Entity Extraction

Entity extraction (also called named entity recognition or NER) is the process of automatically identifying and categorising specific pieces of information in text — such as person names, company names, locations, dates, monetary amounts, and product names.

How entity extraction works

Given a sentence like "Apple announced a $3 billion investment in Munich on Tuesday," entity extraction identifies:

• "Apple" → Organisation
• "$3 billion" → Money
• "Munich" → Location
• "Tuesday" → Date

Modern entity extraction uses pre-trained language models that understand context. The model knows "Apple" is a company here (not a fruit) because of the surrounding words "announced" and "investment."

Common entity types

• Person — individual names
• Organisation — companies, institutions, agencies
• Location — cities, countries, addresses
• Date/Time — dates, times, durations
• Money — monetary values and currencies
• Product — specific products or services
• Custom entities — industry-specific types (drug names, legal citations, part numbers)

Business applications

• Document processing — automatically extracting key information from contracts, invoices, and reports
• Customer support — identifying product names, order numbers, and issue types from support tickets
• Compliance — scanning documents for personally identifiable information (PII)
• Media monitoring — tracking mentions of your company, competitors, or executives
• Research — extracting relationships between entities from scientific or financial literature

Approaches

• Pre-trained NER models — spaCy, Hugging Face models that work out of the box for common entity types
• LLM-based extraction — using prompts to extract entities from text, more flexible for custom types but slower
• Fine-tuned models — training a model on your specific entity types and domain language for maximum accuracy

Challenges

• Ambiguity ("Washington" — person, city, or state?)
• Nested entities ("Bank of America" contains "America")
• Domain-specific entities require custom training data
• Performance varies across languages and domains