Definition
Entity extraction is the process of automatically identifying and classifying named entities — such as people, organisations, dates, monetary amounts, legal references, and jurisdictions — within unstructured text. It is a core natural language processing (NLP) task that transforms raw documents into structured data by tagging each entity mention with a type label. In legal AI, entity extraction powers downstream capabilities like knowledge graph construction, metadata enrichment, cross-referencing between documents, and structured search filtering.
Why it matters
- Structured metadata from unstructured text — legislation and rulings arrive as prose; entity extraction identifies the article numbers, dates, parties, and monetary thresholds embedded within, making them searchable and filterable
- Knowledge graph construction — extracted entities and their relationships form the nodes and edges of a knowledge graph, enabling the system to answer relational queries like “which rulings cite article 215 WIB92?”
- Cross-referencing — when a circular mentions a royal decree by name, entity extraction identifies the reference and links it to the corresponding document in the knowledge base
- Search enrichment — extracted entities become metadata that supports faceted search, allowing users to filter results by jurisdiction, tax type, date range, or authority level
How it works
Entity extraction typically operates in two stages:
Detection identifies the boundaries of entity mentions in the text — determining that “Grondwettelijk Hof” in a sentence is a single entity, not two separate words. This is challenging because entity names can span multiple words, contain common words, or overlap with surrounding text.
Classification assigns a type to each detected entity. Standard types include person, organisation, location, and date, but legal NLP extends this with domain-specific types: legislative references (article numbers, law codes), court identifiers, tax categories, jurisdictional markers, and monetary amounts with currency.
Modern entity extraction uses transformer-based models fine-tuned on labelled legal text. The model processes each token in context and predicts whether it is part of an entity and what type it belongs to (using BIO tagging: Begin, Inside, Outside). Pre-trained legal language models perform significantly better than general-purpose models because they understand legal naming conventions, citation formats, and domain terminology.
For Belgian tax law, entity extraction must handle multilingual documents (Dutch, French, German), recognise references to specific Belgian legal instruments (WIB92, BWHI, KB/AR), and distinguish between federal and regional legislative references.
Common questions
Q: How is entity extraction different from keyword extraction?
A: Keyword extraction identifies the most important terms in a document regardless of type. Entity extraction identifies specific named items and classifies them by category. “Vennootschapsbelasting” might be extracted as a keyword; “article 185 WIB92” would be extracted as a legislative reference entity with structured metadata (article number: 185, law code: WIB92).
Q: How accurate is entity extraction on legal text?
A: General-purpose NER models typically achieve 85-90% F1 on legal text out of the box. Fine-tuning on domain-specific annotated data pushes this to 93-97% F1, depending on entity type. Dates and monetary amounts are easiest; legislative cross-references and jurisdictional markers are harder due to their varied formatting.