Additional pretrained BERT base Japanese finance

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

The codes for the pretraining are available at retarfi/language-pretraining.

Model architecture

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

Training Data

The models are additionally trained on financial corpus from Tohoku University's BERT base Japanese model (cl-tohoku/bert-base-japanese).

The financial corpus consists of 2 corpora:

Summaries of financial results from October 9, 2012, to December 31, 2020
Securities reports from February 8, 2018, to December 31, 2020

The financial corpus file consists of approximately 27M sentences.

Tokenization

You can use tokenizer Tohoku University's BERT base Japanese model (cl-tohoku/bert-base-japanese).

You can use the tokenizer:

tokenizer = transformers.BertJapaneseTokenizer.from_pretrained('cl-tohoku/bert-base-japanese')

Training

The models are trained with the same configuration as BERT base in the original BERT paper; 512 tokens per instance, 256 instances per batch, and 1M training steps.

Citation

@article{Suzuki-etal-2023-ipm,
  title = {Constructing and analyzing domain-specific language model for financial text mining}
  author = {Masahiro Suzuki and Hiroki Sakaji and Masanori Hirano and Kiyoshi Izumi},
  journal = {Information Processing & Management},
  volume = {60},
  number = {2},
  pages = {103194},
  year = {2023},
  doi = {10.1016/j.ipm.2022.103194}
}

Licenses

The pretrained models are distributed under the terms of the Creative Commons Attribution-ShareAlike 4.0.

Acknowledgments

This work was supported by JSPS KAKENHI Grant Number JP21K12010 and JST-Mirai Program Grant Number JPMJMI20B1.