AggregatedTop3: Mean of the three highest non-hard similarity scores

This metric aims to combine the output of multiple similarity methods into a single robust alignment score. It assigns 1 directly when HardMatch returns 1, and otherwise takes the mean of the three highest scores produced by the four non-hard methods (SequenceMatch, LevenshteinDistance, JaroWinklerDistance, SemanticCosineSimilarity).

Example:

Generate_Label: has_name

Gold_Standard_Label: hasName

Per-method scores:

• HardMatch = 0
• SequenceMatch = 0.93
• LevenshteinDistance = 0.88
• JaroWinklerDistance = 0.97
• SemanticCosineSimilarity = 0.90

Result:

Top-3 of non-hard methods = {0.97, 0.93, 0.90}

AggregatedTop3 = (0.97 + 0.93 + 0.90) / 3 ≈ 0.93

What is being measured?

This metric mitigates the noise of any single matcher by averaging multiple independent similarity signals while still trusting an exact match when one is available. Final acceptance also requires that the aggregated score reaches the type-specific threshold (τ_C = 0.6 for classes, τ_P = 0.7 for properties).