Distance Normalization about levenshtein

[maxbachmann]

Levenshtein.ratio is a normalized version of the InDel similarity, which is a modified version of the Levenshtein distance which only allows Insertions + Deletions (or in terms of the Levenshtein distance gives Substitutions a weight of 2).

It uses the following steps for the calculation:

indel_dist = indel_distance(s1, s2)
norm_indel_dist = indel_dist / (len(s1) + len(s2))
norm_indel_sim = 1.0 - norm_indel_dist

It is unclear why the original author considered this a good idea, since this difference in algorithm between Levenshtein.distance and Levenshtein.ratio is very surprising for most users.

In case you actually want the normalized Levenshtein similarity, you can use:

from rapidfuzz.distance import Levenshtein
Levenshtein.normalized_similarity(s1, s2)

instead, which is implemented as:

lev_dist = Levenshtein.distance(s1, s2)
norm_lev_dist = lev_dist / max(len(s1), len(s2))
norm_lev_sim = 1.0 - norm_lev_dist

[kjijakli]

Thank you so much!