Codebase and data are uploaded in progress.

VOLT(-py) is a vocabulary learning codebase that allows researchers and developers to automaticaly generate a vocabulary with suitable granularity for machine translation.
To help more readers understand our work better, I write a blog at this repo.

What's New:

• July 2021: Support vocabulary learning for classification.
• July 2021: Support En-De translation, TED bilingual translation, and multilingual translation.
• July 2021: Support subword-nmt tokenization.
• July 2021: Support sentencepiece tokenization.

What's On-going:

• Support pip usage.

Features:

• Efficient: CPU learning on one machine.
• Easy-to-use: Support widely-used tokenization toolkits, subword-nmt and sentencepiece.

Requirements and Installation

The required environments:

• python 3
• tqdm
• mosedecoder
• subword-nmt
• POT (local POT)

To use VOLT and develop locally:

git clone https://github.com/Jingjing-NLP/VOLT/
cd VOLT
git clone https://github.com/moses-smt/mosesdecoder.git
git clone https://github.com/rsennrich/subword-nmt.git
pip3 install sentencepiece
pip3 install tqdm 
cd POT
pip3 install --editable ./ -i https://pypi.doubanio.com/simple --user
cd ../

Usage

• The first step is to get vocabulary candidates based on tokenized texts. Notice: the tokenized texts should be in charater level. Please do not use segmentation tools to segment your texts. The sub-word vocabulary can be generated by subword-nmt and sentencepiece. Here are two examples.

  • This example shows how to learn a vocabulary for seq2seq tasks ( including source data and target data).

    
    #Assume source_file is the file stroing texts in the source data
    #Assume target_file is the file stroing texts in the target data
    size=30000 # the size of BPE
    cat source_file > training_data
    cat target_file >> training_data 

  subword-nmt style:

  mkdir bpeoutput BPE_CODE=bpeoutput/code # the path to save vocabulary python3 subword-nmt/learn_bpe.py -s $size < training_data > $BPE_CODE python3 subword-nmt/apply_bpe.py -c $BPE_CODE < source_file > bpeoutput/source.file python3 subword-nmt/apply_bpe.py -c $BPE_CODE < target_file > bpeoutput/target.file

  sentencepiece style:

  cd examples mkdir spmout python3 spm/spm_train.py --input=training_data --model_prefix=spm --vocab_size=$size --character_coverage=1.0 --model_type=bpe

  After this step, you will see spm.vocab and spm.model.

  Change spm.vocab to a file where each line is splited via a single space like example "abc 100"

  sed -i 's/\t/ /g' spm.vocab python3 spm/spm_encoder.py --model spm.model --inputs source_file --outputs spmout/source.file --output_format piece python3 spm/spm_encoder.py --model spm.model --inputs target_file --outputs spmout/target.file --output_format piece

  * This example shows how to get a vocabulary from a single file for non-seq2seq tasks.

  Assume source_file is the file stroing your data

  size=30000 # the size of BPE

  subword-nmt style:

  mkdir bpeoutput BPE_CODE=bpeoutput/code # the path to save vocabulary python3 subword-nmt/learn_bpe.py -s $size < source_file > $BPE_CODE python3 subword-nmt/apply_bpe.py -c $BPE_CODE < source_file > bpeoutput/source.file

  sentencepiece style:

  cd examples mkdir spmout python3 spm/spm_train.py --input=source_file --model_prefix=spm --vocab_size=$size --character_coverage=1.0 --model_type=bpe

  After this step, you will see spm.vocab and spm.model

  python3 spm/spm_encoder.py --model spm.model --inputs source_file --outputs spmout/source.file --output_format piece

• The second step is to run VOLT scripts. It accepts the following parameters:

  • --source_file: the file storing source data for seq2seq tasks or the file string all raw texts for non-seq2seq tasks.
  • --token_candidate_file: the file storing token candidates. Each line is splited via a single space like example "abc 100"
  • --tokenizer: which toolkit you use to get token candidates. Only two choices are supported: subword-nmt and sentencepiece.
  • --size_file: the file to store the vocabulary size recommended by VOLT.
  • --target_file: (optional) the file storing target data for seq2seq tasks. None by default.
  • --max_number: (optional) the maximum size of the vocabulary generated by VOLT. 10,000 by default.
  • --interval: (optional) the search granularity in VOLT. 1,000 by default.
  • --loop_in_ot: (optional) the maximum interation loop in the Sinkhorn solution. 500 by default.
  • --threshold: (optional) the threshold to decide which tokens are added into the final vocabulary from the optimal matrix. Small threshold means that the final vocabulary is more like BPE-style vocabulary. 1e-5 by default.

