A method for constructing digital human knowledge graph for the living inheritance of experience of famous traditional Chinese medicine practitioners

Abstract

The present invention relates to the field of knowledge graph technology, specifically to a method for constructing a digital human knowledge graph for the living inheritance of the experience of famous old Chinese medicine practitioners, comprising the following steps: obtaining Chinese medicine literature data, disassembling the literature sentence by sentence, cutting continuous sentences into independent vocabulary units, marking the grammatical role of each vocabulary in the sentence, and identifying multiple entities in the text by analyzing the combination rules between adjacent vocabulary, and generating a document entity set. In the present invention, through grammatical annotation and context pattern analysis, the accuracy and coverage of entity extraction are improved, the co-occurrence frequency and modification relationship derivation are combined, the interpretability and semantic network dimension of entity association are optimized, dynamic weight calculation and attribute fusion are performed, the adaptive optimization ability of the knowledge graph is realized, the matching degree calculation and closed-loop verification mechanism are implemented, the accuracy of clinical contraindication identification is improved, the structured terminology system and the dynamic evolution mechanism are constructed, and a Chinese medicine knowledge system that takes into account deep semantic expression and continuous evolution is constructed.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and in particular to a method for constructing a digital human knowledge graph for the living inheritance of the experience of famous traditional Chinese medicine practitioners. Background Art

[0002] The field of knowledge graph technology includes multiple fields such as computer science, artificial intelligence and data management. It involves representing and processing complex knowledge systems in a graphical way. The core content includes knowledge acquisition, modeling, storage, query, reasoning and visualization. The knowledge graph is composed of nodes and edges, that is, the relationships between entities. It effectively organizes and manages large amounts of data and reveals the intrinsic connections between information through semantic relationships. Application scenarios include intelligent question and answer, recommendation systems and semantic search. In the process of constructing the knowledge graph, various technical means are involved, such as data extraction, knowledge fusion, graph optimization and updating, to support more intelligent and accurate information processing.

[0003] Among them, the method for constructing a digital human knowledge graph for the living inheritance of the experience of famous old Chinese medicine practitioners refers to modeling the medical knowledge and experience of famous old Chinese medicine practitioners through digital means to build a proprietary knowledge graph. The technical matters targeted by the patent subject include extracting knowledge from multiple aspects of the clinical experience, treatment methods, and drug use of famous old Chinese medicine practitioners, and organizing and storing them in the form of a graph. Specifically, natural language processing technology is used to analyze the medical records, literature materials, and various text data of famous old Chinese medicine practitioners to identify the entities and relationships therein, and knowledge extraction and graph construction algorithms are used to present knowledge points in a structured manner, and comprehensive modeling is carried out in combination with the multi-dimensional information in the process of Chinese medicine diagnosis and treatment, including diseases, symptoms, and treatment plans, to construct a knowledge graph, and combined with the update and maintenance of the graph to ensure continuous updating and improvement in the process of living inheritance, providing a more systematic knowledge management and experience inheritance platform for the field of Chinese medicine.

[0004] Existing knowledge graph construction methods have the problem of insufficient integration of multi-source heterogeneous data in the application of traditional Chinese medicine. The split processing mode of text analysis and speech recognition makes it difficult to form an effective knowledge complementation mechanism. There is a lack of in-depth mining of grammatical structure in the document parsing process. Simply relying on word frequency statistics tends to ignore contextual semantic associations, resulting in deviations in entity relationship extraction. Traditional entity recognition methods have not established a cross-document co-occurrence frequency analysis mechanism and cannot effectively capture the complex network relationships between symptoms, drugs, and treatment methods in the process of traditional Chinese medicine diagnosis and treatment. The static graph construction model does not provide sufficient support for the attribute fusion of synonymous entities and is difficult to adapt to the terminology in the traditional Chinese medicine knowledge system. The diverse characteristics of oral experience lead to knowledge redundancy and fragmentation, insufficient consideration of the dynamic characteristics of physician experience inheritance, and no closed-loop feedback mechanism has been formed. Knowledge updating relies on manual intervention and has poor timeliness. Simply relying on literature data can easily miss implicit knowledge in oral experience. The lack of cross-validation of voice entities and text entities may lead to the loss of key treatment plans. The derivation of entity relationships does not quantify the correlation strength index, which affects the accuracy of diagnosis and treatment decisions. It is impossible to effectively identify key clinical information such as drug compatibility contraindications. These defects restrict the actual application value of knowledge graphs in the inheritance of TCM experience and make it difficult to support a dynamic and systematic knowledge management system. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings existing in the prior art and propose a method for constructing a digital human knowledge graph for the living inheritance of the experience of famous old Chinese medicine practitioners.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: a method for constructing a digital human knowledge graph for the living inheritance of the experience of famous traditional Chinese medicine practitioners, comprising the following steps:

