What is the best way to remove accents (normalize) in a Python unicode string?

Question

I have a Unicode string in Python, and I would like to remove all the accents (diacritics).

I found an elegant way to do this (in Java):

1. convert the Unicode string to its long normalized form (with a separate character for letters and diacritics)
2. remove all the characters whose Unicode type is "diacritic".

Do I need to install a library such as pyICU or is this possible with just the Python standard library? And what about Python 3?

I would like to avoid explicitly mapping characters

• Pat Myron: docs.python.org/library/unicodedata.html

Answer 1

Unidecode transliterates any unicode string into the closest possible representation in ascii text:

>>> from unidecode import unidecode
>>> unidecode('kožušček')
'kozuscek'
>>> unidecode('北亰')
'Bei Jing '
>>> unidecode('François')
'Francois'

• kolinko: depends what you're trying to achieve. for example I'm doing a search right now, and I don't want to transliterate greek/russian/chinese, I just want to replace "ą/ę/ś/ć" with "a/e/s/c"
• Karl Bartel: @EOL unidecode works for great for strings like "François", if you pass unicode objects to it. It looks like you tried with a plain byte string.
• Mikhail Korobov: Note that unidecode >= 0.04.10 (Dec 2012) is GPL. Use earlier versions or check github.com/kmike/text-unidecode if you need a more permissive license and can stand a slightly worse implementation.
• Mark Amery: @EOL You'll be pleased to know that in the latest version of the library, 'François' is mapped to 'Francois' as you'd expect.
• Eric Duminil: unidecode replaces ° with deg. It does more than just removing accents.

Answer 2

How about this:

import unicodedata
def strip_accents(s):
 return ''.join(c for c in unicodedata.normalize('NFD', s)
 if unicodedata.category(c) != 'Mn')

This works on greek letters, too:

>>> strip_accents(u"A \u00c0 \u0394 \u038E")
u'A A \u0394 \u03a5'
>>> 

The character category "Mn" stands for Nonspacing_Mark, which is similar to unicodedata.combining in MiniQuark's answer (I didn't think of unicodedata.combining, but it is probably the better solution, because it's more explicit).

And keep in mind, these manipulations may significantly alter the meaning of the text. Accents, Umlauts etc. are not "decoration".

• alexis: These are not composed characters, unfortunately--even though "ł" is named "LATIN SMALL LETTER L WITH STROKE"! You'll either need to play games with parsing unicodedata.name, or break down and use a look-alike table-- which you'd need for Greek letters anyway (Α is just "GREEK CAPITAL LETTER ALPHA").
• alexis: @andi, I'm afraid I can't guess what point you want to make. The email exchange reflects what I wrote above: Because the letter "ł" is not an accented letter (and is not treated as one in the Unicode standard), it does not have a decomposition.
• lenz: @alexis (late follow-up): This works perfectly well for Greek as well – eg. "GREEK CAPITAL LETTER ALPHA WITH DASIA AND VARIA" is normalised into "GREEK CAPITAL LETTER ALPHA" just as expected. Unless you are referring to transliteration (eg. "α" → "a"), which is not the same as "removing accents"...
• Art: Mostly works nice :) But it doesn't transform ß into ascii ss in example. I would still use unidecode to avoid accidents.
• Brambor: + for not requiring installing anything

Answer 3

I just found this answer on the Web:

import unicodedata
def remove_accents(input_str):
 nfkd_form = unicodedata.normalize('NFKD', input_str)
 only_ascii = nfkd_form.encode('ASCII', 'ignore')
 return only_ascii

It works fine (for French, for example), but I think the second step (removing the accents) could be handled better than dropping the non-ASCII characters, because this will fail for some languages (Greek, for example). The best solution would probably be to explicitly remove the unicode characters that are tagged as being diacritics.

Edit: this does the trick:

import unicodedata
def remove_accents(input_str):
 nfkd_form = unicodedata.normalize('NFKD', input_str)
 return u"".join([c for c in nfkd_form if not unicodedata.combining(c)])

unicodedata.combining(c) will return true if the character c can be combined with the preceding character, that is mainly if it's a diacritic.

