SATRE with Data-efficiency and Computational Efficiency

This code introduces Self-attention over Tree for Relation Extraction (SATRE) for the large scale sentence-level relation extraction task (TACRED).

See below for an overview of the model architecture:

Requirements

Our model was trained on two Nvidia GTX 1080Ti graphic cards.

• Python 3 (tested on 3.7.6)

• Pytorch (tested on 1.2.0)

• CUDA （tested on 10.2.89）

• tqdm

• unzip, wget (for downloading only)

Preparation

The code requires that you have access to the TACRED dataset (LDC license required). Once you have the TACRED data, please put the JSON files under the directory dataset/tacred.

First, download and unzip GloVe vectors:

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

Then prepare vocabulary and initial word vectors with:

python3 prepare_vocab.py dataset/tacred dataset/vocab --glove_dir dataset/glove

This will write vocabulary and word vectors as a numpy matrix into the dir dataset/vocab.

Training

To train the SATRE model, run:

bash train.sh 0 1

Model checkpoints and logs will be saved to ./saved_models/1.

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

Evaluation

Our trained model is saved under the dir saved_models/1. To run evaluation on the test set, run:

bash  eval.sh 0 1 test

Related Repo

Codes are adapted from the repo of the EMNLP18 paper Graph Convolution over Pruned Dependency Trees Improves Relation Extraction and the repo of the ACL19 paper Attention Guided Graph Convolutional Networks for Relation Extraction.