homepage

You cannot use Python's lib re for handling Unicode regular expressions because it violates the standard set out for the same in UTS#18 on Unicode Regular Expressions in RL1.2a on compatibility properties. What \w is allowed to match is clearly explained there, but Python has its own idea. Because it is in clear violation of the standard, it is misleading and wrong for Python to claim that the re.UNICODE flag makes \w and friends match Unicode. Here are the failed test cases when the attached file is run under v3.2; there are further failures when run under v2.7.
FAIL lib re found non alphanumeric string café
FAIL lib re found non alphanumeric string K
FAIL lib re found non alphanumeric string ͅ
FAIL lib re found non alphanumeric string ְ
FAIL lib re found non alphanumeric string 0
FAIL lib re found non alphanumeric string 𐍁
FAIL lib re found non alphanumeric string Unicode
FAIL lib re found non alphanumeric string 𐐔𐐯𐑅𐐨𐑉𐐯𐐻
FAIL lib re found non alphanumeric string connector‿punctuation
FAIL lib re found non alphanumeric string Ὰͅ_Στο_Διάολο
FAIL lib re found non alphanumeric string 𐌰𐍄𐍄𐌰‿𐌿𐌽𐍃𐌰𐍂‿𐌸𐌿‿𐌹𐌽‿𐌷𐌹𐌼𐌹𐌽𐌰𐌼
FAIL lib re found all alphanumeric string 123
FAIL lib re found all alphanumeric string 123
FAIL lib re found all alphanumeric string 1⁄41⁄23⁄4
FAIL lib re found all alphanumeric string (3)
Note that Matthew Barnett's regex lib for Python handles all of these cases in comformance with The Unicode Standard.

However desireable it would be, I do not believe there is any claim in the manual that the re module follows the evolving Unicode consortium r.e. standard. If I understand, you are saying that this statement in the doc, "Matches Unicode word characters;" is not now correct and should be revised. Was it once correct? Could we add "by an older definition of 'word' character"?
There has been some discussion of adding regex to the stdlib, possibly as a replacement for re. You posts indicate that regex is more improved then some realized, and hence has more incompatibilities that we realized, and hence is less suitable as a strictly backwards-compatible replacement. So I think it needs to be looked at as a parallel addition. I do not know Mathew's current position on the subject.

