ner-bert-base-cased-pt-lenerbr

What is ner-bert-base-cased-pt-lenerbr?

This is a specialized Named Entity Recognition (NER) model designed for the Portuguese legal domain. Built on BERT architecture, it was fine-tuned using the LeNER-Br dataset to identify and classify legal entities in Portuguese texts. The model demonstrates exceptional performance, particularly in recognizing person names (98.3% F1), temporal expressions (96.6% F1), and organizational entities (89.3% F1).

Implementation Details

The model was developed through a two-stage training process: first specializing the language model on legal texts, then fine-tuning for NER tasks. It uses a BERT base architecture with specialized tokenization for Portuguese legal terminology and achieves its results through careful hyperparameter optimization including a learning rate of 2e-5 and gradient accumulation steps of 2.

• Trained on LeNER-Br dataset with legal domain focus
• Implements transformer architecture with specialized Portuguese tokenization
• Uses AdamW optimizer with linear learning rate scheduling
• Trained for 10 epochs with batch size 4

Core Capabilities

• Recognition of 6 entity types: JURISPRUDENCIA, LEGISLACAO, LOCAL, ORGANIZACAO, PESSOA, TEMPO
• Optimal performance on person name recognition (98.7% precision)
• Strong temporal expression detection (96.6% F1 score)
• Efficient processing of legal documents with context awareness