When reading and writing from various systems, it is not uncommon to encounter encoding issues when the systems have different locales. In this post I show several options for handling such issues.
Example
Say you have a field containing names and there’s a Czech name "Mořic" containing an r with caron, which you have to export to a csv with Windows-12521 encoding. This will fail:
>>> example = 'Mořic' >>> example.encode('WINDOWS-1252') UnicodeEncodeError: 'charmap' codec can't encode character '\u0159' in position 2: character maps to
Unfortunately, Windows-1252 does not support this character and thus an exception is raised, so we need a way to handle such encoding issues.
Encoding options
Since Python 3.32, the str
type is represented in Unicode. Unicode characters have no representation in bytes; this is what character encoding does – a mapping from Unicode characters to bytes. Each encoding handles the mapping differently, and not all encodings supports all Unicode characters, possibly resulting in issues when converting from one encoding to the other. Only the UTF family supports all Unicode characters. The most commonly used encoding is UTF-8, so stick with that whenever possible.
With str.encode
you have several error handling options. The default signature is str.encode(encoding="utf-8", errors="strict")
. Given the example "Mořic", the error options are:
Errors value | Description | Result |
---|---|---|
strict | Encoding errors raise a UnicodeError (default). | Exception |
ignore | Ignore erroneous characters. | Moic |
replace | Replace erroneous characters with ?. | Mo?ic |
xmlcharrefreplace | Replace erroneous characters with XML character reference. | Mořic |
backslashreplace | Replace erroneous characters with backslashed escape sequence. | Mo\\u0159ic |
namereplace | Replace erroneous characters with \N{...} escape sequence. | Mo\\N{LATIN SMALL LETTER R WITH CARON}ic |
Text normalisation
A nice alternative is to normalise the data first with unicodedata.normalize. The unicode standard defines some characters as composed from multiple other characters. For example, "ř" is composed of "r" (Latin small letter r (U+0072)) and "ˇ" (Combining caron (U+030C)). Not all character information websites show this information, but e.g. this page displays the composed characters and normalisation forms: https://chars.suikawiki.org/char/0159.
Normalisation can be applied in four forms:
Normal Form | Full Name |
---|---|
NFD | Normalisation Form Canonical Decomposition |
NFC | Normalisation Form Canonical Composition |
NFKD | Normalisation Form Compatibility Decomposition |
NFKC | Normalisation Form Compatibility Composition |
To understand Unicode normal forms, we need a bit of background information first.
Unicode and composed characters
In Unicode, characters are mapped to so-called code points. Every character in the Unicode universe3 is expressed by a code point written as U+ and four hexadecimal digits; e.g. U+0061 represents lowercase "a".
The Unicode standard provides two ways for specifying composed characters:
- Decomposed: as a sequence of combining characters
- Precomposed: as a single combined character
For example, the character "ã" (lowercase a with tilde) in decomposed form is given as U+0061 (a) U+0303 (˜), or in precomposed form as U+00E3 (ã).
Composition and Decomposition
Composition is the process of combining multiple characters to form a single character, typically a base character and one or more marks4. Decomposition is the reverse; splitting a composed character into multiple characters.
Before diving into normalisation, let’s define a function for printing the Unicode code points for each character in a string:
>>> def unicodes(string): >>> return ' '.join('U+{:04X}'.format(ord(c)) for c in string) >>> >>> example = 'Mořic' >>> print(unicodes(example)) U+004D U+006F U+0159 U+0069 U+0063
Canonical and Compatibility Equivalence
A problem arises when characters have multiple representations. For example the Ångström symbol Å (one Ångström unit equals one ten-billionth of a meter) can be represented in three ways:
U+212B U+00C5 U+0041 U+030A
How can we determine if strings are equal when their decomposed forms are different? Unicode equivalence is defined in two ways:
- Canonical equivalence
- Compatibility equivalence
When a character from different code points has the same appearance and meaning, it is considered canonically equivalent. For example all three representations of the Ångström example above have the same appearance and meaning, and are thus canonically equivalent.
