Difference between revisions of "English and Kazakh"

From Apertium
Jump to navigation Jump to search
Line 211: Line 211:
 
***Where/When + NP or PP, but not for WHAT
 
***Where/When + NP or PP, but not for WHAT
 
***Simple questions Do,Did + NP or PP
 
***Simple questions Do,Did + NP or PP
  +
***how write rule for "Are/Were?", rule for "Noun is/was" has the same pattern
 
** 'which' as a determiner → precedes noun ("which house" → "қайсы үй") or genitive construct ("үйдің қайсысы")
 
** 'which' as a determiner → precedes noun ("which house" → "қайсы үй") or genitive construct ("үйдің қайсысы")
  +
**"which" as adv-itg, "which house do you like?"
 
* relatives (simple relatives: "the book that I wrote", "the book which I wrote" → "Мен жазған кітап"; adverbial relatives "when he came" → "Ол келгенде" [uses locative!])
 
* relatives (simple relatives: "the book that I wrote", "the book which I wrote" → "Мен жазған кітап"; adverbial relatives "when he came" → "Ол келгенде" [uses locative!])
 
**added rules for simple relatives(that,which)
 
**added rules for simple relatives(that,which)
Line 221: Line 223:
 
***lexical selection for "anything" as "yeshnarse" for negative sentences and as "bir narse" for affirmative sentences.
 
***lexical selection for "anything" as "yeshnarse" for negative sentences and as "bir narse" for affirmative sentences.
 
* Choosing auxiliaries for present continuous ("be" → "bol" (default), "zhatyr", "otyr" ,etc.)
 
* Choosing auxiliaries for present continuous ("be" → "bol" (default), "zhatyr", "otyr" ,etc.)
  +
**can't be solved by lexical selection
 
* deciding t1x versus t2x:
 
* deciding t1x versus t2x:
 
** NPs and PPs in t1x as long as possible (hard design choice, tedious work, code repetition, but...)
 
** NPs and PPs in t1x as long as possible (hard design choice, tedious work, code repetition, but...)

Revision as of 17:54, 28 January 2014

Starting work on Apertium English to Kazakh

These notes are basically for Anel, Aizhan and Assem who have started to develop this language pair... And Aida too...

Installing what is needed

Operating System

Install a suitable GNU/Linux system such as Debian, Ubuntu, Mint...

Install build essentials, etc.

Open a terminal window and type

sudo apt-get install subversion build-essential g++ pkg-config gawk libxml2 \
libxml2-dev libxml2-utils xsltproc flex automake autoconf libtool libpcre3-dev \
cmake libicu-dev libboost-dev libgoogle-perftools-dev bison libreadline-dev zlib1g-dev

enter your password and Wait till the packages are downloaded and installed.

If you don't already have a directory for sources, make one in your home directory and enter it:

cd ~
mkdir Source
cd Source

Install HFST

This language pair uses the Helsinki Finite State Toolkit for Kazakh generation, so we need to install it, and its dependencies. (But OpenFST is now included with HFST, so there is no longer a need to install OpenFST separately.)

Install Foma

Main article: Foma
svn checkout http://foma.googlecode.com/svn/trunk/foma/ foma 
cd foma
make
sudo make install
cd ..

Install HFST

Main article: HFST
svn co https://svn.code.sf.net/p/hfst/code/trunk/hfst3
cd hfst3/
./autogen.sh
scripts/generate-cc-files.sh # It's OK if this step fails
./configure --enable-lexc --with-foma --disable-tagger --enable-proc
make
sudo make install
sudo ldconfig
cd ..
Troubleshooting

When doing "make" with old autotools (pre 1.14?)

make[5]: *** No rule to make target `xre_parse.hh', needed by `xre_lex.ll'.  Stop.

Run scripts/generate-cc-files.sh and then make again.

Install VISLCG3

Main article: Apertium and Constraint Grammar
svn co http://beta.visl.sdu.dk/svn/visl/tools/vislcg3/trunk vislcg3
cd vislcg3
./cmake.sh 
make -j3
sudo make install
cd ..

