---
license: apache-2.0
language:
- ja
---






# Model Card for japanese-spoken-language-bert

日本語READMEは[こちら](./README_JA.md)

<!-- Provide a quick summary of what the model is/does. [Optional] -->
These BERT models are pre-trained on written Japanese (Wikipedia) and fine-tuned on Spoken Japanese.
We used CSJ and the Japanese diet record.
CSJ (Corpus of Spontaneous Japanese) is provided by NINJAL (https://www.ninjal.ac.jp/).
We only provide model parameters. You have to download other config files to use these models.

We provide three models down below:
- **1-6 layer-wise** (Folder Name: models/1-6_layer-wise)  
    Fine-Tuned only 1st-6th layers in Encoder on CSJ.

- **TAPT512 60k** (Folder Name: models/tapt512_60k)  
    Fine-Tuned on CSJ.

- **DAPT128-TAPT512** (Folder Name: models/dapt128-tap512)  
    Fine-Tuned on the diet record and CSJ.

#  Table of Contents

- [Model Card for japanese-spoken-language-bert](#model-card-for-japanese-spoken-language-bert)
- [Table of Contents](#table-of-contents)
- [Model Details](#model-details)
  - [Model Description](#model-description)
- [Training Details](#training-details)
  - [Training Data](#training-data)
  - [Training Procedure](#training-procedure)
- [Evaluation](#evaluation)
  - [Testing Data, Factors & Metrics](#testing-data-factors--metrics)
    - [Testing Data](#testing-data)
    - [Factors](#factors)
    - [Metrics](#metrics)
  - [Results](#results)
- [Citation](#citation)
- [More Information](#more-information-optional)
- [Model Card Authors](#model-card-authors-optional)
- [Model Card Contact](#model-card-contact)
- [How to Get Started with the Model](#how-to-get-started-with-the-model)


# Model Details

## Model Description

<!-- Provide a longer summary of what this model is/does. -->
These BERT models are pre-trained on written Japanese (Wikipedia) and fine-tuned on Spoken Japanese.
We used CSJ and the Japanese diet record.
CSJ (Corpus of Spontaneous Japanese) is provided by NINJAL (https://www.ninjal.ac.jp/).
We only provide model parameters. You have to download other config files to use these models.

We provide three models down below:
- 1-6 layer-wise (Folder Name: models/1-6_layer-wise)  
    Fine-Tuned only 1st-6th layers in Encoder on CSJ.

- TAPT512 60k (Folder Name: models/tapt512_60k)  
    Fine-Tuned on CSJ.

- DAPT128-TAPT512 (Folder Name: models/dapt128-tap512)  
    Fine-Tuned on the diet record and CSJ.

**Model Information**
- **Model type:** Language model
- **Language(s) (NLP):** ja
- **License:** Copyright (c) 2021 National Institute for Japanese Language and Linguistics and Retrieva, Inc. Licensed under the Apache License, Version 2.0 (the “License”)


# Training Details

## Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

- 1-6 layer-wise: CSJ
- TAPT512 60K: CSJ
- DAPT128-TAPT512: The Japanese diet record and CSJ


## Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

We continuously train the pre-trained Japanese BERT model ([cl-tohoku/bert-base-japanese-whole-word-masking](https://huggingface.co/cl-tohoku/bert-base-japanese-whole-word-masking); written BERT).

In detail, see [Japanese blog](https://tech.retrieva.jp/entry/2021/04/01/114943) or [Japanese paper](https://www.anlp.jp/proceedings/annual_meeting/2021/pdf_dir/P4-17.pdf).

# Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

## Testing Data, Factors & Metrics

### Testing Data

<!-- This should link to a Data Card if possible. -->

We use CSJ for the evaluation.


### Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

We evaluate the following tasks on CSJ:
- Dependency Parsing
- Sentence Boundary
- Important Sentence Extraction

### Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

- Dependency Parsing: Undirected Unlabeled Attachment Score (UUAS)
- Sentence Boundary: F1 Score
- Important Sentence Extraction: F1 Score

## Results

| | Dependency Parsing | Sentence Boundary | Important Sentence Extraction |
| :--- | ---: | ---: | ---: |
| written BERT | 39.4 | 61.6 | 36.8 |
| 1-6 layer wise | 44.6 | 64.8 | 35.4 |
| TAPT 512 60K | - | - | 40.2 |
| DAPT128-TAPT512 | 42.9 | 64.0 | 39.7 |


# Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

```bibtex
@inproceedings{csjbert2021,
    title = {CSJを用いた日本語話し言葉BERTの作成},
    author = {勝又智 and 坂田大直},
    booktitle = {言語処理学会第27回年次大会},
    year = {2021},
}
```


# More Information

https://tech.retrieva.jp/entry/2021/04/01/114943 (In Japanese)

# Model Card Authors

<!-- This section provides another layer of transparency and accountability. Whose views is this model card representing? How many voices were included in its construction? Etc. -->

Satoru Katsumata

# Model Card Contact

pr@retrieva.jp

# How to Get Started with the Model

Use the code below to get started with the model.

<details>
<summary> Click to expand </summary>

1. Run download_wikipedia_bert.py to download BERT model which is trained on Wikipedia.

```bash
python download_wikipedia_bert.py
```

This script downloads config files and a vocab file provided by Inui Laboratory of Tohoku University from Hugging Face Model Hub.
https://github.com/cl-tohoku/bert-japanese

2. Run sample_mlm.py to confirm you can use our models.

```bash
python sample_mlm.py
```

</details>