CJK support for textwrap

@birkenfeld

BPO	24665
Nosy	@birkenfeld, @terryjreedy, @vstinner, @ezio-melotti, @bitdancer, @methane, @fgallaire, @serhiy-storchaka, @yan12125, @JulienPalard, @xi
PRs	bpo-24665: double-width CJK chars support for textwrap #89 bpo-24665: Add CJK support in textwrap by default. #5649 bpo-24665: Add CJK support in textwrap by default. #5649 gh-56708: support custom len function in textwrap.wrap #28136
Files	CJK.patch CJK+fix.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2018-07-09.08:58:24.285>
created_at = <Date 2015-07-19.02:44:39.415>
labels = ['3.8', 'type-feature', 'library', 'expert-unicode']
title = 'CJK support for textwrap'
updated_at = <Date 2021-09-03.08:13:33.576>
user = 'https://github.com/fgallaire'

bugs.python.org fields:

activity = <Date 2021-09-03.08:13:33.576>
actor = 'xi2'
assignee = 'none'
closed = True
closed_date = <Date 2018-07-09.08:58:24.285>
closer = 'methane'
components = ['Library (Lib)', 'Unicode']
creation = <Date 2015-07-19.02:44:39.415>
creator = 'fgallaire'
dependencies = []
files = ['39949', '40451']
hgrepos = []
issue_num = 24665
keywords = ['patch']
message_count = 24.0
messages = ['246930', '246932', '247001', '247008', '247009', '247016', '247022', '247025', '276275', '284068', '287736', '287753', '287756', '287820', '287821', '287822', '288785', '297563', '312076', '321283', '321284', '321291', '321304', '323733']
nosy_count = 11.0
nosy_names = ['georg.brandl', 'terry.reedy', 'vstinner', 'ezio.melotti', 'r.david.murray', 'methane', 'fgallaire', 'serhiy.storchaka', 'yan12125', 'mdk', 'xi2']
pr_nums = ['89', '5649', '5649', '28136']
priority = 'normal'
resolution = 'rejected'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue24665'
versions = ['Python 3.8']

This is new feature and can be added only in 3.6.

bpo-12499 looks related. See also bpo-12568.

Bad wrapping of CJK chars is a bug.
I don't understand why Python2 should be broken forever!

Because to get proper unicode support, we wrote python3, and because handling anything other than single-character-width characters in textwrap is a new feature.

FUD about Python here is something I wasn't expecting.

Python 2 supports Unicode and is still used a lot by a lot of people.

CJK people are not subhumans, so don't support CJK is something called, wait... a bug ! And it's a shame that it was not fixed earlier.

Python 3 has this bug too, so it's not really what I would call a "proper unicode support".

The problem is (if I'm understanding this correctly, which I may not be, I'm not a unicode expert) is that how you compute and manipulate CJK characters in python2 differs depending on whether you are dealing with a wide build or a narrow build. And the fact that python3 doesn't handle it either is why this would be a new feature (see the referenced issues).

But I could be wrong. I leave it to the unicode experts.

What textwrap does not take into account the width of characters is not the only problem. It also does not take into account combining characters and control codes. Implementing all this will significantly complicate the code and possibly should lie outside the standard library. Such large change obviously is new feature. I believe that we should first provide a common interface to determine the width of the line (bpo-12499) and allow to determine the appropriate algorithm at the application level. Also provide helper functions like in bpo-12568.

If your unicode experts haven't fix this BUG still now, this will never be done (by this experts).

We can say they are not true unicode experts as they have forgotten since a so long time billions of CJK people !

Or perhaps we haven't had a CJK user interested enough in using textwrap to provide the needed enhancements. It seems like there is interest in solving the related problems recently, so perhaps some progress will be made now.

The fact that you view it as a bug does not mean that it is a bug from the point of view of the python project. Every enhancement could be considered a bug, but we have a strict backward compatibility policy for maintenance release (which python 2.7.x is) which follows the semantic versioning principle that the API does not change in a micro release. According to Serhiy, fixing this correctly will require API changes.

CJKwrap a little lib to fix this bug:
https://github.com/fgallaire/cjkwrap

Hello everybody,

This is a Python3 version of my lib CJKwrap:
https://github.com/fgallaire/cjkwrap3

It could be integrated as a new lib in Python 3.7.

People who are using textwrap will have no surprise, and people who wants CJK width support will be happy to use it.

Any remarks welcome.

