Levenshtein realisation counts substitution as 2 edits instead of 1 about levenshtein

[maxbachmann]

The documentation was not super clear on this. ratio calculates a similarity based on the indel distance which only allows insertions and deletions (behaves the same as levenshtein with substitution weight 2).

distance calculates the normal Levenshtein distance, so I understand why this is surprising. This is done for legacy reason, since this has always been the behavior in this library. I would recommend you to use rapidfuzz instead which is used here internally as well:

>> from rapidfuzz.distance import Levenshtein
>> Levenshtein.normalized_similarity("abcde", "abcd")
0.8
>>> Levenshtein.normalized_similarity("abcde", "abcde 2")
0.7142857142857143

while indel gives you:

>>> from rapidfuzz.distance import Indel
>>> Indel.normalized_similarity("abcde", "abcd")
0.8888888888888888
>>> Indel.normalized_similarity("abcde", "abcde 2")
0.8333333333333334

Note that uniform levenshtein is normalized using max(len1, len2), while the indel distance is normalized using len1 + len2. The reason for this is that it uses the maximum possible score for normalization. For Indel that is removing all characters and then inserting all of the other string, while for Levenshtein it's replacing the common part and removing/inserting the length difference.

[4ilin]

Thank you for your prompt reply, Max. That's very clear. As deletion counts for 2, it all adds up.