Found another inconsistency between Unicode and the TM/TMW docs. I've sent e-mail to Tony Duff asking who's right, but I'm putting this in the errata under the assumption that even if Unicode is wrong, Unicode's wrong view will somehow rule the day.
Also, TMW->EWTS now generates \uF021-\uF0FF or \u0F00-\u0FFF escapes when appropriate. A few TMW glyphs still give errors.
Also, there's now a test to be sure that TM<->TMW and TMW->EWTS won't break in the future (except for the one glyph in TMW that isn't in TM, that one isn't tested). The baselines have not been hand-verified, but changes will be detected.
Also, TMW->EWTS now generates \uF021-\uF0FF or \u0F00-\u0FFF escapes when appropriate. A few TMW glyphs still give errors.
Also, there's now a test to be sure that TM<->TMW and TMW->EWTS won't break in the future (except for the one glyph in TMW that isn't in TM, that one isn't tested). The baselines have not been hand-verified, but changes will be detected.
Readded the line for reversed dza; it should never have been deleted, as that breaks TM<->TMW. I tested the whole mapping by hand once; this incident shows that automation is very helpful.
'{' and '}' were swapped...
The Unicode for something was "", not "none".
+R, +W, +Y, R+ now in use (though more testing is needed)
stacks that use full-form subjoined RA and YA consonants.
ACIP {RVA} was converting to the wrong things.
The TMW for {RVA} was converting to the wrong ACIP.
Checked all the 'DLC' tags in the ttt (ACIP->Tibetan) package.
non-reference, publicly available ACIP files (hundreds of megabytes of
them) through the converter. The frequencies of these tsheg bars in
in the file, too.
This mechanism is for Andres (who noticed KAsh=>K+sh in practice) and power users only, and not power users until I document the thing outside of the source code.
Now Jskad has two "Convert selected ACIP to Tibetan" conversions, one with and one without warnings, built in to Jskad proper (not the converter, that is).
Now Jskad has two "Convert selected ACIP to Tibetan" conversions, one with and one without warnings, built in to Jskad proper (not the converter, that is).
The code would be cleaner if I could bear to delete my terrible hack. Maybe in a month, when I don't feel so dumb for coding it up in the first place.
The correct solution for such things is to give the ACIP->Tibetan converters a pre-filter mechanism. This would be before the lexer or part of the lexer (maybe you only want to filter tsheg bars), and it would allow the end user to specify things like "s/SNYAM'AM/S+NYAMA'AMA/g".
Fixed ACIP->Unicode/TMW for BDE, which should be B-DE, not B+DE, because the former is legal Tibetan.
The ACIP->EWTS subroutine has improved.
TMW->Wylie and TMW->ACIP are improved in error cases.
TMW->ACIP has friendly embedded error messages now.
because I don't know which glyphs o and x correspond to. For that
reason, they cause ERRORs.
The proposed THDL Extended Wylie ~X and X is now used for U+0F35 and
U+0F37 respectively.
TMW->Wylie text. All the conversions show you which format they take
as input and which format they give as output.
File filter for ACIP files added.
The GUI converter suggests a file extension wisely.
Fixed newline bug in ACIP->Unicode converter.
that the following would look better with shortened 'dreng-bu also,
but I'm sticking with the TM/TMW docs:
dz+r~137,2~~4,46~1,110~4,120~1,123~1,126~4,106~4,113~f5b,fb2
dz+w~138,2~~4,47~1,110~4,120~1,123~1,126~4,106~4,113~f5b,fad
dz+h~139,2~~4,48~1,110~4,120~1,123~1,126~4,106~4,113~0F5C
dz+h+y~140,2~~4,49~1,110~4,121~1,123~1,126~4,107~4,114~f5c,fb1
dz+h+r~141,2~~4,50~1,110~4,121~1,123~1,126~4,107~4,114~f5c,fb2
dz+h+l~249,2~~4,51~1,110~4,123~1,123~1,126~4,110~4,117~f5c,fb3
dz+h+w~143,2~~4,52~1,110~4,122~1,123~1,126~4,108~4,115~f5c,fad
mappings, so I've put them back, but with the EWTS non-correspondences
\tmwXYYY.
Jskad no longer supports superscribed or subscribed numerals, because
EWTS does not.
Brought <?Other?> closer to EWTS
Removed __TILDE__ (no longer in EWTS)
Changed M^ to ^M per new EWTS draft
Added ai, au, -i from WW tibwn.ini -- they were missing in this version
bugs; it is pre-alpha. It's usable, though, and finds tons of errors
in ACIP input files, with the user deciding just how pedantic to be.