[0007] S1: Obtain TCM literature data, disassemble the literature sentence by sentence, cut continuous sentences into independent vocabulary units, mark the grammatical role of each word in the sentence, identify multiple entities in the text by analyzing the combination rules between adjacent words, and generate a literature entity set;

[0008] S2: Based on the document entity set, the oral speech of the famous traditional Chinese medicine practitioner is obtained, the speech signal is divided into multiple segments and converted into a text symbol sequence, and multiple entities in the speech are extracted by comparing the collocation rules of adjacent words in the text sequence and comparing multiple entities in the document, and a speech recognition entity list is output;

[0009] S3: calling the speech recognition entity list, performing spatial mapping on the entity set of the text and speech sources, analyzing the co-occurrence frequency of entities across documents, identifying the mutual relationship between multiple entities by analyzing the grammatical structure at the sentence level, and generating an entity relationship derivation graph;

[0010] S4: Using the entity relationship derivation graph, entities are converted into nodes and relationships are converted into edges. The association strength is calculated based on the connection frequency between nodes, and the attributes of synonymous entity nodes are fused. Combined with physician feedback, the connection relationship is updated in real time to output a traditional Chinese medicine knowledge graph.

[0011] As a further solution of the present invention, the document entity set specifically includes sentence decomposition results, vocabulary unit extraction results, and medical entity data sets; the speech recognition entity list specifically includes oral speech collection results, text conversion records, and speech entity sets; the entity relationship derivation graph specifically includes entity co-occurrence relationships, before and after modification relationships, and relationships between entities; the traditional Chinese medicine knowledge graph includes entity node conversion results, relationship edge conversion records, and synonymous entity attribute fusion.

[0012] As a further solution of the present invention, the process of obtaining the document entity set is specifically as follows:

[0013] S111: Obtain TCM literature data, disassemble the literature content sentence by sentence, analyze the sentence structure, extract vocabulary units, perform part-of-speech tagging and syntactic role assignment for each vocabulary, identify various types of vocabulary, including nouns, verbs, and adjectives, and establish syntactic structure analysis results;

[0014] S112: Based on the syntactic structure analysis result, the collocation pattern between adjacent words is analyzed, and the stability of the word combination in the sentence is calculated using the formula:

[0015] ;

[0016] Calculate the association strength of word pairs, and obtain a set of high-frequency word combinations by counting the co-occurrence frequencies between multiple words;

[0017] in, represents the association strength of the word combination, Representative words The co-occurrence frequency of Representative words The weight in the syntactic structure, Represents the number of words involved in the calculation, An index of vocabulary;

[0018] S113: calling the high-frequency word combination set, comparing with the existing medical terminology library, identifying medical entities in the text, including symptoms, drugs, treatment methods, causes, counting the distribution of multiple types of entities in the document, and establishing a document entity set.

[0019] As a further solution of the present invention, the process of obtaining the speech recognition entity list is specifically as follows:

[0020] S211: Based on the document entity set, the oral speech of the famous traditional Chinese medicine practitioner is obtained, and the audio signal is preprocessed, including sampling rate adjustment, noise removal and speech enhancement, and the processed audio signal is divided into multiple segments. The speech is converted into time and frequency using short-time Fourier transform, and time domain and frequency domain features are extracted. The speech is denoised to obtain a speech segmentation result;

[0021] S212: Based on the speech segmentation result, convert the audio data into a text symbol sequence to obtain speech text data;

[0022] S213: calling the speech text data, combining multiple entities of the document, comparing the collocation rules of adjacent words in the text, and calculating the similarity between the candidate words and the known medical entities, using the formula:

[0023] ;

[0024] Calculate the matching credibility of each word in the speech text, evaluate the meaning consistency of multiple medical entity words in the speech text of famous traditional Chinese medicine practitioners and literature, detect and extract medical terms in the speech text, and obtain the speech recognition entity list;

[0025] in, Represents the matching confidence of text words, Representing text words The context weight of Represents the semantic matching degree between the word and the medical entity. represents the phoneme matching probability of the vocabulary, represents the contextual deformation error of the word, Represents the number of valid words in the text, The index of the vocabulary in the speech text.

