A knowledge graph-based automatic creation method for a nursing corpus

Abstract

The application discloses a kind of knowledge graph-based nursing corpus automatic creation method, it is related to artificial intelligence technology and medical information processing field, the method includes: multidimensional data acquisition stage, knowledge graph construction stage, corpus automatic generation and update stage, corpus quality control and evaluation stage.The application adopts knowledge graph-driven corpus automatic generation method, by constructing the ontology structure of nursing field, corpus generation template and updating algorithm, the dynamic generation and real-time update of corpus are realized;Through deep learning model and transfer learning technology, the application continuously introduces new nursing knowledge and diversified information structure in corpus, increases the diversity of corpus content.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence technology and medical information processing, in particular to a nursing corpus automatic creation method based on a knowledge graph. BACKGROUND

[0002] With the rapid development of medical informatization, the demand for high-quality, structured nursing corpus in the nursing field has increased significantly, and developing a method for automatically constructing and continuously updating a nursing corpus has become a key issue for research.

[0003] In terms of nursing knowledge graph construction, foreign countries started earlier and have a relatively mature ontology framework. The United States and Japan have constructed a nursing practice ontology based on the International Classification for Nursing Practice (ICNP), providing strong support for the in-depth application of nursing information systems and nursing data mining. In contrast, China's nursing knowledge graph is still in its early stages of development, and further exploration is needed to effectively integrate nursing practice needs and knowledge graph technology to form a field knowledge graph that adapts to diverse nursing scenarios.

[0004] A corpus is a collection of natural language texts containing certain language features, and is an important tool for linguistic and natural language processing research. Corpus research originated in the United Kingdom and the United States in the 1960s, initially focusing on a single language (such as English), and has gone through the following four stages: the first stage (1960s) of small corpus, mainly used for basic linguistic research; the second stage (1980s) of large-scale corpus, with data levels expanded by millions of times; the third stage (late 20th century) of dynamic corpus, focusing on content updates and opening historical data; the fourth stage of open, dynamic corpus based on the Internet, gradually becoming a more extensive language resource. Currently, foreign countries pay more attention to the application of corpus in nursing education and training, and use corpus to improve the professional ability and practice level of nursing staff. In China, corpus research began in the 1990s, mainly from the perspective of linguistics, and its application in the nursing field is still in its infancy. The future development of nursing corpus can be achieved by combining artificial intelligence technology to construct an intelligent decision support system, helping nursing staff make more accurate judgments and decisions in clinical practice.

[0005] In summary, the demand for high-quality, structured nursing corpus in the nursing field has increased significantly. Although foreign countries have accumulated mature experience in nursing knowledge graph and corpus construction, China's nursing field still lacks an automated tool that can efficiently integrate nursing practice needs and knowledge systems. Therefore, it is particularly necessary to develop an automated nursing corpus construction method based on a knowledge graph. SUMMARY

[0006] The present application aims at overcoming the deficiencies of the prior art, and provides a nursing corpus automatic creation method based on a knowledge graph, so as to realize the automatic construction and continuous updating of nursing data resources of the nursing corpus.

[0007] The present application aims at overcoming the deficiencies of the prior art, and provides a nursing corpus automatic creation method based on a knowledge graph, so as to realize the automatic construction and continuous updating of nursing data resources of the nursing corpus.

[0008] A nursing corpus automatic creation method based on a knowledge graph comprises the following steps:

[0009] A multi-dimensional data acquisition stage: data is extracted from professional literature databases, electronic medical record systems and teaching materials and books, data preprocessing operations are performed, and intelligent analysis and entity recognition are performed on the preprocessed data;

[0010] A knowledge graph construction stage: a nursing field ontology is constructed, including extracting and defining entity core concepts, relationships and attributes, mapping the extracted entities, relationships and attributes to the constructed ontology through knowledge representation learning, and constructing and storing a knowledge graph based on entity and relationship vectors;

[0011] A corpus automatic generation and updating stage: corpus automatic generation, including: designing a corpus generation template and decomposing it into key elements, mapping the key elements to the entities and relationships in the knowledge graph, generating new nursing cases and simulated scenarios according to the entities and relationships, and filling in structured data according to entity and attribute information; corpus automatic updating, based on information entropy recognition and updating of low information content or outdated information in the corpus;

[0012] A corpus quality control and evaluation stage: a hierarchical expert review system and quality evaluation model are established, key nodes in the graph are generated, and the latest research is combined for evaluation and reevaluation; active learning strategies and migration techniques are used to automatically correct the corpus.