The biggest outstanding bug is the silent one: treating { }, space, as
tsheg instead of whitespace when we ought to know better.
that say "ya can take a ga prefix" etc.
The ACIP->Unicode converter now gives warnings (optionally, and by
default, inline). This converter now produces output even when
lexical errors occur, but the output has errors and warnings inline.
\, the Sanskrit virama, is not used. Of the 1370-odd ACIP texts I've
got here, about 57% make it through the gauntlet (fewer if you demand
a vowel or disambiguator on every stack of a non-Tibetan tsheg bar).
b, c, d, e, ... do not belong in ACIP, so the scanner rejects them.
This should make it even easier to distinguish automatically between
Tibetan and English texts.
is a correction, that's a comment, this is Tibetan, that's Latin
(English), that's Tibetan inter-tsheg-bar punctuation, etc.) It now
accepts more real-world ACIP files, i.e. it handles illegal
constructs. The error checking is more user-friendly. There are now
tests.
Added some tsheg bars that Peter E. Hauer of Linguasoft sent me to the
tests. Many thanks, Peter. I still need to implement rules that say,
"This is not Tibetan, it must be Sanskrit, because that letter doesn't
take a MA prefix."
up that String into tsheg bars, punctuation, etc., while finding
errors. I've tested it some, but I'm not yet committing the tests.
Next step: a converter that takes an ACIP file as input and outputs
TMW+Latin.
and it has the capability to produce error messages and warnings that
make sense to the user. One can now get the correct parse, if one
exists, for an ACIP tsheg bar.
One could even feed in ACIP and get a list of warnings about things as
innocuous as PADMA, which a dumb converter would have trouble with.
One could then turn ACIP into well-behaved ACIP for that dumb
converter, if you really wanted to.
Still to do:
o Scan ACIP files into tsheg bars.
o Produce TMW/Latin (from which you can get Unicode, etc.).
o E-mail the illegal tsheg bars to the ACIP fellows so they can fix
the affected documents (most of the Kangyur has unparseable
creatures).
Our disambiguation is now perfect, happening when and only when it is
necessary. These are all illegal, so it shouldn't affect many
existing conversions. But if there were typos, it could.
Our disambiguation is now perfect, happening when and only when it is
necessary. These are all illegal, so it shouldn't affect many
existing conversions. But if there were typos, it could.
This way through clicking on the application through the wizard one can choose
to connect to the available on-line dicts, open a local dict or generate a dict database.
(which I found on suigeneris.org, not apache.org) in order to bulletproof the
Tibetan Converter tests. They used to fail due to nondeterminism in the
Java RTF writer; they should no longer fail.
I've also changed it so that the Tibetan Converter tests run in headless
mode, which means that they'll run on the nightly builds server.
or <?Numbers?> commands; it instead hard-codes the appropriate comma-
delimited lists. This is cleaner because WylieWord and Jskad had different
values for these lists.
TMW->Wylie conversions with the new-and-improved TMW->Wylie
algorithm faulty.
Now I'm using it a little more than you need to, e.g. b.lha instead of blha is
generated because bla and b.la are ambiguous.
TMW->Wylie conversions with the new-and-improved TMW->Wylie
algorithm faulty.
Now I'm using it a little more than you need to, e.g. b.lha instead of blha is
generated because bla and b.la are ambiguous.
like \bullet, \emdash, etc., and this fix only works for Windows or OS/2 RTF
files, not for Mac RTF files. So if you want a TM->TMW conversion to work,
use MS Word for Windows, not for the Mac.
'<' and '>'. The current keyboard implementation makes this an either-or
proposition, when fundamentally it need not be.
Added a <?Numbers?> command and an <?Input:Numbers?> command to
tibwn.ini; broke the numbers apart from the consonants. This facilitates the
new-and-improved Tibetan->Wylie conversion.
Tibetan->Wylie is now done by forming legal tsheg-bars. A legal tsheg bar
is converted into perfect THDL Wylie. See code comments to learn what
it thinks is a legal tsheg-bar, but it inlcudes bskyUMbsH minus the trailing
punctuation (H), e.g.
Illegal sequences, such as runs of transliterated Sanskrit, are turned into
unambiguous Wylie; each glyph is followed by a vowel or a disambiguator
('.').
I've made it so that the illegal sequences are as beautiful as possible. You
get 'pad+me', for example, not the equivalent but uglier 'pad+m.e.'.
mouse-clicked on the new Jskad window, you could cause an infinite
regression of requestFocus() operations because the menu would try
to get focus back. I grab focus from the menu now.
work on a Linux console, e.g. The JUnit tests will too, though 'ant
check' still fails because we don't sneak the -Djava.awt.headless=true
into the process early enough.
which means that the command-line tool can finally function with a headless
graphics device. Hopefully it will speed things up, too. It also means that
entering Roman text into the TMW->Unicode conversion and TMW->TM
conversion will be easy.
