Based on the well-known DeBERTa V3 model
Specialized for the Japanese language
Does not use a morphological analyzer during inference
Respects word boundaries to some extent (does not produce tokens spanning multiple words like の都合上 or の判定負けを喫し)

How to use

python

from transformers import AutoTokenizer, AutoModelForTokenClassification

model_name = 'globis-university/deberta-v3-japanese-large'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

Tokenizer

工藤氏によって示された手法で学習しました。

以下のことを意識しています:

推論時の形態素解析器なし
トークンが単語の境界を跨がない (辞書: unidic-cwj-202302)
Hugging Faceで使いやすい
大きすぎない語彙数

本家の DeBERTa V3 は大きな語彙数で学習されていることに特徴がありますが、反面埋め込み層のパラメータ数が大きくなりすぎる (microsoft/deberta-v3-base モデルの場合で埋め込み層が全体の 54%) ことから、本モデルでは小さめの語彙数を採用しています。

注意点として、 xsmall 、 base 、 large の 3 つのモデルのうち、前者二つは unigram アルゴリズムで学習しているが、 large モデルのみ BPE アルゴリズムで学習している。深い理由はなく、 large モデルのみ語彙サイズを増やすために独立して学習を行ったが、なぜか unigram アルゴリズムでの学習がうまくいかなかったことが原因である。原因の探究よりモデルの完成を優先して、 BPE アルゴリズムに切り替えた。

The tokenizer is trained using the method introduced by Kudo.

Key points include:

No morphological analyzer needed during inference
Tokens do not cross word boundaries (dictionary: unidic-cwj-202302)
Easy to use with Hugging Face
Smaller vocabulary size

Although the original DeBERTa V3 is characterized by a large vocabulary size, which can result in a significant increase in the number of parameters in the embedding layer (for the microsoft/deberta-v3-base model, the embedding layer accounts for 54% of the total), this model adopts a smaller vocabulary size to address this.

Note that, among the three models: xsmall, base, and large, the first two were trained using the unigram algorithm, while only the large model was trained using the BPE algorithm. The reason for this is simple: while the large model was independently trained to increase its vocabulary size, for some reason, training with the unigram algorithm was not successful. Thus, prioritizing the completion of the model over investigating the cause, we switched to the BPE algorithm.

Data

Dataset Name	Notes	File Size (with metadata)	Factor
Wikipedia	2023/07; WikiExtractor	3.5GB	x2
Wikipedia	2023/07; cl-tohoku's method	4.8GB	x2
WikiBooks	2023/07; cl-tohoku's method	43MB	x2
Aozora Bunko	2023/07; globis-university/aozorabunko-clean	496MB	x4
CC-100	ja	90GB	x1
mC4	ja; extracted 10%, with Wikipedia-like focus via DSIR	91GB	x1
OSCAR 2023	ja; extracted 10%, with Wikipedia-like focus via DSIR	26GB	x1

Training parameters

Number of devices: 8
Batch size: 8 x 8
Learning rate: 6.4e-5
Maximum sequence length: 512
Optimizer: AdamW
Learning rate scheduler: Linear schedule with warmup
Training steps: 2,000,000
Warmup steps: 100,000
Precision: Mixed (fp16)
Vocabulary size: 48,000

Evaluation

Model	#params	JSTS	JNLI	JSQuAD	JCQA
≤ small
izumi-lab/deberta-v2-small-japanese	17.8M	0.890/0.846	0.880	-	0.737
globis-university/deberta-v3-japanese-xsmall	33.7M	0.916/0.880	0.913	0.869/0.938	0.821
base
cl-tohoku/bert-base-japanese-v3	111M	0.919/0.881	0.907	0.880/0.946	0.848
nlp-waseda/roberta-base-japanese	111M	0.913/0.873	0.895	0.864/0.927	0.840
izumi-lab/deberta-v2-base-japanese	110M	0.919/0.882	0.912	-	0.859
ku-nlp/deberta-v2-base-japanese	112M	0.922/0.886	0.922	0.899/0.951	-
ku-nlp/deberta-v3-base-japanese	160M	0.927/0.891	0.927	0.896/-	-
globis-university/deberta-v3-japanese-base	110M	0.925/0.895	0.921	0.890/0.950	0.886
large
cl-tohoku/bert-large-japanese-v2	337M	0.926/0.893	0.929	0.893/0.956	0.893
nlp-waseda/roberta-large-japanese	337M	0.930/0.896	0.924	0.884/0.940	0.907
nlp-waseda/roberta-large-japanese-seq512	337M	0.926/0.892	0.926	0.918/0.963	0.891
ku-nlp/deberta-v2-large-japanese	339M	0.925/0.892	0.924	0.912/0.959	-
globis-university/deberta-v3-japanese-large	352M	0.928/0.896	0.924	0.896/0.956	0.900

License

CC BY SA 4.0

Acknowledgement

計算リソースに ABCI を利用させていただきました。ありがとうございます。

We used ABCI for computing resources. Thank you.