[0013] Further, the multi-dimensional data acquisition stage specifically comprises:

[0014] In the form of a multi-node cooperative mode crawler system, the abstract, the main text and the conclusion information in the professional literature database are automatically extracted through an API interface or a page parsing-based mode; the nursing records, diagnosis information and treatment plans in the medical records are automatically extracted from the hospital information system and the patient electronic medical record system by accessing an authorized API interface, and the patient personal information is desensitized; the scanned images of paper teaching materials and books are converted into editable text by using scanning and optical character recognition technology, and the digital teaching material book content is obtained from the data provider through platform authorization and API interface;

[0015] The collected text is subjected to data preprocessing, including data format conversion, data cleaning, data filtering and merging, and data denoising;

[0016] The pre-processed text is intelligently analyzed and entity-recognized by using a named entity recognition algorithm, including nursing operation terminology entity recognition and disease or symptom name entity recognition.

[0017] All data is standardized to a unified format, labeled using a keyword extraction algorithm, and stored in a database according to the label and data type.

[0018] Further, the data preprocessing of the collected text includes data format conversion, data cleaning, data filtering and merging, and data denoising, specifically including:

[0019] Data from different data sources is uniformly converted into a structured or semi-structured standard format.

[0020] Invalid or incomplete medical record data information is removed through data cleaning, including useless characters, blank content, and text with format errors.

[0021] For repeated literature, medical records, or other redundant information, multiple deduplication algorithms are used for screening and merging.

[0022] Data denoising is performed, i.e., removing meaningless punctuation marks, advertising information, or other useless content in the text.

[0023] Further, the knowledge graph construction phase specifically includes:

[0024] Constructing a nursing knowledge graph ontology: defining core concepts in the nursing field and using them as basic nodes in the knowledge graph, setting specific attributes and their relationships based on core concepts to form a logical chain; using ontology language to structurally describe the core concepts and relationships;

[0025] Building relationships between nursing knowledge graph ontologies: using deep learning models for knowledge representation learning, modeling simple relationships and multiple relationships between entities, and converting discrete symbols into continuous vector representations;

[0026] Generating and storing the knowledge graph: storing the trained entity and relationship vectors into a graph database, defining the query path of entities and relationships, setting up an index for common queries, and generating a nursing knowledge graph; using the built-in visualization tool of the graph database to generate a visual display of the nursing knowledge graph.

[0027] Further, the relationship construction between the nursing knowledge graph ontology: using deep learning models for knowledge representation learning, modeling simple relationships and multiple relationships between entities, and converting discrete symbols into continuous vector representations, specifically including:

[0028] Triplets are extracted from the knowledge graph as training samples to train and adjust the parameters of the TransE model, and the entities and relationships in the triplets are mapped to low-dimensional vectors to model simple relationships between entities, and the relationship structure is matched by translating entity vectors in a low-dimensional vector space;

[0029] The DistMult model is used to model multiple relationships through a bilinear model, realize multi-dimensional relationship expression of nursing, and optimize the vectorized relationship to make entity vectors with similar relationships closer;

[0030] The relationship of adjacent nodes is captured and the node representation is optimized through the graph convolution network, and the graph attention network is used to assign different attention weights to each node in the graph, so that the core node gets more attention.

[0031] Further, the automatic generation of the corpus specifically includes:

[0032] By analyzing the common cases and teaching material structure in the nursing field, the corpus template is decomposed into key elements and mapped to the entities and relationships in the knowledge graph to generate template formats for different scenarios; the corpus template includes a symptom description template and a complication handling template;

[0033] A generative adversarial network and a variational autoencoder are used respectively to generate new nursing cases and simulated scenarios through entities and relationships in the graph;

[0034] Based on the corpus generation template, structured data is automatically filled by extracting corresponding entity and attribute information from the knowledge graph through semantic analysis.

[0035] Further, the respective use of the generative adversarial network and the variational autoencoder to generate new nursing cases and simulated scenarios through entities and relationships in the graph specifically includes:

[0036] Based on the nursing knowledge graph, a generative adversarial network GAN is trained to generate nursing descriptions, and a nursing case is generated based on the nursing description;

[0037] A variational autoencoder VAE is used to generate variant sentences, and a new simulated scenario is generated using the variant sentences.

[0038] Further, the automatic updating of the corpus specifically includes:

[0039] The TF-IDF algorithm is used to calculate the information amount of each corpus, and the corpus with low information amount is marked as a replaceable item;

[0040] The graph convolution network and graph attention network model are used to automatically analyze and optimize the knowledge graph structure, adjust the outdated or newly added nursing knowledge, regularly update the knowledge graph, and regenerate the nursing corpus for the replaceable items.

