Japanese - Vietnamese Character-based NMT systems

This repository contains all data and recipes for building the systems and replicating the results of our paper How Transformer Revitalizes Character-based Neural Machine Translation: An Investigation on Japanese-Vietnamese Translation Systems, which achieved SOTA results in term of BLEU/RIBES for the Japanese$\leftrightarrow$Vietnamese translation directions.

Dataset

The data is collected from Globse, TED talks, Asian Language Treebank and Tatoeba corpus. More detailed information about the data, please go to https://github.com/ngovinhtn/JaViCorpus

For all experiments, the corpus was split into training, development (dev2010) and test (tst2010) sets:

Dataset	Sentences
Training Data	339523
Development	558
Test	1,215

Recipes

We compared word-based (sub-word-based) and character-based NMT systems using two architectures: Recurrent-based and Transformer. In order to run our experiments with the NMT frameworks you need to install the libraries that each of the framework requires:

1. Recurrent: OpenNMT-py(https://github.com/OpenNMT/OpenNMT-py)
2. Transformer: NMTGMinor (https://github.com/quanpn90/NMTGMinor)

We provide the ready-to-train data, but in case if you want to run the preprocessing steps for Japanese and Vietnamese texts, you need the following tools:

1. KyTea - for Japanese tokenization and word segmentation
2. Pyvi - for Vietnamese word segmentation
3. Moses scripts - for Vietnamese tokenization

The steps:

1. prepare_data_***.sh - to extract vocabularies and convert the preprocessed texts into pytorch tensors prior to the training process
2. train***.sh - train the model
3. translate_***.sh - use the trained model to translate new texts.

You need to modify those scripts a little bit to point to the paths of those toolkits installed/downloaded in your computers/GPU clusters.

Performance

1. Japanese to Vietnamese

Models	BLEU	RIRES
Word2WordRecurrent	11.05	0.663
Word2WordTransformer	11.72	0.681
Char2CharRecurrent	10.06	0.657
Char2CharTransformer	13.34	0.688

2. Vietnamese to Japanese

Models	BLEU	RIRES
Word2WordRecurrent	11.13	0.593
Word2WordTransformer	13.07	0.679
Char2CharRecurrent	9.61	0.566
Char2CharTransformer	15.05	0.691

Citation

If you use the recipes, please cite our paper: Thi-Vinh Ngo, Thanh-Le Ha, Phuong-Thai Nguyen, Le-Minh Nguyen: How Transformer Revitalizes Character-based Neural Machine Translation: An Investigation on Japanese-Vietnamese Translation Systems. Proceedings of the 16th International Workshop on Spoken Language Translation 2019 (IWSLT 2019), Hongkong.

Bibtex:

@inproceedings{Ngo2019How,
  author    = {Thi{-}Vinh Ngo and
               Thanh{-}Le Ha and
               Phuong{-}Thai Nguyen and
               Le{-}Minh Nguyen},
  title     = {{How Transformer Revitalizes Character-based Neural Machine Translation: An Investigation on Japanese-Vietnamese Translation Systems}},
  booktitle = {{Proceedings of the 16th International Workshop on Spoken Language Translation 2019 (IWSLT 2019))}},
  year      = {2019},
  address   = {{Hongkong}},
  url       = {https://arxiv.org/pdf/1910.02238.pdf}
}