Textual Similarity vs. Structural Relationship

Concepts that are textually similar (high cosine similarity in embedding space) are not necessarily structurally related in an ontology. And concepts that are structurally related may have no textual similarity at all.

A semantic similarity search finds concepts whose descriptions match your query. It doesn’t find concepts that are connected in meaningful ways through the ontology’s relationship structure.

This is why embedding-based extraction can identify relevant-sounding terms but miss the actual patterns encoded in class hierarchies, property constraints, and logical axioms.

Structural relationship requires traversing the graph. Textual similarity only requires comparing vectors.

Related: 07-molecule—vectors-vs-graphs