[0041] The duplicate entries are cleaned by using a text deduplication algorithm, so as to ensure that the generated corpus has no duplication.

[0042] Further, the hierarchical expert review system and quality evaluation model are established, and the key nodes of the graph are generated, and the latest research is combined for evaluation and reevaluation, specifically including:

[0043] A hierarchical expert review system is established, a plurality of authoritative experts with 10 years or more of clinical experience and academic background are selected to form a review team, a plurality of review modes are set, including node review, batch review and random sampling; the system automatically distributes the key nodes in the graph establishment and updating process to a plurality of relevant experts for independent review and evaluation, and the review content specifically includes deep checking of node content, definition and logical relationship; the system automatically invites experts to reevaluate regularly according to new research, new technologies and new guidelines in the nursing field;

[0044] A quality evaluation model is established to perform multi-dimensional periodic evaluation on the corpus, including core nursing knowledge point coverage rate checking, low precision area problem diagnosis and quality evaluation report generation; the main indicators of the quality evaluation model include precision, recall and F1 score.

[0045] Further, the active learning strategy and migration technology are used to realize automatic correction of the corpus, specifically including:

[0046] An active learning algorithm model is introduced, a sample selection mechanism is adopted, high information gain or large uncertainty data are preferentially selected for manual labeling, and then the quality evaluation model is trained to gradually optimize the quality evaluation model;

[0047] The migration learning technology is applied, the knowledge similarity detection and the category-based differential migration strategy are introduced, the verified knowledge in the existing corpus is migrated to the new corpus verification process, the verification of the corpus is completed, and the corpus is automatically corrected according to the verification result.

[0048] The beneficial effects of the present application are:

[0049] (1) Improve the construction efficiency and update speed

[0050] With the aid of the automatic creation process, the corpus can quickly collect, organize and generate the knowledge content required in the nursing field, greatly shortening the construction time of the corpus. At the same time, the automatic updating mechanism enables the corpus to integrate the latest research and practical achievements in the nursing industry in real time, ensuring that the corpus always keeps pace with the times. Compared with the traditional manual updating method, the work efficiency is significantly improved, and the corpus can adapt to the changes of new requirements and knowledge in a short time.

[0051] (2) Reducing construction and maintenance costs

[0052] Through the automatic corpus generation and updating algorithm, the invention reduces the workload of manual annotation and review, and reduces the input cost of human resources. Using machine learning and deep learning technology, the system can automatically identify, classify and update content, greatly reducing the need for human review, thereby reducing the long-term maintenance cost of the corpus, making it more economical.

[0053] (3) Double quality assurance

[0054] The quality of the corpus is double-protected by the expert review system and the quality evaluation model. The expert review system verifies the key nodes and relationships in the graph manually to ensure the accuracy and authority of the corpus content; the quality evaluation model periodically detects the comprehensiveness and timeliness of the corpus through precision, recall rate, F1 score and other comprehensive indicators, and timely discovers and corrects potential problems. This double protection mechanism significantly improves the reliability of the corpus and its applicability in nursing teaching, research and practice.

[0055] (4) Improve the diversity and innovation of the corpus

[0056] Through deep learning models and transfer learning technology, the invention continuously introduces new nursing knowledge and diverse information structures into the corpus, increasing the diversity of corpus content. The application of transfer learning technology enables verified knowledge to quickly adapt and expand to new nursing contexts, reducing repetitive work while improving content innovation to meet changing practical needs.

[0057] (5) Enhance practicality and application value

[0058] Based on automated creation, updating and quality assurance, the corpus constructed by the invention has higher practicality. The corpus can provide timely and accurate knowledge support for nursing practitioners and researchers, and effectively assist nursing education and training. With the help of transfer learning and deep learning models, the content of the corpus can be flexibly applied to different nursing scenarios and cases, improving its application and transformation value in nursing teaching and clinical practice. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of a nursing corpus automatic creation method based on knowledge graph provided by the invention. DETAILED DESCRIPTION

[0060] The technical solutions of the present application will be described clearly and completely below in connection with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] Referring to Figure 1 The present application provides a technical solution:

[0062] A nursing corpus automatic creation method based on a knowledge graph, comprising:

[0063] S1, a multi-dimensional data acquisition stage: data is captured from professional literature databases, electronic medical record systems, and teaching materials and books, data preprocessing operations are performed, and intelligent analysis and entity recognition are performed on the preprocessed data.