[0026] As a further solution of the present invention, the process of obtaining the entity relationship derivation graph is specifically as follows:

[0027] S311: calling the speech recognition entity list, comparing the entity sets of the text and speech sources, calculating the distribution of the same entity in multiple sources, analyzing the occurrence frequency of multiple entities in the speech text and the literature text, and obtaining a cross-text entity distribution matrix;

[0028] S312: Based on the cross-text entity distribution matrix, the co-occurrence frequencies of entities from multiple text sources are calculated, the co-occurrence times of the target entity in multiple documents are analyzed, and the association strength of entity pairs is calculated using the formula:

[0029] ;

[0030] Calculate the relevance of entities in multiple documents by calculating the association strength of each entity pair, using rows and columns to represent the two entities in the entity pair, and generate a cross-document entity co-occurrence matrix;

[0031] in, Represents the entity association strength, Representing Entity Normalized frequency of occurrence in the spoken text, Representing Entity The normalized frequency of occurrence in the document text, and are the entity means of the two text sources, Represents the number of entities counted, is the index of the entity;

[0032] S313: calling the cross-document entity co-occurrence matrix, parsing the grammatical structure of the sentence, analyzing the modification relationship and semantic dependency path between entities, evaluating the interaction between multiple entities in the sentence structure, and constructing an entity relationship derivation graph.

[0033] As a further solution of the present invention, the acquisition process of the TCM knowledge graph is specifically as follows:

[0034] S411: calling the entity relationship derivation graph, converting multiple medical entities into nodes of the graph, establishing edge connections according to the semantics, co-occurrence frequency, and dependency relationship between entities, calculating the connection strength of each node, and obtaining an entity network topology structure;

[0035] S412: Based on the entity network topology, the connection strength between multiple nodes is calculated, and synonymous entities are merged, and the formula is set as follows:

[0036] ;

[0037] Calculate the fusion adjustment factors between entities and generate a synonymous entity fusion matrix;

[0038] in, represents the fusion adjustment factor, Representing Entity The structural weights in the network, Representing Entity Frequency in the knowledge base, Represents the number of entities counted, is the index of the entity;

[0039] S413: Call the synonymous entity fusion matrix, and update the connection relationship between nodes in real time based on physician feedback to generate a traditional Chinese medicine knowledge graph.

[0040] As a further embodiment of the present invention, the method further comprises:

[0041] S5: calling the TCM knowledge graph, extracting a set of drug entities associated with multiple disease entities, calculating the matching degree between multiple drugs and diseases by comparing the drug entity attribute data with the disease entity feature data, identifying contraindicated drugs associated with the target disease, and generating a set of drug and disease relationship;

[0042] The drug and symptom relationship set includes a set of drug entities associated with symptoms, a drug and symptom matching degree, and a contraindicated drug identification result.

[0043] As a further solution of the present invention, the process of obtaining the drug and disease relationship set is specifically as follows:

[0044] S511: calling the TCM knowledge graph, extracting a set of drug entities related to multiple diseases, and obtaining a set of disease-related drugs;

[0045] S512: Based on the set of disease-associated drugs, the attribute data of each drug entity, including efficacy, meridians, and contraindications, and the characteristic data of the disease entity, including cause, symptoms, and physiological effects, are matched and analyzed using the formula:

[0046] ;

[0047] Calculate and obtain the drug matching value;

[0048] in, Represents drug matching degree, is the attribute vector of the disease characteristics, is the mean vector corresponding to the drug, is the covariance matrix of the attributes, is the inverse matrix of the covariance matrix, represents the transpose operation;

[0049] S513: Call the drug matching degree value, identify and record contraindicated drugs related to multiple diseases, and form a drug and disease relationship set.

[0050] Compared with the prior art, the advantages and positive effects of the present invention are:

[0051] In the present invention, by analyzing the combination rules between adjacent words, multiple entities in the text are identified and a document entity set is generated. In the present invention, the accuracy and coverage of entity extraction are improved through grammatical annotation and context pattern analysis, and the co-occurrence frequency and modification relationship derivation are combined to optimize the interpretability of entity association and semantic network dimension, dynamic weight calculation and attribute fusion, realize the adaptive optimization ability of knowledge graph, matching degree calculation and closed-loop verification mechanism, improve the accuracy of clinical contraindication identification, structure the terminology system and dynamic evolution mechanism, and build a Chinese medicine knowledge system that takes into account deep semantic expression and continuous evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0053] Figure 2 Obtaining a flow chart for the document entity set of the present invention;