Cheers

After discussion with Haypo, CJK support is now implemented as and option, disabled by default for backward compatibility reasons.

PR on GitHub:
#89

See also
http://www.unicode.org/reports/tr29/
http://www.unicode.org/reports/tr14/

I remove the dependency to bpo-12568 since the current PR uses unicodedata.east_asian_width(), not the C function wcswidth().

I reread bpo-6755, bpo-12485, bpo-12499, and bpo-12568 about the textwrap/char width topic, and went to those conclusions:

It's a hard topic [1][2], so we may not succeed in a single shot.
The work is already done by wcwidth in the libc, which does not exists on windows and may not exist on MacOS.
The work is already partially done for CJK in the Unicode standard.

So I'm OK with adding CJK support to textwrap *as a first step*, which means I'm *not* ok with the CJK parameter to the wrap function, as maybe in the future we'll do more.

I'm also not OK with this being not the default, as many code using textwrap do not know in advance if they need CJK, they should not care, it should be done right by default.

But having CJK support by default also mean we'll have to fallback silently to a non-unicode textwrap if unicodedata is not available, as Victor said in PR-89 "Python requires optparse to compile modules like unicodedata, optparse imports textwrap which now always requires unicodedata", which may or may not lead to surprising behavior.

[1] The unicode standard is not clear about character width (not its role, font dependent), only where it's OK to put the line break.
[2] CJK is not enough, there's plenty of other characters of width not equal to one, like combining characters, tabulations, non-printables, Prepended_Concatenation_Mark having a typical width of zero but sometimes 1, U+00AD being tricky, Hangul Jamo medial vowels and final consonants are "conjoining", and so on... this need a huge effort / attention, this if we're going down this hole it means a lot of maintenance / new issues about this or that character being reported with the "wrong" size.

I'm not expert of this area, but Korean support is still totally broken:

>>> import unicodedata
>>> s = "\u1100\u1161\u11a8"
>>> unicodedata.east_asian_width(s[0])
'W'
>>> unicodedata.east_asian_width(s[1])
'N'
>>> unicodedata.east_asian_width(s[2])
'N'
>>> s
'각'
>>> s[:2]
'가'
>>> print(s[:2] + '\n' + s[2:])
가
ᆨ
>>> print(s[:1] + '\n' + s[1:])
ᄀ
ᅡᆨ

I think "CJK support" is not nice naming for just using east_asian_width.
PR-5649 doesn't use "CJK" word in APIs, but uses in commit log (PR title)
and NEWS entry.

Maybe, it should be "Textwrap now uses unicodedata.east_asian_width()
to calculate text width when unicodedata is available."

Hi Inada, you're right and that's more or less why I not used CJK in the implementation: mainly I don't want to close the door to future enhancements on this topic (char width) of non-CJK languages (like those "invisible hyphens" here only to tell textwrapers where they can wrap inside of words, and so on).

I'll remove CJK mentions (sadly maybe not today).

I think that this issue should be closed, as it is based on some confusions and errors.

Textwrap works in terms of characters. The wrap method "wraps the single paragraph in text (a string) so every line is at most width characters long." When the module was written, 'character' meant "printable ascii (or 'extended ascii') character". It now means 'unicode codepoint'. Both are mentally real abstractions but have no particular correspondence to physical length. Calling textwrap buggy because it works in characters is wrong.

Translating 'character' to 'fixed-width character space', so that one can measure physical length in terms of 'spaces' as a physical unit, is exact if and only if all characters are displayed in the same width space. This is true for fixed pitch output devices that simulate typewriters and text that only used fixed-width characters from a fixed pitch font. For long lines, the translation for a variable-pitch fonts may or may not be good enough for a particular use.

As David noted, textwrap already fails for Ascii control characters. And it does cause problem when they are used in wrapped text. They are coded with 2 or 4 characters on input and may display as 0, 1, (possibly 2), 4, or 5 characters on output, depending on the display code and the display 'device'. As to the latter, "print('x\ax')" displays as 'xx' in a Windows console and as something like x�x in a tkinter Text widget, except that the numbers in the box here on Firefox are not present, so that the tk box is (sort-of) the same width as 'x'.

The particular premise of this issue is that CJK characters are somehow special and that 2.x releases, and now 3.x releases, are particularly broken for CJK text. Not so. If one has text that only uses same-width characters in a fixed-pitch CJK font (including wide spaces so columns line up), then textwrap works as well as it does for any other fixed-pitch text (ie, Ascii or Latin1). If one wants lines of a particular physical width, one passes a character width argument that corresponds to the desired physical width.

