Add unicode script info to the unicode database

[doerwalter]

BPO	6331
Nosy	@malemburg, @loewis, @doerwalter, @vstinner, @benjaminp, @ezio-melotti, @akitada, @berkerpeksag, @gnprice
Files	unicode-script.diff unicode-script-2.diff unicode-script-3.diff

^{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 = None
created_at = <Date 2009-06-23.20:50:57.790>
labels = ['type-feature', '3.7', 'expert-unicode']
title = 'Add unicode script info to the unicode database'
updated_at = <Date 2019-08-28.05:09:27.665>
user = 'https://github.com/doerwalter'

bugs.python.org fields:

activity = <Date 2019-08-28.05:09:27.665>
actor = 'Greg Price'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Unicode']
creation = <Date 2009-06-23.20:50:57.790>
creator = 'doerwalter'
dependencies = []
files = ['14348', '14356', '14418']
hgrepos = []
issue_num = 6331
keywords = ['patch', 'needs review']
message_count = 17.0
messages = ['89642', '89647', '89671', '89675', '89701', '89973', '111040', '177469', '177506', '214204', '214633', '214636', '226266', '251214', '285269', '320090', '320092']
nosy_count = 13.0
nosy_names = ['lemburg', 'loewis', 'doerwalter', 'vstinner', 'benjamin.peterson', 'ezio.melotti', 'akitada', 'berker.peksag', 'PanderMusubi', 'Elizacat', 'Cosimo Lupo', 'Denis Jacquerye', 'Greg Price']
pr_nums = []
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue6331'
versions = ['Python 3.7']

[doerwalter]

This patch adds a function unicodedata.script() that returns information
about the script of the Unicode character.

[loewis]

I think the patch is incorrect: the default value for the script
property ought to be Unknown, not Common (despite UCD.html saying the
contrary; see UTR#24 and Scripts.txt).

I'm puzzled why you use a hard-coded list of script names. The set of
scripts will certainly change across Unicode versions, and I think it
would be better to learn the script names from Scripts.txt.

Out of curiosity: how does the addition of the script property affect
the number of distinct database records, and the total size of the database?

I think a common application would be lower-cases script names, for more
efficient comparison; UCD has also changed the spelling of the script
names over time (from being all-capital before). So I propose that
a) two functions are provided: one with the original script names, and
one with the lower-case script names
b) keep cached versions of interned script name strings in separate
arrays, to avoid PyString_FromString every time.

I'm doubtful that script names need to be provided for old database
versions, so I would be happy to not record the script for old versions,
and raise an exception if somebody tries to get the script for an old
database version - surely applications of the old database records won't
be accessing the script property, anyway.

[doerwalter]

Martin v. Löwis wrote:

Martin v. Löwis <martin@v.loewis.de> added the comment:

I think the patch is incorrect: the default value for the script
property ought to be Unknown, not Common (despite UCD.html saying the
contrary; see UTR#24 and Scripts.txt).

Fixed.

I'm puzzled why you use a hard-coded list of script names. The set of
scripts will certainly change across Unicode versions, and I think it
would be better to learn the script names from Scripts.txt.

I hardcoded the list, because I saw no easy way to get the indexes
consistent across both versions of the database.

Out of curiosity: how does the addition of the script property affect
the number of distinct database records, and the total size of the database?

I'm not exactly sure how to measure this, but the length of
_PyUnicode_Database_Records goes from 229 entries to 690 entries.

If it's any help I can post the output of makeunicodedata.py.

I think a common application would be lower-cases script names, for more
efficient comparison; UCD has also changed the spelling of the script
names over time (from being all-capital before). So I propose that
a) two functions are provided: one with the original script names, and
one with the lower-case script names

It this really neccessary, if we only have one version of the database?

b) keep cached versions of interned script name strings in separate
arrays, to avoid PyString_FromString every time.

Implemented.

I'm doubtful that script names need to be provided for old database
versions, so I would be happy to not record the script for old versions,
and raise an exception if somebody tries to get the script for an old
database version - surely applications of the old database records won't
be accessing the script property, anyway.

OK, I've removed the script_changes info for the old database. (And with
this change the list of script names is no longer hardcoded).

Here's a new version of the patch (unicode-script-2.diff).

[loewis]

> I'm puzzled why you use a hard-coded list of script names. The set of
> scripts will certainly change across Unicode versions, and I think it
> would be better to learn the script names from Scripts.txt.

I hardcoded the list, because I saw no easy way to get the indexes
consistent across both versions of the database.