[0064] In a specific embodiment, the multi-dimensional data acquisition stage specifically comprises:

[0065] S11, in the form of a multi-node cooperative mode crawler system, through an API interface or a page parsing-based mode, abstracts, main text, and conclusion information in professional literature databases are automatically captured; wherein the professional literature databases include but are not limited to PubMed, Wanfang, Weipu, and Dingxiang Garden, etc.

[0066] By accessing an authorized API interface, nursing records, diagnosis information, and treatment plans in medical records are automatically extracted from a hospital information system (HIS) and a patient electronic medical record system (EMR), and patient personal information is desensitized;

[0067] Paper teaching materials and books are scanned and converted into editable text using scanning and optical character recognition technology, and digital teaching material content is obtained from data providers through platform authorization using an API interface. For example, standardized nursing operation processes and nursing methods are obtained from relevant nursing professional books and operation manuals, guidelines, etc.

[0068] The multi-dimensional data acquisition stage analyzes the structure of data from different sources, and is divided into structured data (such as database records) and unstructured data (such as PDF documents), and suitable data extraction methods are determined for data of different structures.

[0069] S12, data preprocessing is performed on the collected text, including data format conversion, data cleaning, data filtering and merging, and data denoising, specifically including:

[0070] S121, data from different data sources is uniformly converted into a structured or semi-structured standard format, such as JSON, XML, CSV, etc.

[0071] S122, removing invalid or incomplete medical record data information through data cleaning, including useless characters, blank content and format error text;

[0072] S123, for repeated literature, medical records or other redundant information, a variety of deduplication algorithm is used for screening and merging;

[0073] S124, data denoising processing, that is, removing meaningless punctuation marks, advertising information or other useless content in the text.

[0074] Through data preprocessing, the availability of data can be improved, ensuring that the information in the corpus is unique and has no repetition, and the purity of the data, laying a foundation for subsequent corpus construction and knowledge graph generation.

[0075] S13, using named entity recognition algorithm (NER) to intelligently analyze and recognize entities in the preprocessed text, including nursing operation term entity recognition and disease or symptom name entity recognition.

[0076] The entity recognition process mainly includes the recognition of nursing operation entities, and the system automatically detects and extracts specific nursing operation terms such as "intravenous infusion", "gastric lavage", "indwelling catheterization", "electrocardiogram monitoring", etc.; and the recognition of disease name entities, through the recognition of disease or symptom names involved in the text, such as "pneumonia", "cirrhosis", "hypertension", etc., to ensure that important medical information is accurately extracted.

[0077] S14, standardizing all data to a unified format, using keyword extraction algorithm to labelize the data content, and storing the preprocessed data in the database according to the label and data type.

[0078] For example, standardize the time format and unit of all data to a unified format, use keyword extraction algorithm to labelize the data content, such as classify and store according to "symptom", "nursing measure", "treatment method" and other labels, to ensure the efficiency of subsequent processing. Then, according to different data types, store the data in corresponding relational database (such as MySQL for structured data) or document database (such as MongoDB for unstructured data), and set up data backup regularly, to ensure data security, support database backup and fast recovery function.

[0079] S2, knowledge graph construction phase: constructing the ontology of nursing field, including extracting and defining entity core concepts, relationships and attributes, mapping the extracted entities, relationships and attributes to the constructed ontology through knowledge representation learning, and constructing and storing the knowledge graph based on entity and relationship vectors.

[0080] In one specific embodiment, the knowledge graph construction phase specifically includes:

[0081] S21, constructing a nursing knowledge graph ontology: defining core concepts in the field of nursing and taking them as basic nodes in the knowledge graph, setting specific attributes and their mutual relationships based on the core concepts to form a logical chain; using ontology language to structurally describe the core concepts and relationships can ensure the accuracy of data association and query.

[0082] In one example, it specifically includes:

[0083] (1) Through literature analysis and multiple rounds of expert consultation, list the core concepts and secondary concepts in the field of nursing. Core concepts include "disease", "symptom", "nursing measure", "effect", etc. The secondary concepts of "nursing measure" can include "monitoring", "rehabilitation", "prevention", etc.

[0084] (2) Specify attributes and types for each concept. The attributes of "nursing measure" include "implementation time", "applicable symptoms", etc.

