Knowledge graph error detection with unsupervised triplet network

Knowledge Graphs (KGs) are essential in various application areas. However, their quality is often compromised due to a large number of errors. Existing KG error detection methods have limitations in versatility and practicality. They mainly rely on supervision information such as entity types or error labels, which is not always easy to obtain in real scenarios. To address this issue, the paper proposes SymNet , an innovative framework for efficiently detecting errors in KG. SymNet utilises a triple embedding strategy, treating each triple as a node, and constructs a dynamic triple network through relational symmetric triples. To overcome the challenges posed by the special data characteristics and label scarcity of KG, a multi-layer information integration design is introduced. The BiLSTM module captures local-level intra-triplet translation information, while the graph attention network collects global-level inter-triplet contextual information. This multi-layer information integration strategy enables SymNet to more comprehensively understand the associated information in the knowledge graph, resulting in a significant improvement in error detection accuracy. The experimental results demonstrate that SymNet outperforms current state-of-the-art error detection algorithms in terms of both accuracy and efficiency when tested on two real-world knowledge graphs. Our research offers novel concepts and efficient solutions for addressing error detection issues in knowledge graphs, providing robust backing for enhancing the quality of knowledge graphs in practical scenarios.