Knowledge Graph Completion via Local Semantic Contexts

摘要

Knowledge graphs are playing an increasingly important role for many search tasks such as entity search, question answering, etc. Although there are millions of entities and thousands of relations in many existing knowledge graphs such as Preebase and DBpedia, they are still far from complete. Previous approaches to complete knowledge graphs are either factor decomposition based methods or machine learning based ones. We propose a complementary approach that estimates the likelihood of a triple existing based on similarity measure of entities and some common semantic patterns of the entities. Such a way of triple estimation is very effective which exploits the semantic contexts of entities. Experimental results demonstrate that our model achieves significant improvements on knowledge graph completion compared with the state-of-art techniques.