Beefy Boxes and Bandwidth Generously Provided by pair Networks
Clear questions and runnable code
get the best and fastest answer
 
PerlMonks  

Re^3: Character encoding woes - unicode or not?

by tchrist (Pilgrim)
on Feb 02, 2012 at 13:13 UTC ( #951427=note: print w/ replies, xml ) Need Help??


in reply to Re^2: Character encoding woes - unicode or not?
in thread Character encoding woes - unicode or not?

Encode::Detect is my new best friend, thanks to Anonymous Monk.
Then you don't know it very well. It’s really rather poor. Here are the failure results on a small sampling of real-world test files:
 Status  Filename Right      E::D::Detector  ==================================================  Wrong!   7843118       ascii       UNABLE TO GUESS           9501430       utf8        UTF-8          10318897       cp1252      windows-1252          10329150       cp1252      windows-1252  Wrong!  10358003       MacRoman    UNABLE TO GUESS  Wrong!  10358042       MacRoman    UNABLE TO GUESS  Wrong!  10429209       MacRoman    UNABLE TO GUESS          10482611       cp1252      windows-1252  Wrong!  10542098       MacRoman    UNABLE TO GUESS          10617571       cp1252      windows-1252  Wrong!  10625668       iso-8859-1  windows-1252          10676968       cp1252      windows-1252          10677497       cp1252      windows-1252          10963661       MacRoman    UNABLE TO GUESS          11042188       macRoman    UNABLE TO GUESS          11212329       utf8        UTF-8          11287402       cp1252      windows-1252  Wrong!  11470876       MacRoman    windows-1252          11842027       iso-8859-1  windows-1252  Wrong!  11940257       ascii       UNABLE TO GUESS  Wrong!  11972335       MacRoman    UNABLE TO GUESS  Wrong!  12091502       iso-8859-1  windows-1252          12169614       utf8        UTF-8          12495435       MacRoman    windows-1252          12736309       MacRoman    windows-1252          14641909       MacRoman    windows-1252          14652344       utf8        UTF-8          14751857       cp1252      windows-1252          15037632       cp1252      windows-1252          15070898       cp1252      windows-1252  Wrong!  15154606       MacRoman    windows-1252          15201223       cp1252      windows-1252  Wrong!  15315962       iso-8859-1  Big5          15328020       cp1252      windows-1252  Wrong!  17298172       MacRoman    windows-1252  Wrong! 116059400       MacRoman    windows-1252
I have a working snapshot of a module that actually does work right on such things called Encode::Guess::Educated. It has no noncore dependencies. It is designed to detect the encoding of English-language biomedical research papers. It can reliably detect not merely ASCII and UTF-{8,16,32}, but also the very conflicting 8-bit encodings.

The reason it can do this is that it works off a training model. I looked at three different corpora to do this: one containing 3½M non-ASCII codepoints, one containing 14M of them, and one containing 29M of them. It makes an educated guest based on conformance to a particular model. And it does very well.

Right now it has only a CLI API and an OO API, no Export-based one. Here’s the easiest way to use the CLI API, via a simple program called gank:

$ gank 011526914.txt cp1252 $ gank 00*.txt Sym*.txt 0115*.txt 001313968.txt: ascii 001328180.txt: utf8 007499277.txt: iso-8859-1 Symbola602.txt: UTF-16 011526914.txt: cp1252 011535589.txt: iso-8859-1 011570876.txt: MacRoman
The underlying class’s default training model derives from the complete PubMed Open Access corpus, and it therefore attains an extremely high measured accuracy of 99.79% when used on English-language biomedical texts. It also does well on other texts using any Latin-based alphabet. I have comparative statistics using two alternate training models, but the PMCOA model is fine for most purposes.

You may also give gank a -s option to give you a short ‘score-card’ of the various encodings it considered:

*91.718532 +2.285393 MacRoman 3.640513 -0.941206 iso-8859-1, iso-8859-15, cp1252 3.639257 -0.941552 cp1250 1.001698 -2.231634 iso-8859-2
EXPLANATION:
  • The first column is all scores normalized to 0..100.
  • The second column is the natural log of the real score.
  • The rest is which encoding have that score, and in the order of preference for breaking ties of the same score. I have it arranged so it says it’s the smallest subset that works; i.e., ascii < latin1 < cp1252, etc.
There’s also a -l option to give you a long report that illustrates what each possible shoice would be if it were in that encoding, with paired lines of literal UTF-8 and \N{...} named characters.
total bytes=15903, high bytes=22, distinct high bytes=8
  *49.582509 +0.909655 cp1252
      => "I–V, Copyright © 2001 Outline • Acknowledgements 12000× g Südhof Marquèze, Llinás. ScienceDirect® is"
      => "I\N{EN DASH}V, Copyright \N{COPYRIGHT SIGN} 2001 Outline \N{BULLET} Acknowledgements 12000\N{MULTIPLICATION SIGN} g S\N{LATIN SMALL LETTER U WITH DIAERESIS}dhof Marqu\N{LATIN SMALL LETTER E WITH GRAVE}ze, Llin\N{LATIN SMALL LETTER A WITH ACUTE}s. ScienceDirect\N{REGISTERED SIGN} is"
   49.557280 +0.909146 cp1250
      => "I–V, Copyright © 2001 Outline • Acknowledgements 12000× g Südhof Marqučze, Llinás. ScienceDirect® is"
      => "I\N{EN DASH}V, Copyright \N{COPYRIGHT SIGN} 2001 Outline \N{BULLET} Acknowledgements 12000\N{MULTIPLICATION SIGN} g S\N{LATIN SMALL LETTER U WITH DIAERESIS}dhof Marqu\N{LATIN SMALL LETTER C WITH CARON}ze, Llin\N{LATIN SMALL LETTER A WITH ACUTE}s. ScienceDirect\N{REGISTERED SIGN} is"
    0.860211 -3.144560 MacRoman
      => "IñV, Copyright © 2001 Outline ï Acknowledgements 12000◊ g S¸dhof MarquËze, Llin·s. ScienceDirectÆ is"
      => "I\N{LATIN SMALL LETTER N WITH TILDE}V, Copyright \N{COPYRIGHT SIGN} 2001 Outline \N{LATIN SMALL LETTER I WITH DIAERESIS} Acknowledgements 12000\N{LOZENGE} g S\N{CEDILLA}dhof Marqu\N{LATIN CAPITAL LETTER E WITH DIAERESIS}ze, Llin\N{MIDDLE DOT}s. ScienceDirect\N{LATIN CAPITAL LETTER AE} is"
I need to do more work on its API — this is just a proof of concept, although it does comes with a halfway decent test suite — and of course document it, but I’m hunkered down right now correcting page-proofs on Camel4, so I probably won’t get to sprucing up the module for another 7–10 days.

--tom


Comment on Re^3: Character encoding woes - unicode or not?
Select or Download Code

Log In?
Username:
Password:

What's my password?
Create A New User
Node Status?
node history
Node Type: note [id://951427]
help
Chatterbox?
and the web crawler heard nothing...

How do I use this? | Other CB clients
Other Users?
Others romping around the Monastery: (17)
As of 2014-09-17 14:30 GMT
Sections?
Information?
Find Nodes?
Leftovers?
    Voting Booth?

    How do you remember the number of days in each month?











    Results (82 votes), past polls