Most existing Levenshtein libraries are not very flexible: all edit operations have cost 1.
However, sometimes not all edits are created equal. For instance, if you are doing OCR correction, maybe substituting '0' for 'O' should have a smaller cost than substituting 'X' for 'O'. If you are doing human typo correction, maybe substituting 'X' for 'Z' should have a smaller cost, since they are located next to each other on a QWERTY keyboard.
This library supports all theses use cases, by allowing the user to specify different weights for edit operations involving every possible combination of letters. The core algorithms are written in Cython, which means they are blazing fast to run.
The Levenshtein distance function supports setting different costs for inserting characters, deleting characters, and substituting characters. Thus, Levenshtein distance is well suited for detecting OCR errors.
The Damerau-Levenshtein distance function supports setting different costs for inserting characters, deleting characters, substituting characters, and transposing characters. Thus, Damerau-Levenshtein distance is well suited for detecting human typos, since humans are likely to make transposition errors, while OCR is not.
Levenshtein distance: https://en.wikipedia.org/wiki/Levenshtein_distance and https://en.wikipedia.org/wiki/Wagner%E2%80%93Fischer_algorithm
Optimal String Alignment: https://en.wikipedia.org/wiki/Damerau%E2%80%93Levenshtein_distance#Optimal_string_alignment_distance
Damerau-Levenshtein distance: https://en.wikipedia.org/wiki/Damerau%E2%80%93Levenshtein_distance#Distance_with_adjacent_transpositions
pip install weighted-levenshtein
import numpy as np
from weighted_levenshtein import lev, osa, dam_lev
insert_costs = np.ones(128, dtype=np.float64) # make an array of all 1's of size 128, the number of ASCII characters
insert_costs[ord('D')] = 1.5 # make inserting the character 'D' have cost 1.5 (instead of 1)
# you can just specify the insertion costs
# delete_costs and substitute_costs default to 1 for all characters if unspecified
print(lev('BANANAS', 'BANDANAS', insert_costs=insert_costs)) # prints '1.5'
delete_costs = np.ones(128, dtype=np.float64)
delete_costs[ord('S')] = 0.5 # make deleting the character 'S' have cost 0.5 (instead of 1)
# or you can specify both insertion and deletion costs (though in this case insertion costs don't matter)
print(lev('BANANAS', 'BANANA', insert_costs=insert_costs, delete_costs=delete_costs)) # prints '0.5'
substitute_costs = np.ones((128, 128), dtype=np.float64) # make a 2D array of 1's
substitute_costs[ord('H'), ord('B')] = 1.25 # make substituting 'H' for 'B' cost 1.25
print(lev('HANANA', 'BANANA', substitute_costs=substitute_costs)) # prints '1.25'
# it's not symmetrical! in this case, it is substituting 'B' for 'H'
print(lev('BANANA', 'HANANA', substitute_costs=substitute_costs)) # prints '1'
# to make it symmetrical, you need to set both costs in the 2D array
substitute_costs[ord('B'), ord('H')] = 1.25 # make substituting 'B' for 'H' cost 1.25 as well
print(lev('BANANA', 'HANANA', substitute_costs=substitute_costs)) # now it prints '1.25'
transpose_costs = np.ones((128, 128), dtype=np.float64)
transpose_costs[ord('A'), ord('B')] = 0.75 # make swapping 'A' for 'B' cost 0.75
# note: now using dam_lev. lev does not support swapping, but osa and dam_lev do.
# See Wikipedia links for difference between osa and dam_lev
print(dam_lev('ABNANA', 'BANANA', transpose_costs=transpose_costs)) # prints '0.75'
# like substitution, transposition is not symmetrical either!
print(dam_lev('BANANA', 'ABNANA', transpose_costs=transpose_costs)) # prints '1'
# you need to explicitly set the other direction as well
transpose_costs[ord('B'), ord('A')] = 0.75 # make swapping 'B' for 'A' cost 0.75
print(dam_lev('BANANA', 'ABNANA', transpose_costs=transpose_costs)) # now it prints '0.75'
lev
, osa
, and dam_lev
are aliases for levenshtein
,
optimal_string_alignment
, and damerau_levenshtein
, respectively.
http://weighted-levenshtein.readthedocs.io/
- All string lookups are case sensitive.
- The costs parameters only accept numpy arrays, since the underlying Cython implementation relies on this for fast lookups. The numpy arrays are indexed using the
ord()
value of the characters. Thus, only the first 128 ASCII letters are accepted, anddict
andlist
are not accepted. Consequently, the strings must be strictlystr
objects, notunicode
. - This library is compatible with both Python 2 and Python 3 (see
tox.ini
for tested versions).
from weighted_levenshtein.clev cimport c_levenshtein as lev, c_optimal_string_alignment as osa, c_damerau_levenshtein as dam_lev
import numpy as np
a = np.ones(128, dtype=np.float64)
b = np.ones((128, 128), dtype=np.float64)
print(lev("BANANA", 4, "BANANAS", 5, a, a, b))
For the Cython API, functions are prefixed with a c_
with respect to the Python API. Also, the string parameters are followed by their length. The data types of the numpy arrays specifying the costs still need to be np.float64
, consistent with the Python API.
Function signatures below:
cdef double c_damerau_levenshtein(
unsigned char* str_a,
Py_ssize_t len_a,
unsigned char* str_b,
Py_ssize_t len_b,
double[::1] insert_costs,
double[::1] delete_costs,
double[:,::1] substitute_costs,
double[:,::1] transpose_costs) nogil
cdef double c_optimal_string_alignment(
unsigned char* word_m,
Py_ssize_t m,
unsigned char* word_n,
Py_ssize_t n,
double[::1] insert_costs,
double[::1] delete_costs,
double[:,::1] substitute_costs,
double[:,::1] transpose_costs) nogil
cdef double c_levenshtein(
unsigned char* word_m,
Py_ssize_t m,
unsigned char* word_n,
Py_ssize_t n,
double[::1] insert_costs,
double[::1] delete_costs,
double[:,::1] substitute_costs) nogil