What NER turns unstructured text into
Named Entity Recognition is the bridge between prose a human reads and data a program can use. It scans free text and tags the real-world things it mentions:
| Label | Captures | Typical use |
|---|---|---|
| PER | People’s names | Redaction, contact extraction |
| LOC | Cities, countries, places | Geotagging, logistics |
| ORG | Companies, institutions | Competitive analysis, CRM enrichment |
| MISC | Nationalities, events, products | Tagging, search facets |
Feed it a news article and you get a structured list of who, where, and which organizations — the raw material for knowledge graphs, search indexes, résumé parsers, and automated tagging.
Why capitalization makes or breaks results
This model leans heavily on capitalization and context as signals, because that’s how proper nouns appear in well-formed English. Write “apple released a phone” in all lowercase and the model may miss the company entirely; write “Apple released a phone” and it confidently tags ORG. If you’re extracting from lowercased data (chat logs, transcripts, search queries), expect lower recall — and consider restoring capitalization first if accuracy matters.
Reading the color-coded output
Entities are highlighted by type so you can scan a document at a glance: people, locations, organizations, and miscellaneous entities each get their own color. This visual pass is often the fastest way to audit a document — you can immediately see whether the names you expected were caught and whether anything sensitive slipped through.
Where it runs and what that means
The BERT-based model (~40–60MB) downloads once and runs entirely in your browser via Transformers.js, so confidential text stays local. The trade-off is the one-time download and a dependence on your device’s speed for very long documents — split large texts into sections if processing stalls. For pulling out patterns rather than named things (dates, SKUs, phone numbers), pair this with the Regex Tester, which catches the structured identifiers NER isn’t designed to find.