Text::Unidecode -- US-ASCII transliterations of Unicode text

  use utf8;
  use Text::Unidecode;
  print unidecode(
    "\x{5317}\x{4EB0}\n"
     # those are the Chinese characters for Beijing
  );

  # That prints: Bei Jing

It often happens that you have non-Roman text data in Unicode, but you can't display it usually because you're trying to show it to a user via an application that doesn't support Unicode, or because the fonts you need aren't accessible. You could represent the Unicode characters as ??????? or \15BA\15A0\1610..., but that's nearly useless to the user who actually wants to read what the text says.

What Text::Unidecode provides is a function, CWunidecode(...) that takes Unicode data and tries to represent it in US-ASCII characters (i.e., the universally displayable characters between 0x00 and 0x7F). The representation is almost always an attempt at transliteration i.e., conveying, in Roman letters, the pronunciation expressed by the text in some other writing system. (See the example in the synopsis.)

Unidecode's ability to transliterate is limited by two factors:

* The amount and quality of data in the original

So if you have Hebrew data that has no vowel points in it, then Unidecode cannot guess what vowels should appear in a pronounciation. S f y hv n vwls n th npt, y wn't gt ny vwls n th tpt. (This is a specific application of the general principle of Garbage In, Garbage Out.)

* Basic limitations in the Unidecode design

Writing a real and clever transliteration algorithm for any single language usually requires a lot of time, and at least a passable knowledge of the language involved. But Unicode text can convey more languages than I could possibly learn (much less create a transliterator for) in the entire rest of my lifetime. So I put a cap on how intelligent Unidecode could be, by insisting that it support only context-insensitive transliteration. That means missing the finer details of any given writing system, while still hopefully being useful.

Unidecode, in other words, is quick and dirty. Sometimes the output is not so dirty at all: Russian and Greek seem to work passably; and while Thaana (Divehi, AKA Maldivian) is a definitely non-Western writing system, setting up a mapping from it to Roman letters seems to work pretty well. But sometimes the output is very dirty: Unidecode does quite badly on Japanese and Thai.

If you want a smarter transliteration for a particular language than Unidecode provides, then you should look for (or write) a transliteration algorithm specific to that language, and apply it instead of (or at least before) applying Unidecode.

In other words, Unidecode's approach is broad (knowing about dozens of writing systems), but shallow (not being meticulous about any of them).

Text::Unidecode provides one function, CWunidecode(...), which is exported by default. It can be used in a variety of calling contexts: This returns a copy of CW$in, transliterated. This is the same as CW$out = unidecode(join '', @in); This returns a list consisting of copies of CW@in, each transliterated. This is the same as CW@out = map scalar(unidecode($_)), @in; Each item on input is replaced with its transliteration. This is the same as CWfor(@bar, $foo, @baz) { $_ = unidecode($_) }

You should make a minimum of assumptions about the output of CWunidecode(...). For example, if you assume an all-alphabetic (Unicode) string passed to CWunidecode(...) will return an all-alphabetic string, you're wrong some alphabetic Unicode characters are transliterated as strings containing punctuation (e.g., the Armenian letter at 0x0539 currently transliterates as CWT`.

However, these are the assumptions you can make:

•

Each character 0x0000 - 0x007F transliterates as itself. That is, CWunidecode(...) is 7-bit pure.

•

The output of CWunidecode(...) always consists entirely of US-ASCII characters i.e., characters 0x0000 - 0x007F.

•

All Unicode characters translate to a sequence of (any number of) characters that are newline (\n) or in the range 0x0020-0x007E. That is, no Unicode character translates to \x01, for example. (Altho if you have a \x01 on input, you'll get a \x01 in output.)

•

Yes, some transliterations produce a \n but just a few, and only with good reason. Note that the value of newline (\n) varies from platform to platform see perlport in perlport.

•

Some Unicode characters may transliterate to nothing (i.e., empty string).

•

Very many Unicode characters transliterate to multi-character sequences. E.g., Han character 0x5317 transliterates as the four-character string Bei .