    
    #For seq2seq tasks with source file and target file, you can use the following commands:
    #subword-nmt style
    python3 ../ot_run.py --source_file bpeoutput/source.file --target_file bpeoutput/target.file \
        --token_candidate_file $BPE_CODE \
        --vocab_file bpeoutput/vocab --max_number 10000 --interval 1000  --loop_in_ot 500 --tokenizer subword-nmt --size_file bpeoutput/size 
    #sentencepiece style
    python3 ../ot_run.py --source_file spmoutput/source.file --target_file spmoutput/target.file \
        --token_candidate_file spm.vocab \
        --vocab_file spmoutput/vocab --max_number 10000 --interval 1000  --loop_in_ot 500 --tokenizer sentencepiece --size_file spmoutput/size 

  For non-seq2seq tasks with one source file, you can use the following commands:

  subword-nmt style

  python3 ../ot_run.py --source_file bpeoutput/source.file \ --token_candidate_file $BPE_CODE \ --vocab_file bpeoutput/vocab --max_number 10000 --interval 1000 --loop_in_ot 500 --tokenizer subword-nmt --size_file bpeoutput/size

  sentencepiece style

  BPE_CODE=spm.vocab python3 ../ot_run.py --source_file spmoutput/source.file \ --token_candidate_file spm.vocab \ --vocab_file spmoutput/vocab --max_number 10000 --interval 1000 --loop_in_ot 500 --tokenizer sentencepiece --size_file spmoutput/size

• The third step is to use the generated vocabulary to segment your texts:

  #subword-nmt style
  echo "#version: 0.2" > bpeoutput/vocab.seg # add version info
  echo bpeoutput/vocab >> bpeoutput/vocab.seg
  python3 $BPEROOT/apply_bpe.py -c bpeoutput/vocab.seg < source_file > bpeoutput/source.file
  python3 $BPEROOT/apply_bpe.py -c bpeoutput/vocab.seg < target_file > bpeoutput/source.file #optional if your task does not contain target texts
  
  #sentencepiece style
  #for sentencepiece toolkit, here we only keep the optimal size
  best_size=$(cat spmoutput/size)
  #training_data contains source data and target data (optional if target data is provided)
  python3 spm/spm_train.py --input=training_data --model_prefix=spm --vocab_size=$best_size --character_coverage=1.0 --model_type=bpe
  python3 spm/spm_encoder.py --model spm.model --inputs source_file --outputs spmout/source.file --output_format piece
  python3 spm/spm_encoder.py --model spm.model --inputs target_file --outputs spmout/target.file --output_format piece #optional if your task does not contain target texts

• The last step is to use the segmented texts for downstream tasks. You can use the repo Fairseq for training and evaluation. We also upload the training and evaluation code in path "examples/". Notice: For a comparison of BLEU, you need to do "remove-bpe" operations for the generated texts.

Examples

We have given several examples in path "examples/", including En-De translation, En-Fr translation, multilingual translation, and En-De translation without joint vocabularies.

• En-De translation: run_ende.sh
• En-De translation without joint vocabularies: run_ende_withoutjoint.sh
• En-Fr translation: run_enfr.sh
• TED bilingual translation: run_ted_bilingual.sh
• TED bilingual translation with sentencepiece: run_ted_bilingual_senencepiece.sh
• TED many-to-one translation: run_ted_multilingual.sh

Datasets

The WMT-14 En-de translation data can be downloaed via the running scripts.

For TED X-EN data, you can download at X-EN. For TED EN-X data, you can download at EN-X

Citation

Please cite as:

@inproceedings{volt,
  title = {Vocabulary Learning via Optimal Transport for Neural Machine Translation},
  author= {Jingjing Xu and
               Hao Zhou and
               Chun Gan and
               Zaixiang Zheng and
               Lei Li},
  booktitle = {Proceedings of ACL 2021},
  year = {2021},
}