JST_LOGO.JPG NICT_LOGO.JPG KYOTO-U_LOGO.JPG

WAT 2014

The 1st Workshop on Asian Translation
Baseline Systems
Data preparation for the JE and EJ subtasks

[BASELINE SYSTEMS TOP] | [SETUP] | [FOR PHRASE-BASED AND HIERARCHICAL PHRASE-BASED SMT] | [FOR STRING-TO-TREE AND TREE-TO-STRING SMT]

Setup

We assume that there are ASPEC text files (test.txt, dev.txt, train-1.txt, train-2.txt, and train-3.txt) in corpus.org/.
MOSES_SCRIPT=${path}/mosesdecoder-RELEASE-2.1.1/scripts
SCRIPT_DIR=${path}/script.converter.distribution
cd corpus.org/

  • Extracting sentences
  • for name in dev test; do
      perl -ne 'chomp; @a=split/ \|\|\| /; print $a[2], "\n";' < ${name}.txt > ${name}.ja.txt
      perl -ne 'chomp; @a=split/ \|\|\| /; print $a[3], "\n";' < ${name}.txt > ${name}.en.txt
    done
    for name in train-1 train-2 train-3; do
      perl -ne 'chomp; @a=split/ \|\|\| /; print $a[3], "\n";' < ${name}.txt > ${name}.ja.txt
      perl -ne 'chomp; @a=split/ \|\|\| /; print $a[4], "\n";' < ${name}.txt > ${name}.en.txt
    done

  • (Removing date expressions at EOS in Japanese in the training and development data to reduce noise)
  • for file in train-1 train-2 train-3 dev; do
      mv ${file}.ja.txt ${file}.ja.txt.org
      cat ${file}.ja.txt.org | perl -Mencoding=utf8 -pe 's/(.)[[0-9.]+]$/${1}/;' > ${file}.ja.txt
    done

    cd ..
    Back to top

    For Phrase-based and Hierarchical Phrase-based SMT

    mkdir corpus.tok
    cd corpus.tok

  • Tokenizing sentences in Japanese
  • for file in train-1 train-2 train-3 dev test; do
    cat ../corpus.org/${file}.ja.txt | \
      perl -Mencoding=utf8 -pe 's/ / /g;' | \
        juman -b | \
        perl -ne 'chomp; if($_ eq "EOS"){print join(" ",@b),"\n"; @b=();} else {@a=split/ /; push @b, $a[0];}' | \
        perl -pe 's/^ +//; s/ +$//; s/ +/ /g;' | \
        perl -Mencoding=utf8 -pe 'tr/\|[]/|[]/; ' \
        > ${file}.ja
    done

  • Tokenizing sentences in English
  • for file in train-1 train-2 train-3 dev test; do
      cat ../corpus.org/${file}.en.txt | \
        perl ${SCRIPT_DIR}/z2h-utf8.pl | \
        perl ${MOSES_SCRIPT}/tokenizer/tokenizer.perl -l en \
        > ${file}.tok.en
    done

  • Training truecaser for English
  • cat train-1.tok.en train-2.tok.en train-3.tok.en dev.tok.en > train_dev.tok.en
    ${MOSES_SCRIPT}/recaser/train-truecaser.perl --model truecase-model.en --corpus train_dev.tok.en

  • Truecasing English sentences
  • for file in train-1 train-2 train-3 dev test; do
      ${MOSES_SCRIPT}/recaser/truecase.perl --model truecase-model.en < ${file}.tok.en > ${file}.en
    done

  • Building training data for language models
  • cat train-1.ja train-2.ja train-3.ja > train-all.ja
    cat train-1.en train-2.en train-3.en > train-all.en

  • Cleaning training data for translation models
  • perl ${MOSES_SCRIPT}/training/clean-corpus-n.perl train-1 ja en train-clean 1 40

    cd ..

    Back to top

    For String-to-Tree and Tree-to-String SMT

    mkdir corpus.tree
    cd corpus.tree
    for file in train-1 dev test; do
      ln -s ../corpus.tok/${file}.ja
    done

  • Tokenizing sentences in English
  • for file in train-1 train-2 train-3 dev test; do
      cat ../corpus.org/${file}.en.txt | \
        perl ${SCRIPT_DIR}/z2h-utf8.pl | \
        perl ${MOSES_SCRIPT}/tokenizer/tokenizer.perl -l en -penn \
        > ${file}.tok.en
    done

  • Training truecaser for English
  • cat train-1.tok.en train-2.tok.en train-3.tok.en dev.tok.en > train_dev.tok.en
    ${MOSES_SCRIPT}/recaser/train-truecaser.perl --model truecase-model.en --corpus train_dev.tok.en

  • Truecasing English sentences
  • for file in train-1 train-2 train-3; do
      ${MOSES_SCRIPT}/recaser/truecase.perl --model truecase-model.en < ${file}.tok.en > ${file}.en
    done

  • Building training data for a language model
  • cat train-1.en train-2.en train-3.en | perl -pe 's/\-LRB\-/\(/g; s/\-RRB\-/\)/g;' > train-all.en

  • Parsing English sentences
  • ln -s train-1.ja train-1.tok.ja
    perl ${MOSES_SCRIPT}/training/clean-corpus-n.perl train-1.tok ja en train.reduced.tok 1 40

    for file in train.reduced dev test; do
      cat ${file}.tok.en | \
        perl ${MOSES_SCRIPT}/training/wrappers/parse-de-berkeley.perl \
        -binarize \
        -ja ${path}/BerkeleyParser-1.7/BerkeleyParser-1.7.jar \
        -gr ${path}/BerkeleyParser-1.7/eng_sm6.gr \
        > ${file}.tok.xml.en
    done

  • Cleaning training data for translation models
  • ln -s train.reduced.tok.ja train.reduced.tok.xml.ja
    perl ${MOSES_SCRIPT}/training/clean-corpus-n.perl \
      train.reduced.tok.xml ja en \
      train-clean.tok.xml 1 40 \
      --ignore-xml

  • Truecasing English sentences
  • for file in train-clean dev test; do
      ${MOSES_SCRIPT}/recaser/truecase.perl --model truecase-model.en < ${file}.tok.xml.en > ${file}.en
    done

    ln -s train-clean.tok.xml.ja train-clean.ja

    cd ..
    Back to top

    JST (Japan Science and Technology Agency)
    NICT (National Institute of Information and Communications Technology)
    Kyoto University
    Last Modified: 2014-07-07