KEnS

Resources and code for paper "Multilingual Knowledge Graph Completion via Ensemble Knowledge Transfer"

Install

Make sure your local environment has the following installed:

python==3.6
pandas
tensorflow==1.10.0
numpy==1.16.2
pandas==1.0.3

Install the dependents using:

pip install -r requirements.txt

Run the experiments

To train the model, use:

python ./run.py --knowledge_model rotate --target_language ja --use_default

You can use --knowledge_model transe to switch from KEnS(RotatE) to the KEnS(TransE).
--target_language could be set as ja, es, el, en, fr.
--use_default means to use the preset hyper-parameter combinations.
By default, the trained models are saved in $PROJECT_DIR$/trained_model/kens-$KNOWLEDGE_MODEL$-$DIM$/$TARGET_LANGUAGE$.

To set your own hyper-parameters:

python ./run.py --knowledge_model rotate --target_language ja -d 400 -b 2048 -lr 1e-2 --rotate_gamma 24 --reg_scale 1e-4 --base_align_step 5 --knowledge_step_ratio 20 --align_lr 1e-3

Download the pre-trained KEnS(RotatE) model (dimension=400) for Japanese KG: https://drive.google.com/file/d/1GJJmkStYuRVfTYXi1OvtuCwVflkKaqD0/view?usp=sharing

To test the trained model with the ensemble techniques, use:

python ./test.py --knowledge_model rotate --target_language ja --model_dir $TRAINED_MODEL_DIR$  -d $YOUR_MODEL_DIM$

Reference

Please refer to our paper:

Xuelu Chen, Muhao Chen, Changjun Fan, Ankith Uppunda, Yizhou Sun, Carlo Zaniolo. Multilingual Knowledge Graph Completion via Ensemble Knowledge T. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, 2020

@inproceedings{chen2020multilingual,
  title={Multilingual Knowledge Graph Completion via Ensemble Knowledge Transfer},
  author={Chen, Xuelu and Chen, Muhao and Fan, Changjun and Uppunda, Ankith and Sun, Yizhou and Zaniolo, Carlo},
  booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings},
  pages={3227--3238},
  year={2020}
}

stasl0217 / KEnS