LevenshteinDistance: Character-level edit distance approximation

This metric aims to evaluate character-level similarity between a generated label and a gold-standard label using the normalised Levenshtein edit distance. The raw distance counts the minimum number of single-character insertions, deletions, and substitutions needed to transform one string into the other; the score is normalised to [0, 1] as 1 − distance / max(len_a, len_b).

Example:

Generate_Label_1: Wine

Generate_Label_2: Wines

Generate_Label_3: Liquor

Gold_Standard_Label: Wine

Result:

1. LevenshteinDistance(Generate_Label_1, Gold_Standard_Label) = 1.00
2. LevenshteinDistance(Generate_Label_2, Gold_Standard_Label) ≈ 0.80
3. LevenshteinDistance(Generate_Label_3, Gold_Standard_Label) ≈ 0.17

What is being measured?

This metric captures fine-grained character-level proximity between labels. It is well suited to detecting typos, plurals, and small spelling drifts, but is purely lexical and cannot detect semantic equivalence.