This project provides a fine-tuned ALBERT model for Named Entity Recognition (NER) on Chinese text, aiming to offer comparable or superior performance to BERT with significantly reduced model size and training time. It is intended for researchers and practitioners working with Chinese NLP tasks.

How It Works

The project leverages the ALBERT architecture, specifically a Chinese pre-trained ALBERT base model, and fine-tunes it on NER tasks. ALBERT's parameter reduction techniques, such as parameter sharing and factorized embedding parameterization, contribute to its smaller footprint and faster training compared to BERT, while aiming to maintain high accuracy.

Quick Start & Requirements

• Install/Run: python albert_ner.py --task_name ner --do_train true --do_eval true --data_dir data --vocab_file ./albert_config/vocab.txt --bert_config_file ./albert_base_zh/albert_config_base.json --max_seq_length 128 --train_batch_size 64 --learning_rate 2e-5 --num_train_epochs 3 --output_dir albert_base_ner_checkpoints
• Prerequisites: TensorFlow > 1.13 (specifically tested with 1.15), ALBERT Chinese base model files.
• Setup: Requires downloading and placing the ALBERT Chinese base model.

Highlighted Details

• Achieved an evaluation F1 score of 0.928 on Chinese NER after 3 epochs of training.
• Demonstrates comparable or improved performance over BERT with a significantly smaller model size and reduced training time.

Maintenance & Community

No specific community channels, roadmap, or contributor information is provided in the README.

Licensing & Compatibility

The README does not specify a license. Compatibility with TensorFlow 2.0 is explicitly stated as not supported.

Limitations & Caveats

The project is explicitly stated to not support TensorFlow 2.0. The README does not detail the specific dataset used for fine-tuning or provide extensive documentation beyond the execution command.