[0085] (3) Define the relationships between core concepts to form a logical chain, so that the logical association of different elements in the nursing process can be intuitively displayed. For example, "disease-symptom" association, such as hypertension-headache, "nursing measure-effect" association, such as drug therapy-blood pressure reduction, etc.

[0086] (4) Use ontology language (such as OWL, Web Ontology Language) to structurally describe concepts and relationships, ensuring the consistency and logic of concept expression, so that the nursing knowledge graph can be standardized in computer systems and stored, providing a foundation for further processing.

[0087] S22, constructing the relationship between nursing knowledge graph ontologies: using deep learning models for knowledge representation learning, modeling simple relationships and multiple relationships between entities, and converting discrete symbols into continuous vector representations; specifically, it includes:

[0088] S221, extracting triples (such as <fever, applicable, cooling care>) from the knowledge graph as training samples, training and adjusting the TransE model, mapping entities and relationships in triples to low-dimensional vectors for simple relationship modeling between entities, and matching relationship structures by translating entity vectors in low-dimensional vector space; adjusting model parameters such as learning rate and vector dimension can ensure that the trained vectors can effectively represent the semantic association between concepts.

[0089] S222, multi-relation modeling is performed by a bilinear model using the DistMult model to express multi-dimensional relations of nursing (such as "nursing implementation-effect-nursing evaluation"), and the vectorized relations are optimized so that entity vectors with similar relations are closer, thereby improving the query effect of the knowledge graph.

[0090] S223, the relations of adjacent nodes are captured by a graph convolution network, and the node representation is optimized, and a graph attention network is used to assign different attention weights to each node in the graph, so that the core nodes are given higher attention, the graph has a higher priority in the retrieval and reasoning process, and accurate modeling is achieved.

[0091] S23, generating and storing the knowledge graph: storing the trained entity and relation vectors into a graph database (such as Neo4j), and defining the query path of the entity and the relation, setting the index for the common query, generating the nursing knowledge graph; using the visualization tool built-in the graph database to generate the visualization display of the nursing knowledge graph. Index setting can improve the graph query speed, and visualization display provides support for subsequent checking and verifying the graph structure.

[0092] S3, automatic generation and updating of the corpus: automatic generation of the corpus, including: designing a corpus generation template and decomposing it into key elements, mapping the key elements to the entities and relations in the knowledge graph, generating new nursing cases and simulated scenarios according to the entities and relations, and filling in the structured data according to the entity and attribute information; automatic updating of the corpus, identifying and updating the low information amount or outdated information in the corpus based on the information entropy of the corpus;

[0093] Among them, the entities in the knowledge graph include symptoms, nursing intervention plans, treatment methods, etc., and the relations between entities include causal relations, dependency relations, etc. The generated corpus will include multiple case scenarios, increasing the diversity of cases and providing rich materials for the application of the corpus.

[0094] In a specific embodiment, the automatic generation of the corpus specifically includes:

[0095] S311, by analyzing the structure of common cases and teaching materials in the nursing field, the corpus template is decomposed into key elements and mapped to the entities and relations in the knowledge graph to generate a template format for different scenarios; the corpus template includes a symptom description template and a complication handling template.

[0096] For example: in the symptom description template, such as "the patient has a high fever symptom and needs to take ice bag cooling and other nursing measures"; in the complication handling template, "diabetic patients need to be alert to the occurrence of diabetic ketoacidosis, and the recommended nursing measures are timely fluid replacement".

[0097] S312, respectively, using generative adversarial networks and variational autoencoder, generating new nursing cases and simulated scenarios through entities and relationships in the graph; Specifically, it includes:

[0098] With the nursing knowledge graph as the condition, the generative adversarial network GAN is trained to generate nursing descriptions, and the nursing cases are generated according to the nursing descriptions. The generated description text is, for example, "The patient has symptoms of Staphylococcus aureus infection, and it is recommended to change the dressing and medicine every day."

[0099] The variational autoencoder VAE is used to generate variant sentences, and the new simulated scenarios are generated using the variant sentences.

[0100] S313, based on the corpus generation template, the corresponding entity and attribute information is extracted from the knowledge graph through semantic analysis to automatically fill in the structured data.

[0101] The corpus generation template is designed based on the entities and relationships in the knowledge graph to create a framework that can generate new corpus that meets the needs of a specific field, which can improve the diversity and coverage of the data.

