Skip to content

Semantic similarity and relatedness

Semantic similarity measures how closely related two concepts or entities are. In knowledge graphs, similarity can be computed from:

Path-based measures: - Shortest path: Distance between entities via the shortest connecting path - Transitive closure: Considering all reachable paths and their properties - Weighted metrics: Accounting for edge types, entity specificity, or rarity

Structural measures: - Common neighbors: Overlap in direct connections - Degrees: How general/specific entities are based on graph degree - Betweenness: Importance in connecting other entities

Vector/embedding-based measures: - Learned representations: Neural embeddings from random walks, graph neural networks - Distributional similarity: Computing relatedness from co-occurrence patterns

In the misinformation domain, semantic similarity informs: - Fact checking: Assessing if a claim's subject and object are sufficiently related for the claim to be plausible - Claim clustering: Grouping similar false claims to detect variations of the same misinformation - Knowledge graph reasoning: Finding alternative paths that support or refute a statement

Key papers in this wiki