Edit 2: remove_accents expects a unicode string, not a byte string. If you have a byte string, then you must decode it into a unicode string like this:

encoding = "utf-8" # or iso-8859-15, or cp1252, or whatever encoding you use
byte_string = b"café" # or simply "café" before python 3.
unicode_string = byte_string.decode(encoding)

• Jabba: I had to add 'utf8' to unicode: nkfd_form = unicodedata.normalize('NFKD', unicode(input_str, 'utf8'))
• MestreLion: @Jabba: , 'utf8' is a "safety net" needed if you are testing input in terminal (which by default does not use unicode). But usually you don't have to add it, since if you're removing accents then input_str is very likely to be utf8 already. It doesn't hurt to be safe, though.
• MiniQuark: @rbp: you should pass a unicode string to remove_accents instead of a regular string (u"é" instead of "é"). You passed a regular string to remove_accents, so when trying to convert your string to a unicode string, the default ascii encoding was used. This encoding does not support any byte whose value is >127. When you typed "é" in your shell, your O.S. encoded that, probably with UTF-8 or some Windows Code Page encoding, and that included bytes >127. I'll change my function in order to remove the conversion to unicode: it will bomb more clearly if a non-unicode string is passed.
• rbp: @MiniQuark that worked perfectly >>> remove_accents(unicode('é'))
• s29: This answer gave me the best result on a large data set, the only exception is "ð"- unicodedata wouldn't touch it!

Answer 4

Actually I work on project compatible python 2.6, 2.7 and 3.4 and I have to create IDs from free user entries.

Thanks to you, I have created this function that works wonders.

import re
import unicodedata
def strip_accents(text):
 """
 Strip accents from input String.
 :param text: The input string.
 :type text: String.
 :returns: The processed String.
 :rtype: String.
 """
 try:
 text = unicode(text, 'utf-8')
 except (TypeError, NameError): # unicode is a default on python 3 
 pass
 text = unicodedata.normalize('NFD', text)
 text = text.encode('ascii', 'ignore')
 text = text.decode("utf-8")
 return str(text)
def text_to_id(text):
 """
 Convert input text to id.
 :param text: The input string.
 :type text: String.
 :returns: The processed String.
 :rtype: String.
 """
 text = strip_accents(text.lower())
 text = re.sub('[ ]+', '_', text)
 text = re.sub('[^0-9a-zA-Z_-]', '', text)
 return text

result:

text_to_id("Montréal, über, 12.89, Mère, Françoise, noël, 889")
>>> 'montreal_uber_1289_mere_francoise_noel_889'

• Daniel Reis: With Py2.7, passing an already unicode string errors at text = unicode(text, 'utf-8'). A workaround for that was to addexcept TypeError: pass
• Toothpick Anemone: Is there someway to make sure than if the letter M is large (capitalized) in the input then M is also large (capitalized) in the output? The input string "Montréal" becomes the output string "Montreal".

Answer 5

This handles not only accents, but also "strokes" (as in ø etc.):

import unicodedata as ud
def rmdiacritics(char):
 '''
 Return the base character of char, by "removing" any
 diacritics like accents or curls and strokes and the like.
 '''
 desc = ud.name(char)
 cutoff = desc.find(' WITH ')
 if cutoff != -1:
 desc = desc[:cutoff]
 try:
 char = ud.lookup(desc)
 except KeyError:
 pass # removing "WITH ..." produced an invalid name
 return char

This is the most elegant way I can think of (and it has been mentioned by alexis in a comment on this page), although I don't think it is very elegant indeed. In fact, it's more of a hack, as pointed out in comments, since Unicode names are – really just names, they give no guarantee to be consistent or anything.

There are still special letters that are not handled by this, such as turned and inverted letters, since their unicode name does not contain 'WITH'. It depends on what you want to do anyway. I sometimes needed accent stripping for achieving dictionary sort order.

EDIT NOTE:

Incorporated suggestions from the comments (handling lookup errors, Python-3 code).

