JHL-HUST / TITer

TITer

This is the data and coded for our EMNLP 2021 paper TimeTraveler: Reinforcement Learning for Temporal Knowledge Graph Forecasting

Qucik Start

Data preprocessing

This is not necessary, but can greatly shorten the experiment time.

python3 preprocess_data.py --data_dir data/ICEWS14

Dirichlet parameter estimation

If you use the reward shaping module, you need to do this step.

python3 mle_dirichlet.py --data_dir data/ICEWS14 --time_span 24

Train

you can run as following:

python3 main.py --data_path data/ICEWS14 --cuda --do_train --reward_shaping --time_span 24

Test

you can run as following:

python3 main.py --data_path data/ICEWS14 --cuda --do_test --IM --load_model_path xxxxx

Acknowledgments

model/dirichlet.py is from https://github.com/ericsuh/dirichlet

Cite

@inproceedings{Haohai2021TITer,
	title={TimeTraveler: Reinforcement Learning for Temporal Knowledge Graph Forecasting},
	author={Haohai Sun, Jialun Zhong, Yunpu Ma, Zhen Han, Kun He.},
	booktitle={EMNLP},
	year={2021}
}