[0054] Figure 3 A flow chart for obtaining a speech recognition entity list of the present invention;

[0055] Figure 4 Obtaining a flow chart for the entity relationship derivation diagram of the present invention;

[0056] Figure 5 Obtaining a flow chart for the TCM knowledge graph of the present invention;

[0057] Figure 6 A flow chart is obtained for the drug and condition relationship set of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0060] See also Figure 1The present invention provides a technical solution: a method for constructing a digital human knowledge graph for the living inheritance of the experience of famous old Chinese medicine practitioners, comprising the following steps:

[0061] S1: Obtain TCM literature data, disassemble the literature sentence by sentence, cut continuous sentences into independent vocabulary units, mark the grammatical role of each word in the sentence, and identify multiple entities in the text by analyzing the combination rules between adjacent words, including symptoms, drugs, treatment methods, and causes of disease. According to the frequency of repeated appearance of words in the paragraph and the contextual matching pattern, a corresponding relationship table for each category of medical terms is established to generate a document entity set;

[0062] S2: Based on the document entity set, the oral speech of famous traditional Chinese medicine practitioners is obtained, the speech signal is divided into multiple segments and converted into a text symbol sequence. By comparing the collocation rules of adjacent words in the text sequence and comparing multiple entities in the document, multiple entities in the speech are extracted and a speech recognition entity list is output;

[0063] S3: Call the speech recognition entity list, perform spatial mapping on the entity sets of text and speech sources, analyze the co-occurrence frequency of entities across documents, extract the pre- and post-modification relationships between entities in the sentence by analyzing the grammatical structure at the sentence level, identify the relationships between multiple entities, and generate an entity relationship derivation graph;

[0064] S4: Using the entity relationship derivation graph, entities are converted into nodes and relationships into edges. The association strength is calculated based on the connection frequency between nodes. The attributes of synonymous entity nodes are fused. Combined with physician feedback, the connection relationship is updated in real time to output the TCM knowledge graph.

[0065] S5: Call the TCM knowledge graph to extract a set of drug entities associated with multiple disease entities. By comparing the drug entity attribute data with the disease entity feature data, calculate the matching degree between multiple drugs and diseases, identify the contraindicated drugs associated with the target disease, and generate a set of drug and disease relationship.

[0066] The document entity set specifically includes sentence decomposition results, vocabulary unit extraction results, and medical entity data sets. The speech recognition entity list specifically includes oral speech collection results, text conversion records, and speech entity sets. The entity relationship derivation graph specifically includes entity co-occurrence relationships, before and after modification relationships, and relationships between entities. The traditional Chinese medicine knowledge graph includes entity node conversion results, relationship edge conversion records, and synonymous entity attribute fusion. The drug and disease relationship set includes the drug entity set associated with the disease, the drug and disease matching degree, and the contraindicated drug identification results.

[0067] See also Figure 2 , the specific steps for obtaining the document entity set are:

[0068] S111: Obtain traditional Chinese medicine literature data, disassemble the literature content sentence by sentence, analyze the sentence structure, extract lexical units, assign part-of-speech tags and syntactic roles to each word, identify various types of words, including nouns, verbs, adjectives, and establish the analysis results of the syntactic structure;

[0069] Obtain traditional Chinese medicine literature data, disassemble it based on the literature content, extract sentence-level text, and parse it according to morphological and syntactic rules. First, convert traditional Chinese medicine literature into standardized text data and perform sentence-level splitting on the literature content so that each sentence becomes an independent analysis unit. Subsequently, perform morphological analysis on each sentence, break down the basic lexical units therein, such as nouns, verbs, adjectives, etc., and mark the specific roles of these words in the sentence, including subject, predicate, object, etc. Through such analysis, the internal structure of the text can be understood more accurately. For example, in the sentence "Wind-cold colds can be relieved by ginger soup", "wind-cold colds" can be extracted as a disease entity, "ginger soup" as a drug entity, and "relieved" as a descriptive word related to the treatment method.

[0070] After completing the basic lexical parsing, the next step is to calculate the occurrence frequency of each word and perform normalization to measure the importance of each word in the overall literature. To calculate the normalized word frequency of a certain word, the following formula is used:

[0071] ;

[0072] where is the normalized word frequency of the word , is the number of occurrences of the word in the literature, is the total number of words counted. Suppose in a medical literature, the number of occurrences of "wind-cold colds" is , the number of occurrences of "ginger soup" is , the number of occurrences of "relieved" is , and the number of occurrences of "etiology" is .