•

Within these constraints, I may change the transliteration of characters in future versions. For example, if someone convinces me that the Armenian letter at 0x0539, currently transliterated as T`, would be better transliterated as D, I may well make that change.

Text::Unidecode is meant to be a transliterator-of-last resort, to be used once you've decided that you can't just display the Unicode data as is, and once you've decided you don't have a more clever, language-specific transliterator available. It transliterates context-insensitively that is, a given character is replaced with the same US-ASCII (7-bit ASCII) character or characters, no matter what the surrounding character are.

The main reason I'm making Text::Unidecode work with only context-insensitive substitution is that it's fast, dumb, and straightforward enough to be feasable. It doesn't tax my (quite limited) knowledge of world languages. It doesn't require me writing a hundred lines of code to get the Thai syllabification right (and never knowing whether I've gotten it wrong, because I don't know Thai), or spending a year trying to get Text::Unidecode to use the ChaSen algorithm for Japanese, or trying to write heuristics for telling the difference between Japanese, Chinese, or Korean, so it knows how to transliterate any given Uni-Han glyph. And moreover, context-insensitive substitution is still mostly useful, but still clearly couldn't be mistaken for authoritative.

Text::Unidecode is an example of the 80/20 rule in action you get 80% of the usefulness using just 20% of a real solution.

A real approach to transliteration for any given language can involve such increasingly tricky contextual factors as these

The previous / preceding character(s)

What a given symbol X means, could depend on whether it's followed by a consonant, or by vowel, or by some diacritic character.

Syllables

A character X at end of a syllable could mean something different from when it's at the start which is especially problematic when the language involved doesn't explicitly mark where one syllable stops and the next starts.

Parts of speech

What X sounds like at the end of a word, depends on whether that word is a noun, or a verb, or what.

Meaning

By semantic context, you can tell that this ideogram X means shoe (pronounced one way) and not time (pronounced another), and that's how you know to transliterate it one way instead of the other.

Origin of the word

X means one thing in loanwords and/or placenames (and derivatives thereof), and another in native words. X normally makes the /X/ sound, except for this list of seventy exceptions (and words based on them, sometimes indirectly). Or: you never can tell which of the three ways to pronounce X this word actually uses; you just have to know which it is, so keep a dictionary on hand!

Language

The character X is actually used in several different languages, and you have to figure out which you're looking at before you can determine how to transliterate it.

Out of a desire to avoid being mired in any of these kinds of contextual factors, I chose to exclude all of them and just stick with context-insensitive replacement.

Things that need tending to are detailed in the TODO.txt file, included in this distribution. Normal installs probably don't leave the TODO.txt lying around, but if nothing else, you can see it at http://search.cpan.org/search?dist=Text::Unidecode

The Text::Unidecode motto is:

  It's better than nothing!

...in both meanings: 1) seeing the output of CWunidecode(...) is better than just having all font-unavailable Unicode characters replaced with ?'s, or rendered as gibberish; and 2) it's the worst, i.e., there's nothing that Text::Unidecode's algorithm is better than.

If you get really implausible nonsense out of CWunidecode(...), make sure that the input data really is a utf8 string. See perlunicode in perlunicode.

Thanks to Harald Tveit Alvestrand, Abhijit Menon-Sen, and Mark-Jason Dominus.

Unicode Consortium: http://www.unicode.org/

Geoffrey Sampson. 1990. Writing Systems: A Linguistic Introduction. ISBN: 0804717567

Randall K. Barry (editor). 1997. ALA-LC Romanization Tables: Transliteration Schemes for Non-Roman Scripts. ISBN: 0844409405 [ALA is the American Library Association; LC is the Library of Congress.]

Rupert Snell. 2000. Beginner's Hindi Script (Teach Yourself Books). ISBN: 0658009109

Copyright (c) 2001 Sean M. Burke. All rights reserved.

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself.

This program is distributed in the hope that it will be useful, but without any warranty; without even the implied warranty of merchantability or fitness for a particular purpose.

Much of Text::Unidecode's internal data is based on data from The Unicode Consortium, with which I am unafiliated.

Sean M. Burke CWsburke@cpan.org