| Property | Value |
|---|---|
| Parameters | 110M |
| Architecture | 12-layer, 768-hidden, 12-heads BERT |
| Training Data | 260,773 SEC 10-K filings (1993-2019) |
| Paper | FiNER: Financial Numeric Entity Recognition for XBRL Tagging |
What is SEC-BERT-NUM?
SEC-BERT-NUM is a specialized BERT model designed specifically for financial domain natural language processing. Its unique feature is the uniform handling of numerical expressions by replacing all number tokens with a [NUM] pseudo-token, preventing fragmentation of numeric values. The model was trained on a massive dataset of SEC filings, making it particularly effective for financial text analysis tasks.
Implementation Details
The model builds upon the BERT-BASE architecture but incorporates several domain-specific optimizations. It uses a custom 30k subword vocabulary trained from scratch on financial documents and follows the same training setup as BERT-BASE with 1 million training steps.
- Trained using Google's official BERT repository
- Optimized for both PyTorch and TF2 compatibility
- Implements special numeric token handling through pre-processing
- Trained on Google Cloud TPU v3-8
Core Capabilities
- Superior performance in financial text understanding
- Consistent handling of numerical expressions
- Enhanced masked token prediction for financial contexts
- Specialized vocabulary for financial domain
Frequently Asked Questions
Q: What makes this model unique?
SEC-BERT-NUM's distinctive feature is its uniform handling of numerical expressions through the [NUM] token, which helps maintain consistency in financial text processing and improves performance on financial NLP tasks.
Q: What are the recommended use cases?
The model is particularly well-suited for financial text analysis tasks, including financial numeric entity recognition, sentiment analysis of financial documents, and processing of SEC filings and other financial reports.