KoRean based Bert pre-trained (KR-BERT)

This is a release of Korean-specific, small-scale BERT models with comparable or better performances developed by Computational Linguistics Lab at Seoul National University, referenced in KR-BERT: A Small-Scale Korean-Specific Language Model.

Vocab, Parameters and Data

	Mulitlingual BERT (Google)	KorBERT (ETRI)	KoBERT (SKT)	KR-BERT character	KR-BERT sub-character
vocab size	119,547	30,797	8,002	16,424	12,367
parameter size	167,356,416	109,973,391	92,186,880	99,265,066	96,145,233
data size	- (The Wikipedia data for 104 languages)	23GB 4.7B morphemes	- (25M sentences, 233M words)	2.47GB 20M sentences, 233M words	2.47GB 20M sentences, 233M words

Model	Masked LM Accuracy
KoBERT	0.750
KR-BERT character BidirectionalWordPiece	0.779
KR-BERT sub-character BidirectionalWordPiece	0.769

Sub-character

Korean text is basically represented with Hangul syllable characters, which can be decomposed into sub-characters, or graphemes. To accommodate such characteristics, we trained a new vocabulary and BERT model on two different representations of a corpus: syllable characters and sub-characters.

In case of using our sub-character model, you should preprocess your data with the code below.

import torch
from transformers import BertConfig, BertModel, BertForPreTraining, BertTokenizer
from unicodedata import normalize

tokenizer_krbert = BertTokenizer.from_pretrained('/path/to/vocab_file.txt', do_lower_case=False)

# convert a string into sub-char
def to_subchar(string):
    return normalize('NFKD', string)

sentence = '토크나이저 예시입니다.'
print(tokenizer_krbert.tokenize(to_subchar(sentence)))

Tokenization

BidirectionalWordPiece Tokenizer

We use the BidirectionalWordPiece model to reduce search costs while maintaining the possibility of choice. This model applies BPE in both forward and backward directions to obtain two candidates and chooses the one that has a higher frequency.

	Mulitlingual BERT	KorBERT character	KoBERT	KR-BERT character WordPiece	KR-BERT character BidirectionalWordPiece	KR-BERT sub-character WordPiece	KR-BERT sub-character BidirectionalWordPiece
냉장고 nayngcangko "refrigerator"	냉#장#고 nayng#cang#ko	냉#장#고 nayng#cang#ko	냉#장#고 nayng#cang#ko	냉장고 nayngcangko	냉장고 nayngcangko	냉장고 nayngcangko	냉장고 nayngcangko
춥다 chwupta "cold"	[UNK]	춥#다 chwup#ta	춥#다 chwup#ta	춥#다 chwup#ta	춥#다 chwup#ta	추#ㅂ다 chwu#pta	추#ㅂ다 chwu#pta
뱃사람 paytsalam "seaman"	[UNK]	뱃#사람 payt#salam	뱃#사람 payt#salam	뱃#사람 payt#salam	뱃#사람 payt#salam	배#ㅅ#사람 pay#t#salam	배#ㅅ#사람 pay#t#salam
마이크 maikhu "microphone"	마#이#크 ma#i#khu	마이#크 mai#khu	마#이#크 ma#i#khu	마이크 maikhu	마이크 maikhu	마이크 maikhu	마이크 maikhu

Models

	TensorFlow		PyTorch
	character	sub-character	character	sub-character
WordPiece tokenizer	WP char	WP subchar	WP char	WP subchar
Bidirectional WordPiece tokenizer	BiWP char	BiWP subchar	BiWP char	BiWP subchar

Requirements

• transformers == 2.1.1
• tensorflow < 2.0

Downstream tasks

Naver Sentiment Movie Corpus (NSMC)

• If you want to use the sub-character version of our models, let the subchar argument be True.

• And you can use the original BERT WordPiece tokenizer by entering bert for the tokenizer argument, and if you use ranked you can use our BidirectionalWordPiece tokenizer.

• tensorflow: After downloading our pretrained models, put them in a models directory in the krbert_tensorflow directory.

• pytorch: After downloading our pretrained models, put them in a pretrained directory in the krbert_pytorch directory.

# pytorch
python3 train.py --subchar {True, False} --tokenizer {bert, ranked}

# tensorflow
python3 run_classifier.py \
  --task_name=NSMC \
  --subchar={True, False} \
  --tokenizer={bert, ranked} \
  --do_train=true \
  --do_eval=true \
  --do_predict=true \
  --do_lower_case=False\
  --max_seq_length=128 \
  --train_batch_size=128 \
  --learning_rate=5e-05 \
  --num_train_epochs=5.0 \
  --output_dir={output_dir}

The pytorch code structure refers to that of https://github.com/aisolab/nlp_implementation .

NSMC Acc.

	multilingual BERT	KorBERT	KoBERT	KR-BERT character WordPiece	KR-BERT character Bidirectional WordPiece	KR-BERT sub-character WordPiece	KR-BERT sub-character Bidirectional WordPiece
pytorch	-	89.84	89.01	89.34	89.38	89.20	89.34
tensorflow	87.08	85.94	n/a	89.86	90.10	89.76	89.86

Citation

If you use these models, please cite the following paper:

@article{lee2020krbert,
    title={KR-BERT: A Small-Scale Korean-Specific Language Model},
    author={Sangah Lee and Hansol Jang and Yunmee Baik and Suzi Park and Hyopil Shin},
    year={2020},
    journal={ArXiv},
    volume={abs/2008.03979}
  }

Contacts

[email protected]