Edit model card

BERT base Japanese (IPA dictionary)

This is a BERT model pretrained on texts in the Japanese language.

This version of the model processes input texts with word-level tokenization based on the IPA dictionary, followed by the WordPiece subword tokenization.

The codes for the pretraining are available at cl-tohoku/bert-japanese.

Model architecture

The model architecture is the same as the original BERT base model; 12 layers, 768 dimensions of hidden states, and 12 attention heads.

Training Data

The model is trained on Japanese Wikipedia as of September 1, 2019. To generate the training corpus, WikiExtractor is used to extract plain texts from a dump file of Wikipedia articles. The text files used for the training are 2.6GB in size, consisting of approximately 17M sentences.

Tokenization

The texts are first tokenized by MeCab morphological parser with the IPA dictionary and then split into subwords by the WordPiece algorithm. The vocabulary size is 32000.

Training

The model is trained with the same configuration as the original BERT; 512 tokens per instance, 256 instances per batch, and 1M training steps.

Licenses

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

Acknowledgments

For training models, we used Cloud TPUs provided by TensorFlow Research Cloud program.

Downloads last month
2,491,734
Inference API
Examples
Mask token: [MASK]

Model tree for tohoku-nlp/bert-base-japanese

Finetunes
7 models

Dataset used to train tohoku-nlp/bert-base-japanese

Spaces using tohoku-nlp/bert-base-japanese 15