> Terry J. Reedy <tjreedy@udel.edu> added the comment:
> However desireable it would be, I do not believe there is any claim in the =
> manual that the re module follows the evolving Unicode consortium r.e. stan=
My from the hip thought is that if re cannot be fixed to follow
the Unicode Standard, it should be deprecated in favor of code
that can if such is available, because you cannot process Unicode
text with regular expressions otherwise.
> dard. If I understand, you are saying that this statement in the doc, "Matc=
> hes Unicode word characters;" is not now correct and should be revised. Was=
> it once correct? Could we add "by an older definition of 'word' character"=
> ?
Yes, your hunch is exactly correct. They once had a lesser definition that
they have now. It is very very old. I had to track this down for Java
once. There is some discussion of a "word_character class" at least 
as far back as tr18v3 from back in 1998.
 http://www.unicode.org/reports/tr18/tr18-3.html
By the time tr18v5 rolled around just a year later in 1999, the overall
document has changed substantially, and you can clearly see its current
shape there. Word characters are supposed to include all code points with
the Alphabetic property, for example. 
 http://www.unicode.org/reports/tr18/tr18-5.html
However, the word "alphabetic" has *never* been synonymous in 
Unicode with 
 \p{gc=Lu}
 \p{gc=Ll}
 \p{gc=Lt}
 \p{gc=Lm}
 \p{gc=Lo}
as many people incorrectly assume, nor certainly to 
 \p{gc=Lu}
 \p{gc=Ll}
 \p{gc=Lt}
let alone to 
 \p{gc=Lu}
 \p{gc=Ll}
Rather, it has since its creation included code points that are not
letters, such as all GC=Nl and also certain GC=So code points. And,
notoriously, U+0345. Indeed it is here I first noticed that that Python had
already broken with the Standard, because U+0345 COMBINING GREEK
YPOGEGRAMMENI is GC=Mn, but Alphabetic=True, yet I have shown that 
Python's title method is messing up there. 
I wouldn't spend too much in archaeological digs, though, because lots of
stuff has changed since the less millennium. It was in tr18v7 from 2003-05
that we hit paydirt, because this is when the famous Annex C of RL1.2a 
fame first appeared:
 http://www.unicode.org/reports/tr18/tr18-7.html#Compatibility_Properties
Notice how it defines \w to be nothing more than \p{alpha}, \p{digit}, and
\p{gc=Pc}. It does not yet contain the requirement that all Marks be
counted as part of the word, just the few that are alphas -- which the
U+0345 counts for, since it has an uppercase map of a capital iota!
That particular change did not occur until tr18v8 in 2003-08, barely
a scant three months later.
 http://www.unicode.org/reports/tr18/tr18-8.html#Compatibility_Properties
Now at last we see word characters defined in the modern way that we 
have become used to. They must match any of:
 \p{alpha}
 \p{gc=Mark}
 \p{digit}
 \p{gc=Connector_Punctuation}
BTW, Python is matching all of 
 \p{GC=N}
meaning
 \p{GC=Nd}
 \p{GC=Nl}
 \p{GC=No}
instead of the required 
 \p{GC=Nd}
which is a synonym for \p{digit}.
I don't know had that happened, because \w has never included
all number code points in Unicode, only the decimal number ones.
That all goes to show why, when citing conformance to some aspect of 
The Unicode Standard, one must be exceedingly careful just how one 
does so!
The Unicode Consortium recognizes this is an issue, and I am pretty
sure I can hear it in your own subtext as well. 
Kindly bear with and forgive me for momentarily sounding like a standard
lawyer. I do this because to show not just why it is important to get
references to the Unicode Standard correct, but indeed, how to do so.
After I have given the formal requirements, I will then produce
illustrations of various purported claims, some of which meet the
citation requirements, and others which do not.
=======================================================================
To begin with, there is an entire technical report on conformance.
It includes:
 http://unicode.org/reports/tr33/
 The Unicode Standard [Unicode] is a very large and complex standard.
 Because of this complexity, and because of the nature and role of the
 standard, it is often rather difficult to determine, in any particular
 case, just exactly what conformance to the Unicode Standard means.
...
 Conformance claims must be specific to versions of the Unicode
 Standard, but the level of specificity needed for a claim may vary
 according to the nature of the particular conformance claim. Some
 standards developed by the Unicode Consortium require separate
 conformance to a specific version (or later), of the Unicode Standard.
 This version is sometimes called the base version. In such cases, the
 version of the standard and the version of the Unicode Standard to
 which the conformance claim refers must be compatible.
However, you don't need to read tr33, really, because *the* most important
thing bits about conformance are to be found on pp. 57-58 of Chapter 3 of
the published Unicode Standard:
 http://www.unicode.org/versions/Unicode6.0.0/ch03.pdf
 References to the Unicode Standard
 The documents associated with the major, minor, and update versions are called the major
 reference, minor reference, and update reference, respectively. For example, consider Uni-
 code Version 3.1.1. The major reference for that version is The Unicode Standard, Version
 3.0 (ISBN 0-201-61633-5). The minor reference is Unicode Standard Annex #27, "The Uni-
 code Standard, Version 3.1." The update reference is Unicode Version 3.1.1. The exact list of
 contributory files, Unicode Standard Annexes, and Unicode Character Database files can
 be found at Enumerated Version 3.1.1.
 The reference for this version, Version 6.0.0, of the Unicode Standard, is
 The Unicode Consortium. The Unicode Standard, Version 6.0.0, defined
 by: The Unicode Standard, Version 6.0 (Mountain View, CA: The Uni-
 code Consortium, 2011. ISBN 978-1-936213-01-6)
 References to an update (or minor version prior to Version 5.2.0) include a reference to
 both the major version and the documents modifying it. For the standard citation format
 for other versions of the Unicode Standard, see "Versions" in Section B.6, Other Unicode
 Online Resources.
 Precision in Version Citation
 Because Unicode has an open repertoire with relatively frequent updates, it is important
 not to over-specify the version number. Wherever the precise behavior of all Unicode char-
 acters needs to be cited, the full three-field version number should be used, as in the first
 example below. However, trailing zeros are often omitted, as in the second example. In such
