Sangah Lee  Hansol Jang  Yunmee Baik  Suzi Park  Hyopil Shin

10.5626/JOK.2020.47.7.682

BERT language modeling embedding model Korean language modeling Vocabulary tokenizer

Recent models for the sentence embedding use huge corpus and parameters. They have massive data and large hardware and it incurs extensive time to pre-train. This tendency raises the need for a model with comparable performance while economically using training data. In this study, we proposed a Korean-specific model KR-BERT, using sub-character level to character-level Korean dictionaries and BidirectionalWordPiece Tokenizer. As a result, our KR-BERT model performs comparably and even better than other existing pre-trained models using one-tenth the size of training data from the existing models. It demonstrates that in a morphologically complex and resourceless language, using sub-character level and BidirectionalWordPiece Tokenizer captures language-specific linguistic phenomena that the Multilingual BERT model missed.