FinBERTNER<T>
Models & Types · AiDotNet.NER.TransformerBased
FinBERT-NER: Financial domain BERT for Named Entity Recognition in financial text.
Inheritance: NeuralNetworkBase<T> → NERNeuralNetworkBase<T> → SequenceLabelingNERBase<T> → TransformerNERBase<T> → FinBERTNER<T>
For Beginners
FinBERT is BERT trained on financial documents like SEC filings, earnings reports, and financial news. It understands financial jargon, company names, and financial metrics better than general BERT. Use FinBERT-NER for extracting entities from financial text: company names, ticker symbols, financial figures, regulatory terms.
How It Works
FinBERT-NER (Araci, 2019 - “FinBERT: Financial Sentiment Analysis with Pre-trained Language Models”; Yang et al., IJCAI 2020 - “FinBERT: A Pretrained Language Model for Financial Communications”) is BERT further pre-trained on large-scale financial corpora for domain-specific NLP tasks including financial NER.
Pre-training Data:
- Financial news articles (Reuters, Bloomberg)
- SEC filings (10-K, 10-Q, 8-K reports)
- Financial analyst reports
- Earnings call transcripts
- ~4.9B tokens of financial text
Financial NER Entity Types:
- Company: Apple Inc., Goldman Sachs, Tesla
- Person: CEOs, CFOs, board members, analysts
- Financial Metric: Revenue, EBITDA, P/E ratio, market cap
- Currency/Amount: $1.5 billion, EUR 200 million
- Date/Period: Q3 2023, fiscal year 2024, YoY
- Regulation: Dodd-Frank, Basel III, SOX
- Index/Ticker: S&P 500, NASDAQ, AAPL
Why Financial NER Needs Domain Models: Financial text has unique challenges: “Apple” is a company (not fruit), “bearish” describes market sentiment (not an animal), and “spread” refers to yield difference (not butter). FinBERT’s financial pre-training resolves these ambiguities that confuse general BERT.
Performance:
- Financial NER: ~90-92% F1 (vs general BERT ~85-87% on financial text)
- Financial sentiment analysis: ~88% accuracy (state-of-the-art)
Example
var architecture = new NeuralNetworkArchitecture<float>(inputFeatures: 512, numClasses: 9);
var finBERTNER = new FinBERTNER<float>(architecture);
Constructors
| Constructor | Summary |
|---|---|
FinBERTNER(NeuralNetworkArchitecture<T> architecture, TransformerNEROptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer) |
Creates a FinBERT-NER model in native training mode. |
FinBERTNER(NeuralNetworkArchitecture<T> architecture, string modelPath, TransformerNEROptions options) |
Creates a FinBERT-NER model in native training mode. |