The model architecture is the same as ELECTRA small in the original ELECTRA implementation; 12 layers, 256 dimensions of hidden states, and 4 attention heads.
Training Data
The models are trained on the Japanese version of Wikipedia.
The training corpus is generated from the Japanese version of Wikipedia, using Wikipedia dump file as of June 1, 2021.
The corpus file is 2.9GB, consisting of approximately 20M sentences.
Tokenization
The texts are first tokenized by MeCab with IPA dictionary and then split into subwords by the WordPiece algorithm.
The vocabulary size is 32768.
Training
The models are trained with the same configuration as ELECTRA small in the original ELECTRA paper except size; 128 tokens per instance, 128 instances per batch, and 1M training steps.
The size of the generator is the same of the discriminator.
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}
}