 a case, writing 3.1 is in all respects equivalent to writing 3.1.0.
 1. The Unicode Standard, Version 3.1.1
 2. The Unicode Standard, Version 3.1
 3. The Unicode Standard, Version 3.0 or later
 4. The Unicode Standard
 Where some basic level of content is all that is important, phrasing such as in the third
 example can be used. Where the important information is simply the overall architecture
 and semantics of the Unicode Standard, the version can be omitted entirely, as in example 4.
 References to Unicode Character Properties
 Properties and property values have defined names and abbreviations, such as
	 Property: General_Category (gc)
	 Property Value: Uppercase_Letter (Lu)
 To reference a given property and property value, these aliases are used, as in this example:
	 The property value Uppercase_Letter from the General_Category prop-
	 erty, as specified in Version 6.0.0 of the Unicode Standard.
 Then cite that version of the standard, using the standard citation format that is provided
 for each version of the Unicode Standard.
 When referencing multi-word properties or property values, it is permissible to omit the
 underscores in these aliases or to replace them by spaces.
 When referencing a Unicode character property, it is customary to prepend the word "Uni-
 code" to the name of the property, unless it is clear from context that the Unicode Standard
 is the source of the specification.
 References to Unicode Algorithms
 A reference to a Unicode algorithm must specify the name of the algorithm or its abbrevia-
 tion, followed by the version of the Unicode Standard, as in this example:
 The Unicode Bidirectional Algorithm, as specified in Version
 6.0.0 of the Unicode Standard.
 See Unicode Standard Annex #9, "Unicode Bidirectional Algorithm,"
 (http://www.unicode.org/reports/tr9/tr9-23.html)
=======================================================================
Now for some concrete citation examples, both correct and dubious.
In the JDK7 documentation for on the Character class we find:
 Character information is based on the Unicode Standard, version 6.0.0.
That one is a perfectly good conformance citation, even if there seems 
a bit of wiggle in "is based on", but no matter. It is short and does
everything it needs to.
However, in the JDK7 documentation for the Pattern class we
somewhat problematically find:
 Unicode support 
 This class is in conformance with Level 1 of Unicode Technical
 Standard #18: Unicode Regular Expression, plus RL2.1 Canonical
 Equivalents.
And similarly, in the JDK7 documentation for the Normalizer class we find:
 This class provides the method normalize which transforms Unicode
 text into an equivalent composed or decomposed form, allowing for
 easier sorting and searching of text. The normalize method supports
 the standard normalization forms described in Unicode Standard
 Annex #15 — Unicode Normalization Forms.
The problem with those second two Java refs is that they to my reading
appear to be in technical violation, for they give neither a full
version number nor a date of publication. 
You *have* to give one or the other, or both. 
Java got themselves into a heap of trouble (so to speak) over
this once before because it turned out that the version of the
document they were actually in conformance with was quite
literally from the previous millennium!!
That's why you need to give versions and publication dates.
Here are some other citations.
First, from the perldelta manpage that the Perl 5.14 release ships with:
 Perl comes with the Unicode 6.0 data base updated with Corrigendum
 #8 <http://www.unicode.org/versions/corrigendum8.html>, with one
 exception noted below. See <http://unicode.org/versions/Unicode6.0.0/> 
 for details on the new release. Perl does not support any Unicode 
 provisional properties, including the new ones for this release.
That is quite complete, as it even includes the specific which
corrigenda we follow and explains the matter of properties.
Or this from the perlunicode manpage of that same release:
 Unicode Regular Expression Support Level
 The following list of Unicode supported features for
 regular expressions describes all features currently
 directly supported by core Perl. The references to "Level
 N" and the section numbers refer to the Unicode Technical
 Standard #18, "Unicode Regular Expressions", version 13,
 from August 2008.
See all that? Notice how it gives the name of the document, its revision
number, and its publication date. You don't have to do all that for the
main Unicode release, but you really ought to when referring to individual
technical reports BECAUSE THESE GET UPDATED ASYNCRONOUSLY.
I would suggest you pick a version of tr18 that you conform to,
and state which of its requirements you do and do not meet.
However, I cannot find any version of tr18 that has existed during the
present millennium that Python comes even close to meeting more than one
or two requirements for. Given that, it may be better to no longer make
any claims regarding Unicode at all. That seems like back-peddaling to
me, not future-thinking.
Matthew's regex module, however, does *almost* everything right that re
does wrong. It may be that as with Java's io vs nio classes (and now
heaven forbid nio2!), you actually can't fix the only module and must
create a wholly new namespace. I cannot answer that.
For RL1.2 proper, the first properties requirement, Java was only missing a
few, so they went and added the missing properties. I strongly urge you to
do so because you cannot handle Unicode without properties. Rl1.2 requires
only 11 of them, so it isn't too hard. Matthew supports many many more.
However, because the \w&c issues are bigger, Java addressed the tr18 RL1.2a
issues differently, this time by creating a new compilation flag called
UNICODE_CHARACTER_CLASSES (with corresponding embedded "(?U)" regex flag.)
Truth be told, even Perl has secret pattern compilation flags to govern
this sort of thing (ascii, locale, unicode), but we (well, I) hope you
never have to use or even notice them. 
That too might be a route forward for Python, although I am not quite sure
how much flexibility and control of your lexical scope you have. However,
the "from __future_" imports suggest you may have enough to do something
slick so that only people who ask for it get it, and also importantly that
they get it all over the place so don't have to add an extra flag or u'...'
or whatever every single time. 
This isn't something I've looked much into, however.
Hope this clarifies things.
--tom