[0102] In one specific embodiment, the corpus automatic updating specifically includes:

[0103] S321, the information amount of each corpus is calculated using the TF-IDF algorithm, and the corpus with low information amount is marked as a replaceable item;

[0104] S322, the knowledge graph structure is automatically analyzed and optimized using a graph convolution network and a graph attention network model, the outdated or newly added nursing knowledge is adjusted, the knowledge graph is updated regularly, and the replaceable item is regenerated as a nursing corpus;

[0105] S323, the duplicate items are cleaned up using a text deduplication algorithm to ensure that the generated corpus has no duplicates.

[0106] To identify and update the low information amount or outdated information in the corpus, the application designs an information entropy-based analysis algorithm to identify the nodes with low information amount or sparse data by evaluating the information entropy of each node in the graph, and to apply corpus update to these nodes first to enrich and optimize the graph content. The automatic updating and timeliness maintenance of the corpus ensure that the knowledge graph content is accurate, up-to-date, and can reflect the changes in the field in a timely manner.

[0107] S4, corpus quality control and evaluation stage: a hierarchical expert review system and quality evaluation model are established to evaluate and reevaluate the key nodes in the graph in combination with the latest research; active learning strategies and migration techniques are used to automatically correct the corpus.

[0108] Further, the hierarchical expert review system and quality evaluation model are established to evaluate and re-evaluate the key nodes in the atlas generation combined with the latest research, which specifically includes:

[0109] A hierarchical expert review system is established, and a number of authoritative experts with more than 10 years of clinical experience and academic background are selected to form a review team. Multiple review modes are set, including node review, batch review, and random sampling. The system automatically assigns key nodes in the atlas establishment and update process to multiple relevant experts for independent review and evaluation. The review content specifically includes in-depth checking of node content, definition, and logical relationship. The system automatically invites experts to re-evaluate regularly based on new research, new technologies, and new guidelines in the nursing field;

[0110] A quality evaluation model is established to conduct multi-dimensional periodic evaluation of the corpus, including core nursing knowledge point coverage check, low precision area problem diagnosis, and quality evaluation report generation. The main indicators of the quality evaluation model include precision, recall, and F1 score.

[0111] Coverage analysis is conducted on common nursing situations, symptoms, diagnoses, and treatment plans, etc. to detect areas that are not included or lack information. The evaluation model ensures that core nursing knowledge points are fully covered through coverage check to optimize the comprehensiveness of the corpus. The model diagnoses problems in low precision areas, identifies possible causes of errors such as data missing and classification errors, and helps the development team to make precise corrections. The quality evaluation report generated through regular evaluation cycles will cover the trends of various indicators of the corpus and be visualized through charts for easy understanding by the development team and relevant personnel. Through regular detection and feedback mechanisms, the research and development team can timely discover potential problems and make adjustments to ensure the stability and scientificity of the corpus content.

[0112] Further, the active learning strategy and transfer technology are used to realize automatic correction of the corpus, which specifically includes:

[0113] An active learning algorithm model is introduced, and a sample selection mechanism is adopted to preferentially select data with high information gain or high uncertainty for manual labeling. Then the quality evaluation model is trained to gradually optimize the quality evaluation model. Through active learning, the dependence on a large number of labeled samples is reduced, and model quality is improved with a small amount of manual verification, reducing manual workload and cost while ensuring the accuracy and authority of the corpus, gradually improving its operability and reliability in practical applications.

[0114] By applying transfer learning techniques, introducing knowledge similarity detection and category-based differential transfer strategies, the verified knowledge in existing corpora is transferred to the new corpus validation process, completing the corpus validation, and the corpus is automatically corrected according to the validation results. Transfer learning can effectively improve the validation efficiency, especially in new knowledge fields, which can significantly reduce the time cost of initial manual validation. Through this technology, the corpus validation process can achieve efficient updating while ensuring accuracy, which helps to quickly adapt to new nursing practice needs.

[0115] The main differences between the present invention and the prior art are as follows:

[0116] 1. Automatic creation and update mechanism: The present invention uses a knowledge graph-driven corpus automatic generation method, which realizes the dynamic generation and real-time update of the corpus by constructing the ontology structure of the nursing field, the corpus generation template and the update algorithm. In contrast, the existing technology usually relies on manual or semi-automatic methods to create and update the corpus, which is low in efficiency and difficult to quickly adapt to changes in domain knowledge.

[0117] 2. Structured corpus generation based on knowledge graph: The present invention can generate diversified corpus that meets the needs of the nursing field by using structured knowledge graph and corpus generation template, combined with GAN or VAE generation model. The existing technology is mostly unstructured data, lacking systematic template design and knowledge graph support, and the generated corpus content is easily influenced by human factors and has low consistency.

[0118] 3. Intelligent corpus update and system maintenance: The present invention identifies information with low information density or outdated content through information entropy or TF-IDF algorithm, and regularly updates the corpus to ensure the timeliness of the content. This intelligent update mechanism improves the applicability and timeliness of the corpus, so that the generated content can always meet the latest nursing guidelines and standards. The existing technology lacks automatic timeliness maintenance means, and the update frequency is not high and the content is easily outdated.

[0119] 4. Domain-specific knowledge representation and relationship modeling: Through ontology construction of knowledge graph and relationship modeling based on TransE or DistMult algorithm, the present invention can structure the core concepts, attributes and their associations in the nursing field, realize accurate association and hierarchical of corpus content. The corpus of the prior art usually does not model and structure the domain knowledge specifically, and it is difficult to reflect the professional concept association in the nursing field, and the accuracy and logic of the corpus are relatively insufficient.

[0120] 5. Extensibility and applicability: Since the present invention uses knowledge graph as a structured representation tool, the corpus is highly extensible and adaptable to different nursing scenarios and user needs. The existing technology lacks deep adaptation to nursing scenarios, and has limited extensibility and application scenarios.

[0121] The foregoing is considered as illustrative only of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and operation described. Accordingly, all such variations are intended to be included within the scope of the present application as defined in the claims below and their equivalents.

Claims

1. A method for automatically creating a nursing corpus based on knowledge graph, characterized in that: include: Multi-dimensional data collection stage: Capture data from professional literature databases, electronic medical record systems, and textbooks, perform data preprocessing operations, and conduct intelligent analysis and entity recognition on the preprocessed data; Knowledge graph construction phase: Constructing the nursing domain ontology, including extracting and defining the core concepts, relationships, and attributes of entities, mapping the extracted entities, relationships, and attributes to the constructed ontology through knowledge representation learning, and constructing and storing the knowledge graph based on entity and relationship vectors; Corpus automatic generation and update stage: Corpus automatic generation, including: designing corpus generation templates and breaking them down into key elements, mapping key elements to entities and relationships in the knowledge graph, generating new nursing cases and simulation scenarios based on entities and relationships, and filling structured data based on entity and attribute information; Corpus automatic update, identifying and updating low-information or outdated information in the corpus based on the information entropy of the corpus. The entropy-based analysis algorithm, by evaluating the information entropy of each node in the graph, identifies nodes with low information content or sparse data, and preferentially applies corpus updates to these nodes to enrich and optimize the graph content; Corpus quality control and evaluation stage: Establish a hierarchical expert review system and quality assessment model, and the system will establish a map The key nodes in the updating process are automatically assigned to multiple relevant experts for independent review and evaluation. When the graph generates key nodes, they are evaluated and re-evaluated in combination with the latest research. Active learning strategies and transfer technologies are used to achieve automatic correction of the corpus.

2. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 1, characterized in that: The multi-dimensional data collection stage specifically includes: Using a multi-node collaborative crawler system, the system automatically captures abstracts, main texts, and conclusions from professional literature databases through API interfaces or page-based parsing. It automatically extracts nursing records, diagnostic information, and treatment plans from hospital information systems and patient electronic medical record systems by accessing authorized API interfaces, while desensitizing patients' personal identity information. It uses scanning and optical character recognition technology to convert scanned images of paper textbooks into editable text, and obtains digital textbook content from data providers through API interfaces authorized by the platform. Perform data preprocessing on the collected text, including data format conversion, data cleaning, data screening and merging, and data denoising; Use named entity recognition algorithms to perform intelligent analysis and entity recognition on the preprocessed text, including entity recognition of nursing operation terms and entity recognition of disease or symptom names; All data are standardized into a unified format, the data content is labeled using a keyword extraction algorithm, and the pre-processed data is classified and stored in the database according to the label and data type.

3. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 2, characterized in that: The data preprocessing of the collected text includes: data format conversion, data cleaning, data screening and merging, and data denoising, specifically including: Convert data from different data sources into a standard structured or semi-structured format; Data cleaning is used to remove invalid or incomplete medical record data information, including useless characters, blank content, and incorrectly formatted text; Multiple deduplication algorithms are used to screen and merge duplicate documents, medical records, or other redundant information; Perform data denoising, that is, remove meaningless punctuation, advertising information or other useless content in the text.

4. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 1, characterized in that: The knowledge graph construction stage specifically includes: Constructing a nursing knowledge graph ontology: defining core concepts in the nursing field and using them as basic nodes in the knowledge graph. Setting specific attributes and their relationships based on the core concepts to form a logical chain. Using ontology language to provide a structured description of the core concepts and relationships. Constructing the relationship between nursing knowledge graphs and ontologies: Using deep learning models to learn knowledge representation, modeling simple and multi-relationships between entities, and converting discrete symbols into continuous vector representations; Generate and store the knowledge graph: store the trained entity and relationship vectors in the graph database, define the query path of entities and relationships, set indexes for common queries, and generate a nursing knowledge graph; use the built-in visualization tools of the graph database to generate a visual display of the nursing knowledge graph.

5. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 4 is characterized in that: The construction of the relationship between the nursing knowledge graph ontology: using the deep learning model to learn knowledge representation, perform simple and multi-relationship modeling between entities, and convert discrete symbols into continuous vector representations, specifically including: Extract triplets from the knowledge graph as training samples, train the TransE model and adjust its parameters. Map the entities and relationships in the triplets into low-dimensional vectors to model simple relationships between entities. Match the relationship structure by translating the entity vectors in the low-dimensional vector space. The DistMult model is used to perform multi-relationship modeling through a bilinear model to achieve the expression of multi-dimensional nursing relationships. At the same time, the vectorized relationships are optimized to make the entity vectors with similar relationships closer. The graph convolutional network is used to capture the relationship between neighboring nodes and optimize node representation. The graph attention network is used to assign different attention weights to each node in the graph, so that core nodes receive higher attention.

6. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 1, characterized in that: The automatic generation of the corpus specifically includes: By analyzing common cases and teaching material structures in the nursing field, the corpus templates are decomposed into key elements and connected with the entity and relationship mapping in the knowledge graph to generate template formats for different scenarios; the corpus templates include symptom description templates and complication treatment templates; Generate new nursing cases and simulated scenarios using entities and relationships in the graph using generative adversarial networks and variational autoencoders, respectively; Based on the corpus generation template, corresponding entities and attribute information are extracted from the knowledge graph through semantic analysis to automatically fill in structured data.

7. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 6, characterized in that: The method uses generative adversarial networks and variational autoencoders to generate new nursing cases and simulation scenarios through entities and relationships in the graph, including: Based on the nursing knowledge graph, the generative adversarial network (GAN) is trained to generate nursing descriptions, and then nursing cases are generated based on the nursing descriptions. Use variational autoencoder (VAE) to generate mutated sentences, and use the mutated sentences to generate new simulated scenarios.

8. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 1, characterized in that: The automatic updating of the corpus specifically includes: The TF-IDF algorithm is used to calculate the information content of each corpus, and the corpus with lower information content is marked as replaceable items; Utilize graph convolutional networks and graph attention network models to automatically analyze and optimize the knowledge graph structure, adjust outdated or newly added nursing knowledge, regularly update the knowledge graph, and regenerate nursing corpus for replaced entries; Use text deduplication algorithms to clean up duplicate entries and ensure that the generated corpus is free of duplication.

9. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 1, characterized in that: The aforementioned establishment of a hierarchical expert review system and quality assessment model will evaluate and re-evaluate key nodes of the graph generation process in combination with the latest research, specifically including: A hierarchical expert review system was established, with a review team composed of authoritative experts with at least 10 years of clinical experience and academic background. Multiple review modes were set up, including node review, batch review, and random sampling. The system automatically assigned key nodes in the map creation and update process to multiple relevant experts for independent review and evaluation. The review content specifically included in-depth inspection of node content, definitions, and logical relationships. The system automatically invited experts to conduct re-evaluations regularly based on new research, new technologies, and new guidelines in the nursing field. A quality assessment model is established to conduct regular multi-dimensional assessments of the corpus, including coverage checks on core nursing knowledge points, problem diagnosis in low-precision areas, and generation of quality assessment reports. The main indicators of the quality assessment model include precision, recall, and F1 score.

10. The method for automatically creating a nursing corpus based on a knowledge graph according to claim 9, characterized in that: The active learning strategy and transfer technology are used to achieve automatic corpus correction, specifically including: An active learning algorithm model is introduced, and a sample selection mechanism is adopted to give priority to data with high information gain or large uncertainty for manual labeling, thereby training the quality assessment model and gradually optimizing the quality assessment model. By applying transfer learning technology, introducing knowledge similarity detection and category-based differentiated transfer strategy, the verified knowledge in the existing corpus is transferred to the new corpus verification process, the corpus verification is completed, and the corpus is automatically corrected according to the verification results.

Images