OpenNRE Documentation

To use OpenNRE, clone the repository and install the required dependencies and the package itself. Ensure you have an appropriate PyTorch version installed for your machine's CUDA version before proceeding.

1. Clone the repository:

   git clone https://github.com/thunlp/OpenNRE.git

   (Use --depth 1 for a faster clone if needed).
2. Install requirements:

   pip install -r requirements.txt

3. Install the package:

   python setup.py install

   If you intend to modify the source code, use python setup.py develop instead.

OpenNRE does not include data or pretrained files in the repository to keep the deployment fast. You must download them manually using the provided scripts in the benchmark and pretrain folders.

Example: To download the FewRel dataset, run:

bash benchmark/download_fewrel.sh

OpenNRE provides example scripts in the example folder for training both supervised and bag-level relation extraction models. You can use provided datasets or your own.

Train a Bag-Level CNN-ATT Model

Use example/train_bag_cnn.py to train on datasets like nyt10m:

python example/train_bag_cnn.py \
    --metric auc \
    --dataset nyt10m \
    --batch_size 160 \
    --lr 0.1 \
    --weight_decay 1e-5 \
    --max_epoch 100 \
    --max_length 128 \
    --seed 42 \
    --encoder pcnn \
    --aggr att

Train a Supervised BERT Model

Use example/train_supervised_bert.py to train on the wiki80 dataset:

python example/train_supervised_bert.py \
    --pretrain_path bert-base-uncased \
    --dataset wiki80

You can perform sentence-level relation extraction by loading a pretrained model using opennre.get_model() and calling the .infer() method. The .infer() method requires a dictionary containing the text and the character positions of the head (h) and tail (t) entities.

To use a GPU, call .cuda() on the model instance before inference.

The following models are currently available for immediate use via opennre.get_model():

• wiki80_cnn_softmax: Trained on wiki80 dataset with a CNN encoder.
• wiki80_bert_softmax: Trained on wiki80 dataset with a BERT encoder.
• wiki80_bertentity_softmax: Trained on wiki80 dataset with a BERT encoder (using entity representation concatenation).
• tacred_bert_softmax: Trained on TACRED dataset with a BERT encoder.
• tacred_bertentity_softmax: Trained on TACRED dataset with a BERT encoder (using entity representation concatenation).