Difference between revisions of "User:Kamush/GSoC2021ProgresReport"

From Apertium
Jump to navigation Jump to search
Line 126: Line 126:
| -
| -
|
|
* Some additions to apertium-kaz & apertium-uzb
* Some transfer rules
* Some transfer rules
* Some more lexical selection rules
* Some more lexical selection rules & bidix
* Some more bidix
|-
|-
|Week 9
|Week 9

Revision as of 17:10, 1 August 2021

Progress Report

Time Period Goal Bidix Coverage WER,PER Details/Comments
kaz-uzb kaz-uzb kaz-uzb uzb-kaz
Community Bonding Period

May 17-June 5

  • Installed Apertium
  • Initialize kaz-uzb pair
  • Collect data in both languages
426

(+426)

43.80 % - -
  • Installed Apertium and necessary tools;
  • Cloned Apertium-kaz and apertium-uzb, initialized the kaz-uzb pair
  • Translated a small sample text;
  • Extracted Uzbek and Kazakh wiki corpus;
  • Collected Kazakh-Uzbek dictionary and parallel corpora;
Week 1

June 6-12

Make Uzbek better 2220

(+1794)

52.11 % - -
  • Went through all Uzbek and Kazakh stems;
  • Initialized the pair with apertium-recursive;
  • Collected dictionaries from other pairs for crossdic;
  • Obtained crossdic results from two ways.
Week 2

June 13-19

Expand bilingual dictionary 5262

(+3042)

77.03 % 74.77% / 67.57% 64.23% / 54.37%
  • Started adding bilingual dictionary elements;
Week 3

June 20-26

More on .dix and .lrx 8543

(+3281)

81.55 % 74.77% / 67.57% 64.23% / 54.37%
  • Expanded bilingual dictionary;
  • Started sample Lexical selection rules;
Week 4

June 27-July 3

Focus on transfer rules 9432

(+889)

81.85 % 74.77% / 67.57% 64.23% / 54.37%
  • Expanded bilingual dictionary more;
Week 5

July 4-10

Test translator and expand more 11008

(+1576)

82.81% 74.77% / 67.57% 64.23% / 54.37%
  • Expanded bilingual dictionary;
  • Collected texts for lexical selection rules, tried a small script;
  • Translated a Big Kazkh text into Uzbek for better WER/PER calculation.
Week 6

July 11-17

Focus more on transfer rules +191 84.46 % 71.83% / 63.52% 67.24% / 60.31%
  • Addeed more stems to the bidix to reach 85% coverage
  • Added 1500+ Lexical Selection Rules to kaz-uzb.lrx for ambiguous translations
  • Added 1200+ Lexical Selection Rules to uzb-kaz.lrx for ambiguous translations
  • Prepared 200 sentences of parallel corpora.
Week 7

July 18-24

Test the kaz-uzb translator +219 84.58 % 69.21% / 60.69% 67.10% / 59.89%
  • Started the transfer rules
  • Some more bidix
  • Some Lexical selection rules
Week 8

July 25-31

Focus on transfer rules +500 85.17 % - -
  • Some additions to apertium-kaz & apertium-uzb
  • Some transfer rules
  • Some more lexical selection rules & bidix
Week 9

August 1-7

Focus on testvoc - - - - -
Week 10

August 8-14

Finalize work - - - - -

TODO

  • More transfer rules
  • Testvoc

ONGOING

  • Transfer rules
  • Some lexical selection rules
  • Collecting more bidix
    • Adding only most-frequent words from kazakh wiki hitparade until 85% coverage is reached

DONE (+Notes & Comments)