Download apertium, lttoolbox and eng-kaz data from SVN

Main article: Minimal installation from SVN
cd ~/Source
svn co https://svn.code.sf.net/p/apertium/svn/trunk/lttoolbox
svn co https://svn.code.sf.net/p/apertium/svn/trunk/apertium
svn co https://svn.code.sf.net/p/apertium/svn/trunk/apertium-lex-tools
svn co https://svn.code.sf.net/p/apertium/svn/trunk/apertium-tools
svn co https://svn.code.sf.net/p/apertium/svn/languages/apertium-kaz
svn co https://svn.code.sf.net/p/apertium/svn/incubator/apertium-eng-kaz

Compile and install lttoolbox

cd lttoolbox/
PKG_CONFIG_PATH=/usr/local/lib/pkgconfig ./autogen.sh
make
sudo make install
sudo ldconfig

Compile and install apertium

cd ..
cd apertium/
PKG_CONFIG_PATH=/usr/local/lib/pkgconfig ./autogen.sh
make
sudo make install
sudo ldconfig

Compile and install apertium-lex-tools

cd ..
cd apertium-lex-tools
PKG_CONFIG_PATH=/usr/local/lib/pkgconfig ./autogen.sh
make
sudo make install
sudo ldconfig

Install Kazakh language

cd ..
cd apertium-kaz
./autogen.sh
make

Install English--Kazakh language pair data from incubator

cd ..
cd apertium-eng-kaz/
PKG_CONFIG_PATH=/usr/local/lib/pkgconfig ./autogen.sh --with-lang2=$HOME/Source/apertium-kaz
make

Troubleshooting

If you get:

lt-comp: error while loading shared libraries: liblttoolbox3-3.2.so.0: cannot open shared object file: No such file or directory

Then you should do:

sudo ldconfig

Browse SVN

Here you can look at changes that have been made:

http://sourceforge.net/p/apertium/svn/HEAD/tree/incubator/apertium-eng-kaz/

Contact

IRC

Open up XChat (normally "Programs -> Internet -> XChat IRC") and type:

/server irc.freenode.net
/join #apertium
/join #hfst

To install xchat:

sudo apt-get install xchat 

In Windows:

