Source code for PAKDD 2021 paper Multi-Grained Dependency Graph Neural Network for Chinese Open Information Extraction.
Download character embeddings gigaword_chn.all.a2b.uni.ite50.vec (Google Drive or Baidu Pan) and word embeddings sgns.merge.word (Google Drive or Baidu Pan).
Change utils/config.py line 34 and 35 to your word and character embedding file path.
Run the following command to train a predicate extraction model:
CUDA_VISIBLE_DEVICES=0 python train_predicate_span.py --train_file data/train.json --dev_file data/dev.json --test_file data/test.json --gat_nhead 5 --gat_layer 3 --strategy n --batch_size 50 --lr 0.001 --lr_decay 0.01 --use_clip False --optimizer Adam --droplstm 0 --dropout 0.6 --dropgat 0.3 --gaz_dropout 0.4 --norm_char_emb True --norm_gaz_emb False --param_stored_directory ./logs/predicate --lstm_layer 1 --gat_nhidden 60 --data_stored_directory ./logs/generated_data_predicate/ --positive_weight 3
To train an argument extraction model, run the following command:
CUDA_VISIBLE_DEVICES=0 python train_argument.py --train_file data/train.json --dev_file data/dev.json --test_file data/test.json --gat_nhead 5 --gat_layer 3 --strategy n --batch_size 50 --lr 0.001 --lr_decay 0.01 --use_clip False --optimizer Adam --droplstm 0 --dropout 0.6 --dropgat 0.3 --gaz_dropout 0.4 --norm_char_emb True --norm_gaz_emb False --param_stored_directory ./logs/argument --lstm_layer 1 --gat_nhidden 60 --data_stored_directory ./logs/generated_data_argument/
Change predict_argument.py line 132 to best argument model path, then run the following commands:
CUDA_VISIBLE_DEVICES=0 python predict_argument.py --gold_file data/test.json --rel_pred_file logs/evaluate/detailed_predicate_predictions.json --gat_nhead 5 --gat_layer 2 --strategy n --batch_size 50 --norm_char_emb True --norm_gaz_emb False --param_stored_directory ./logs/argument --lstm_layer 1 --gat_nhidden 60 --data_stored_directory ./logs/generated_data_argument/
python utils/eval_joint.py --gold_file data/test.json --pred_file ./logs/evaluate/detailed_argument_predictions.json --output_dir ./logs/evaluate
If the code helps you, please cite our paper:
@InProceedings{lyu2021MGDGNN,
author="Lyu, Zhiheng
and Shi, Kaijie
and Li, Xin
and Hou, Lei
and Li, Juanzi
and Song, Binheng",
title="Multi-Grained Dependency Graph Neural Network for Chinese Open Information Extraction",
booktitle="Advances in Knowledge Discovery and Data Mining",
year="2021",
publisher="Springer International Publishing",
address="Cham",
pages="155--167",
}This repo is adapted from Graph4CNER. We thank them for their work.