What is named entity recognition (NER)?

How NER works

An NER model reads a sentence and tags spans of words with a type. The usual core types are PERSON, ORGANIZATION and LOCATION; many models add dates, times, money amounts, products and events. Modern NER uses language models trained on large sets of hand-labeled sentences, so it relies on context rather than a fixed list of names. That is how it can tell that “Jordan” is a person in “Jordan said yes” and a place in “we flew to Jordan.”

NER only finds and types the name. Deciding which real-world person or company it refers to, say that “Apple” and “Apple Inc.” and “the Cupertino company” are the same thing, is a further step called entity linking.

An example on a transcript

Take the transcribed line: “Maria from Northwind said the Denver office signs on March 3 for about forty thousand dollars.” An NER pass would tag Maria as a person, Northwind as an organization, Denver as a location, March 3 as a date and forty thousand dollars as money. Run that across a year of sales calls and you can list every call where Northwind came up, without anyone having written a note.

Counted across a library, entities also give a quick map of what it covers: which clients come up most, which places a documentary keeps returning to, which names appear together. That is useful even before you search for anything specific.

Where it goes wrong

• Transcription errors cascade. If speech recognition hears a company name as ordinary words, NER has no name to find.
• Weak capitalization cues. Some transcripts are all lowercase or loosely punctuated, and capital letters are one of the signals NER uses.
• Ambiguous names. Words like Apple, Amazon or Paris can be ordinary words, companies, places or people.
• Unusual names. Names that are rare in the training data, including many non-English names, are missed more often.
• Spoken references. People say “she,” “that company” or a nickname far more than the full name, and NER tags only the name itself.

In MediaFind

MediaFind picks people, places and organizations out of every transcript on your computer, and Named entities is one of its search modes, so you can find each time someone or something came up across your library. Because entities come from the transcript, spelling matters: MediaFind's names & terms list keeps the people and products you care about spelled right during transcription, which gives entity search real names to work with. Read how entity search works, or see it applied to journalism and sales calls.

Frequently asked questions

What are the main entity types in NER?

Most systems find people, organizations and locations, and many also find dates, times, amounts of money, products and events.

Is NER the same as keyword search for names?

No. Keyword search needs you to know the exact name. NER discovers names you did not know to look for and labels what kind of thing each one is.

Does NER work on audio?

Indirectly. Audio is transcribed first, and NER runs on the transcript, so transcription quality limits what it can find.