English | 简体中文

DeepKE is a knowledge extraction toolkit based on PyTorch, supporting low-resource, document-level and multimodal scenarios for entity, relation and attribute extraction. DeepKE-cnSchema is an off-the-shelf version. Users can download the model to realize entity and relation knowledge extraction directly without training.

DeepKE is a knowledge extraction toolkit supporting low-resource, document-level and multimodal scenarios for entity, relation and attribute extraction. We provide documents, Google Colab tutorials, online demo, and slides for beginners.

To promote efficient Chinese knowledge graph construction, we provide DeepKE-cnSchema, a specific version of DeepKE, containing off-the-shelf models based on cnSchema. DeepKE-cnSchema supports multiple tasks such as Chinese entity extraction and relation extraction. It can extract 50 relation types and 28 entity types, including common entity types such as person, location, city, institution, etc and the common relation types such as ancestral home, birthplace, nationality and other types.

For entity extraction and relation extraction tasks, we provide models based on RoBERTa-wwm-ext, Chinese and BERT-wwm, Chinese respectively.

It is recommended to use Baidu Netdisk download in Chinese Mainland, and Google download for overseas users.

As for the entity extraction model, take pytoch version DeepKE(RE), RoBERTa-wwm-ext, Chinese as an example. After downloading, files of the model are obtained:

checkpoints_robert
    |- added_tokens.json          # added tokens
    |- config.json                # config
    |- eval_results.txt           # evaluation results
    |- model_config.json          # model config
    |- pytorch_model.bin          # model
    |- special_tokens_map.json    # special tokens map
    |- tokenizer_config.bin       # tokenizer config
    |- vocab.txt                  # vocabulary

where config.json and vocab.txt is completely consistent with the original Google RoBERTa-wwm-ext, Chinese. PyTorch version contains pytorch_model. bin, config. json, vocab. txt file.

As for the relation extraction model, take pytoch version DeepKE(RE), RoBERTa-wwm-ext, Chinese as an example. The model is pth file after downloading.

After downloading the model, users can directly quick-load it to extract entity and relation.

We have conduct experiments on Chinese named entity recognition and relation extraction datasets. The experimental results are as follows:

DeepKE leverageschinese-bert-wwmandchinese-roberta-wwm-extto train and obtain the DeepKE-cnSchema(NER) model. Hyper-parameters used in the model are predefined. Finally, we can obtain the following results after training:

DeepKE leverageschinese-bert-wwmandchinese-roberta-wwm-extto train and obtain the DeepKE-cnschema(RE) model. Hyper-parameters used in the model are predefined. Finally, we can obtain the following results after training:

DeepKE-cnSchema is an off-the-shelf version that supports the Chinese knowledge graphs construction. CnSchema is developed for Chinese information processing, which uses advanced knowledge graphs, natural language processing and machine learning technologies. It integrates structured text data, supports rapid domain knowledge modeling and open data automatic processing across data sources, domains and languages, and provides schema-level support and services for emerging application markets such as intelligent robots, semantic search and intelligent computing. Currently, the Schema types supported by DeepKE-cnSchema are as follows:

Users can directly download the model for usage. The details are as follows：

1. Create the downloaded folder as checkpoints

2. Set the parameter text in predict.yaml as the sentence to be predicted To use the trained model, just set the input sentence "《星空黑夜传奇》是连载于起点中文网的网络小说，作者是啤酒的罪孽". After running python oredict.py, results can be obtained which show that the entity type "星空黑夜传奇" is "网络小说", "起点中文网" is "网站" and "啤酒的罪孽" is "人物".

   text="《星空黑夜传奇》是连载于起点中文网的网络小说，作者是啤酒的罪孽"

3. Predict

   python predict.py

   Finally, output the results:

   NER句子：
   《星空黑夜传奇》是连载于起点中文网的网络小说，作者是啤酒的罪孽
   NER结果：
   [('星','B-UER'),('空','I-UER'),('黑','I-UER'),('夜','I-UER'),('传','I-UER'),('奇','I-UER'),('起','B-ZFM'),('点','I-ZFM'),('中','I-ZFM'),('文','I-ZFM'),('网','I-ZFM'),('啤','B-YAS'),('酒','I-YAS'),('的','I-YAS'),('罪','I-YAS'),('孽','I-YAS')]

Users can directly download the model for usage. The details are as follows：

1. Modify the parameter fpin predict.yamlto the path of downloaded file, num_relationsin embedding.yamlto 51(relation nums) and model in config.yamlto lm

2. Predict. The text and entity pairs to be predicted are fed to the program through the terminal.

   python predict.py

   To use the trained model, run python predict.py and input the sentence "歌曲《人生长路》出自刘德华国语专辑《男人的爱》，由李泉作词作曲，2001年出行发版". The given entity pair are "男人的爱" and "人生长路". Finally, the extracted relation is "所属专辑".

