WAT 2016

The 3rd Workshop on Asian Translation

Baseline Systems

Hierarchical Phrase-based SMT

for all subtasks

Setup

LANG_F=ja
LANG_E=en
CORPUS_LM=../corpus.tok/train
(For JE and EJ of ASPEC, "CORPUS_LM=../corpus.tok/train-all")
CORPUS=../corpus.tok/train-clean
DEV_F=../corpus.tok/dev.${LANG_F}
DEV_E=../corpus.tok/dev.${LANG_E}
TEST=../corpus.tok/test.${LANG_F}
REF=../corpus.tok/test.${LANG_E}
LM_ORDER=5
JOBS=16

MOSES_SCRIPT=${path}/mosesdecoder-RELEASE-2.1.1/scripts
MOSES_BIN_DIR=${path}/mosesdecoder-RELEASE-2.1.1/bin
EXT_BIN_DIR=${path}/giza-pp/bin

WORK_DIR=work.${LANG_F}-${LANG_E}
TRAINING_DIR=${WORK_DIR}/training
MODEL_DIR=${WORK_DIR}/training/model



mkdir hierModel
cd hierModel/
mkdir -p ${TRAINING_DIR}/lm


Training Language Model

LM_FILE=`pwd`/${TRAINING_DIR}/lm/lm.${LANG_E}.arpa.gz

${MOSES_BIN_DIR}/lmplz --order ${LM_ORDER} -S 80% -T /tmp < ${CORPUS_LM}.${LANG_E} | gzip > ${LM_FILE}


Training Translation Model

${MOSES_SCRIPT}/training/train-model.perl \
  --root-dir `pwd`/${TRAINING_DIR} \
  --model-dir `pwd`/${MODEL_DIR} \
  --corpus ${CORPUS} \
  --f ${LANG_F} \
  --e ${LANG_E} \
  --parallel \
  --alignment grow-diag-final-and \
  --score-options "--GoodTuring" \
  --hierarchical \
  --glue-grammar \
  --lm 0:${LM_ORDER}:${LM_FILE}:8 \
  --cores ${JOBS} \
  --sort-buffer-size 10G \
  --parallel \
  >& ${TRAINING_DIR}/training_TM.log


Tuning

${MOSES_SCRIPT}/training/filter-model-given-input.pl \
  ${MODEL_DIR}.filtered/dev \
  ${MODEL_DIR}/moses.ini \
  ${DEV_F} \
  --Hierarchical

mkdir -p ${WORK_DIR}/tuning

${MOSES_SCRIPT}/training/mert-moses.pl \
  ${DEV_F} \
  ${DEV_E} \
  ${MOSES_BIN_DIR}/moses_chart \
  `pwd`/${MODEL_DIR}.filtered/dev/moses.ini \
  --mertdir ${MOSES_BIN_DIR} \
  --working-dir `pwd`/${WORK_DIR}/tuning/mert \
  --threads ${JOBS} \
  --no-filter-phrase-table \
  --decoder-flags "-threads ${JOBS} -max-chart-span 1000" \
  --predictable-seeds \
  >& ${WORK_DIR}/tuning/mert.log



Insert weights into the configuration file.

perl ${MOSES_SCRIPT}/ems/support/substitute-weights.perl \
  ${MODEL_DIR}/moses.ini \
  ${WORK_DIR}/tuning/mert/moses.ini \
  ${MODEL_DIR}/moses-tuned.ini


Translating

OUTPUT_DIR=${WORK_DIR}/output
mkdir ${OUTPUT_DIR}

${MOSES_SCRIPT}/training/filter-model-given-input.pl \
  ${MODEL_DIR}.filtered/test \
  ${MODEL_DIR}/moses-tuned.ini \
  ${TEST} \
  --Hierarchical

outfile=${OUTPUT_DIR}/test.out

${MOSES_BIN_DIR}/moses_chart -config ${MODEL_DIR}.filtered/test/moses.ini -max-chart-span 1000 -threads ${JOBS} < ${TEST} > ${outfile} 2> ${outfile}.log



Recase the output

For English and Indonesian

${MOSES_SCRIPT}/recaser/detruecase.perl < ${outfile} > ${outfile}.tok


Detokenize the output

For English and Indonesian

${MOSES_SCRIPT}/tokenizer/detokenizer.perl -l en < ${outfile}.tok > ${outfile}.detok


For Japanese and Chinese

cat ${outfile} | \
  perl -Mencoding=utf8 -pe 's/([^Ａ-Ｚａ-ｚA-Za-z]) +/${1}/g; s/ +([^Ａ-Ｚａ-ｚA-Za-z])/${1}/g; ' \
  > ${outfile}.detok

JST (Japan Science and Technology Agency)
NICT (National Institute of Information and Communications Technology)
Kyoto University
Last Modified: 2016-06-11