KAEG

Requirements

Packages listed below are required.

• python==3.10
• CUDA==12.1
• PyTorch==2.1.0
• spacy==3.7.5
• tqdm==4.65.0
• requests==2.32.3
• beautifulsoup4==4.12.3
• wikidata==0.8.1
• dgl-cu121
• transformers==4.45.2
• rich==13.7.1
• omegaconf==2.3.0
• hydra-core==1.3.2
• axial-attention==0.6.1
• opt_einsum==3.4.0

Datasets

Our experiments include the DocRED and Re-DocRED datasets. The expected file structure is as follows:

KEGA
├── dataset
│   ├── DocRED
│   |   ├── kg
│   |   │   ├── dev_graph.json
│   |   │   ├── test_graph.json
│   |   │   └── train_annotated_graph.json
│   │   ├── meta
│   │   ├── ref
│   │   ├── dev.json
│   │   ├── dev_coref.json
│   │   ├── test.json
│   │   ├── test_coref.json
│   │   ├── train_annotated.json
│   │   ├── train_annotated_coref.json
│   │   └── train_distant.json
│   └── Re-DocRED
│       ├── kg
│       │   ├── dev_revised_graph.json
│       │   ├── test_revised_graph.json
│       │   └── train_revised_graph.json
│       ├── meta
│       ├── ref
│       ├── dev_revised.json
│       ├── dev_revised_coref.json
│       ├── test_revised.json
│       ├── test_revised_coref.json
│       ├── train_distant.json
│       ├── train_revised.json
│       ├── train_revised_coref.json
│   ├── gen_coref.py
│   ├── gen_graph.py
│   ├── README.md
│   └── requirements.txt

Training

If your dataset folder does not contain the kg and xxx-coref.json files, we recommend that you first run the gen_coref. py and gen_graph. py files under the dataset file, which generate corresponding reference files and knowledge graphs based on the dataset file. After obtaining these files, start executing the training code.

DocRED

The corresponding parameter configuration for the model can be found in configs/train_docred.yaml. After configuring the model parameters, you only need to execute the following command:

python train_docred.py

Re-DocRED

The corresponding parameter configuration for the model can be found in configs/train.yaml. After configuring the model parameters, you only need to execute the following command:

python train.py

Evaluation

You need to adjust the corresponding parameter configuration file configs/xxx. yaml for the dataset, such as the model path. After completing the adjustment, you can execute the following command:

# Re-DocREd
python train.py

# DocRED
python train_docred.py