/PRGE

Error detection in Knowledge Graphs: Path Ranking, Embeddings or both?

Primary LanguageHTML

PRGE

Error detection in Knowledge Graphs: Path Ranking, Embeddings or both?

alt text

Introduction

Here we provide the code for the Path Ranking Guided Embeddings (PRGE) framework. This is a hybrid approach that utilizes confidence scores from a Path Ranking method to Train Knowledge Graph Embeddings.

Scripts

  • PRGE training script in src folder
    • Need to provide triples file and confidence scores file for training. Confidence scores can be obtained from any PRA method, here, we used PaTyBRED by Melo et al. (see here and here for details)
  • Evaluation notebooks for results (already run with provided results)

Data

We provide already trained embeddings (download from here) and additional files for 3 different datasets: 2 benchmark datasets (FB15K, WN18) and one custom dataset (Dementia, see here for more info). These embeddings can be directly used for the evaluation scripts. The additional files (triples/confidence scores) can be used to train the embeddings from scratch using the cpp script provided.