The following is based on what I see in IDLE's Settings dialog Font page font sample for Windows 10 'Source Code Pro'. It includes samples from 12 'alphabet's To view it, run 'python -m idlelib', and on the top menu click Options => Configure IDLE. When the selected font is not a full BMP unicode font, Tk and Windows use other fonts, scaled to the same height, to synthesize a fairly complete BMP 'font'. The [Help] text says a bit more but has a mistake. What I see:

Font size corresponds to physical height. Hence, the lines are very close to the same height. Some fonts look smaller or larger because they specify more or less blank space between lines. One factor is the use of descenders, as in Arabic.

Character width for a fixed height varies. 20 characters in Greek, IPA, Hebrew, and Arabic take progressive less physical space than 20 Ascii or Latin1 characters. 20 characters in Devanagari, Cyrillic, and Tamil take progressively more. (The Tamil line only has 14 chars.) None of these are obviously fixed pitch.

The Chinese, Korean, and Japanese samples have a fixed pitch. The characters are *not* actually 'double-wide', at least no relative to most other languages. The 10 CJK characters are as wide as 16 Source Code Pro characters. To match the physical width of 72 Ascii spaces, one should pass 'width=45'.

But note that the exact ratio for Ascii depends on the font. It is a little higher for Courier and Lucida Console. It ranges from about 1 (for Arabic) to 2 (for Tamil) for other languages. The first 10 Tamil characters are slightly wider than the 10 CJK characters, so counting each CJK character as two average Tamil character is completely wrong.

My conclusion: the proposal is unnecessary for pure CJK text; it is wrong in hard-coding a fix only for CJK; the CJK fix is wrong in hard-coding a particular ratio, in particular, one that is at the extreme end of the range of possibilities. Therefore, I think the open PR should be closed. I also think this issue should be closed in favor of bpo-12499, which proposed to allow users to pass a transform function suitable for their particular use case. If that is implemented, and we decide to then add some sample functions, or rather, function factories, and to include one specifically for CJK*, then a new PR will be needed, and a new issue would be appropriate.

A more generic function factory for text with characters of two width classes might have as inputs a condition to identify '2nd language characters' and their fixed or average width relative relative to the 'first' language.

Thanks, Terry.

I have doubt about "east asian width" support should be merged.
While I agree it is "better than nothing", it may make textwrap much slower.
But I didn't have courage to reject the PR.

Textwrap focused on ASCII and English-like (space separated) language.
"support unicode" is very hard problem. We should consider about grapheme cluster (UAX29), east asian width (UAX11) (But utf8proc is better than UAX11), and line breaking algorithm (UAX14).

For wrapping text on terminal, some terminal emulators and multiplexers (e.g. tmux) implements much nicer algorithm.

I think 3rd party C extension based on algorithm used by tmux is best solution.
If someone really want this feature, please try it on PyPI.

I understand "want it in stdlib!". But text wrapping is very hard, complicated problem. Since stdlib grows slower, and backward compatibility restrict us, I think it should be implemented in 3rd party library first.

My msg323731 for bpo-12568 refers to my msg321291 above. I did some new experiments with column spacing for European characters in Windows console (as opposed to tk Text) and discovered that some, including some Latin1 characters used in English text, may also become double-width in certain fonts. The problem of calculating physical line lengths is even harder than I thought before, and does not require non-English text.

Navigation Menu

Repository navigation

CJK support for textwrap #68853

5 remaining items

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

CJK support for textwrap #68853

Description

Activity

serhiy-storchaka commented on Jul 19, 2015

fgallaire commented on Jul 19, 2015

bitdancer commented on Jul 20, 2015

fgallaire commented on Jul 21, 2015

bitdancer commented on Jul 21, 2015

serhiy-storchaka commented on Jul 21, 2015

fgallaire commented on Jul 21, 2015

bitdancer commented on Jul 21, 2015

fgallaire commented on Sep 13, 2016

fgallaire commented on Dec 27, 2016

fgallaire commented on Feb 14, 2017

5 remaining items

methane commented on Mar 2, 2017

vstinner commented on Jul 3, 2017

JulienPalard commented on Feb 12, 2018

methane commented on Jul 8, 2018

JulienPalard commented on Jul 8, 2018

terryjreedy commented on Jul 8, 2018

methane commented on Jul 9, 2018

terryjreedy commented on Aug 18, 2018

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions