Research on the extraction method of acupuncture and moxibustion prescription based on natural language processing technology
10.3760/cma.j.cn115398-20230808-00079
- VernacularTitle:基于自然语言处理技术的针灸处方提取方法研究
- Author:
Ying LI
1
;
Yuebo JIANG
;
Ling GUAN
Author Information
1. 解放军总医院第一医学中心中医（针灸）科，北京 100853
- Keywords:
Acupuncture and moxibustion prescription;
Natural language processing;
Pre-trained language model
- From:
International Journal of Traditional Chinese Medicine
2024;46(11):1506-1510
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To study a method for automatically extracting acupuncture and moxibustion prescriptions from clinical literature to assist data mining of acupuncture and moxibustion prescriptions; To support clinical research and decision-making in acupuncture and moxibustion treatment.Methods:The Chinese journal articles on clinical trials of acupuncture and moxibustion in CNKI from January 1, 1992 to December 31, 2022 were searched. The titles and abstracts of 750 articles were randomly selected and manually labeled. The three main entities of disease name, acupuncture and moxibustion method and acupoint of acupuncture and moxibustion prescription were selected. From the data set, 70% was selected as the training set, 15% as the validation set, and 15% as the test set for the experiment. The extraction of acupuncture prescriptions was considered a sequence labeling task. A model for automatic extraction of acupuncture prescriptions was built using a pretrained language model (PLM), and four different PLMs were selected to compare their entity recognition effects. The impact of negative sampling and label smoothing training techniques on the model was further investigated.Results:The model based on eHealth had the highest F1 scores (92.84). During training, if only negative sampling technology was used, F1 value increased to 93.53; if only label smoothing was used, F1 value increased to 93.64; if negative sampling and label smoothing were used simultaneously, F1 value increased to 94.28, an increase of 1.55%. Conclusions:This study proposes a fast and accurate model for extracting acupuncture and moxibustion prescriptions. The research shows that the model recognition effect based on eHealth in the biomedical field is the best, and the recognition effect of the model can be further improved by using negative sampling and label smoothing training techniques.