Compatibility equivalence is defined as a sequence of code points which only have the same meaning, but are not equal visually. For example fractions are considered compatible equivalent: ¼ (U+00BC) and 1⁄4 (U+0031 U+2044 U+0034) do not have the same visual appearance, but do have the same meaning and are thus compatibility equivalent.
Compatibility equivalence is considered a weaker equivalence form and a subset of canonical equivalence. When a character is canonically equivalent, it is also compatibility equivalent, but not vice versa.
Applying Unicode normalisation forms
Now, with this background information, we can get to the Unicode normal forms. Given the example "Mořic" at the start, we can apply normalisation before encoding this string with Windows-1252:
>>> import unicodedata >>> >>> def unicodes(string): >>> return ' '.join('U+{:04X}'.format(ord(c)) for c in string) >>> >>> example = "Mořic" >>> >>> print(unicodes(example)) U+004D U+006F U+0159 U+0069 U+0063 # 5 Unicode code points, so the ř is given in precomposed form >>> example.encode("WINDOWS-1252") UnicodeEncodeError: 'charmap' codec cant encode character '\u0159' in position 2: character maps to undefined> # Windows-1252 cannot encode U+0159 (ř) >>> nfd_example = unicodedata.normalize("NFD", example) >>> print(unicodes(nfd_example)) U+004D U+006F U+0072 U+030C U+0069 U+0063 # 6 Unicode code points, so the ř is given in decomposed form >>> print(nfd_example) Mořic # Python shell with UTF-8 encoding still displays the r with caret >>> nfd_example.encode("WINDOWS-1252") UnicodeEncodeError: 'charmap' codec cant encode character '\u030c' in position 3: character maps to undefined> # Windows-1252 can now encode U+0072 (r), but not U+030C (ˇ) >>> print(nfd_example.encode('WINDOWS-1252', 'ignore')) Moric # Successfully encoded Windows-1252 and ignored U+030C (ˇ)
That’s it! With unicodedata.normalize("NFD", "Mořic").encode('WINDOWS-1252', 'ignore')
we can normalise first and then encode Windows-1252, ignoring the unknown characters for Windows-1252, resulting in Moric
. I like this alternative, usually people are okay with this since it doesn’t mingle the data too much and keeps it readable.
Improve your Python skills, learn from the experts!
At GoDataDriven we offer a host of Python courses from beginner to expert, taught by the very best professionals in the field. Join us and level up your Python game:
- Python Essentials – Great if you are just starting with Python.
- Data Science with Python Foundation – Want to make the step up from data analysis and visualization to true data science? This is the right course.
- Advanced Data Science with Python – Learn to productionize your models like a pro and use Python for machine learning.
References
- https://www.joelonsoftware.com/2003/10/08/the-absolute-minimum-every-software-developer-absolutely-positively-must-know-about-unicode-and-character-sets-no-excuses/
- https://medium.com/concerning-pharo/an-implementation-of-unicode-normalization-7c6719068f43
- http://unicode.org/reports/tr15/#Canon_Compat_Equivalence
- https://www.b-list.org/weblog/2017/sep/05/how-python-does-unicode
- Windows-1252 was the first default character set in Microsoft Windows and thus you will encounter it in lots of legacy Windows systems. The default for Windows systems nowadays is UTF-16. ↩
- Pre-Python 3.3,
str
was a byte string (sequence of bytes in certain encoding, default ASCII) andunicode
, a Unicode string. ↩ - More precisely, the „Unicode universe“ is the Unicode Character Database (UCD), which contains all Unicode characters and its properties and metadata. ↩
- Unicode categorises characters. Each category is denoted by an abbreviation of two letters, first an uppercase and second a lowercase letter. The uppercase letters shows the major category, the lowercase the minor category. The major category for marks is „M“. ↩