Couldn't you have a global cache, something like

scripts = ['Unknown']
def findscript(script):
  try:
    return scripts.index(script)
  except ValueError:
    scripts.append(script)
    return len(scripts)-1

> Out of curiosity: how does the addition of the script property affect
> the number of distinct database records, and the total size of the database?

I'm not exactly sure how to measure this, but the length of
_PyUnicode_Database_Records goes from 229 entries to 690 entries.

I think this needs to be fixed, then - we need to study why there are
so many new records (e.g. what script contributes most new records),
and then look for alternatives.

One alternative could be to create a separate Trie for scripts.

I'd also be curious if we can increase the homogeneity of scripts
(i.e. produce longer runs of equal scripts) if we declare that
unassigned code points have the script that corresponds to the block
(i.e. the script that surrounding characters have), and then only
change it to "Unknown" at lookup time if it's unassigned.

If it's any help I can post the output of makeunicodedata.py.

I'd be interested in "size unicodedata.so", and how it changes.
Perhaps the actual size increase isn't that bad.

> a) two functions are provided: one with the original script names, and
> one with the lower-case script names

It this really neccessary, if we only have one version of the database?

I don't know what this will be used for, but one application is
certainly regular expressions. So we need an efficient test whether
the character is in the expected script or not. It would be bad if
such a test would have to do a .lower() on each lookup.

[doerwalter]

I was comparing apples and oranges: The 229 entries for the trunk where
for an UCS2 build (the patched version was UCS4), with UCS4 there are
317 entries for the trunk.

size unicodedata.o gives:

__TEXT __DATA __OBJC others dec hex
13622 587057 0 23811 624490 9876a

for trunk

and

__TEXT __DATA __OBJC others dec hex
17769 588817 0 24454 631040 9a100

for the patched version.

[doerwalter]

Here is a new version that includes a new function scriptl() that
returns the script name in lowercase.

[BreamoreBoy]

Could someone with unicode knowledge take this review on, given that comments have already been made and responded to?

[PanderMusubi]

Please, also consider reviewing functionality offered by:
http://pypi.python.org/pypi/unicodescript/
and
http://pypi.python.org/pypi/unicodeblocks/
which could be used to improve and extend the proposed patch.

[PanderMusubi]

The latest version of the respective sources can be found here:
https://github.com/ConradIrwin/unicodescript
and here:
https://github.com/simukis/unicodeblocks

[loewis]

Pander: In what way would this extend or improve the current patch?

[PanderMusubi]

I see the patch support Unicode scripts https://en.wikipedia.org/wiki/Script_%28Unicode%29 but I am also interested in support for Unicode blocks https://en.wikipedia.org/wiki/Unicode_block

Code for support for the latter is at https://github.com/nagisa/unicodeblocks

I could ont quiet make out of the patch also supports Unicode blocks. If not, shoudl that be requested in a separete issue?

Furthermore, support for Unicode scripts and blocks should be updated each time a new version of Unicode standard is published. Someone should check of the latest patch should be updated to the latest version of Unicode. Not only for this issue but for each release of PYthon.

[loewis]

Adding support for blocks should indeed go into a separate issue. Your code for that is not suitable, as it should integrate with the existing make_unicodedata.py script, which your code does not.

And yes, indeed, of course, we automatically update (nearly) all data in Python automatically from the files provided by the Unicode consortium.

[Elizacat]

I think this needs to be fixed, then - we need to study why there are
so many new records (e.g. what script contributes most new records),
and then look for alternatives.

The "Common" script appears to be very fragmented and may be the cause of the issues.

One alternative could be to create a separate Trie for scripts.

Not having seen the one in C yet, I have one written in Python, custom-made for storing the script database, based on the general idea of a range tree. It stores ranges individually straight out of Scripts.txt. The general idea is you take the average of the lower and upper bounds of a given range (they can be equal). When searching, you compare the codepoint value to the average in the present node, and use that to find which direction to search the tree in.

Without coalescing neighbouring ranges that are the same script, I have 1,606 nodes in the tree (for Unicode 7.0, which added a lot of scripts). After coalescing, there appear to be 806 nodes.

If anyone cares, I'll be more than happy to post code for inspection.

I don't know what this will be used for, but one application is
certainly regular expressions. So we need an efficient test whether
the character is in the expected script or not. It would be bad if
such a test would have to do a .lower() on each lookup.

This is actually required for restriction-level detection as described in Unicode TR39, for all levels of restriction above ASCII-only (http://www.unicode.org/reports/tr39/#Restriction_Level_Detection).