• janek37: You should catch the exception if the new symbol doesn't exist. For example there's SQUARE WITH VERTICAL FILL ▥, but there's no SQUARE. (not to mention that this code transforms UMBRELLA WITH RAIN DROPS ☔ into UMBRELLA ☂).
• matanox: This looks elegant in harnessing the semantic descriptions of characters that are available. Do we really need the unicode function call in there with python 3 though? I think a tighter regex in place of the find would avoid all the trouble mentioned in the comment above, and also, memoization would help performance when it's a critical code path.
• lenz: @matanster no, this is an old answer from the Python-2 era; the unicode typecast is no longer appropriate in Python 3. In any case, in my experience there is no universal, elegant solution to this problem. Depending on the application, any approach has its pros and cons. Quality-thriving tools like unidecode are based on hand-crafted tables. Some resources (tables, algorithms) are provided by Unicode, eg. for collation.
• mirek: I just repeat, what is above (py3): 1) unicode(char)->char 2) try: return ud.lookup(desc) except KeyError: return char
• lenz: @mirek you are right: since this thread is so popular, this answer deserves some updating/improving. I edited it.

Answer 6

In my view, the proposed solutions should NOT be accepted answers. The original question is asking for the removal of accents, so the correct answer should only do that, not that plus other, unspecified, changes.

Simply observe the result of this code which is the accepted answer. where I have changed "Málaga" by "Málagueña:

accented_string = u'Málagueña'
# accented_string is of type 'unicode'
import unidecode
unaccented_string = unidecode.unidecode(accented_string)
# unaccented_string contains 'Malaguena'and is of type 'str'

There is an additional change (ñ -> n), which is not requested in the OQ.

A simple function that does the requested task, in lower form:

def f_remove_accents(old):
 """
 Removes common accent characters, lower form.
 Uses: regex.
 """
 new = old.lower()
 new = re.sub(r'[àáâãäå]', 'a', new)
 new = re.sub(r'[èéêë]', 'e', new)
 new = re.sub(r'[ìíîï]', 'i', new)
 new = re.sub(r'[òóôõö]', 'o', new)
 new = re.sub(r'[ùúûü]', 'u', new)
 return new

• Martin Thoma: "correct answer should only do that, not that plus other, unspecified, changes" -> you make capital letters in lower case
• Just a nice guy: Well... your answer convert "M" into "m" (not requested by OQ)

Answer 7

gensim.utils.deaccent(text) from Gensim - topic modelling for humans:

'Sef chomutovskych komunistu dostal postou bily prasek'

Another solution is unidecode.

Note that the suggested solution with unicodedata typically removes accents only in some character (e.g. it turns 'ł' into '', rather than into 'l').

• lcieslak: deaccent still gives ł instead of l.
• Nuno André: You needn't to install NumPy and SciPy to get accents removed.
• Etienne Kintzler: thanks for gensim reference ! how does it compare to unidecode (in terms of speed or accuracy) ?
• Martin: Changes the "ñ" for "n" which you wouldn't want, at least if you're looking for removing accents in Spanish

Answer 8

In response to @MiniQuark's answer:

I was trying to read in a csv file that was half-French (containing accents) and also some strings which would eventually become integers and floats. As a test, I created a test.txt file that looked like this:

Montréal, über, 12.89, Mère, Françoise, noël, 889

I had to include lines 2 and 3 to get it to work (which I found in a python ticket), as well as incorporate @Jabba's comment:

import sys 
reload(sys) 
sys.setdefaultencoding("utf-8")
import csv
import unicodedata
def remove_accents(input_str):
 nkfd_form = unicodedata.normalize('NFKD', unicode(input_str))
 return u"".join([c for c in nkfd_form if not unicodedata.combining(c)])
with open('test.txt') as f:
 read = csv.reader(f)
 for row in read:
 for element in row:
 print remove_accents(element)

The result:

Montreal
uber
12.89
Mere
Francoise
noel
889

(Note: I am on Mac OS X 10.8.4 and using Python 2.7.3)

• MiniQuark: remove_accents was meant to remove accents from a unicode string. In case it's passed a byte-string, it tries to convert it to a unicode string with unicode(input_str). This uses python's default encoding, which is "ascii". Since your file is encoded with UTF-8, this would fail. Lines 2 and 3 change python's default encoding to UTF-8, so then it works, as you found out. Another option is to pass remove_accents a unicode string: remove lines 2 and 3, and on the last line replace element by element.decode("utf-8"). I tested: it works. I'll update my answer to make this clearer.
• aseagram: Nice edit, good point. (On another note: The real problem I've realised is that my data file is apparently encoded in iso-8859-1, which I can't get to work with this function, unfortunately!)
• MiniQuark: aseagram: simply replace "utf-8" with "iso-8859-1", and it should work. If you're on windows, then you should probably use "cp1252" instead.
• PM 2Ring: BTW, reload(sys); sys.setdefaultencoding("utf-8") is a dubious hack sometimes recommended for Windows systems; see stackoverflow.com/questions/28657010/… for details.

