bert-ancient-chinese

Introduction

With the current wave of Artificial Intelligence and Digital Humanities sweeping the world, the automatic analysis of modern Chinese has achieved great results. However, the automatic analysis and research of ancient Chinese is relatively weak, and it is difficult to meet the actual needs of Sinology, history, philology, Chinese history and the education of Sinology and traditional culture. There are many controversies about characters, words and parts of speech in ancient Chinese, and there are many difficulties in resource construction. Digital Humanities research requires large-scale corpora and high-performance ancient natural language processing tools. In view of the fact that pre-trained language models have greatly improved the accuracy of text mining in English and modern Chinese texts, there is an urgent need for pre-trained models for the automatic processing of ancient texts.

In 2022, we took part in EvaHan 2022, the first NLP tool evaluation competition in the field of ancient Chinese. bert-ancient-chinese is trained to further optimize the model effect in open environment.

If you want to refer to our work, you can refer to this paper：

@inproceedings{wang2022uncertainty,
  title={The Uncertainty-based Retrieval Framework for Ancient Chinese CWS and POS},
  author={Wang, Pengyu and Ren, Zhichen},
  booktitle={Proceedings of the Second Workshop on Language Technologies for Historical and Ancient Languages},
  pages={164--168},
  year={2022}
}

You can view the introduction of the Chinese version through this link.

Further Pre-training

Compared with the previous pre-trained models, bert-ancient-chinese mainly has the following characteristics:

Ancient Chinese texts mostly appear in traditional Chinese characters and contain a large number of uncommon Chinese characters, which makes the vocab table (vocabulary) of the pre-trained model without some uncommon Chinese characters. bert-ancient-chinese further expands the vocab (dictionary) of the pre-trained model by learning in a large-scale corpus. The final vocab table size is 38208, compared to bert-base-chinese vocabulary size of 21128, siku-bert vocabulary size of 29791, bert-ancient-chinese has a larger vocabulary, and also includes more uncommon vocabulary word, which is more conducive to improving the performance of the model in downstream tasks. The vocab table is the vocabulary table, which is included in the vocab.txt in the pre-trained model.
bert-ancient-chinese uses a larger training set. Compared with siku-bert only using "Siku Quanshu" as training dataset, we use a larger-scale dataset (about six times that of "Siku Quanshu"), covering from the Ministry of Cong, the Ministry of Taoism, the Ministry of Buddhism, the Ministry of Confucianism, the Ministry of Poetry, the Ministry of History, the Ministry of Medicine, the Ministry of Art, the Ministry of Yi, and the Ministry of Zi, are richer in content and wider in scope than the "Siku Quanshu".
Based on the idea of Domain-Adaptive Pretraining, bert-ancient-chinese was trained on the basis of bert-base-chinese and was combined with ancient Chinese corpus to obtain a pre-trained model for the field of automatic processing of ancient Chinese.

How to use

Huggingface Transformers

The from_pretrained method based on Huggingface Transformers can directly obtain bert-ancient-chinese model online.

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("Jihuai/bert-ancient-chinese")

model = AutoModel.from_pretrained("Jihuai/bert-ancient-chinese")

Download PTM

The model we provide is the PyTorch version.

From Huggingface

Download directly through Huggingface's official website, and the model on the official website has been updated to the latest version simultaneously:

bert-ancient-chinese:Jihuai/bert-ancient-chinese · Hugging Face

From Cloud Disk

Download address:

Model	Link
bert-ancient-chinese	Link Extraction code: qs7x

Evaluation & Results

We tested and compared different pre-trained models on the training and test sets provided by the competition EvaHan 2022. We compare the performance of the models by fine-tuning them on the downstream tasks of Chinese Word Segmentation(CWS) and part-of-speech tagging(POS Tagging).

We use BERT+CRF as the baseline model to compare the performance of siku-bert, siku-roberta and bert-ancient-chinese on downstream tasks. To fully utilize the entire training dataset, we employ K-fold cross-validation, while keeping other hyperparameters the same. The evaluation index is the F1 value.

<table> <tr> <td></td> <td colspan="2" align="center"> Zuozhuan </td> <td colspan="2" align="center"> Shiji </td> </tr> <tr> <td></td> <td align="center">CWS</td> <td align="center">POS</td> <td align="center">CWS</td> <td align="center">POS</td> </tr> <tr> <td align="center">siku-bert</td> <td align="center">96.0670%</td> <td align="center">92.0156%</td> <td align="center">92.7909%</td> <td align="center">87.1188%</td> </tr> <tr> <td align="center">siku-roberta</td> <td align="center">96.0689%</td> <td align="center">92.0496%</td> <td align="center">93.0183%</td> <td align="center">87.5339%</td> </tr> <tr> <td align="center">bert-ancient-chinese</td> <td align="center"> 96.3273% </td> <td align="center"> 92.5027% </td> <td align="center"> 93.2917% </td> <td align="center"> 87.8749% </td> </tr> </table>

Citing

If our content is helpful for your research work, please quote it in the paper.

Disclaim

The experimental results presented in the report only show the performance under a specific data set and hyperparameter combination, and cannot represent the essence of each model. The experimental results may change due to random number seeds and computing equipment. Users can use the model arbitrarily within the scope of the license, but we are not responsible for the direct or indirect losses caused by using the content of the project.

Acknowledgment

bert-ancient-chinese is based on bert-base-chinese to continue training.

Thanks to Prof. Xipeng Qiu and the Natural Language Processing Laboratory of Fudan University.

Contact us

Pengyu Wang：wpyjihuai@gmail.com