New Method Measures Semantic Similarity Between Knowledge Graphs

Knowledge Graphs (KGs) are powerful tools for organizing relational data, but current embedding methods primarily focus on individual entities or triples, neglecting the semantic similarity between entire graphs or subgraphs. This new research introduces a method to assess graph-to-graph semantic similarity, which is crucial for understanding if two KGs represent the same underlying information, even if their structures differ. To achieve this, the researchers developed a semantic matching dataset by modifying text documents and extracting KGs, creating reliable ground-truth correspondences. They then evaluated various approaches, including text-based, structure-based, and KG embedding-based methods. Their proposed KG embedding-based scoring function, EmbPairSim, which uses maximal pairwise entity similarity, demonstrated superior performance. EmbPairSim achieved significantly higher Mean Reciprocal Rank (MRR) compared to Sentence-BERT, while using fewer parameters. This indicates that KG embedding representations can effectively and compactly capture graph-level semantic information, offering a more robust way to compare and understand complex knowledge structures.