TUNiB-Electra

We release several new versions of the ELECTRA model, which we name TUNiB-Electra. There are two motivations. First, all the existing pre-trained Korean encoder models are monolingual, that is, they have knowledge about Korean only. Our bilingual models are based on the balanced corpora of Korean and English. Second, we want new off-the-shelf models trained on much more texts. To this end, we collected a large amount of Korean text from various sources such as blog posts, comments, news, web novels, etc., which sum up to 100 GB in total.

You can use TUNiB-Electra with the Hugging Face transformers library.

What's New:

• Sep 19, 2021 Released a tech blog
• Sep 17, 2021 Released TUNiB-Electra.

How to use

You can use this model directly with transformers library:

from transformers import AutoModel, AutoTokenizer

# Small Model (Korean-English bilingual model)
tokenizer = AutoTokenizer.from_pretrained('tunib/electra-ko-en-small')
model = AutoModel.from_pretrained('tunib/electra-ko-en-small')

# Base Model (Korean-English bilingual model)
tokenizer = AutoTokenizer.from_pretrained('tunib/electra-ko-en-base')
model = AutoModel.from_pretrained('tunib/electra-ko-en-base')

# Small Model (Korean-only model)
tokenizer = AutoTokenizer.from_pretrained('tunib/electra-ko-small')
model = AutoModel.from_pretrained('tunib/electra-ko-small')

# Base Model (Korean-only model)
tokenizer = AutoTokenizer.from_pretrained('tunib/electra-ko-base')
model = AutoModel.from_pretrained('tunib/electra-ko-base')

Tokenizer example

>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained('tunib/electra-ko-en-base')
>>> tokenizer.tokenize("tunib is a natural language processing tech startup.")
['tun', '##ib', 'is', 'a', 'natural', 'language', 'processing', 'tech', 'startup', '.']
>>> tokenizer.tokenize("튜닙은 자연어처리 테크 스타트업입니다.")
['튜', '##닙', '##은', '자연', '##어', '##처리', '테크', '스타트업', '##입니다', '.']

Results on Korean downstream tasks

Small Models

	# Params	Avg.	NSMC (acc)	Naver NER (F1)	PAWS (acc)	KorNLI (acc)	KorSTS (spearman)	Question Pair (acc)	KorQuaD (Dev) (EM/F1)	Korean-Hate-Speech (Dev) (F1)
TUNiB-Electra-ko-small	14M	81.29	89.56	84.98	72.85	77.08	78.76	94.98	61.17 / 87.64	64.50
TUNiB-Electra-ko-en-small	18M	81.44	89.28	85.15	75.75	77.06	77.61	93.79	80.55 / 89.77	63.13
KoELECTRA-small-v3	14M	82.58	89.36	85.40	77.45	78.60	80.79	94.85	82.11 / 91.13	63.07

Base Models

	# Params	Avg.	NSMC (acc)	Naver NER (F1)	PAWS (acc)	KorNLI (acc)	KorSTS (spearman)	Question Pair (acc)	KorQuaD (Dev) (EM/F1)	Korean-Hate-Speech (Dev) (F1)
TUNiB-Electra-ko-base	110M	85.99	90.95	87.63	84.65	82.27	85.00	95.77	64.01 / 90.32	71.40
TUNiB-Electra-ko-en-base	133M	85.34	90.59	87.25	84.90	80.43	83.81	94.85	83.09 / 92.06	68.83
KoELECTRA-base-v3	110M	85.92	90.63	88.11	84.45	82.24	85.53	95.25	84.83 / 93.45	67.61
KcELECTRA-base	124M	84.75	91.71	86.90	74.80	81.65	82.65	95.78	70.60 / 90.11	74.49
KoBERT-base	90M	81.92	89.63	86.11	80.65	79.00	79.64	93.93	52.81 / 80.27	66.21
KcBERT-base	110M	79.79	89.62	84.34	66.95	74.85	75.57	93.93	60.25 / 84.39	68.77
XLM-Roberta-base	280M	83.03	89.49	86.26	82.95	79.92	79.09	93.53	64.70 / 88.94	64.06

Results on English downstream tasks

Small Models

	# Params	Avg.	CoLA (MCC)	SST (Acc)	MRPC (Acc)	STS (Spearman)	QQP (Acc)	MNLI (Acc)	QNLI (Acc)	RTE (Acc)
TUNiB-Electra-ko-en-small	18M	80.44	56.76	88.76	88.73	86.12	88.66	79.03	87.26	68.23
ELECTRA-small	13M	79.71	55.6	91.1	84.9	84.6	88.0	81.6	88.3	63.6
BERT-small	13M	74.06	27.8	89.7	83.4	78.8	87.0	77.6	86.4	61.8

Base Models

	# Params	Avg.	CoLA (MCC)	SST (Acc)	MRPC (Acc)	STS (Spearman)	QQP (Acc)	MNLI (Acc)	QNLI (Acc)	RTE (Acc)
TUNiB-Electra-ko-en-base	133M	85.2	65.36	92.09	88.97	90.61	90.91	85.32	91.51	76.53
ELECTRA-base	110M	85.7	64.6	96.0	88.1	90.2	89.5	88.5	93.1	75.2
BERT-base	110M	80.8	52.1	93.5	84.8	85.8	89.2	84.6	90.5	66.4

Pre-training data

• The Pile
• Reddit
• OpenWebText
• OpenSubtitles
• Modu Corpus
• Namuwiki
• KcBERT
• KoWiki
• Korean Petitions
• AI Hub Translation
• Etc.

Acknowledgement

The project was created with Cloud TPU support from the Tensorflow Research Cloud (TFRC) program.

Citation

If you find this code/model useful, please consider citing:

@misc{tunib-electra,
  author       = {Ha, Sangchun and Kim, Soohwan and Ryu, Myeonghyeon and
                  Keum, Bitna and Oh, Saechan and Ko, Hyunwoong and Park, Kyubyong},
  title        = {TUNiB-Electra},
  howpublished = {\url{https://github.com/tunib-ai/tunib-electra}},
  year         = {2021},
}

License

TUNiB-Electra is licensed under the terms of the Apache 2.0 License.

Copyright 2021 TUNiB Inc. http://www.tunib.ai All Rights Reserved.