Abstract

Discovering precise and specific rules from knowledge graphs is regarded asan essential challenge, which can improve the performances of many downstreamtasks and even provide new ways to approach some Natural Language Processingresearch topics. In this paper, we provide a fundamental theory for knowledgegraph reasoning based on the ending anchored rules. Our theory provides precisereasons explaining why or why not a triple is correct. Then, we implement ourtheory by what we call the EARDict model. Results show that our EARDict modelsignificantly outperforms all the benchmark models on three large datasets ofknowledge graph completion. Especially, our model achieves a Hits@10 score of96.6 percent on WN18RR.

 

Quick Read (beta)