Latent-Forests

Learning Latent Forests for Medical Relation Extraction (authors' PyTorch implementation for the IJCAI20 paper)

Learning Latent Forests for Medical Relation Extraction

This paper/code introduces the Latent Forests Graph Convolutional graph convolutional networks (LFGCNs) for the medical relation extraction task.

You can find the paper here

Requirements

Our model was trained on GPU GeForce RTX 2080 Ti P100-SXM2 of Ubuntu 16.04.6 LTS

• Python 3 (tested on 3.6.11)

• PyTorch (tested on 1.5.0)

• CUDA (tested on 10.2)

• tqdm

• unzip, wget (for downloading only)

We have released our trained model in this repo. You can find the trained model under the saved_models directory. There is no guarantee that the model is the same as we released and reported if you run the code on different environments (including hardware and software).

Preparation

The code includes three datasets: CPR, PGR and Semeval, all of them under the directory dataset.

First, download and unzip GloVe vectors:

chmod +x download.sh; ./download.sh

Then prepare vocabulary and initial word vectors for different datasets (cpr/pgr/semeval). Take CPR as an walking example. You can repalce the cpr with other datasets:

python3 prepare_vocab.py dataset/cpr dataset/cpr --glove_dir dataset/glove

This will write vocabulary and word vectors as a numpy matrix into their corresponding dir dataset/cpr.

Training

To train the LFGCN model, run:

chmod +x train_cpr.sh; ./train_cpr.sh

Model checkpoints and logs will be saved to ./saved_models/cpr.

For details on the use of other parameters, please refer to train.py.

Evaluation

Our pretrained model is saved under the dir saved_models/cpr. To run evaluation on the test set, run:

python3 eval.py saved_model/cpr --data_dir dataset/cpr

Use --model checkpoint_epoch_100.pt to specify a model checkpoint file.

Related Repo

The paper uses the model DCGCN, for detail architecture please refer to the TACL19 paper Densely Connected Graph Convolutional Network for Graph-to-Sequence Learning. Codes are adapted from the repo of the AGGCN paper Attention Guided Graph Convolutional Networks for Relation Extraction.

Citation

@inproceedings{Guo2020LearningLF,
  title={Learning Latent Forests for Medical Relation Extraction},
  author={Zhijiang Guo and Guoshun Nan and Wei Lu and Shay B. Cohen},
  booktitle={Proc. of IJCAI},
  year={2020}
}