CGRE

This repository provides the implemention for the paper Distantly-Supervised Long-Tailed Relation Extraction Using Constraint Graphs.

Please cite our paper if our datasets or code is helpful to you ~ 😊

@article{liang2022distantly,
  title={Distantly-Supervised Long-Tailed Relation Extraction Using Constraint Graphs},
  author={Liang, Tianming and Liu, Yang and Liu, Xiaoyan and Sharma, Gaurav and Guo, Maozu},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2022},
  publisher={IEEE}
}

Requirements

• Python 3.6+
• Pytorch 1.4.0+
• PyTorch Geometric (see https://github.com/rusty1s/pytorch_geometric for detail)
• Transformers

Dataset

We provide three processed datasets: NYT-520K, NYT-570K and GDS. Download the datasets and pretrained word embeddings from here, and extract them in data/.

Training & Evaluation

Vanilla CGRE consists of PCNN and GCN, but we also provide some different choices of backbone models: CNN, PCNN and Bert for sentence encoding, and GCN, GAT and SAGE for graph encoding.

For example, you can try CNN+GAT on NYT-520K by the following command:

python train.py --config 520K_CNN_GAT.yaml

and

python eval.py --config 520K_CNN_GAT.yaml

Please see configuration files in config/ for more options.

Results

PR curves in our paper are stored in Curves/.

Data Format

train.json & test.json

[
    {
        "text": "he is a son of vera and william lichtenberg of belle_harbor , queens .",
        "sub": {"id": "m.0ccvx", "name": "queens", "type": "GPE"},
        "obj": {"id": "m.05gf08", "name": "belle_harbor", "type": "GPE"},
        "rel": "/location/location/contains"
    },
    ...
]

rel2id.json

{
    "NA": 0,
    "/location/neighborhood/neighborhood_of": 1,
    ...
}

type2id.json

{
    "NONE": 0,
    "CARDINAL": 1,
    ...
}

constraint_graph.json

{
    relation_1: [[head_type_1, tail_type_1], [head_type_2, tail_type_2], ...],
    ...
}