of text that name things and label each with a type — so that free text can be searched, filtered, joined with databases and acted on.

What counts as an entity

Common entity types are people, organizations, locations, dates, amounts of money and products. Real projects add their own: drug names in medical notes, part numbers in maintenance logs, clause types in contracts. Turn the types on and off below, and switch to the token view to see what a model actually predicts.

Entity

A span of text that refers to a specific thing, such as “Northwind Labs” or “3 March 2025”.

Entity type

The category of an entity: person, organization, location, date, money, product and so on.

Span

A run of consecutive tokens, identified by its start and end positions.

Entity linking

The next step after NER: connecting “Acme Corp” to the right record in a database or knowledge base.

NER as sequence labeling

Models do not draw boxes around text. They solve an easier problem: give every token one tag. The widely used BIO scheme marks the first token of an entity with B- (begin), the following tokens of the same entity with I- (inside), and everything else with O (outside). Adjacent B-/I- tags are then merged back into spans.

Three ways to build an extractor

Rules and dictionaries. Regular expressions catch well-formatted things — amounts, dates, e-mail addresses, ticket numbers — and word lists catch known names. Rules are fast, predictable and need no training data, but they cannot see context and break on anything they have not been told about.

Statistical sequence models. Conditional random fields (CRFs) learn from labeled examples which features predict each tag — capitalization, word shape, neighboring words. For a decade they were the standard approach.

Transformers. Fine-tuned models such as BERT read the whole sentence at once and give each token a contextual embedding, then classify it. With a few thousand labeled sentences they usually outperform everything else — and pretrained models for the common types are free to download.

import spacy

nlp = spacy.load("en_core_web_trf")              # a pretrained transformer pipeline
doc = nlp("Acme Corp agreed to pay $4.2 million to Globex by 30 June.")
for ent in doc.ents:
    print(ent.text, ent.label_, ent.start_char, ent.end_char)
# Acme Corp ORG 0 9
# $4.2 million MONEY 24 36
# Globex ORG 40 46
# 30 June DATE 50 57

Evaluating an extractor

NER is evaluated on entities, not tokens, because a system that gets most tokens right can still produce broken spans. Each predicted span is compared with the human “gold” labels:

• Precision — of the spans the model predicted, how many were correct?
• Recall — of the real entities, how many did the model find?
• F1 — the harmonic mean of the two, a single score that punishes a weakness in either.

The deciding question is what counts as correct. Strict scoring demands exact boundaries and the right type; lenient scoring accepts any overlap with the right type. Always report which one you used.

Notice how much the rule-based system gains under lenient scoring: its main weakness is boundaries (“Priya” instead of “Priya Sharma”), while its type errors — “June” as a person, “Skyline” as a place — are exactly the context mistakes described above.

Entity extraction turns documents into facts. The last lecture of this module turns a whole collection of documents into something people can ask questions of: semantic search.