Added support for two more oddballs.
Deprecated the oddball lookup method because it drops up to 30 glyphs in
TibetanMachine. The correct solution is to transform the RTF before Java's
busted RTF readers ever see it. \'97 becomes \u151, e.g.
beginning of the document as they should and as they are documented to.
They now do, and they bracket the bad characters with the TM or TMW for
U+0F3C on the left and the TM or TMW for U+0F3D on the right.
Some cleanup.
the troublesome glyphs are now put at the beginning of the document
AFTER AN ACHEN. This makes a glyph like \tmw7095 visible atop the
achen.
Major fix to the handling of paragraphs in conversion; we were (for
whatever reason) dropping paragraphs before.
faster than TMW->Unicode etc.; this is because many fewer replacements
are made (i.e., more text is replaced each time a replacement is
performed).
I must find a way to still preserve formatting but do many fewer
replacements in TMW->{Unicode,TM} and TM->TMW.
inserting 5 characters at a time and then skipping ahead just one
position. I don't think this affected correctness.
I believe there's still a terrible (exponential?) slowdown as the
input file gets bigger, however. Perhaps not -- but we run through
the first 1000 TMW glyphs in 6 seconds, the 20th thousand takes at
least 60 seconds. Is TMW->Wylie faster than TMW->Unicode? If so,
why?
Thought: don't use a DuffPane within TibetanConverter -- it can only
add overhead, right? My hprof profile said that the conversion was
taking just a couple of percent of the work; the rest was going to
display-related stuff that you should only see if you were displaying
the document. I'm not!
Is the EWTS '_' to be represented as U+0020, or is it a wider space?
Does TMW9.42, Dza, map to U+0F5F,U+0F39?
Does TMW6.60, r+y, map to U+0F62,U+0FBB or to U+0F6A,U+0FBB? (Likewise with r+w, TMW6.61, TMW6.62, etc.)
Is U+0F7E a bindu? What Unicode does TMW7.96 map to, for example? What does TMW7.91 map to?
Should TMW8.97 and TMW8.98 map to swastiskas elsewhere in Unicode? If so, which codepoints? Likewise with TMW9.60, a Chinese character.
Does TMW7.68 map to U+0F39?
Does TMW7.74, the ITHI secret sign, have a Unicode mapping? f68,fa0,f80,f72 comes close, but fa0 would be too large, wouldn't it?
What Unicode does TMW9.61 map to? Is it for sequences like f40,f7c,f60,f72? Or is it for f60,f72,f7c?
and then tried to convert that TMW to Wylie. I swear it's Java's
problem (see the ugly stack trace in the code and decide for
yourself), and I tried replacing rather than
inserting-and-then-removing, but it didn't work. I've left these
things as options.
default, all oddballs to appear once in the resulting document.
This'll help me find the correct glyphs for the oddballs, and it'll
prevent the average user from converting a document with oddballs.
I've fixed that.
I've also added a couple of Unicode mappings to give a flavor for how
multi-codepoint mappings will be represented.
TM->TMW conversion takes about 1 second per thousand glyphs on my
PIII-550.
noticed that formatting is mostly OK but sometimes gets bungled slightly.
I tried everything I could think of, and now I'm passing the buck to Java's
RTF support.
TMW_RTF_TO_THDL_WYLIE (now misnamed) support TMW->TM
conversion (but not TM->TMW). There is an automated test case for a
TMW->TM conversion.
I have full confidence in this conversion. Even the smallest glitch in the core
functionality (not formatting) would surprise me.
Note that the JUnit test TMW_RTF_TO_THDL_WYLIETest sometimes fails
due to one- or two-line diffs between the actual and expected outputs. This
is because Java's RTF support is not deterministic, I'm guessing, and is not
a real failure. I'm too lazy to make a more elaborate sed/diff mechanism
that works on all platforms, and that would complicate the build anyway.
verified this extensively and have full confidence that these mappings
agree with Tony Duff's Tibetan! 5.1 documentation (except as described
below).
To get them, I had to disregard Tony Duff's tables for a few glyphs: the
characters with ordinal 32 and 45 (space and hyphen in Roman ASCII,
space and tsheg in Tibetan). For these glyphs, we must have mappings
from TibetanMachineSkt4.32 to something, etc., and those mappings were
not present. I've normalized the mapping for these glyphs, as it is arbitrary
because the same two glyphs just appear fifteen times each.