http://www.silverex.org/download/

Chat logs/archives: http://alpha.visl.sdu.dk/~tino/pisg/freenode/logs/

Mailing list

Email: apertium-turkic@lists.sourceforge.net

http://blog.gmane.org/gmane.science.linguistics.turkic.mt

November 2013 to-do list

  • Check the constraint-grammar file for strange rules and also rules that may not be correct. Try to understand the rules we have.
  • Write .t1x and .t2x code to deal with case (capitalization)
  • Make sure all new rules have the correct superblank management
    • That is, where a rule reorders items, all superblanks in the reordered area should go before the reordered area
  • The copula, t2x rules: past copula with PP ("I was from Kazakhstan"), negative copula, adverbial adjective phrases in copula "The man is very large"
    • added rule for past copula with PP
    • rule with negative copula for PP in past and present
    • solved by adding rule "preadv + adj" as AdjP
  • NPs with adverbs, particularly "very" ("Three very beautiful children")
    • solved by adding rule "preadv + adj" and "preadv"
  • 1-word chunks to always provide a translation for any English word
    • solved partly by adding pronouns: before,behind,below,in,on,after,towards,through, under as adverbs, and rules in constraint grammar.
    • even for stranded prepositions, so that we have a translation for "in" similar to "inside" etc., as if they were adverbs
  • noun-noun compounds in NPs and PPs: add the most frequent ones to t1x (hard to do in t2x as prepositions are solved in t1x)
    • Solved by adding rules to t1x:
      • noun1 noun2
      • adjec noun1 noun2
      • det num noun1 noun2
      • num noun1 noun2
      • prep num noun1 noun2
      • num adjec noun1 noun2
      • det num adjec noun1 noun2
      • prep num adjec noun1 noun2
      • prep det num adjec noun1 noun2
  • interrogative sentences: Yes/no (-ba) and informative ("Where is my Kazakh dictionary"?) → probably work for t2x
    • special questions were done by adding rules:
      • Where/When + NP or PP, but not for WHAT
      • Simple questions Do,Did + NP or PP
      • how write rule for "Are/Were?", rule for "Noun is/was" has the same pattern
    • 'which' as a determiner → precedes noun ("which house" → "қайсы үй") or genitive construct ("үйдің қайсысы")
    • "which" as adv-itg, "which house do you like?"
  • relatives (simple relatives: "the book that I wrote", "the book which I wrote" → "Мен жазған кітап"; adverbial relatives "when he came" → "Ол келгенде" [uses locative!])
    • added rules for simple relatives(that,which)
    • added rule for "when"
  • check some changes made to punctuation regular expressions in the English dictionary to solve mismatches with the Kazakh dictionary
  • "-ing" is hard (check the appropriate section in Tagging_guidelines_for_English. This gives problems in "I like playing football" vs. "I like flying birds" and will be hard. Transitivity could be a clue? What to do in t1x and what in t2x? Also "Flying planes can be dangerous", famous ambiguity). Try to get as much as possible done with CG rules.
  • Negative pronous ("yesh" forms)
    • write lexical selection to generate "yesh-" forms from "any-" forms in negatives or "bir" forms in questions, e.g. "do" "not" vblex.inf "anything" (dictionaries should be populated with alternatives)
      • lexical selection for "anything" as "yeshnarse" for negative sentences and as "bir narse" for affirmative sentences.
  • Choosing auxiliaries for present continuous ("be" → "bol" (default), "zhatyr", "otyr" ,etc.)
    • can't be solved by lexical selection
  • deciding t1x versus t2x:
    • NPs and PPs in t1x as long as possible (hard design choice, tedious work, code repetition, but...)
  • Some adjective phrases like "num "years old"".
  • Comparative constructs (more ADJ than NP → NP-dat karaganda ADJ-comp)
    • is done, by 4 rules in t2x
    • added comparative and superlative adj, in the biggest city as PP
  • Adverbial phrases: think on how to treat them similarly to PP in .t2x ("very quickly" is not that different from "in the park" when it comes to t2x reordering)
  • (Partly done) Pseudo-modals "finish" "start" "love" "hate" "enjoy", "like" which take -ing and sometimes "to".... Three possibilities to deal with them: (1) a long def-cat, (2) def-list and tests in rules, and (3) Jim's <exception> (dangerous!). Route: change new-gen-simple-verb macro with a deflist, to generate VP_psmod, translate "-ing" into NPs (as they take case,etc.) and write .t2x rules. Careful: -ing desambiguation not too goodd
  • Collect parallel kaz-eng corpora!

What to take care of when writing rules

What are these systems going to be used for and how does this affect design?

It is quite unlikely that a system like this will ever be used for postediting the output into publishable text. The output is quite unlikely to be useful, particularly for sentences longer than a few words, as it would be very difficult to get the right word order.

It might be used, however, for:

  • (1) interactive MT as in http://www.dlsi.ua.es/bbcat/?slang=eng&tlang=kaz ?
  • (2) fuzzy-match repair (when a translator using a computer-aided translation system gets a very good fuzzy match from a translaton memory, MT output can be intelligently used to find which parts of the target side need to be changed an actually change them (a thesis at the Universitat d'Alacant). This is because short segments may get very good translations.
  • (3) assimilation or gisting (understanding what a text is about); the evaluation of this may be tricky but some Apertiumers have had interesting ideas: http://journal.sepln.org/sepln/ojs/ojs/index.php/pln/article/view/4867

Indeed, evaluation may be tricky in general.

Development should take these applications into account:

  • (1) and (2): getting good translations for short segments (2, 3, 4 words) can be very helpful here
  • (3): the idea here would be not to pay attention to features that do not impair understanding (e.g, English articles can be deleted; "of Kazakh constituents order acceptable may be", etc.). good translations for short phrases (linguistically motivated segments) could be the key here

Questions we had, open issues

  • morphology of reflexives ("Öz") → Mikel has to talk to Jonathan and Ilnar to make it work as in apertium-kir or as bir-bir reciprocals. Make morphology describe the real morphotactics of these forms.
  • why do we have gender in Kazakh morphologies when gender is not represented? Make morphology describe the real morphotactics of these forms??? (Fran gave reasons for not doing so, check apertium-turkic)

