Easy Start

April 22, 2025 · View on GitHub

Model

This project implements extraction models for NER tasks in the Standard scenario. The corresponding paths are:

BiLSTM-CRF
Bert (Tip: If using the dataset provided below, we recommend setting learning_rate to 2e-5 and num_train_epochs to 10)
W2NER

Experimental Results

Model	Accuracy	Recall	F1 Score	Inference Speed (People's Daily)
BERT	91.15	93.68	92.40	106s
BiLSTM-CRF	92.11	88.56	90.29	39s
W2NER	96.76	96.11	96.43	-

Clone the Repository

git clone https://github.com/zjunlp/DeepKE.git
cd DeepKE/example/ner/standard

Environment Setup

1. Create a Python virtual environment and activate it:

conda create -n deepke python=3.8
conda activate deepke

2. Install dependencies:

pip install -r requirements.txt

Parameter Configuration

1. Model Parameters

Model-specific configurations (e.g., model path, hidden layer dimensions, case sensitivity) can be found in the conf/hydra/model/*.yaml directory.

2. Other Parameters

Settings for environment paths and other hyperparameters during training are located in train.yaml and custom.yaml.

Note: Vocabulary usage during training:

For the Bert model, the vocabulary is derived from the pre-trained weights on Hugging Face.

For BiLSTM-CRF, the vocabulary must be built based on the training dataset and saved in a .pkl file for prediction and evaluation (configured in the model_vocab_path attribute of lstmcrf.yaml).

For model downloads with network error, we recommend using the Hugging Face mirror site . After downloading, modify the model path in conf/hydra/model/*.yaml.

Training with Dataset

1. Supported Data Formats

The model supports json, docx, and txt formats. For details, refer to the data folder. The default dataset is the People's Daily (Chinese NER) with text data in {word, label} pairs.

Note for English datasets: Update lan in config.yaml before prediction, and install nltk with nltk.download('punkt').

2. Prepare Data

Download the dataset:

wget 121.41.117.246:8080/Data/ner/standard/data.tar.gz
tar -xzvf data.tar.gz

Place the following files in the data folder:

train.txt: Training dataset
valid.txt: Validation dataset
test.txt: Test dataset

3. Start Training

Choose the appropriate model for your target scenario:

Bert
```
python run_bert.py
```
- Update hydra/model in config.yaml to bert. Hyperparameters for BERT are in bert.yaml.
- Multi-GPU training is supported by setting use_multi_gpu to True in train.yaml. Specify GPUs with os.environ['CUDA_VISIBLE_DEVICES'].
- For the default dataset downloaded above: Set learning_rate to 2e-5 and num_train_epochs to 10 in train.yaml.
- Tip: If the dataset is small, reduce the learning rate to avoid fast parameter updates. For large datasets, try increasing the learning rate to speed up convergence. You can also monitor the training process using wandb (set config.yaml's use_wandb parameter to True) and adjust the learning rate and epochs for optimal performance.
BiLSTM-CRF
```
python run_lstmcrf.py
```
Configure BiLSTM-CRF hyperparameters in lstmcrf.yaml. Modify other training parameters in the conf folder.
W2NER
```
cd w2ner
python run.py
```
Hyperparameters are in model.yaml. Specify the GPU index using the device parameter (set to 0 for single GPU setups).

Training Output

Logs and Results
- Training logs are saved in the logs folder.
- Model checkpoints are stored in the checkpoints folder.
Batch Size
- For BERT training, a batch size of 64 or more is recommended.

Prediction

Run predictions using:

python predict.py

Prepare weak_supervised data

If you only have text data and corresponding dictionaries, but no canonical training data.

You can get weakly supervised formatted training data through automated labeling methods.

Please make sure that:

Provide high-quality dictionaries
Enough text data

prepare-data