SemanticCosineSimilarity: Semantic distance over dense vector representations

This metric aims to evaluate semantic similarity between a generated label and a gold-standard label by encoding each label with a pretrained sentence-transformer and computing the cosine similarity between their vector representations. The score lies in [-1, 1], with values close to 1 indicating high semantic equivalence.

Example:

Generate_Label_1: Carnivore

Generate_Label_2: madeFromGrape

Generate_Label_3: Software

Gold_Standard_Label_1: Predator

Gold_Standard_Label_2: usedToMake

Gold_Standard_Label_3: Wine

Result:

1. SemanticCosineSimilarity(Generate_Label_1, Gold_Standard_Label_1) ≈ 0.85
2. SemanticCosineSimilarity(Generate_Label_2, Gold_Standard_Label_2) ≈ 0.78
3. SemanticCosineSimilarity(Generate_Label_3, Gold_Standard_Label_3) ≈ 0.10

What is being measured?

This metric captures semantic equivalence between labels even when their surface forms differ. It is the strongest single contributor to property-level matching in pilot evaluation, because it correctly identifies paraphrases and synonyms that purely string-based methods miss.