Pretrained Japanese BERT models

This is a repository of pretrained Japanese BERT models. The models are available in Transformers by Hugging Face.

Model hub: https://huggingface.co/cl-tohoku

For information on the previous versions of our pretrained models, see the v1.0 tag of this repository.

Model Architecture

The architecture of our models are the same as the original BERT models proposed by Google.

BERT-base models consist of 12 layers, 768 dimensions of hidden states, and 12 attention heads.
BERT-large models consist of 24 layers, 1024 dimensions of hidden states, and 16 attention heads.

Training Data

The models are trained on the Japanese version of Wikipedia. The training corpus is generated from the Wikipedia Cirrussearch dump file as of August 31, 2020.

The generated corpus files are 4.0GB in total, consisting of approximately 30M sentences. We used the MeCab morphological parser with mecab-ipadic-NEologd dictionary to split texts into sentences.

$WORK_DIR/corpus/jawiki-20200831/corpus_sampled.txt">

$ WORK_DIR="$HOME/work/bert-japanese"

$ python make_corpus_wiki.py \
--input_file jawiki-20200831-cirrussearch-content.json.gz \
--output_file $WORK_DIR/corpus/jawiki-20200831/corpus.txt \
--min_text_length 10 \
--max_text_length 200 \
--mecab_option "-r $HOME/local/etc/mecabrc -d $HOME/local/lib/mecab/dic/mecab-ipadic-neologd-v0.0.7"

# Split corpus files for parallel preprocessing of the files
$ python merge_split_corpora.py \
--input_files $WORK_DIR/corpus/jawiki-20200831/corpus.txt \
--output_dir $WORK_DIR/corpus/jawiki-20200831 \
--num_files 8

# Sample some lines for training tokenizers
$ cat $WORK_DIR/corpus/jawiki-20200831/corpus.txt|grep -v '^$'|shuf|head -n 1000000 \
> $WORK_DIR/corpus/jawiki-20200831/corpus_sampled.txt

Tokenization

For each of BERT-base and BERT-large, we provide two models with different tokenization methods.

For wordpiece models, the texts are first tokenized by MeCab with the Unidic 2.1.2 dictionary and then split into subwords by the WordPiece algorithm. The vocabulary size is 32768.
For character models, the texts are first tokenized by MeCab with the Unidic 2.1.2 dictionary and then split into characters. The vocabulary size is 6144.

We used fugashi and unidic-lite packages for the tokenization.

$WORK_DIR/tokenizers/jawiki-20200831/character/vocab.txt">

$ WORK_DIR="$HOME/work/bert-japanese"

# WordPiece (unidic_lite)
$ TOKENIZERS_PARALLELISM=false python train_tokenizer.py \
--input_files $WORK_DIR/corpus/jawiki-20200831/corpus_sampled.txt \
--output_dir $WORK_DIR/tokenizers/jawiki-20200831/wordpiece_unidic_lite \
--tokenizer_type wordpiece \
--mecab_dic_type unidic_lite \
--vocab_size 32768 \
--limit_alphabet 6129 \
--num_unused_tokens 10

# Character
$ head -n 6144 $WORK_DIR/tokenizers/jawiki-20200831/wordpiece_unidic_lite/vocab.txt \
> $WORK_DIR/tokenizers/jawiki-20200831/character/vocab.txt

Training

The models are trained with the same configuration as the original BERT; 512 tokens per instance, 256 instances per batch, and 1M training steps. For training of the MLM (masked language modeling) objective, we introduced whole word masking in which all of the subword tokens corresponding to a single word (tokenized by MeCab) are masked at once.

For training of each model, we used a v3-8 instance of Cloud TPUs provided by TensorFlow Research Cloud program. The training took about 5 days and 14 days for BERT-base and BERT-large models, respectively.