November 2013 work done

Aida will complete and document this list

  • Regression test is completed with new sentences 426/426
  • Structural transfer
    • Reported speech sentences
    • Conditionals(First and Second)
    • "be" + adjective in present and past : "You are/were beautiful"
    • "be" + PP: "I am from Kazakhstan"
    • Rule for demonstrative pronouns
    • Rule for negative pronouns("nothing,nobody, anything(only for negative sentences)") changing verb to negative
    • Def-list of pseudo-mod verbs(I LIKE/ENJOY/LOVE/HATE/START/FINISH playing) int .t1x and choosing them in gen-simple-verb macro
    • Rule for "-ing" words as NP<subst> in .t1x,for example, I like playing.
    • "ing" + NP in .t2x
    • Rule for "would"(.t1x) + NP(.t2x)
  • Lexical selection
    • One rule for "residence"
    • Rules in CG
  • Dictionary work
    • Put some country and city names into apertium-eng-kaz.eng-kaz.dix as NP-TOP
    • Added missing pronouns to bilingual dictionary
    • Corrected verbs which iv to tv(tv to iv)
    • Changed "would" <inf> to <past> in eng.dix

Old stuff scheduled for removal

Some of this information is outdated and needs work, but make sure that everything is there before removing this part.

Postpositions

Apparently Kazakh has 5 kinds of postpositions, according to the case of the NP they follow. Some of those following genitive may be interpreted as "nouns" with a case, such as

бақшаның астында

garden-of bottom-in

garden.gen bottom.loc

"under the garden"

where астын is roughly the noun "bottom", much as in Basque "ortu-a-ren azpi-an" "azpi" is a noun.

With nominative (or base form)

Check this list:

  • арқылы through
  • туралы about
  • секілді similarly to
  • жөнінде about

With genitive

  • астынан from below
  • астында above (top-its-in)
  • жанынан from beside (side-its-from)
  • жанында beside (side-its-in)

With dative

  • қарай (towards)
  • арналған (intended for)

With ablative

  • кейін behind, after

With instrumental

  • қатар beside
  • бірге together with

Starting work on Apertium Kazakh to English

General ideas

Try to translate as literally as possible in the first prototypes (do not have too many .t2x rules)

Make the most of existing CG-based PoS tagging (wait for instructions on how to use the apertium-kaz.kaz.rlx in apertium-kaz)

Detecting NPs and PPs

There is a lot of stuff in apertium-eng-kaz.kaz-eng.t1x already! We have to study, and check the following.

Main kinds of NPs:

  • accusative and nominative → no preposition
  • genitive → two solutions: N's N or N of N (attention genitive chains)
  • dative → what should one do? (tricky)
  • locative, ablative (make list) → PPs
    • what to do with possessives (particularly 1st and 2nd person) to avoid double possessives in sentences with "mening", etc.
      • Менің бақшам → my garden of me (!)

Composition of NPs: n, adj n, num n, num adj n, ...

Things to take care of:

  • Decide if noun-based postpositions: artynda, ustinde, keyin, etc. will be detected in t1x. A list of lemmas would be necessary, or changes to bilingual dictionaries
  • Make sure we generate plurals for numbers
  • articles (use third-person possessives as a hint to generate definite articles)
    • kalaning baqshasythe garden of city

Detecting VPs

  • simple verbs: decide on reasonable equivalents
    • some may be hard to decide, such as generating future simple in English different from present
    • present or past perfect to generate. Only present perfect
  • compound forms based on zhatyr, otyr, etc.
  • generating negatives (have negative VPs detected separately or use logic (choose) inside t1x rules
  • gender in third-person pronouns (including 'öz' reflexives)

Loose list of problems

  • constructions based on infinite verbs (participles, etc.) (the problem of generating tense)
  • reinserting the verb to be when the copula is missing

November 2013 work done

Aida will complete this part

  • Regression test added
  • Transfer
    • Continuous tenses, simple tenses, negatives
    • Subject pronouns (gender still an open issue)
    • Nouns and adjectives
      • Deleted n.attr from adjective definition
  • Dictionaties
    • changed "жатқан жоқ" in kaz.lexc and eng-kaz.dix to vaux-negative to catch negative present continuous
    • added some words