> However, because the \w&c issues are bigger, Java addressed the tr18 RL1.2a
> issues differently, this time by creating a new compilation flag called
> UNICODE_CHARACTER_CLASSES (with corresponding embedded "(?U)" regex flag.)
> 
> Truth be told, even Perl has secret pattern compilation flags to govern
> this sort of thing (ascii, locale, unicode), but we (well, I) hope you
> never have to use or even notice them. 
> 
> That too might be a route forward for Python, although I am not quite sure
> how much flexibility and control of your lexical scope you have. However,
> the "from __future_" imports suggest you may have enough to do something
> slick so that only people who ask for it get it, and also importantly that
> they get it all over the place so don't have to add an extra flag or u'...'
> or whatever every single time. 
If the current behaviour is buggy or sub-optimal, I think we should
simply fix it (which might be done by replacing "re" with "regex" if
someone wants to shepherd its inclusion in the stdlib).
By the way, thanks for the detailed explanations, Tom.

If the regex module works fine here, I think it's better to leave the re module alone and include the regex module in 3.3.

Really? The re module cannot be salvaged and we should add regex but keep the (buggy) re? That does not make a lot of sense to me. I think it should just be fixed in the re module. Or the re module should be *replaced* by the code from the regex module (but renamed to re, and with certain backwards compatibilities restored, probably). But I really hope the re module (really: the _sre extension module) can be fixed. We should also make a habit in our docs of citing specific versions of the Unicode standard, and specific TR numbers and versions where they apply. (And hopefully we can supply URLs to the Unicode consortium's canonical copies of those documents.)

> Or the re module should be *replaced* by the code from the regex module
> (but renamed to re, and with certain backwards compatibilities
> restored, probably).
This is what I meant.
> But I really hope the re module (really: the _sre extension module)
> can be fixed.
Start fixing these issues from scratch doesn't make much sense IMHO. We could "extract" the fixes from regex and merge them in re, but then again it's probably easier to just replace the whole module.
> We should also make a habit in our docs of citing specific versions
> of the Unicode standard, and specific TR numbers and versions where 
> they apply.
While this is a good thing it's not always doable. Usually someone reports a bug related to something specified in some standard and only that part gets fixed. Sometimes everything else is also updated to follow the whole standard, but often this happens incrementally, so we can't say, e.g., "the re module supports Unicode x.y" unless we go through the whole standard and fix/implements everything.

> But I really hope the re module (really: the _sre extension module)
> can be fixed.
If you mean on 2.7/3.2, then I guess we could extract the fixes from regex, but we have to see if it's doable and someone will have to do it.
Also consider that the regex module is available for 2.7/3.2, so we could suggest the users to use it if they have problems with the re bugs (even if that means having an additional dependency).
ISTM that current plan is:
 * replace re with regex (and rename it) on 3.3 and fix all these bugs;
 * leave 2.7 and 3.2 with the old re and its bugs;
 * let people use the external regex module on 2.7/3.2 if they need to.
If this is not ok, maybe it should be discussed on python-dev.

