MGRE

Code for paper "Multi-Granularity Neural Networks for Document-Level Relation Extraction".

Requirements

• python=3.7
• CUDA=10.0
• torch=1.7.0
• stanford-parser=4.2.0
• pytorch_transformers=1.2.0
• numpy=1.16.0
• nltk=3.6.1
• tqdm=4.59.0
• matplotlib=3.0.2
• scikit_learn=0.21.2

Preprocessing Data

• Download DocRED dataset
• Put the train_annotated.json, dev.json, test.json into the directory data/
• Put the vec.npy,word2id.json,rel2id.json,ner2id.json, char2id.json into the directory prepro_data/

>> python gen_data_tree.py
>> python gen_data.py         # for Glove
>> python gen_data_bert.py    # for BERT

Training

For Golve:

>> python train.py --model_name MRGE --save_name checkpoint_MRGE --train_prefix dev_train --test_prefix dev_dev

For BERT:

>> python train_bert.py --model_name MRGE_bert --save_name checkpoint_MRGE_bert --train_prefix dev_train --test_prefix dev_dev

Testing

For Golve:

>> python test.py --model_name MRGE --save_name checkpoint_MRGE --train_prefix dev_train --test_prefix dev_test

For BERT:

>> python test_bert.py --model_name MRGE_bert --save_name checkpoint_MRGE_bert --train_prefix dev_train --test_prefix dev_test

You will get json file named result.json, and then you can submit it to CodaLab.

Acknowledgement

We refer to the code of DocRED. Thanks for their contributions.