GraphIE: This is the code for word-level GCN

This repo contains the word-level GCN code and data of the paper:

GraphIE: A Graph-Based Framework for Information Extraction. Yujie Qian, Enrico Santus, Zhijing Jin, Jiang Guo, and Regina Barzilay.

This code is implemented by Zhijing Jin. The original framework is accreditted to Di Jin.

For any questions, please contact her via email. (Another part of the code in the paper, i.e. sentence-level GCN is in this github repo.

Get Started

1) Install packages

• Python>=3.6
• PyTorch 0.4.0 # install it according to your cuda version. e.g. conda install pytorch=0.4.0 torchvision cuda80 -c pytorch

conda create -n graphie_env python=3.6
conda activate graphie_env
pip install -r requirements.txt
# install pytorch 0.4.0 in your own way

2) Download data

./download_data.sh # (1) download and preprocess Conll2003; (2) download Glove embeddings

Note: If you need to preprocess new data sources, please see

• Small data samples are provided.
• If have any questions regarding preprocess.py, you can contact the author by email.

Run

python examples/multi_runs_conll.py --gpu_id 0

Model Outputs

Overview

To ensure the files are correct, I did the following checks on Apr 2, 2019.

Output Files

Specs

• You can evaluate the outputs by running the official perl script to calculate NER F1 scores:

code/examples/eval/conll03eval.v2 < outputs/conll03_SeqIE/11141152r09_dev

• In each output file, the format is index \t word \t gold_tag \t predicted_tag. The start of a document is signified by 1 -DOCSTART- O O.
• The dev output is saved in the last epoch in the training.
• The test output is saved in the epoch with the highest F1 score on the dev data.
• To get a better understanding of how the outputs are saved, please refer to the main code file.

Appendix: Full experiment details

conll03 - GraphIE

To show a robust result, we run the code multiple times and gathered the following outputs:

conll03 - SeqIE

To show a robust result, we run the code multiple times and gathered the following outputs: