LegalBERTNER<T>

Models & Types · AiDotNet.NER.TransformerBased

Legal-BERT-NER: Legal domain BERT for Named Entity Recognition in legal documents.

Inheritance: NeuralNetworkBase<T> → NERNeuralNetworkBase<T> → SequenceLabelingNERBase<T> → TransformerNERBase<T> → LegalBERTNER<T>

For Beginners

Legal-BERT is BERT trained on court opinions, contracts, and legislation. It understands legal jargon, citation formats, and legal entity types that general BERT struggles with. Use Legal-BERT-NER for processing contracts, court filings, regulations, or any legal documents where entity extraction is needed.

How It Works

Legal-BERT-NER (Chalkidis et al., EMNLP 2020 Findings - “LEGAL-BERT: The Muppets straight out of Law School”) is BERT pre-trained on 12GB of diverse English legal text for domain-specific legal NLP tasks including NER.

Pre-training Data:

  • EU legislation (EU Acquis, treaties, regulations, directives)
  • US court opinions (case law from all federal courts)
  • US contracts (EDGAR-sourced commercial contracts)
  • Legal academic papers (selected law review articles)
  • ~12GB total, ~2.5B tokens of legal text

Legal NER Entity Types:

  • Court: Supreme Court, District Court, European Court of Justice
  • Judge: Justice Roberts, Judge Smith
  • Legislation: Article 5(1), Section 230, Title VII
  • Citation: Brown v. Board of Education, 347 U.S. 483 (1954)
  • Party: Plaintiff, defendant, appellant names
  • Legal Concept: Due process, habeas corpus, res judicata
  • Jurisdiction: State of New York, European Union, federal

Why Legal NER Needs Domain Models: Legal text has unique structure: nested citations (“see also 42 U.S.C. 1983”), Latin terms (habeas corpus, stare decisis), and specific entity patterns (case citations like “548 F.3d 290 (2d Cir. 2008)”). Legal-BERT understands these patterns.

Performance:

  • Legal NER: ~88-91% F1 (vs general BERT ~82-85% on legal text)
  • Contract entity extraction: ~86-89% F1

Example

var architecture = new NeuralNetworkArchitecture<float>(inputFeatures: 512, numClasses: 9);
var legalBERTNER = new LegalBERTNER<float>(architecture);

Constructors

Constructor Summary
LegalBERTNER(NeuralNetworkArchitecture<T> architecture, TransformerNEROptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) Creates a Legal-BERT-NER model in native training mode.
LegalBERTNER(NeuralNetworkArchitecture<T> architecture, string modelPath, TransformerNEROptions options) Creates a Legal-BERT-NER model in native training mode.