Multilingual Knowledge Graph Completion with Self-Supervised Adaptive Graph Alignment (SS-AGA)

SS-AGA is a multilingual knowledge graph completion framework that transfers knowledge among multiple KGs sources based on limited seed entity alignment.

You can see our ACL 2022 paper “Multilingual Knowledge Graph Completion with Self-Supervised Adaptive Graph Alignment” for more details.

This implementation of SS-AGA is based on Pytorch Geometric API. Our code is built upon the github KEnS and we thank the authors' effort for making it public.

Data

DBP-5L: A Public dataset from https://github.com/stasl0217/KEnS.

E-PKG: A new industrial multilingual E-commerce product KG dataset.

Data format: Each dataset contains the following files and folders:

• entity: Folder that contains the entity list for each KG.
• kg: Folder that contains the KG triple list (head_entity_index, relation_index, tail_entity_index) for each kg.
• seed_alignlinks: Folder that contains seed entity alignment pair list between two KGs.
• relation.txt: File that contains relation that is shared across all KGs.
• entity_embeddings.npy: The numpy file of mbert embedding for each entity from all KGs. Size of [Num_entity_all, 768]. We use the BERT-Base, Multilingual Cased from https://github.com/google-research/bert/blob/master/multilingual.md to generate it. You can download the entity_embeddings.npy for ther DBP-5L dataset from here and for thhe E-PKG data from here .

To run the code, create the folders "dataset/dbp5l", "dataset/epkg" and download the two datasets respectively.

Setup

To run the code, you need the following dependencies:

• Python 3.6.10

• Pytorch 1.10.0

• pytorch_geometric 2.0.4

  • torch-cluster==1.6.0
  • torch-scatter==2.0.9
  • torch-sparse==0.6.13

• numpy 1.16.1

Usage

Execute the following scripts to train the model on the targeted japanese KG:

python run_model.py --target_language ja --use_default

There are some key options of this scrips:

• --target_language: The targeted KG to conduct the KG completion task.
• --num_hop: Number of hops for sampling neighbors for each node.
• --preserved_ratio : How many align links to preserve in learning alignment embeddings. The rest are served as masked alignments and we ask the model to recover them.
• --generation_freq: How many epochs to conduct new pair generation once.
• --use_default: Use the preset hyper-parameter combinations.

The details of other optional hyperparameters can be found in run_model.py.

Citation

Please consider citing the following paper when using our code for your application.

@inproceedings{SS-AGA,
  title={Multilingual Knowledge Graph Completion with Self-Supervised
Adaptive Graph Alignment},
  author={Zijie Huang and Zheng Li and Haoming Jiang and Tianyu Cao and Hanqing Lu and Bing Yin and Karthik Subbian and Yizhou Sun and Wei Wang},
  booktitle={Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2022}
}

License

This project is licensed under the Apache-2.0 License and CDLA-Permissive 2.0 license.