Research On Knowledge Graph Construction And Semantic Retrieval Of Acupuncture Effect

1 ObjectiveBased on clinical randomized controlled trial literature of acupuncture,this study used literature review and expert consultation to establish a acupuncture effect knowledge graph model.By utilizing natural language processing techniques such as named entity recognition,an entity extraction model was constructed for clinical randomized controlled trial literature on acupuncture.With the help of the Knowledge Organization System of Traditional Chinese Medicine(KOS),a knowledge graph of acupuncture effect based on clinical randomized controlled trial literature was built.Finally,an application scenario was designed to apply the knowledge graph to semantic retrieval and visualization of search results.This study solved the time-consuming and laborious problems in the process of mining massive clinical randomized controlled trial literature on acupuncture through entity extraction,making it easier for researchers and clinicians to conduct secondary studies on clinical randomized controlled trial literature on acupuncture.Through semantic retrieval and visualization,this study provides intuitive and convenient knowledge services for clinical acupuncture research.2 Contents and methods2.1 knowledge model construction of acupuncture effectBased on the literature review of randomized controlled trials(RCTs)on acupuncture and the existing knowledge graphs related to acupuncture and RCT literature,and with the help of expert consultation,the elements of acupuncture effects in RCT literature,entity classification system,entity relationships,entity attributes,and constraints were determined.On this basis,the acupuncture effects knowledge graph model was developed,which guided the construction of the entity extraction model’s labels and laid the foundation for the establishment of the knowledge graph.2.2 Research on named entity identification method of acupuncture randomized controlled trialThe RCT literature on acupuncture was manually annotated to form a BIO annotation set.Three entity extraction models were proposed using a method based on dictionaries and deep learning,including word2vec-BiLSTM-CRF,ALBERT-BiLSTM-CRF,and ALBERT-BiLSTM-CRF+dictionary.The entity extraction effects of the three models were compared using indicators such as precision P,recall R,and F1 scores,and the ALBERT-BiLSTM-CRF+dictionary model with the best entity recognition performance was selected and embedded in the automatic database construction system developed by the research group earlier.2.3 Construction of knowledge map of acupuncture effectThe PDF document data is converted into TXT format data by OCR recognition technology,and the related elements of acupuncture effect in TXT text are automatically extracted by using the knowledge extraction model obtained in the above research,and manual proofreading is carried out on the self-built library platform.EXCEL data set is established,elements,relationships between elements and attributes of elements are established in KOS platform according to the knowledge model of acupuncture effect,and the data set is imported into KOS platform to construct the knowledge map of acupuncture effect.2.4 Design and implementation of semantic retrieval system for acupuncture effectBased on the knowledge map of acupuncture effect,the semantic retrieval system of acupuncture effect is designed from two aspects:scene design and function design.After the system design is successful,the documents are extracted,and the results of extraction are compared with those of manual extraction in the database to test the accuracy of the model in the second research content.3 Results3.1 The elements,relations and attributes of acupuncture effect are defined,and the knowledge model of acupuncture effect is constructed.There are four main categories and fourteen types of entities in the acupuncture effect ontology.The literature-related entities include literature and experiments;acupuncture treatment-related entities include acupuncture points,needling techniques,and manipulation techniques;disease-related entities include diseases,syndromes,and symptoms;acupuncture effect-related entities include functional rating scales,physical examination indicators,and blood biochemical indicators.eleven types of relationships have been established between entities,including influence,act on,diagnose as,exhibit,have acupuncture points,have needling techniques,have experiments,observe diseases,have indicators,have functional rating items,have manipulation techniques,observe syndromes,observe symptoms,have blood biochemical items,and have physical examination items.Five entity attributes have been established,including literature ID,literature source,intervention time,total sample size,and total effective rate.3.2 The knowledge extraction model of acupuncture effect is constructed and the extraction platform is built.(1)This study used two methods,word2vec and ALBERT,for word vector transformation.The resulting vectors were input into a BiLSTM-CRF model for training.The entity recognition performance of the "ALBERT-BiLSTM-CRF+dictionary" model was the best,with P value of 92.57%,R value of 91.42%,and F1 value of 91.85%.The ALBERT-BiLSTM-CRF model performed slightly worse,with P value of 83.10%,R value of 81.14%,and F1 value of 81.98%.The word2vec-BiLSTM-CRF model had the worst performance,with P value of 81.82%,R value of 70.76%,and F1 value of 75.48%.Looking at entity categories,the top three entities with the highest precision were acupuncture techniques,needling methods,and acupuncture points,with precision rates of 98%,97%,and 97%,respectively.The three entities with the lowest precision rates were symptoms,disease names,and indicators,with precision rates of 92%,89%,and 79%,respectively.(2)Based on the ALBERT-BiLSTM-CRF+dictionary knowledge extraction model,this study designed and completed an automatic database system for RCT literature knowledge extraction module,which supports functions such as literature import,knowledge extraction,manual verification,and data export.This study used this knowledge extraction module to complete the structured processing of acupuncture RCT literature.3.3 The knowledge map of acupuncture effect is constructed.This study used the KOS platform to construct the acupuncture effect knowledge graph model by adding entity types,relationships,and attributes for acupuncture effects.Then,the Excel batch import module was used to import the structured processed acupuncture RCT literature using a configuration file.Finally,a total of 14,708 acupuncture RCT literature from 1956 to 2020 were included,with 54,338 entities,226,393 relationship properties,and 73,955 data properties.3.4 A semantic retrieval system is constructed.The platform’s two application scenarios were designed according to actual needs.The first is to discover the acupoint selection rules for acupuncture treatment of a certain disease through similarity and combination calculation.The second is to study the inherent connections and differences between different diseases but with the same syndrome,such as differences and similarities in acupuncture points,needling methods,detection indicators,functional score indicators,blood biochemical indicators,and physical examination indicators,in order to assist in disease diagnosis.Four search modes,including one-to-many,many-to-many,many-to-one,and combined search,were designed,and the calculation and visualization display of acupuncture effect knowledge were implemented.4 ConclusionsThis study successfully constructed a knowledge graph of acupuncture effects based on RCTs,using a combination of human and machine approaches,providing a new knowledge service for acupuncture clinical practice.This knowledge graph includes information on acupuncture points,needling techniques,acupuncture methods,and detection indicators,among others,which can help clinical doctors and researchers better understand the mechanism and effects of acupuncture treatment and assist in exploring the treatment principles and theoretical foundations for treating diseases with acupuncture.Additionally,the semantic retrieval system based on the knowledge graph of acupuncture effects can provide more flexible and personalized retrieval methods,providing users with a quick and accurate way to obtain knowledge.This study provides valuable references and support for acupuncture clinical practice and acupuncture research,and has certain theoretical and practical application value.