Answer 9

perfplot

import unicodedata
from random import choice
import perfplot
import regex
import text_unidecode
def remove_accent_chars_regex(x: str):
 return regex.sub(r'\p{Mn}', '', unicodedata.normalize('NFKD', x))
def remove_accent_chars_join(x: str):
 # answer by MiniQuark
 # https://stackoverflow.com/a/517974/7966259
 return u"".join([c for c in unicodedata.normalize('NFKD', x) if not unicodedata.combining(c)])
perfplot.show(
 setup=lambda n: ''.join([choice('Málaga François Phút Hơn 中文') for i in range(n)]),
 kernels=[
 remove_accent_chars_regex,
 remove_accent_chars_join,
 text_unidecode.unidecode,
 ],
 labels=['regex', 'join', 'unidecode'],
 n_range=[2 ** k for k in range(22)],
 equality_check=None, relative_to=0, xlabel='str len'
)

• mike rodent: Haha... amazing. All these bits and pieces did actually install. The script did actually run. The graph actually displayed. And it is very similar to yours. unidecode actually handles the Chinese characters. And none of the three comes up with the hilarious "FranASSois".

Answer 10

Here is a short function which strips the diacritics, but keeps the non-latin characters. Most cases (e.g., "à" -> "a") are handled by unicodedata (standard library), but several (e.g., "æ" -> "ae") rely on the given parallel strings.

Code

from unicodedata import combining, normalize
LATIN = "ä æ ǽ đ ð ƒ ħ ı ł ø ǿ ö œ ß ŧ ü "
ASCII = "ae ae ae d d f h i l o o oe oe ss t ue"
def remove_diacritics(s, outliers=str.maketrans(dict(zip(LATIN.split(), ASCII.split())))):
 return "".join(c for c in normalize("NFD", s.lower().translate(outliers)) if not combining(c))

NB. The default argument outliers is evaluated once and not meant to be provided by the caller.

Intended usage

As a key to sort a list of strings in a more "natural" order:

sorted(['cote', 'coteau', "crottez", 'crotté', 'côte', 'côté'], key=remove_diacritics)

Output:

['cote', 'côte', 'côté', 'coteau', 'crotté', 'crottez']

If your strings mix texts and numbers, you may be interested in composing remove_diacritics() with the function string_to_pairs() I give elsewhere.

Tests

To make sure the behavior meets your needs, take a look at the pangrams below:

examples = [
 ("hello, world", "hello, world"),
 ("42", "42"),
 ("你好,世界", "你好,世界"),
 (
 "Dès Noël, où un zéphyr haï me vêt de glaçons würmiens, je dîne d’exquis rôtis de bœuf au kir, à l’aÿ d’âge mûr, &cætera.",
 "des noel, ou un zephyr hai me vet de glacons wuermiens, je dine d’exquis rotis de boeuf au kir, a l’ay d’age mur, &caetera.",
 ),
 (
 "Falsches Üben von Xylophonmusik quält jeden größeren Zwerg.",
 "falsches ueben von xylophonmusik quaelt jeden groesseren zwerg.",
 ),
 (
 "Љубазни фењерџија чађавог лица хоће да ми покаже штос.",
 "љубазни фењерџија чађавог лица хоће да ми покаже штос.",
 ),
 (
 "Ljubazni fenjerdžija čađavog lica hoće da mi pokaže štos.",
 "ljubazni fenjerdzija cadavog lica hoce da mi pokaze stos.",
 ),
 (
 "Quizdeltagerne spiste jordbær med fløde, mens cirkusklovnen Walther spillede på xylofon.",
 "quizdeltagerne spiste jordbaer med flode, mens cirkusklovnen walther spillede pa xylofon.",
 ),
 (
 "Kæmi ný öxi hér ykist þjófum nú bæði víl og ádrepa.",
 "kaemi ny oexi her ykist þjofum nu baedi vil og adrepa.",
 ),
 (
 "Glāžšķūņa rūķīši dzērumā čiepj Baha koncertflīģeļu vākus.",
 "glazskuna rukisi dzeruma ciepj baha koncertfligelu vakus.",
 )
]
for (given, expected) in examples:
 assert remove_diacritics(given) == expected

