xlm-roberta-ner-japanese
(Japanese caption : 日本語の固有表現抽出のモデル)
This model is a fine-tuned version of xlm-roberta-base (pre-trained cross-lingual RobertaModel
) trained for named entity recognition (NER) token classification.
The model is fine-tuned on NER dataset provided by Stockmark Inc, in which data is collected from Japanese Wikipedia articles.
See here for the license of this dataset.
Each token is labeled by :
Label id | Tag | Tag in Widget | Description |
---|---|---|---|
0 | O | (None) | others or nothing |
1 | PER | PER | person |
2 | ORG | ORG | general corporation organization |
3 | ORG-P | P | political organization |
4 | ORG-O | O | other organization |
5 | LOC | LOC | location |
6 | INS | INS | institution, facility |
7 | PRD | PRD | product |
8 | EVT | EVT | event |
Intended uses
from transformers import pipeline
model_name = "tsmatz/xlm-roberta-ner-japanese"
classifier = pipeline("token-classification", model=model_name)
result = classifier("鈴井は4月の陽気の良い日に、鈴をつけて北海道のトムラウシへと登った")
print(result)
Training procedure
You can download the source code for fine-tuning from here.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 12
- eval_batch_size: 12
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 5
Training results
Training Loss | Epoch | Step | Validation Loss | F1 |
---|---|---|---|---|
No log | 1.0 | 446 | 0.1510 | 0.8457 |
No log | 2.0 | 892 | 0.0626 | 0.9261 |
No log | 3.0 | 1338 | 0.0366 | 0.9580 |
No log | 4.0 | 1784 | 0.0196 | 0.9792 |
No log | 5.0 | 2230 | 0.0173 | 0.9864 |
Framework versions
- Transformers 4.23.1
- Pytorch 1.12.1+cu102
- Datasets 2.6.1
- Tokenizers 0.13.1
- Downloads last month
- 1,226,691