japanese token-classification pos wikipedia dependency-parsing

bert-base-japanese-unidic-luw-upos

Model Description

This is a BERT model pre-trained on Japanese Wikipedia texts for POS-tagging and dependency-parsing, derived from bert-base-japanese-v2. Every long-unit-word is tagged by UPOS (Universal Part-Of-Speech).

How to Use

import torch
from transformers import AutoTokenizer,AutoModelForTokenClassification
tokenizer=AutoTokenizer.from_pretrained("KoichiYasuoka/bert-base-japanese-unidic-luw-upos")
model=AutoModelForTokenClassification.from_pretrained("KoichiYasuoka/bert-base-japanese-unidic-luw-upos")
s="国境の長いトンネルを抜けると雪国であった。"
t=tokenizer.tokenize(s)
p=[model.config.id2label[q] for q in torch.argmax(model(tokenizer.encode(s,return_tensors="pt"))["logits"],dim=2)[0].tolist()[1:-1]]
print(list(zip(t,p)))

or

import esupar
nlp=esupar.load("KoichiYasuoka/bert-base-japanese-unidic-luw-upos")
print(nlp("国境の長いトンネルを抜けると雪国であった。"))

fugashi, unidic-lite and pytokenizations are required.

Reference

安岡孝一: Transformersと国語研長単位による日本語係り受け解析モデルの製作, 情報処理学会研究報告, Vol.2022-CH-128, No.7 (2022年2月), pp.1-8.

See Also

esupar: Tokenizer POS-tagger and Dependency-parser with BERT/RoBERTa/DeBERTa models