Home

Awesome

Ruleformer: Context-aware Rule Mining over Knowledge Graph

This repository provides the source code & data of our paper: Ruleformer: Context-aware Rule Mining over Knowledge Graph (COLING 2022).

Requirement

The dataset should contain train.txt, valid.txt, test.txt in the format of h r t for each line, entities.txt listing all the entities shown in KG, relations.txt listing all the relations shown in KG.

Preprocess

Run the following command to generate subgraph on your dataset:

python transformer/dataset.py -data=DATA -maxN=MAXN -padding=PADDING -jump=JUMP

-data: string, the relative/absolute path of dataset

-maxN: integer, filter nodes whose degree exceed it

-padding: integer, cutting off too long sequence

-jump: integer, length of rule

Train the model

python translate.py -data=DATASET/DATA -jump=JUMP -padding=PADDING -batch_size=BATCH_SIZE -desc=DESC

-d_v, -n_head, -n_layers: integer, hyper parameters of transformers

-subgraph: [OPT] integer, to select another subgraph while train JUMP hops rule.

Decode the rules

python translate.py -data=DATASET/DATA -jump=JUMP -padding=PADDING -batch_size=BATCH_SIZE -desc=DESC -ckpt=CKPT -decode_rule

-ckpt: string, select which checkpoint to decode rules

-the_rel: float, relative threshold of the next relation

-the_rel_min: float, absolute threshold of the next relation

-the_all: float, absolute threshold of the whole rule

Demonstration

# preprocess
python transformer/dataset.py -data=umls -maxN=40 -padding=140 -jump=3

# train
python translate.py -data=DATASET/umls -jump=3 -padding=140 -batch_size=5 -epoch=50 -n_head=6 -d_v=64 -desc=umls -savestep=5

# decode rule
python translate.py -data=DATASET/umls -jump=3 -padding=140 -batch_size=5 -epoch=50 -n_head=6 -d_v=64 -desc=umls-rule -ckpt=EXPS/umls-j3-mul-XX/TranslatorXX.ckpt -decode_rule

Citation

If you find this code useful, please cite the following paper.

@inproceedings{xu-etal-2022-ruleformer,
    title = "Ruleformer: Context-aware Rule Mining over Knowledge Graph",
    author = "Xu, Zezhong  and
      Ye, Peng  and
      Chen, Hui  and
      Zhao, Meng  and
      Chen, Huajun  and
      Zhang, Wen",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.225",
    pages = "2551--2560",
}

Acknowledgement

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