This is a release code for paper "Multi-Relation Extraction via A Global-Local Graph Convolutional Network".
- Python 3.8
- Pytorch 1.4
- Spacy
train.json: link_train
test.json: link_test
Please place it in ./data/nyt/ and modify the parameters of train.py and eval.py.
Please ensure that Spacy is installed before starting.
pip install -U pip setuptools wheel
pip install -U spacy
python -m spacy download en_core_web_md
You can also consider installing the GPU version of Spacy, please refer to official website of Spacy.
The dataset we use are NYT(public) and TACRED, and the data format is as follows:
{
"sentence": ["Not", "many", "people", "have", "cooler", "family", "closets", "to", "raid", "than", "Theodora", "Richards", ",", "the", "daughter", "of", "the", "Rolling", "Stones", "guitarist", "Keith", "Richards", "and", "the", "70", "'s", "supermodel", "Patti", "Hansen", "."],
"subj_start": 27,
"subj_end": 28,
"obj_start": 10,
"obj_end": 11,
"relation": "/people/person/children",
"edges": [[2, 0], [2, 1], [3, 2], [3, 6], [3, 7], [3, 9], [3, 29], [6, 4], [6, 5], [7, 8], [9, 11], [11, 10], [11, 12], [11, 14], [14, 13], [14, 15], [15, 21], [18, 17], [19, 16], [19, 18], [21, 19], [21, 20], [21, 22], [21, 26], [24, 23], [24, 25], [26, 24], [26, 28], [28, 27]]
}
~~~~~
{
"sentence": Original sentence after participle,
"subj_start": The start position of the first entity in a relation,
"subj_end": The end position of the first entity in a relation,
"obj_start": The start position of the second entity in a relation,
"obj_end": The end position of the second entity in a relation,
"relation": The relation type (including no_relation),
"edges": The edges in adjancy matrix.
}
If a sentence has multiple relations, it will appear multiple times. This preprocessing method is mainly for the convenience of training, you can also customize the preprocessing (by modifying the file in /GCN_RE/utils).
Because TACRED is not a fully public dataset (access fee is required), we will not upload pre-processed TACRED dataset. While we will upload the pre-processing results of NYT.
Scripts for training and testing:
#train
python train.py
#dev&test
python eval.py