IWSLT 2022 Evaluation Campaign: Simultaneous Translation Baseline (Engilsh-to-Japanese Text-to-Text)

Table of Contents

• IWSLT 2022 Evaluation Campaign: Simultaneous Translation Baseline (Engilsh-to-Japanese Text-to-Text)
  • Table of Contents
  • Requirements
  • Setup
    • Clone a repository and install required packages
    • Setup fairseq for Wait-k (if needed)
    • (Not avaiable now) Setup fairseq for MMA-IL (if needed)
    • Setup SimulEval
  • Data preparation
  • Setting Environment Variables
  • Preprocessing
  • Wait-K model
    • Training
    • Test
  • (Not avaiable now) MMA-IL (Monotonic Multihead Attention with Infinite Lookback) model
    • Training
    • Test

---

Requirements

• Linux-based system
  • The scripts were tested with Ubuntu 18.04 LTS but would work on 20.04 LTS
• Bash
• Python >= 3.7.0
• (CUDA; not mandatory but highly recommended)
• PyTorch (the following command installs 1.10.1 working with CUDA 11.3)

  $ pip3 install torch==1.10.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Setup

Clone a repository and install required packages

$ git clone --recursive https://github.com/ksudoh/IWSLT2022_simul_t2t_baseline_enja.git
$ cd IWSLT2022_simul_t2t_baseline_enja
$ pip3 install -r requirements.txt

Setup fairseq for Wait-k (if needed)

$ pushd fairseq-wait-k/fairseq
$ python3 setup.py build_ext --inplace
$ popd

(Not avaiable now) Setup fairseq for MMA-IL (if needed)

$ pushd fairseq-mma-il/fairseq
$ python3 setup.py build_ext --inplace
$ popd

Setup SimulEval

$ pushd SimulEval
$ python3 setup.py install --prefix ./
$ popd

Data preparation

• Download MuST-C v2.0 and extract the package
  • Suppose you put the extracted directory en-ja in /path/to/MuST-C/.
• (If needed) Download WMT21 En-Ja data and extract the packages
  • Suppose you put the extracted files/directories in /path/to/WMT-train/.
    • JParaCrawl
    • News Commentary v16
    • WikiTitles v3
    • WikiMatrix v1
    • JESC
    • KFTT
  • Data preparation helper modules for the datasets above are included in this repository

Setting Environment Variables

• The baseline system scripts use the following environment variables.
  • WORKDIR specifies the directory to be used to store the data and models.
  • You may change the setting of TMPDIR if you would like to use the temporary space other than /tmp. The scripts

$ export SRC=en
$ export TRG=ja
$ export MUSTC_ROOT=/path/to/MuST-C
$ export WMT_DATA_ROOT=/path/to/WMT-train
$ export WORKDIR=/path/to/work

Preprocessing

• The wrapper script 10-preprocess.sh conducts the following preprocessings:
  • Extract bilingual sentences from the datasets
  • Train a unigram subword model using SentencePiece (shared across languages)
  • Tokenize the bilingual sentences
  • Binarize the training and development datasets using fairseq-preprocess

$ bash ./10-preprocess.sh

Wait-K model

• Set variables K and CUDA_VISIBLE_DEVICES.
• You may use multiple GPUs, but the batch size becomes larger accordingly.

Training

$ env K=20 CUDA_VISIBLE_DEVICES=0 bash ./20-train-wait-k.sh

Test

$ env K=20 CUDA_VISIBLE_DEVICES=0 bash ./21-test-wait-k.sh

(Not avaiable now) MMA-IL (Monotonic Multihead Attention with Infinite Lookback) model

• Set variable CUDA_VISIBLE_DEVICES.
• You may use multiple GPUs, but the batch size becomes larger accordingly.

Training

$ env CUDA_VISIBLE_DEVICES=0 bash ./30-train-mma-il.sh

Test

$ env CUDA_VISIBLE_DEVICES=0 bash ./31-test-mma-il.sh