Paper link: https://arxiv.org/abs/2204.13957

The results of PIE powered TransE model is as follows.

It's recommended to use a machine with over 400G memory to reproduce the results.

The model implementation is based on PathCon. Thanks for their contributions.

cd entity_typing && sh run.sh DATA_PATH

The ${DTA_PATH} should conatin train_hrt.npy.

After model training and inference, the code will output the distributuon p(r|e) for each entity. As the dataset is too large, we save the results with 15 sparse matrices. Please use src/cat_npz.py to concatenate all the sparse matrices. cd src && python/cat_npz.py $SAVE_PATH This will save the $p(r|e)$ matrix for all entities at ${SAVE_PATH}/e2r_scores.npz.

The sampling implementation is based on dgl-ke. Thanks for their contributions.

cd candidate && sh run.sh DATA_PATH

The ${DATA_PATH} should contain folders, such as 'raw' and 'processed'. As we utilize multi processer to sampling the candidates and each processor saves the result individually, please use script src/cat.py to concatenate all the candidates to one file.

The KGE implementation is based on the official baseline code of WikiKG90M OGB. Thanks for their contributions.

cd wikikg90m-v2
sh install_dgl.sh
sh run.sh DATA_PATH SAVE_PATH
sh run_test.sh SAVE_PATH DATA_PATH VAL_CANDIDATE_PATH TEST_CANDIDATE_PATH