BERT-BiLSTM-CRF NER

BERT-BiLSTM-CRF-NER is a widely referenced open-source TensorFlow solution for named entity recognition that stacks three proven components: Google’s BERT for contextual embeddings, a BiLSTM layer for sequence modeling, and a CRF layer for consistent label decoding. With roughly 4,900 stars on GitHub, it has served as a starting point and teaching reference for countless Chinese and English NER projects.

Overview

Named entity recognition (finding people, organizations, locations and other entities in text) is a building block for search, knowledge graphs and document processing. This project provides the complete workflow: data preprocessing for sequence-labeling formats, fine-tuning BERT jointly with the BiLSTM-CRF head, training and evaluation scripts, and a deployable server component so a trained model can answer prediction requests over an API. It supports both BERT and simpler embedding setups for comparison. The architecture it implements (BERT for representations, BiLSTM for context, CRF for label consistency) was the state-of-the-art recipe for sequence labeling in the early transformer era, and the repository’s clear structure is why it remains a go-to citation for the pattern.

Key Features

• Full BERT + BiLSTM + CRF architecture implemented in TensorFlow
• Data loading and preprocessing for standard NER sequence-labeling datasets
• Joint fine-tuning of BERT with the task-specific sequence head
• Prediction and serving code for deploying a trained model as a service
• Support for Chinese and English entity recognition tasks
• Extensively forked and documented through community blog posts and issues
• Configurable label schemes, so the same code adapts to different entity types and datasets

💳 Pricing

Free and open source on GitHub. There is no paid tier; you train and host models with your own compute.

Who Is It For

NLP engineers and students who need a concrete, end-to-end reference implementation of transformer-based NER, especially for Chinese text extraction in search, finance and document workflows.