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.
How to use
You can use this model directly with transformers library:
from transformers import AutoModel, AutoTokenizer
# Base Model (Korean-English bilingual model)
tokenizer = AutoTokenizer.from_pretrained('tunib/electra-ko-en-base')
model = AutoModel.from_pretrained('tunib/electra-ko-en-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("νλμ μμ°μ΄μ²λ¦¬ ν
ν¬ μ€ννΈμ
μ
λλ€.")
['ν', '##λ', '##μ', 'μμ°', '##μ΄', '##μ²λ¦¬', 'ν
ν¬', 'μ€ννΈμ
', '##μ
λλ€', '.']
Results on Korean downstream tasks
| # 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 | 84.17 | 89.63 | 86.11 | 80.65 | 79.00 | 79.64 | 93.93 | 52.81 / 80.27 | 66.21 |
| KcBERT-base | 110M | 81.37 | 89.62 | 84.34 | 66.95 | 74.85 | 75.57 | 93.93 | 60.25 / 84.39 | 68.77 |
| XLM-Roberta-base | 280M | 85.74 | 89.49 | 86.26 | 82.95 | 79.92 | 79.09 | 93.53 | 64.70 / 88.94 | 64.06 |
Results on English downstream tasks
| # 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 |