  • Made a script to calculate WER for each line of the kaz-uzb.txt file to see what sentences are causing the most problems.
  • Translating big Kaz text into Uzb
    • Aside from those small texts obtained from parallel corpora
    • For better WER/PER calculation
    • For checking transfer rules
    • Started with readily available small JaM Story
      • Azmat and Oygul story in our case
      • Added 47 sentences from that.
    • Chose Nur-Sultan(capital city) article of Kazakh Wiki for that.
      • Split the article into sentences, translated into uzb manually.
      • Added 112 sentences out of Nur-Sultan text.
    • Used QED parallel corpus form internet.
      • Manually selected sentences and corrected their translations.
      • Added 90 sentences from this.
    • Made 250 sentences in total, stopping this here.
  • Tried finding parallel corpora:
  • Lexical selection rules for kaz-uzb
    • Created a script to analyse bidix and find ambiguous stems.
    • Created a script that generates lexical selection rules from manually chosen stems
    • Rules added in kaz-uzb.lrx for 1565 words.
    • Rules added in uzb-kaz.lrx for 1295 words.
  • Apertium-dixtools:
   apertium-dixtools fix apertium-kaz-uzb.kaz-uzb.dix apertium-kaz-uzb.kaz-uzb.dix.fixed
    • Sorting bidix(aligning too):
   apertium-dixtools sort -alignBidix -ignorecase apertium-kaz-uzb.uz-uzb.dix apertium-kaz-uzb.kaz-uzb.dix.sorted
  • Sorting+Deduplicating the bidix for future purposes:
    • Bidix count before deduplication: 11008
    • Fixing the bidix:
      • Removed 1501 entries in section main
    • Bidix count after deduplication: 9507
    • Sorted+Regrouped the dix.
  • Made a script to calculate DixCount, Coverage, WER/PER at once
  • Calculating WER/PER:
   $ cat kaz-big.txt | apertium -d ../ kaz-uzb > kaz-uzb.txt
   $ cat uzb-big.txt | apertium -d ../ uzb-kaz > uzb-kaz.txt
   $ apertium-eval-translator -ref uzb-big.txt -test kaz-uzb.txt
   $ apertium-eval-translator -ref kaz-big.txt -test uzb-kaz.txt
    • Parallel text:
      • JaM Story:
      • “Azamat va Oygul” in our case;
      • kaz-uzb/texts/[kaz|uzb].txt
    • Astana article from Kazakh Wiki
  • Calculated dix Coverage:
    • (kaz-uzb/texts): bash ../../coverage-ltproc-new.sh ../docs/kaz-wiki.txt ../kaz-uzb.automorf.bin
    • coverage: 26752095 / 32305875 (~0.82808761564266561423)
    • remaining unknown forms: 5553780
    • kaz-wiki.txt Sun Jul 11 11:55:25 CEST 2021
  • Counting Dix elements:
    • Apertium-Eval: dixcounter.py:
   python3 ../dixcounter.py apertium-kaz-uzb.kaz-uzb.dix
   python3 WikiExtractor.py --infn kkwiki-20210501-pages-articles-multistream.xml.bz2
    • Saved resulting corpus in docs/kaz-wiki.txt, ~24M tokens
  • Big Crossdix action:
    • Crossing dictionaries: kaz-kaa and kaa-uzb
    • Crossing dictionaries: kaz-tur and tur-uzb
    • Merge the obtained crossdix outputs from two
    • Sort the merged file, removing duplicates
    • Align the result for better visibility
    • Manually check every translation, remove if bad, correct/add if necessary
    • Added 5000+ words from this.
  • Start collecting the kaz-uzb bilingual dictionary
  • Convert the pair to apertium-recursive
    • Done!!!, but I had to remove the entire bilingual repo and recreate it from scratch.
  • Write some lexical selection rules in kaz-uzb.
  • Translate small text (James & Mary story)
  • Translating a text file:
   cat texts/kaz.txt | apertium -d . -f line kaz-uzb
  • Translation of a sentence:
   echo 'Сәлем Әлем' | apertium -d . kaz-uzb
  • Bootstrapping a new language pair apertium-kaz-uzb
    • Installed Apertium-init from pip
      • Had a problem, solved it (tanks to @popcorndude).
    • Downloaded apertium-init.py
      • Did not work
   python3 apertium-init.py kaz-uzb --analyser=hfst --no-prob1 --no-prob2
   cd apertium-kaz-uzb
   ./autogen.sh --with-lang1=../apertium-kaz --with-lang2=../apertium-uzb & make
    • I had to convert it to apertium-recursive, so further flags to be added:
      • -t rtx
  • Final initialization command:
   python3 apertium-init.py kaz-uzb --analyser=hfst --no-prob1 --no-prob2 -t rtx
  • Forked(&Installed) necessary repos on GitHub:
    • Apertium-kaz
      • git clone git@github.com:kamush901/apertium-kaz.git
      • Works well
    • Apertium-uzb
      • git@github.com:kamush901/apertium-uzb.git
      • Works well
  • Installed Apertium and necessary tools:
    • Installed Apertium core using packaging
   wget https://apertium.projectjj.com/apt/install-nightly.sh -O - | sudo bash
   sudo apt-get -f install apertium-all-dev