ku-accms/bert-base-japanese-ssuw
Model description
This is a pre-trained Japanese BERT base model for super short unit words (SSUW).
Pre-processing
The input text should be converted to full-width (zenkaku) characters and segmented into super short unit words in advance (e.g., by KyTea).
How to use
You can use this model directly with a pipeline for masked language modeling:
>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='ku-accms/bert-base-japanese-ssuw')
>>> unmasker("京都 大学 で [MASK] を 専攻 する 。")
[{'sequence': '京都 大学 で 文学 を 専攻 する 。',
'score': '0.1464807540178299',
'token': '14603',
'token_str': '文学'}
{'sequence': '京都 大学 で 哲学 を 専攻 する 。',
'score': '0.08064978569746017',
'token': '15917',
'token_str': '哲学'}
{'sequence': '京都 大学 で 演劇 を 専攻 する 。',
'score': '0.0800977498292923',
'token': '16772',
'token_str': '演劇'}
{'sequence': '京都 大学 で 法学 を 専攻 する 。',
'score': '0.04579947143793106',
'token': '16255',
'token_str': '法学'}
{'sequence': '京都 大学 で 英語 を 専攻 する 。',
'score': '0.045536939054727554',
'token': '14592',
'token_str': '英語'}
Here is how to use this model to get the features of a given text in PyTorch:
import zenhan
import Mykytea
kytea_model_path = "somewhere"
kytea = Mykytea.Mykytea("-model {} -notags".format(kytea_model_path))
def preprocess(text):
return " ".join(kytea.getWS(zenhan.h2z(text)))
from transformers import BertTokenizer, BertModel
tokenizer = BertTokenizer.from_pretrained('ku-accms/bert-base-japanese-ssuw')
model = BertModel.from_pretrained("ku-accms/bert-base-japanese-ssuw")
text = "京都大学で自然言語処理を専攻する。"
encoded_input = tokenizer(preprocess(text), return_tensors='pt')
output = model(**encoded_input)
Training data
We used a Japanese Wikipedia dump (as of 20230101, 3.3GB).
Training procedure
We first segmented the texts into words by KyTea and then tokenized the words into subwords using WordPiece with a vocabulary size of 32,000. We pre-trained the BERT model using transformers library. The training took about 8 days using 4 NVIDIA A100-SXM4-80GB GPUs.
The following hyperparameters were used for the pre-training.
- learning_rate: 2e-4
- weight decay: 1e-2
- per_device_train_batch_size: 80
- num_devices: 4
- gradient_accumulation_steps: 3
- total_train_batch_size: 960
- max_seq_length: 512
- optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-06
- lr_scheduler_type: linear schedule with warmup
- training_steps: 500,000
- warmup_steps: 10,000