Search to Pass Messages for Temporal Knowledge Graph Completion

---

Overview

This repository contains the code for Search to Pass Messages for Temporal Knowledge Graph Completion in Findings of EMNLP 2022.

Requirements

python=3.8
torch==1.9.0+cu111
dgl+cu111==0.6.1

Instructions to run the experiment

Search process

# Random baseline
python main.py --train_mode search --search_mode random --encoder SPATune --max_epoch 200

# SPA supernet training
python main.py --train_mode search --search_mode spos --encoder SPASPOSSearch --search_max_epoch 800

# SPA architecture search
python main.py --train_mode search --search_mode spos_search --encoder SPASPOSSearch --arch_sample_num 1000 --weight_path <xx.pt>

Fine-tuning process

python main.py --train_mode tune --encoder SPATune --search_res_file <xxx.json>

Citation

Readers are welcomed to fork this repository to reproduce the experiments and follow our work. Please kindly cite our paper

@inproceedings{wang2022search,
    title={Search to Pass Messages for Temporal Knowledge Graph Completion},
    author={Wang, Zhen and Du, Haotong and Yao, Quanming and Li, Xuelong},
    booktitle={Findings of the Association for Computational Linguistics: EMNLP 2022},
    pages={6160--6172},
    year={2022}
}

Contact

If you have any questions, feel free to contact me at [email protected].

Acknowledgement

The codes of this paper are partially based on the codes of SANE and TeMP. We thank the authors of above work.