Creation of the pretraining data

$ WORK_DIR="$HOME/work/bert-japanese"

# WordPiece (unidic_lite)
$ mkdir -p $WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/pretraining_data
# It takes 3h and 420GB RAM, producing 43M instances
$ seq -f %02g 1 8|xargs -L 1 -I {} -P 8 python create_pretraining_data.py \
--input_file $WORK_DIR/corpus/jawiki-20200831/corpus_{}.txt \
--output_file $WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/pretraining_data/pretraining_data_{}.tfrecord.gz \
--vocab_file $WORK_DIR/tokenizers/jawiki-20200831/wordpiece_unidic_lite/vocab.txt \
--tokenizer_type wordpiece \
--mecab_dic_type unidic_lite \
--do_whole_word_mask \
--gzip_compress \
--max_seq_length 512 \
--max_predictions_per_seq 80 \
--dupe_factor 10

# Character
$ mkdir $WORK_DIR/bert/jawiki-20200831/character/pretraining_data
# It takes 4h10m and 615GB RAM, producing 55M instances
$ seq -f %02g 1 8|xargs -L 1 -I {} -P 8 python create_pretraining_data.py \
--input_file $WORK_DIR/corpus/jawiki-20200831/corpus_{}.txt \
--output_file $WORK_DIR/bert/jawiki-20200831/character/pretraining_data/pretraining_data_{}.tfrecord.gz \
--vocab_file $WORK_DIR/tokenizers/jawiki-20200831/character/vocab.txt \
--tokenizer_type character \
--mecab_dic_type unidic_lite \
--do_whole_word_mask \
--gzip_compress \
--max_seq_length 512 \
--max_predictions_per_seq 80 \
--dupe_factor 10

Training of the models

Note: all the necessary files need to be stored in a Google Cloud Storage (GCS) bucket.

# BERT-base, WordPiece (unidic_lite)
$ ctpu up -name tpu01 -tpu-size v3-8 -tf-version 2.3
$ cd /usr/share/models
$ sudo pip3 install -r official/requirements.txt
$ tmux
$ export PYTHONPATH="$PYTHONPATH:/usr/share/tpu/models"
$ WORK_DIR="gs://
   
    /bert-japanese
    "
   
$ python3 official/nlp/bert/run_pretraining.py \
--input_files="$WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/pretraining_data/pretraining_data_*.tfrecord" \
--model_dir="$WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/bert-base" \
--bert_config_file="$WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/bert-base/config.json" \
--max_seq_length=512 \
--max_predictions_per_seq=80 \
--train_batch_size=256 \
--learning_rate=1e-4 \
--num_train_epochs=100 \
--num_steps_per_epoch=10000 \
--optimizer_type=adamw \
--warmup_steps=10000 \
--distribution_strategy=tpu \
--tpu=tpu01

# BERT-base, Character
$ ctpu up -name tpu02 -tpu-size v3-8 -tf-version 2.3
$ cd /usr/share/models
$ sudo pip3 install -r official/requirements.txt
$ tmux
$ export PYTHONPATH="$PYTHONPATH:/usr/share/tpu/models"
$ WORK_DIR="gs://
   
    /bert-japanese
    "
   
$ python3 official/nlp/bert/run_pretraining.py \
--input_files="$WORK_DIR/bert/jawiki-20200831/character/pretraining_data/pretraining_data_*.tfrecord" \
--model_dir="$WORK_DIR/bert/jawiki-20200831/character/bert-base" \
--bert_config_file="$WORK_DIR/bert/jawiki-20200831/character/bert-base/config.json" \
--max_seq_length=512 \
--max_predictions_per_seq=80 \
--train_batch_size=256 \
--learning_rate=1e-4 \
--num_train_epochs=100 \
--num_steps_per_epoch=10000 \
--optimizer_type=adamw \
--warmup_steps=10000 \
--distribution_strategy=tpu \
--tpu=tpu02

# BERT-large, WordPiece (unidic_lite)
$ ctpu up -name tpu03 -tpu-size v3-8 -tf-version 2.3
$ cd /usr/share/models
$ sudo pip3 install -r official/requirements.txt
$ tmux
$ export PYTHONPATH="$PYTHONPATH:/usr/share/tpu/models"
$ WORK_DIR="gs://
   
    /bert-japanese
    "
   
$ python3 official/nlp/bert/run_pretraining.py \
--input_files="$WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/pretraining_data/pretraining_data_*.tfrecord" \
--model_dir="$WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/bert-large" \
--bert_config_file="$WORK_DIR/bert/jawiki-20200831/wordpiece_unidic_lite/bert-large/config.json" \
--max_seq_length=512 \
--max_predictions_per_seq=80 \
--train_batch_size=256 \
--learning_rate=5e-5 \
--num_train_epochs=100 \
--num_steps_per_epoch=10000 \
--optimizer_type=adamw \
--warmup_steps=10000 \
--distribution_strategy=tpu \
--tpu=tpu03

