BaryGraph Unveiled, Embedding Relationships Themselves

Existing vector search methods have a limitation in that they treat the relationship between two nodes simply as the proximity of two points, resulting in information loss. BaryGraph solves this problem by representing relationships (Edges) through an independent vector document called BaryEdge.

Core Mechanism:

• BaryEdge Embedding: Combines the quality of the relationship ($q$) and the contextual embedding of the relationship type ($v(type)$) to turn the relationship itself into a searchable document.
• MetaBary Hierarchical Structure: Recursively stacks two BaryEdges to form a higher-level abstraction stage called a MetaBary triad, enabling the discovery of structural connections between distantly related concepts.
• Efficient Traversal: By adopting a tree structure (Forest), traversal to the root is possible with a single $graphLookup, and there are no cycle-handling issues.

Validation Results:

• Tests on the SimLex-999 and WordSim-353 benchmarks showed that a structural metric (Relational Neighborhood Overlap) had a higher correlation with human similarity judgments ($\rho \text{ ≈ } 0.32\text{--}0.53$) than simple cosine similarity.