Pytorch-based implementation of CAGED, and the description of the model and the results can be found in the paper: "Contrastive Knowledge Graph Error Detection".
All the required packages can be installed by running pip install -r requirements.txt
.
Data pre-processing:
Following the previous study[1,2], we employ three real-world datasets that are constructed with noisy triples to be 5%, 10% and 15% of the whole KGs based on the popular benchmarks, i.e. FB15k, WN18RR and NELL-995.
To replicate the experiments from our paper:
Train:
python Our_TopK%_RankingList.py --dataset "WN18RR" --mode train --anomaly_ratio 0.05 --mu 0.005 --lam 0.1
Test:
python Our_TopK%_RankingList.py --dataset "WN18RR" --mode test --anomaly_ratio 0.05 --mu 0.005 --lam 0.1
This repo is built upon the following work:
[1] "Triple trustworthiness measurement for knowledge graph".
https://github.com/TJUNLP/TTMF.git
[2] "Does william shakespeare REALLY write hamlet? knowledge representation learning with confidence".
https://github.com/thunlp/CKRL.git