[0073] Calculate the normalized word frequency of "wind-cold colds" as follows:

[0074] ;

[0075] The normalized word frequency of "ginger soup" is:

[0076] ;

[0077] where indicates that "wind-cold colds" is more important in the medical literature dataset and should be given priority in keyword screening.

[0078] S112: Based on the syntactic structure analysis results, analyze the collocation pattern between adjacent words and calculate the stability of the word combination in the sentence using the formula:

[0079] ;

[0080] Calculate the association strength of word pairs, and obtain a set of high-frequency word combinations by counting the co-occurrence frequencies between multiple words;

[0081] in, represents the association strength of the word combination, Representative words The co-occurrence frequency of Representative words The weight in the syntactic structure, Represents the number of words involved in the calculation, An index of vocabulary;

[0082] Based on the results of syntactic structure analysis, the collocation patterns between adjacent words are further calculated to determine which words have strong semantic associations. First, the co-occurrence frequency of word pairs is counted, that is, the frequency of appearance in the same document paragraph or adjacent sentences. Then, the semantic weight of the word pair is calculated, and the importance of different words is adjusted by normalization method. It is assumed that "cold and cold" and "ginger soup" appear in multiple sentences at the same time, which shows that the two have a strong collocation, while "relief" may be collocated with different medicines in different contexts, which requires further calculation of the collocation stability of each word.

[0083] In order to more accurately evaluate the strong correlation of word combinations, the following formula is used to calculate the correlation strength of word pairs:

[0084] ;

[0085] in, is the association strength of the word combination, For vocabulary The co-occurrence frequency of For vocabulary The weight in the syntactic structure, The number of words involved in the calculation. , , , , , , , , the association strength is calculated as follows:

[0086] ;

[0087] in, This indicates that the vocabulary combination has a strong correlation in the literature and can be used for medical term extraction and text mining later.

[0088] S113: calling a high-frequency word combination set, comparing it with an existing medical terminology library, identifying medical entities in the text, including symptoms, drugs, treatment methods, and causes, counting the distribution of multiple types of entities in the literature, and establishing a literature entity set;

[0089] Call the high-frequency word combination set, combine with the existing medical terminology library, match the medical entities that may be involved in the text, and classify and organize them. In the processing process, the candidate words are first screened, and the categories such as symptoms, drugs, treatment methods, and causes are classified according to the existing medical vocabulary. Then, the distribution of these entities in different documents is calculated to measure the importance of a certain medical entity. For example, in the sentence "Ginger soup is suitable for colds and flu", "colds and flu" is identified as a symptom and "ginger soup" is a drug, and they are classified. The distribution of symptoms, drugs, treatment methods, and causes in the literature is counted, and an index list of medical entities is formed.

[0090] In order to measure the importance of a medical entity in the entire literature, the Inverse Document Frequency (IDF) is used for calculation:

[0091] ;

[0092] in, For Entity The inverse document frequency of is the total number of documents, The number of documents containing this entity, plus 1 to prevent the denominator from being zero. , , , , , the IDF value of "cold" is calculated as follows:

[0093] ;

[0094] Calculate the IDF value of "ginger soup":

[0095] ;

[0096] Calculate the IDF value of "cause":

[0097] ;

[0098] in, This indicates that “ginger soup” appears less frequently in the literature and is therefore more discriminatory, whereas “wind-cold cold” and “cause of disease” are relatively common and have lower weights.

[0099] See also Figure 3 , the specific process of obtaining the speech recognition entity list is as follows:

[0100] S211: Based on the document entity set, the oral speech of famous traditional Chinese medicine practitioners is obtained, and the audio signal is preprocessed, including sampling rate adjustment, noise removal and speech enhancement. The processed audio signal is divided into multiple segments, and the short-time Fourier transform is used to perform time-frequency conversion on the speech, and the time domain and frequency domain features are extracted. The speech is denoised to obtain the speech segmentation result;