Case-preserving variant

LATIN = "ä æ ǽ đ ð ƒ ħ ı ł ø ǿ ö œ ß ŧ ü Ä Æ Ǽ Đ Ð Ƒ Ħ I Ł Ø Ǿ Ö Œ ẞ Ŧ Ü "
ASCII = "ae ae ae d d f h i l o o oe oe ss t ue AE AE AE D D F H I L O O OE OE SS T UE"
def remove_diacritics(s, outliers=str.maketrans(dict(zip(LATIN.split(), ASCII.split())))):
 return "".join(c for c in normalize("NFD", s.translate(outliers)) if not combining(c))

• John J. Camilleri: This looks like a nice solution, but on Python 3 I get the error message ValueError: string keys in translate table must be of length 1. This limitation is clearly stated in the Python docs: docs.python.org/3/library/stdtypes.html#str.maketrans. Maybe you only tested with Python 2?
• Aristide: You're right, there was a "SS" in the (untested) LATIN string for the case-preserving variant. I have replaced it by "ẞ", namely LATIN CAPITAL LETTER SHARP S. Thanks!

Answer 11

Some languages have combining diacritics as language letters and accent diacritics to specify accent.

I think it is more safe to specify explicitly what diactrics you want to strip:

def strip_accents(string, accents=('COMBINING ACUTE ACCENT', 'COMBINING GRAVE ACCENT', 'COMBINING TILDE')):
 accents = set(map(unicodedata.lookup, accents))
 chars = [c for c in unicodedata.normalize('NFD', string) if c not in accents]
 return unicodedata.normalize('NFC', ''.join(chars))

Answer 12

There are already many answers here, but this was not previously considered: using sklearn

from sklearn.feature_extraction.text import strip_accents_ascii, strip_accents_unicode
accented_string = u'Málagueña®'
print(strip_accents_unicode(accented_string)) # output: Malaguena®
print(strip_accents_ascii(accented_string)) # output: Malaguena

This is particularly useful if you are already using sklearn to process text. Those are the functions internally called by classes like CountVectorizer to normalize strings: when using strip_accents='ascii' then strip_accents_ascii is called and when strip_accents='unicode' is used, then strip_accents_unicode is called.

More details

Finally, consider those details from its docstring:

Signature: strip_accents_ascii(s)
Transform accentuated unicode symbols into ascii or nothing
Warning: this solution is only suited for languages that have a direct
transliteration to ASCII symbols.

and

Signature: strip_accents_unicode(s)
Transform accentuated unicode symbols into their simple counterpart
Warning: the python-level loop and join operations make this
implementation 20 times slower than the strip_accents_ascii basic
normalization.

• mike rodent: This works ... but the sklearn module is absolutely huge. If there is a core Python solution (accepted answer), the only possible case for advocating this instead would be if you can prove it is MUCH faster. You don't even attempt to do that.
• Rodrigo Laguna: Agree. However, if you already have scikit as a dependency it is not a problem (in fact, it was my case)

Answer 13

If you are hoping to get functionality similar to Elasticsearch's asciifolding filter, you might want to consider fold-to-ascii, which is [itself]...

A Python port of the Apache Lucene ASCII Folding Filter that converts alphabetic, numeric, and symbolic Unicode characters which are not in the first 127 ASCII characters (the "Basic Latin" Unicode block) into ASCII equivalents, if they exist.

Here's an example from the page mentioned above:

from fold_to_ascii import fold
s = u'Astroturf® paté'
fold(s)
> u'Astroturf pate'
fold(s, u'?')
> u'Astroturf? pate'

EDIT: The fold_to_ascii module seems to work well for normalizing Latin-based alphabets; however unmappable characters are removed, which means that this module will reduce Chinese text, for example, to empty strings. If you want to preserve Chinese, Japanese, and other Unicode alphabets, consider using @mo-han's remove_accent_chars_regex implementation, above.