markuskhoa / mtet

MTet: Multi-domain Translation for English and Vietnamese

MTet: Multi-domain Translation for English-Vietnamese

Release

• v2.0 (Current)
• v1.0

Introduction

We are excited to introduce a new larger and better quality Machine Translation dataset, MTet, which stands for Multi-domain Translation for English and VieTnamese. In our new release, we extend our previous dataset (v1.0) by adding more high-quality English-Vietnamese sentence pairs on various domains. In addition, we also show our new larger Transformer models can achieve state-of-the-art results on multiple test sets.

Get data and model at Google Cloud Storage

Visit our blog post for more details.

Using the code

This code is build on top of vietai/dab:

To prepare for training, generate tfrecords from raw text files:

python t2t_datagen.py \
--data_dir=$path_to_folder_contains_vocab_file \
--tmp_dir=$path_to_folder_that_contains_training_data \
--problem=$problem

To train a Transformer model on the generated tfrecords

python t2t_trainer.py \
--data_dir=$path_to_folder_contains_vocab_file_and_tf_records \
--problem=$problem \
--hparams_set=$hparams_set \
--model=transformer \
--output_dir=$path_to_folder_to_save_checkpoints

To run inference on the trained model:

python t2t_decoder.py \
--data_dir=$path_to_folde_contains_vocab_file_and_tf_records \
--problem=$problem \
--hparams_set=$hparams_set \
--model=transformer \
--output_dir=$path_to_folder_contains_checkpoints \
--checkpoint_path=$path_to_checkpoint

In this colab, we demonstrated how to run these three phases in the context of hosting data/model on Google Cloud Storage.

Dataset

Our data contains roughly 4.2 million pairs of texts, ranging across multiple different domains such as medical publications, religious texts, engineering articles, literature, news, and poems. A more detail breakdown of our data is shown in the table below.

	v1	v2 (MTet)
Fictional Books	333,189	473,306
Legal Document	1,150,266	1,134,813
Medical Publication	5,861	13,410
Movies Subtitles	250,000	721,174
Software	79,912	79,132
TED Talk	352,652	303,131
Wikipedia	645,326	1,094,248
News	18,449	18,389
Religious texts	124,389	48,927
Educational content	397,008	213,284
No tag	5,517	63,896
Total	3,362,569	4,163,710

Data sources is described in more details here.

Acknowledgment

We would like to thank Google for the support of Cloud credits and TPU quota!