SECBertNER<T>

Models & Types · AiDotNet.NER.TransformerBased

SEC-BERT-NER: Securities and Exchange Commission domain BERT for NER in regulatory filings.

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

For Beginners

SEC-BERT is a specialized version of BERT trained on SEC regulatory filings (10-K, 10-Q reports that public companies must file). It excels at extracting entities from these highly structured financial documents. Use SEC-BERT-NER when processing SEC filings, XBRL documents, or regulatory financial text.

How It Works

SEC-BERT-NER (Loukas et al., EMNLP 2022 - “FiNER: Financial Numeric Entity Recognition for XBRL Tagging”) and related work uses BERT pre-trained specifically on SEC EDGAR filings for financial regulatory NER.

Pre-training Data:

  • SEC EDGAR filings: 10-K, 10-Q, 8-K, proxy statements, prospectuses
  • ~250K filing documents spanning 1993-2023
  • Financial regulatory language with strict formatting conventions

SEC-Specific NER Entity Types:

  • XBRL Tags: us-gaap:Revenue, us-gaap:NetIncomeLoss, dei:EntityRegistrantName
  • Filing Entity: Registrant names, CIK numbers
  • Monetary Values: Amounts in SEC-mandated formats ($X,XXX)
  • Date References: Filing dates, fiscal periods, reporting dates
  • Regulatory References: Item numbers, exhibit references, rule citations
  • Financial Statements: Balance sheet items, income statement items

Why SEC-specific NER matters: SEC filings follow strict formatting rules and use specialized terminology. A filing might reference “Item 7” (MD&A section) or “Exhibit 31.1” (SOX certification). SEC-BERT understands these domain-specific patterns that general financial models miss.

Performance:

  • XBRL entity recognition: ~85-88% F1 (vs general BERT ~75-78%)
  • Filing entity extraction: ~91-93% F1

Example

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

Constructors

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