cranooooooo / kgrefiner

Codes and Results of KGRefiner paper.

KGRefiner

An Open-source Framework for Knowledge Graph Refinement.
The Paper: KGRefiner: Knowledge Graph Refinement for Improving Accuracy of Translational Link Prediction Methods

Section 1 : KGRefiner

To reproduce result look at the KGRefiner folder

Section 2 : Datasets

Datasets are suitable to run on OpenKE framework. However, you can find triples in train2id, valid2id, and test2id files.
We made the FB15K237-Refined from FB15K237 and WN18RR-Refined from WN18RR by our KGRefiner.

Section 3 : Reproducing paper's results

Section 3.1 : FB15k237

Base Line	H@10	MR	MRR	Link for reproduction
TransE	45.6	347	29.4	ConvKB paper
TransE + Shang et al	47.6	221	28.8	Notebook
TransE + KGRefiner	47	203	29.1	Notebook
TransD	45.3	256	28.6	HRS paper
TransD + Shang et al	48.2	227	28.5	Notebook
TransD + KGRefiner	43.7	227	24	Notebook
RotatE	47.4	185	29.7	Notebook
RotatE + Shang et al	43.8	218	27.3	Notebook
RotatE + KGRefiner	43.9	226	27.9	Notebook
TransH	36.6	311	21.1	HRS paper
TransH + Shang et al	47.7	237	28.2	Notebook
TransH + KGRefiner	48.9	221	30.2	Notebook

Section 3.2 : WN18RR

Base Line	H@10	MR	MRR	Link for reproduction
TransE	50.1	3384	22.6	Nguyen et al
TransE + KGRefiner	53.7	1125	22.2	Notebook
TransH	42.4	5875	18.6	HRS paper
TransH + KGRefiner	51.4	1534	20.8	Notebook
TransD	42.8	5482	18.5	HRS paper
TransD + KGRefiner	52.3	1348	21.4	Notebook
RotatE	54.7	4274	47.3	OpenKE paper
RotatE + KGRefiner	57.0	683	44.8	Notebook

completing codes (notebooks) for table is still on going.