CBio-NAMER
CBioNAMER (Nested nAMed Entity Recognition for Chinese Biomedical Text) is our method used in CBLUE (Chinese Biomedical Language Understanding Evaluation), a benchmark of Nested Named Entity Recognition. We got the 2nd price of the benchmark by 2021/12/07. Single model CBioNAMER also achieves top20 in CBLUE. The score of CBioNAMER has surpassed human(67.0 in F1-score).
Result
Results of our method:
Results of our single model CBioNAMER:
Approach
CBioNAMER is a sub-model in our result, which is based on GlobalPointer (a powerful open-source model, thanks for author, we rewrite it with Pytorch) and MacBert.
Usage
First, install PyTorch>=1.7.0. There's no restriction on GPU or CUDA.
Then, install this repo as a Python package:
$ pip install CBioNAMER
Python package transformers==4.6.1 would be automatically installed as well.
API
The CBioNAMER package provides the following methods:
CBioNAMER.load_NER(model_save_path='./checkpoint/macbert-large_dict.pth', maxlen=512, c_size=9, id2c=_id2c, c2c=_c2c)
Returns the pretrained model. It will download the model as necessary. The model would use the first CUDA device if there's any, otherwise using CPU instead.
The model_save_path argument specifies the path of the pretrained model weight.
The maxlen argument specifies the max length of input sentences. The sentences longer than maxlen would be cut off.
The c_size argument specifies the number of entity class. Here is 9 for CBLUE.
The id2c argument specifies the mapping between id and entity class. By default, the id2c argument for CBLUE is:
_id2c = {0: 'dis', 1: 'sym', 2: 'pro', 3: 'equ', 4: 'dru', 5: 'ite', 6: 'bod', 7: 'dep', 8: 'mic'}
The c2c argument specifies the mapping between entity class and its Chinese meaning. By default, the c2c argument for CBLUE is:
_c2c = {'dis': "疾病", 'sym': "临床表现", 'pro': "医疗程序", 'equ': "医疗设备", 'dru': "药物", 'ite': "医学检验项目", 'bod': "身体", 'dep': "科室", 'mic': "微生物类"}
The model returned by CBioNAMER.load_NER() supports the following methods:
model.recognize(text: str, threshold=0)
Given a sentence, returns a list of dictionaries with recognized entity, the format of the dictionary is {'start_idx': entity's starting index, 'end_idx': entity's ending index, 'type': entity class, 'Chinese_type': Chinese meaning of entity class, 'entity': recognized entity}. The threshold argument specifies that the returned list only contains the recognized entity with confidence score higher than threshold.
model.predict_to_file(in_file: str, out_file: str)
Given input and output .json file path, the model would do inference according in_file, and the recognized entity would be saved in out_file. The output file can be submitted to CBLUE. The format of input file is like:
[
{
"text": "该技术的应用使某些遗传病的诊治水平得到显著提高。"
},
...
{
"text": "There is a sentence."
}
]
Examples
import CBioNAMER
NER = CBioNAMER.load_NER()
in_file = './CMeEE_test.json'
out_file = './CMeEE_test_answer.json'
NER.predict_to_file(in_file, out_file)
import CBioNAMER
NER = CBioNAMER.load_NER()
text = "该技术的应用使某些遗传病的诊治水平得到显著提高。"
recognized_entity = NER.recognize(text)
print(recognized_entity)
# output:[{'start_idx': 9, 'end_idx': 11, 'type': 'dis', 'Chinese_type': '疾病', 'entity': '遗传病'}]
1.4k Feb 17, 2021
1.5k Feb 17, 2021
1.1k Dec 25, 2022
9 Nov 17, 2022
0 Feb 13, 2022
1.2k Dec 26, 2022
14 Nov 15, 2022
9 Nov 7, 2022
3 Feb 27, 2022
319 Dec 20, 2022
1 Oct 05, 2021
1 Oct 15, 2021
4 Oct 24, 2021
26 Nov 27, 2022
13 Dec 13, 2022
20 Dec 25, 2022
16 Dec 11, 2022
150 Dec 23, 2022
3.1k Jan 08, 2023
14 Nov 01, 2022
41 Nov 18, 2022
342 Nov 21, 2022
419 Dec 20, 2022
15.3k Dec 30, 2022
66 Dec 29, 2022
0 Jul 29, 2021
212 Dec 17, 2022
2.7k Dec 27, 2022
0 Feb 08, 2022