[me]
>> But I really hope the re module (really: the _sre extension module)
>> can be fixed.
[Ezio]
> Start fixing these issues from scratch doesn't make much sense IMHO. We could "extract" the fixes from regex and merge them in re, but then again it's probably easier to just replace the whole module.
I have changed my mind at least half-way. I am open to having regex
(with some changes, details TBD) replace re in 3.3. (I am not yet 100%
convinced, but I'm not rejecting it as strongly as I was when I wrote
that comment in this bug. See the ongoing python-dev discussion on
this topic.)
>> We should also make a habit in our docs of citing specific versions
>> of the Unicode standard, and specific TR numbers and versions where
>> they apply.
>
> While this is a good thing it's not always doable. Usually someone reports a bug related to something specified in some standard and only that part gets fixed. Sometimes everything else is also updated to follow the whole standard, but often this happens incrementally, so we can't say, e.g., "the re module supports Unicode x.y" unless we go through the whole standard and fix/implements everything.
Hm. I think that for Unicode it may actually be important enough to be
consistent in following the whole standard that we should attempt to
be consistent and not just chase bug reports. Now, we may consciously
decide not to implement a certain recommendation of the standard. E.g.
I'm not going to require that IronPython or Jython have string objects
that support O(1) indexing of code points, even (assuming PEP 393 gets
accepted) CPython will have them. But these decisions should be made
explicitly, and documented clearly.
Ideally, we need a "Unicode czar" -- a core developer whose job it is
to keep track of Python's compliance with various parts and versions
of the Unicode standard and who can nudge other developers towards
fixing bugs or implementing features, or update the documentation in
case things don't get added. (I like Tom's approach to Java 1.7, where
he submitted proposed doc fixes explaining the deviations from the
standard. Perhaps a bit passive-aggressive, but it was effective. :-)

> Ideally, we need a "Unicode czar" -- a core developer whose job it is
> to keep track of Python's compliance with various parts and versions
> of the Unicode standard and who can nudge other developers towards
> fixing bugs or implementing features, or update the documentation in
> case things don't get added.
We should first do a full review of the latest Unicode standard and see what's missing. I think there might be parts of older Unicode versions (even < Unicode 5) that are not yet implemented. Chapter 3 is a good place where to start, but I'm not sure that's enough -- there are a few TRs that should be considered as well.
If we manage to catch up with Unicode 6, then it shouldn't be too difficult to review the changes that every new version will introduce and open an issue for each (or a single issue if the changes are limited).
FWIW I'm planning to look at the conformance of the UTF codecs and fix them (if necessary) whenever I'll have time.

The failing re tests after PEP 393 are:
FAIL lib re found non alphanumeric string 'cafe'
FAIL lib re found non alphanumeric string 'K'
FAIL lib re found non alphanumeric string ''
FAIL lib re found non alphanumeric string ''
FAIL lib re found non alphanumeric string 'connector‿punctuation'
FAIL lib re found non alphanumeric string 'Ὰ_Στο_Διάολο'
FAIL lib re found non alphanumeric string '𐌰𐍄𐍄𐌰‿𐌿𐌽𐍃𐌰𐍂‿𐌸𐌿‿𐌹𐌽‿𐌷𐌹𐌼𐌹𐌽𐌰𐌼'
FAIL lib re found all alphanumeric string '123'
FAIL lib re found all alphanumeric string '123'
FAIL lib re found all alphanumeric string '1⁄41⁄23⁄4'
FAIL lib re found all alphanumeric string '(3)'

Whatever I may have said before, I favor supporting the Unicode standard for \w, which is related to the standard for identifiers.
This is one of 2 issues about \w being defined too narrowly. I am somewhat arbitrarily closing #1693050 as a duplicate of this (fewer digits ;-).
There are 3 issues about tokenize.tokenize failing on valid identifiers, defined as \w sequences whose first char is an identifier itself (and therefore a start char). In msg313814 of #32987, Serhiy indicates which start and continue identifier characters are matched by \W for re and regex. I am leaving #24194 open as the tokenizer name issue.

This issue is also implicated in a failure of isalpha and friends.
Easy way to see this is to compare
>>> isalpha('İ')
True
>>> isalpha('İ'.lower())
False
This results from the use of a combining character to encode lower-case Turkish dotted i:
>>> len('İ'.lower())
2
>>> unicodedata.category('İ'.lower()[1])
'Mn'

[One year and 2 days later... :-[
Is this fixed in 3.9? If not, the Versions list above should be updated.
The failure of lower() to preserve 'alpha-ness' is a serious bug, it causes significant failures in e.g. Turkish NLP, and it's _not_ just a failure of the documentation!
Please can we move this to category Unicode and get at least this aspect of the problem fixed? Should I raise a separate issue on isalpha() etc.?