   To change the text to be predicted, modify the _get_predict_instancefunction in predict.py to the following example:

   def _get_predict_instance(cfg):
       flag = input('是否使用范例[y/n]，退出请输入: exit .... ')
       flag = flag.strip().lower()
       if flag == 'y' or flag == 'yes':
   	sentence = '歌曲《人生长路》出自刘德华国语专辑《男人的爱》，由李泉作词作曲，2001年出行发版'
   	head = '男人的爱'
   	tail = '人生长路'
   	head_type = '所属专辑'
   	tail_type = '歌曲'
       elif flag == 'n' or flag == 'no':
   	sentence = input('请输入句子：')
   	head = input('请输入句中需要预测关系的头实体：')
   	head_type = input('请输入头实体类型（可以为空，按enter跳过）：')
   	tail = input('请输入句中需要预测关系的尾实体：')
   	tail_type = input('请输入尾实体类型（可以为空，按enter跳过）：')
       elif flag == 'exit':
   	sys.exit(0)
       else:
   	print('please input yes or no, or exit!')
   	_get_predict_instance()
   
       instance = dict()
       instance['sentence'] = sentence.strip()
       instance['head'] = head.strip()
       instance['tail'] = tail.strip()
       if head_type.strip() == '' or tail_type.strip() == '':
   	cfg.replace_entity_with_type = False
   	instance['head_type'] = 'None'
   	instance['tail_type'] = 'None'
       else:
   	instance['head_type'] = head_type.strip()
   	instance['tail_type'] = tail_type.strip()
   
       return instance

   Finally, output the results:

   “男人的爱”和“人生长路”在句中关系为“所属专辑”，置信度为0.99

After aforementioned trained models are downloaded, entites and their relations in a text can be extracted together. If there are more than two entities in one sentence, some predicted entity pairs may be incorrect because these entity pairs are not in training sets and need to be exracted further. The detailed steps are as follows:

1. In conf, modify text in predict.yaml as the sentence to be predicted, nerfp as the directory of the trained NER model and refp as the directory of the trained RE model.

2. Predict

   python predict.py

   Many results will be output. Take the input text 此外网易云平台还上架了一系列歌曲，其中包括田馥甄的《小幸运》等 as example.

   (1) Output the result of NER: [('田', 'B-YAS'), ('馥', 'I-YAS'), ('甄', 'I-YAS'), ('小', 'B-QEE'), ('幸', 'I-QEE'), ('运', 'I-QEE')]

   (2) Output the processed result: {'田馥甄': '人物', '小幸运': '歌曲'}

   (3) Output the result of RE: "田馥甄" 和 "小幸运" 在句中关系为："歌手"，置信度为0.92。

   (4) Output the result as jsonld

      {
        "@context": {
      "歌手": "https://cnschema.openkg.cn/item/%E6%AD%8C%E6%89%8B/16693#viewPageContent"
        },
        "@id": "田馥甄",
        "歌手": {
      "@id": "小幸运"
        }
      }

If you need to use customized dataset for training, follow the steps bellow:

1. Download customized dataset and put it into the data folder.
2. Modify the parameter bert_modelin train.yamlof the conffolder to the specify model. Users can choose different models to train by modifying the yamlfile.
3. Modify labels in train.yaml as the labels in data/type.txt
4. Train.

   python run.py

If you need to use other models for training, follow the steps bellow:

1、Download the customized dataset and rename it to data.

2、Modify the parameter model_namein train.yamlof the conffolder to lm, num_relationsin embedding_yamlto the number of relations(eg: 51). Users can choose different models to train by modifying the yamlfile.

3、Train.

```bash
python run.py
```

Q: How to use this model? A: It is off-the-shelf. After downloading the model, follow the instructions and you can extract the knowledge contained in the predefined cnSchema. If you want to extract knowledge other than cnSchema, you can use the advanced version of customized data for training

Q: Is there any other cnSchema extraction model available? A: Unfortunately, we can only support part of knowledge extraction of cnSchema for the time being. More knowledge extraction models will be published in the future.

Q: I trained better result than you! A: Congratulations!

Q: Embedding error for customized dataset. A: The Chinese data may contain invisible special characters, which cannot be encoded and thus an error is reported. You can preprocess the Chinese data through the editor or other tools to solve this problem.

If the resources or technologies in this project are helpful to your research work, you are welcome to cite the following papers in your thesis:

@inproceedings{DBLP:conf/emnlp/ZhangXTYYQXCLL22,
  author    = {Ningyu Zhang and
               Xin Xu and
               Liankuan Tao and
               Haiyang Yu and
               Hongbin Ye and
               Shuofei Qiao and
               Xin Xie and
               Xiang Chen and
               Zhoubo Li and
               Lei Li},
  editor    = {Wanxiang Che and
               Ekaterina Shutova},
  title     = {DeepKE: {A} Deep Learning Based Knowledge Extraction Toolkit for Knowledge
               Base Population},
  booktitle = {Proceedings of the The 2022 Conference on Empirical Methods in Natural
               Language Processing, {EMNLP} 2022 - System Demonstrations, Abu Dhabi,
               UAE, December 7-11, 2022},
  pages     = {98--108},
  publisher = {Association for Computational Linguistics},
  year      = {2022},
  url       = {https://aclanthology.org/2022.emnlp-demos.10},
  timestamp = {Thu, 23 Mar 2023 16:56:00 +0100},
  biburl    = {https://dblp.org/rec/conf/emnlp/ZhangXTYYQXCLL22.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

The contents of this project are only for technical research reference and shall not be used as any conclusive basis. Users can freely use the model within the scope of the license, but we are not responsible for the direct or indirect losses caused by the use of the project.

If you have any questions, please submit them in GitHub issue.