EmbeddingDiagnostic: Soft similarity over normalized textual serializations

This metric aims to evaluate near-miss similarity between generated and gold standard structural elements that fail strict matching. Each element (a triple or a TBox axiom) is rendered as a normalised text serialisation, encoded with a pretrained sentence-transformer, and compared by cosine similarity. The score is reported separately and does not affect the strict Precision, Recall, or F1 of the dimension.

Example:

Gold_Axiom: Animal SubClassOf eats some Plant

Generated_Axiom: Animal SubClassOf consumes some Vegetable

Result:

1. Strict_match = 0 (different property and class names)
2. EmbeddingDiagnostic ≈ 0.84 (close in meaning, flagged as a near miss)

What is being measured?

This metric helps diagnose how close a model came to a correct structural answer when it did not produce an exact match. It is a diagnostic for systematic deviations rather than a scoring signal, used to investigate why a generated element was rejected by strict comparison.