A C extension module for fast computation of:

Levenshtein has a some overlap with difflib (SequenceMatcher). It supports only strings, not arbitrary sequence types, but on the other hand it's much faster.

It supports both normal and Unicode strings, but can't mix them, all arguments to a function (method) have to be of the same type (or its subclasses).


apply_edit(edit_operations, source_string, destination_string)

Apply a sequence of edit operations to a string.

In the case of editops, the sequence can be arbitrary ordered subset of the edit sequence transforming source_string to destination_string.


>>> e = editops('man', 'scotsman')
>>> apply_edit(e, 'man', 'scotsman')
>>> apply_edit(e[:3], 'man', 'scotsman')

The other form of edit operations, opcodes, is not very suitable for such a tricks, because it has to always span over complete strings, subsets can be created by carefully replacing blocks with `equal' blocks, or by enlarging `equal' block at the expense of other blocks and adjusting the other blocks accordingly.


>>> a, b = 'spam and eggs', 'foo and bar'
>>> e = opcodes(a, b)
>>> apply_edit(inverse(e), b, a)
'spam and eggs'
>>> e[4] = ('equal', 10, 13, 8, 11)
>>> apply_edit(e, a, b)
'foo and ggs'
distance(string1, string2)

Compute absolute Levenshtein distance of two strings.

Examples (it's hard to spell Levenshtein correctly):

>>> distance('Levenshtein', 'Lenvinsten')
>>> distance('Levenshtein', 'Levensthein')
>>> distance('Levenshtein', 'Levenshten')
>>> distance('Levenshtein', 'Levenshtein')

Yeah, we've managed it at last.

editops(source_string, destination_string)
editops(edit_operations, source_length, destination_length)

Find sequence of edit operations transforming one string to another.

The result is a list of triples (operation, spos, dpos), where operation is one of `equal', `replace', `insert', or `delete'; spos and dpos are position of characters in the first (source) and the second (destination) strings. These are operations on signle characters. In fact the returned list doesn't contain the `equal', but all the related functions accept both lists with and without `equal's.


>>> editops('spam', 'park')
[('delete', 0, 0), ('insert', 3, 2), ('replace', 3, 3)]

The alternate form editops(opcodes, source_string, destination_string) can be used for conversion from opcodes (5-tuples) to editops (you can pass strings or their lengths, it doesn't matter).

hamming(string1, string2)

Compute Hamming distance of two strings.

The Hamming distance is simply the number of differing characters. That means the length of the strings must be the same.


>>> hamming('Hello world!', 'Holly grail!')
>>> hamming('Brian', 'Jesus')

Invert the sense of an edit operation sequence.

In other words, it returns a list of edit operations transforming the second (destination) string to the first (source). It can be used with both editops and opcodes.


>>> inverse(editops('spam', 'park'))
[('insert', 0, 0), ('delete', 2, 3), ('replace', 3, 3)]
>>> editops('park', 'spam')
[('insert', 0, 0), ('delete', 2, 3), ('replace', 3, 3)]
jaro(string1, string2)

Compute Jaro string similarity metric of two strings.

The Jaro string similarity metric is intended for short strings like personal last names. It is 0 for completely different strings and 1 for identical strings.


>>> jaro('Brian', 'Jesus')
>>> jaro('Thorkel', 'Thorgier')
>>> jaro('Dinsdale', 'D')
jaro_winkler(string1, string2[, prefix_weight])

Compute Jaro string similarity metric of two strings.

The Jaro-Winkler string similarity metric is a modification of Jaro metric giving more weight to common prefix, as spelling mistakes are more likely to occur near ends of words.

The prefix weight is inverse value of common prefix length sufficient to consider the strings `identical'. If no prefix weight is specified, 1/10 is used.


>>> jaro_winkler('Brian', 'Jesus')
>>> jaro_winkler('Thorkel', 'Thorgier')
>>> jaro_winkler('Dinsdale', 'D')
>>> jaro_winkler('Thorkel', 'Thorgier', 0.25)
matching_blocks(edit_operations, source_length, destination_length)

Find identical blocks in two strings.

The result is a list of triples with the same meaning as in SequenceMatcher's get_matching_blocks() output. It can be used with both editops and opcodes. The second and third arguments don't have to be actually strings, their lengths are enough.


>>> a, b = 'spam', 'park'
>>> matching_blocks(editops(a, b), a, b)
[(1, 0, 2), (4, 4, 0)]
>>> matching_blocks(editops(a, b), len(a), len(b))
[(1, 0, 2), (4, 4, 0)]

The last zero-length block is not an error, but it's there for compatibility with difflib which always emits it.

One can join the matching blocks to get two identical strings:

>>> a, b = 'dog kennels', 'mattresses'
>>> mb = matching_blocks(editops(a,b), a, b)
>>> ''.join([a[x[0]:x[0]+x[2]] for x in mb])
>>> ''.join([b[x[1]:x[1]+x[2]] for x in mb])
median(string_sequence[, weight_sequence])

Find an approximate generalized median string using greedy algorithm.