# BERT-large, Character
$ ctpu up -name tpu04 -tpu-size v3-8 -tf-version 2.3
$ cd /usr/share/models
$ sudo pip3 install -r official/requirements.txt
$ tmux
$ export PYTHONPATH="$PYTHONPATH:/usr/share/tpu/models"
$ WORK_DIR="gs://
   
    /bert-japanese
    "
   
$ python3 official/nlp/bert/run_pretraining.py \
--input_files="$WORK_DIR/bert/jawiki-20200831/character/pretraining_data/pretraining_data_*.tfrecord" \
--model_dir="$WORK_DIR/bert/jawiki-20200831/character/bert-large" \
--bert_config_file="$WORK_DIR/bert/jawiki-20200831/character/bert-large/config.json" \
--max_seq_length=512 \
--max_predictions_per_seq=80 \
--train_batch_size=256 \
--learning_rate=5e-5 \
--num_train_epochs=100 \
--num_steps_per_epoch=10000 \
--optimizer_type=adamw \
--warmup_steps=10000 \
--distribution_strategy=tpu \
--tpu=tpu04

Licenses

The pretrained models are distributed under the terms of the Creative Commons Attribution-ShareAlike 3.0.

The codes in this repository are distributed under the Apache License 2.0.

Related Work

Original BERT model by Google Research Team
- https://github.com/google-research/bert
- https://github.com/tensorflow/models/tree/master/official/nlp/bert (for TensorFlow 2.0)
Juman-tokenized Japanese BERT model
- Author: Kurohashi-Kawahara Laboratory, Kyoto University
- http://nlp.ist.i.kyoto-u.ac.jp/index.php?BERT日本語Pretrainedモデル
MeCab-Jumandic-tokenized Japanese BERT model (trained with a large mini-batch size)
- Author: National Institute of Information and Communications Technology (NICT)
- https://alaginrc.nict.go.jp/nict-bert/index.html
Sentencepiece Japanese BERT model
- Author: Yohei Kikuta
- https://github.com/yoheikikuta/bert-japanese
Sentencepiece Japanese BERT model, trained on SNS corpus
- Author: Hottolink, Inc.
- https://github.com/hottolink/hottoSNS-bert

Acknowledgments

The models are trained with Cloud TPUs provided by TensorFlow Research Cloud program.

Pretrained Japanese BERT models

Related tags

Overview

Pretrained Japanese BERT models

Model Architecture

Training Data

Tokenization

Training

Creation of the pretraining data

Training of the models

Licenses

Related Work

Acknowledgments

Owner

Inui Laboratory

Officile code repository for "A Game-Theoretic Perspective on Risk-Sensitive Reinforcement Learning"

This is a project built for FALLABOUT2021 event under SRMMIC, This project deals with NLP poetry generation.

💬 Open source machine learning framework to automate text- and voice-based conversations: NLU, dialogue management, connect to Slack, Facebook, and more - Create chatbots and voice assistants

This repository consists of a complete guide on natural language processing (NLP) in Python where we'll learn various techniques for implementing NLP including parsing & text processing and understand how to use NLP for text feature engineering.

Simplified diarization pipeline using some pretrained models - audio file to diarized segments in a few lines of code

An official repository for tutorials of Probabilistic Modelling and Reasoning (2021/2022) - a University of Edinburgh master's course.

code for modular summarization work published in ACL2021 by Krishna et al

Finetune gpt-2 in google colab

BiNE: Bipartite Network Embedding

Comprehensive-E2E-TTS - PyTorch Implementation

Twitter-Sentiment-Analysis - Analysis of twitter posts' positive and negative score.

A spaCy wrapper of OpenTapioca for named entity linking on Wikidata

Reproducing the Linear Multihead Attention introduced in Linformer paper (Linformer: Self-Attention with Linear Complexity)

To be a next-generation DL-based phenotype prediction from genome mutations.

An example project using OpenPrompt under pytorch-lightning for prompt-based SST2 sentiment analysis model

Topic Modelling for Humans

Auto-researching tool generating word documents.

Connectionist Temporal Classification (CTC) decoding algorithms: best path, beam search, lexicon search, prefix search, and token passing. Implemented in Python.

Simple virtual assistant using pyttsx3 and speech recognition optionally with pywhatkit and pther libraries.

A modular Karton Framework service that unpacks common packers like UPX and others using the Qiling Framework.