cl-tohoku commited on
Commit
f58f6fc
1 Parent(s): 686890e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +50 -0
README.md CHANGED
@@ -1,3 +1,53 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ datasets:
4
+ - cc100
5
+ - wikipedia
6
+ language:
7
+ - ja
8
+ widget:
9
+ - text: 東北大学で[MASK]の研究をしています。
10
  ---
11
+
12
+ # BERT base Japanese (unidic-lite with whole word masking, CC-100 and jawiki-20230102)
13
+
14
+ This is a [BERT](https://github.com/google-research/bert) model pretrained on texts in the Japanese language.
15
+
16
+ This version of the model processes input texts with word-level tokenization based on the Unidic 2.1.2 dictionary (available in [unidic-lite](https://pypi.org/project/unidic-lite/) package), followed by the WordPiece subword tokenization.
17
+ Additionally, the model is trained with the whole word masking enabled for the masked language modeling (MLM) objective.
18
+
19
+ The codes for the pretraining are available at [cl-tohoku/bert-japanese](https://github.com/cl-tohoku/bert-japanese/).
20
+
21
+ ## Model architecture
22
+
23
+ The model architecture is the same as the original BERT base model; 12 layers, 768 dimensions of hidden states, and 12 attention heads.
24
+
25
+ ## Training Data
26
+
27
+ The model is trained on the Japanese portion of [CC-100 dataset](https://data.statmt.org/cc-100/) and the Japanese version of Wikipedia.
28
+ For Wikipedia, we generated a text corpus from the [Wikipedia Cirrussearch dump file](https://dumps.wikimedia.org/other/cirrussearch/) as of January 2, 2023.
29
+ The corpus files generated from CC-100 and Wikipedia are 74.3GB and 4.9GB in size and consist of approximately 392M and 34M sentences, respectively.
30
+
31
+ For the purpose of splitting texts into sentences, we used [fugashi](https://github.com/polm/fugashi) with [mecab-ipadic-NEologd](https://github.com/neologd/mecab-ipadic-neologd) dictionary (v0.0.7).
32
+
33
+ ## Tokenization
34
+
35
+ The texts are first tokenized by MeCab with the Unidic 2.1.2 dictionary and then split into subwords by the WordPiece algorithm.
36
+ The vocabulary size is 32768.
37
+
38
+ We used [fugashi](https://github.com/polm/fugashi) and [unidic-lite](https://github.com/polm/unidic-lite) packages for the tokenization.
39
+
40
+ ## Training
41
+
42
+ We trained the model first on the CC-100 corpus for 1M steps and then on the Wikipedia corpus for another 1M steps.
43
+ For training of the MLM (masked language modeling) objective, we introduced whole word masking in which all of the subword tokens corresponding to a single word (tokenized by MeCab) are masked at once.
44
+
45
+ For training of each model, we used a v3-8 instance of Cloud TPUs provided by [TPU Research Cloud](https://sites.research.google/trc/about/).
46
+
47
+ ## Licenses
48
+
49
+ The pretrained models are distributed under the Apache License 2.0.
50
+
51
+ ## Acknowledgments
52
+
53
+ This model is trained with Cloud TPUs provided by [TPU Research Cloud](https://sites.research.google/trc/about/) program.