The source code of Relational Siamese Network (EMNLP2019)
Future up-to-date versions will be available in:
https://github.com/thunlp/RSN
You only need TensorFlow(>=1.12.0) to run this code.
conda create -n rsn python=3.6
pip install networkx python-louvain tensorflow-gpu==1.12
You can download data by:
bash download.sh
The data is already preprocessed from the original FewRel and Glove data.
A standard CNN for supervised RE can be trained by
cd CNN
python train_CNN.py
The Relational Siamese Network for OpenRE can be trained by
cd RSN
python train_RSN.py
By default it will be trained as a semi-supervised RSN.
A supervised RSN can be trained by
python train_RSN.py --trainset_loss_type cross --testset_loss_type none
The FewRel-distant dataset and more models will be released in the future.