BERT with SentencePiece for Japanese text.

This is a repository of Japanese BERT model with SentencePiece tokenizer.

To clone this repository together with the required BERT and WikiExtractor:

git clone --recurse-submodules https://github.com/yoheikikuta/bert-japanese

Pretrained models

We provide pretrained BERT model and trained SentencePiece model for Japanese text. Training data is the Japanese wikipedia corpus from Wikimedia Downloads.
Please download all objects in the following google drive to model/ directory.

• Pretrained BERT model and trained SentencePiece model

Loss function during training is as below (after 1M steps the loss function massively changes because max_seq_length is changed from 128 to 512.):

***** Eval results *****
  global_step = 1400000
  loss = 1.3773012
  masked_lm_accuracy = 0.6810424
  masked_lm_loss = 1.4216621
  next_sentence_accuracy = 0.985
  next_sentence_loss = 0.059553143

Finetuning with BERT Japanese

We also provide a simple Japanese text classification problem with livedoor ニュースコーパス.
Try the following notebook to check the usability of finetuning.
You can run the notebook on CPU (too slow) or GPU/TPU environments.

• finetune-to-livedoor-corpus.ipynb

The results are the following:

• BERT with SentencePiece

                  precision    recall  f1-score   support

  dokujo-tsushin       0.98      0.94      0.96       178
    it-life-hack       0.96      0.97      0.96       172
   kaden-channel       0.99      0.98      0.99       176
  livedoor-homme       0.98      0.88      0.93        95
     movie-enter       0.96      0.99      0.98       158
          peachy       0.94      0.98      0.96       174
            smax       0.98      0.99      0.99       167
    sports-watch       0.98      1.00      0.99       190
      topic-news       0.99      0.98      0.98       163

       micro avg       0.97      0.97      0.97      1473
       macro avg       0.97      0.97      0.97      1473
    weighted avg       0.97      0.97      0.97      1473

• sklearn GradientBoostingClassifier with MeCab

                    precision    recall  f1-score   support

  dokujo-tsushin       0.89      0.86      0.88       178
    it-life-hack       0.91      0.90      0.91       172
   kaden-channel       0.90      0.94      0.92       176
  livedoor-homme       0.79      0.74      0.76        95
     movie-enter       0.93      0.96      0.95       158
          peachy       0.87      0.92      0.89       174
            smax       0.99      1.00      1.00       167
    sports-watch       0.93      0.98      0.96       190
      topic-news       0.96      0.86      0.91       163

       micro avg       0.92      0.92      0.92      1473
       macro avg       0.91      0.91      0.91      1473
    weighted avg       0.92      0.92      0.91      1473

Cautions when using the model as a sentence generation model

The model expects lowercase input and the tokenizer is assumed to be used with do_lower_case=True option, but the special tokens such as [CLS] are registered in uppercase characters.
Therefore, when we put "[CLS] I am ..." it into the tokenizer as a raw string, the tokenizer first makes it lowercase ("[cls] i am ...") and then cannot interpret "[cls]" as the special token, which causes problems.
If you wanna use the model as a sentence generation model, follow from these steps (sorry, it's a little bit confusing):

• keep special tokens (such as [CLS] or [SEP]) uppercase
• make original input sentences lowercase manually (e.g., "i am ...")
• join them together (e.g., "[CLS] i am ...") and put it into the tokenizer with do_lower_case=False option
• put the obtained tokens into the model

Pretraining from scratch

All scripts for pretraining from scratch are provided. Follow the instructions below.

Environment set up

Build a docker image with Dockerfile and create a docker container.

docker build -t bert-ja .
docker run -it --rm -v $PWD:/work -p 8888:8888 bert-ja

Data preparation

Data downloading and preprocessing. It takes about a few hours on GCP n1-standard-16 (16CPUs, 60GB memories) instance.

python3 src/data-download-and-extract.py
bash src/file-preprocessing.sh

The above scripts use the latest jawiki data and wikiextractor module, which are different from those used for the pretrained model. If you wanna prepare the same situation, use the following information:

• bert-japanese: commit 074fe20f33a020769091e1e5552b33867ccbd750
• dataset: jawiki-20181220-pages-articles-multistream.xml.bz2 in the Google Drive
• wikiextractor: commit 1e4236de4237d0a89d0ad7241505d73ee7e23517

Training SentencePiece model

Train a SentencePiece model using the preprocessed data. It takes about two hours on the instance.

python3 src/train-sentencepiece.py

Creating data for pretraining

Create .tfrecord files for pretraining. For longer sentence data, replace the value of max_seq_length with 512.

for DIR in $( find /work/data/wiki/ -mindepth 1 -type d ); do 
  python3 src/create_pretraining_data.py \
    --input_file=${DIR}/all.txt \
    --output_file=${DIR}/all-maxseq128.tfrecord \
    --model_file=./model/wiki-ja.model \
    --vocab_file=./model/wiki-ja.vocab \
    --do_lower_case=True \
    --max_seq_length=128 \
    --max_predictions_per_seq=20 \
    --masked_lm_prob=0.15 \
    --random_seed=12345 \
    --dupe_factor=5 \
    --do_whole_word_mask=False
done

Pretraining

You need GPU/TPU environment to pretrain a BERT model.
The following notebook provides the link to Colab notebook where you can run the scripts with TPUs.

• pretraining.ipynb

How to cite this work in papers

We didn't publish any paper about this work.
Please cite this repository in publications as the following:

@misc{bertjapanese,
  author = {Yohei Kikuta},
  title = {BERT Pretrained model Trained On Japanese Wikipedia Articles},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/yoheikikuta/bert-japanese}},
}