You can optionally pass a weight for each string as the second argument. The weights are interpreted as item multiplicities, although any non-negative real numbers are accepted. Use them to improve computation speed when strings often appear multiple times in the sequence.


>>> median(['SpSm', 'mpamm', 'Spam', 'Spa', 'Sua', 'hSam'])
>>> fixme = ['Levnhtein', 'Leveshein', 'Leenshten',
...          'Leveshtei', 'Lenshtein', 'Lvenstein',
...          'Levenhtin', 'evenshtei']
>>> median(fixme)

Hm. Even a computer program can spell Levenshtein better than me.

median_improve(string, string_sequence[, weight_sequence])

Improve an approximate generalized median string by perturbations.

The first argument is the estimated generalized median string you want to improve, the others are the same as in median(). It returns a string with total distance less or equal to that of the given string.

Note this is much slower than median(). Also note it performs only one improvement step, calling median_improve() again on the result may improve it further, though this is unlikely to happen unless the given string was not very similar to the actual generalized median.


>>> fixme = ['Levnhtein', 'Leveshein', 'Leenshten',
...          'Leveshtei', 'Lenshtein', 'Lvenstein',
...          'Levenhtin', 'evenshtei']
>>> median_improve('spam', fixme)
>>> median_improve(median_improve('spam', fixme), fixme)

It takes some work to change spam to Levenshtein.

opcodes(source_string, destination_string)
opcodes(edit_operations, source_length, destination_length)

Find sequence of edit operations transforming one string to another.

The result is a list of 5-tuples with the same meaning as in SequenceMatcher's get_opcodes() output. But since the algorithms differ, the actual sequences from Levenshtein and SequenceMatcher may differ too.


>>> for x in opcodes('spam', 'park'):
...     print x
('delete', 0, 1, 0, 0)
('equal', 1, 3, 0, 2)
('insert', 3, 3, 2, 3)
('replace', 3, 4, 3, 4)

The alternate form opcodes(editops, source_string, destination_string) can be used for conversion from editops (triples) to opcodes (you can pass strings or their lengths, it doesn't matter).

quickmedian(string[, weight_sequence])

Find a very approximate generalized median string, but fast.

See median() for argument description.

This method is somewhere between setmedian() and picking a random string from the set; both speedwise and quality-wise.


>>> fixme = ['Levnhtein', 'Leveshein', 'Leenshten',
...          'Leveshtei', 'Lenshtein', 'Lvenstein',
...          'Levenhtin', 'evenshtei']
>>> quickmedian(fixme)
ratio(string1, string2)

Compute similarity of two strings.

The similarity is a number between 0 and 1, it's usually equal or somewhat higher than difflib.SequenceMatcher.ratio(), becuase it's based on real minimal edit distance.


>>> ratio('Hello world!', 'Holly grail!')
>>> ratio('Brian', 'Jesus')

Really? I thought there was some similarity.

seqratio(string_sequence1, string_sequence2)

Compute similarity ratio of two sequences of strings.

This is like ratio(), but for string sequences. A kind of ratio() is used to to measure the cost of item change operation for the strings.


>>> seqratio(['newspaper', 'litter bin', 'tinny', 'antelope'],
...          ['caribou', 'sausage', 'gorn', 'woody'])
setmedian(string[, weight_sequence])

Find set median of a string set (passed as a sequence).

See median() for argument description.

The returned string is always one of the strings in the sequence.


>>> setmedian(['ehee', 'cceaes', 'chees', 'chreesc',
...            'chees', 'cheesee', 'cseese', 'chetese'])

You haven't asked me about Limburger, sir.

setratio(string_sequence1, string_sequence2)

Compute similarity ratio of two strings sets (passed as sequences).

The best match between any strings in the first set and the second set (passed as sequences) is attempted. I.e., the order doesn't matter here.


>>> setratio(['newspaper', 'litter bin', 'tinny', 'antelope'],
...          ['caribou', 'sausage', 'gorn', 'woody'])

No, even reordering doesn't help the tinny words to match the woody ones.

subtract_edit(edit_operations, subsequence)

Subtract an edit subsequence from a sequence.

The result is equivalent to editops(apply_edit(subsequence, s1, s2), s2), except that is constructed directly from the edit operations. That is, if you apply it to the result of subsequence application, you get the same final string as from application complete edit_operations. It may be not identical, though (in amibuous cases, like insertion of a character next to the same character).

The subtracted subsequence must be an ordered subset of edit_operations.

Note this function does not accept difflib-style opcodes as no one in his right mind wants to create subsequences from them.


>>> e = editops('man', 'scotsman')
>>> e1 = e[:3]
>>> bastard = apply_edit(e1, 'man', 'scotsman')
>>> bastard
>>> apply_edit(subtract_edit(e, e1), bastard, 'scotsman')