Named Entity Recognition (NER)

Named entity recognition (NER) is a natural language processing task that scans text and identifies mentions of specific entities — people, organisations, locations, dates, monetary values, and other defined categories. It is one of the foundational building blocks of text analytics.

How NER works

Given the text "Satya Nadella announced that Microsoft will invest $10 billion in OpenAI's San Francisco headquarters in January 2025," a NER system identifies:

• "Satya Nadella" → Person
• "Microsoft" → Organisation
• "$10 billion" → Money
• "OpenAI" → Organisation
• "San Francisco" → Location
• "January 2025" → Date

NER approaches

• Rule-based — handcrafted patterns and dictionaries. Precise but inflexible and expensive to maintain.
• Statistical models — trained on annotated text using algorithms like conditional random fields (CRFs). The traditional ML approach.
• Deep learning models — transformer-based models fine-tuned on NER datasets. Current state of the art, especially for complex and ambiguous text.
• LLM-based — using large language models with prompting for entity extraction. Most flexible but slower and more expensive per query.

Standard entity types

The most widely used NER categories (from the OntoNotes standard) include: Person, Organisation, Location, Date, Time, Money, Percentage, Product, Event, and Law.

Custom NER

Many business applications require custom entity types:

• Insurance: policy numbers, claim types, coverage categories
• Healthcare: drug names, symptoms, procedures, anatomical terms
• Legal: case citations, statute references, party names
• Finance: ticker symbols, fund names, regulatory references

Custom NER requires annotated training data specific to your domain.

NER in the LLM era

Large language models have made NER more accessible. Instead of training a custom model, you can prompt an LLM to extract entities from text. This works well for prototyping and low-volume use cases, but dedicated NER models are faster and cheaper at scale.