[0101] After obtaining the oral speech data of famous traditional Chinese medicine practitioners, the sampling rate of the speech signal is first adjusted to ensure the stability of data processing. In general, the sampling rate of the speech signal is set to 16kHz or 44.1kHz to meet different speech analysis needs. Secondly, the adaptive filtering method is used to remove the background noise, and the bandpass filter is used to retain the speech components between 500Hz and 4kHz to improve the clarity of the speech signal. After noise removal, the short-time Fourier transform (STFT) is applied to convert the speech signal to the frequency domain to extract its features. In order to further optimize the processing, the energy threshold detection method is used to identify the effective part of the speech and remove the silent section in the audio. In this process, the background noise threshold is set to 30dB, and the speech signal strength should be kept above 60dB. In order to evaluate the signal quality, the signal-to-noise ratio (SNR) is calculated using the following formula:

[0102] ;

[0103] in, is the signal-to-noise ratio, is the voice signal power, is the noise power. Assume that the signal power of a speech segment is , the noise power is , then calculate:

[0104] ;

[0105] This result shows that the speech signal quality is good and can be used for subsequent speech-to-text processing to finally obtain the speech segmentation result.

[0106] S212: Based on the speech segmentation result, convert the audio data into a text symbol sequence to obtain speech text data;

[0107] Based on the speech segmentation results, the automatic speech recognition (ASR) technology is used to convert the audio data into a sequence of text symbols. In this process, the Mel-frequency cepstral coefficient (MFCC) features of the speech are first extracted, and its feature vector is calculated to match the existing acoustic model. Then, the time series characteristics of the speech are modeled using the hidden Markov model (HMM), and the optimal path search is performed in combination with the language model to predict the most likely text output. Since mistranslation may occur during the speech conversion process, the dynamic time warping (DTW) algorithm is used to optimize the match between the text output and the speech features. During the processing, the text error correction threshold is set to 5%, that is, the error rate of the converted text is allowed to not exceed 5%. In order to measure the accuracy of speech recognition, the word error rate (WER) is calculated using the following formula:

[0108] ;

[0109] in, is the word error rate, To replace the error number, To delete the error number, is the number of insertion errors, is the total number of words. Suppose the number of words with incorrect recognition in a test text is as follows: , , , the total number of words is , then calculate:

[0110] ;

[0111] That is, the WER is 10%, indicating that the speech recognition accuracy is high, and the final output is a text symbol sequence.

[0112] S213: Call the speech text data, combine multiple entities in the literature, compare the collocation rules of adjacent words in the text, calculate the similarity between the candidate words and the known medical entities, and use the formula:

[0113] ;

[0114] Calculate the matching credibility of each word in the speech text, evaluate the meaning consistency of multiple medical entity words in the speech text of famous traditional Chinese medicine practitioners and literature, detect and extract medical terms in the speech text, and obtain the speech recognition entity list;

[0115] in, Represents the matching confidence of text words, Representing text words The context weight of Represents the semantic matching degree between the word and the medical entity. represents the phoneme matching probability of the vocabulary, represents the contextual deformation error of the word, Represents the number of valid words in the text, An index of words in the spoken text;

[0116] The text symbol sequence is called, and based on the document entity set, the collocation rules of adjacent words in the text are compared to calculate the similarity between the candidate words and the known medical entities. During the processing, the context weight of the vocabulary is first calculated, and high-frequency medical vocabulary is given a higher weight value to enhance the matching accuracy of professional terms. Then, the semantic matching algorithm is used to calculate the similarity score between each vocabulary and the medical entity, and the pronunciation similarity of the converted text is evaluated in combination with the phoneme matching probability. In addition, considering the context error in speech recognition, the context deformation error of each vocabulary is calculated to correct the recognition bias caused by accent or connected reading. The following formula is used to calculate the matching credibility:

[0117] ;

[0118] in, Represents the matching confidence of text words, Representing text words The context weight of Represents the semantic matching degree between the word and the medical entity. represents the phoneme matching probability of the vocabulary, represents the contextual deformation error of the word, Represents the number of valid words in the text. Assuming the sample data, the context weight of "cold" is 0.85, the semantic matching degree is 0.92, the phoneme matching probability is 0.90, and the context deformation error is 0.05. The matching reliability of the word is calculated as follows:

[0119] ;

[0120] Calculated:

[0121] ;

[0122] The result shows that the matching reliability of medical-related vocabulary reaches 0.649, which has a high text matching accuracy in the medical context. Finally, medical terms are screened out and the speech recognition entity list is obtained.

[0123] See also Figure 4 , the specific process of obtaining the entity relationship derivation diagram is as follows:

[0124] S311: Call the speech recognition entity list, compare the text with the entity sets from the speech sources, analyze the distribution of the same entities across multiple sources by calculating their occurrences in the speech text and the literature text, and obtain the cross-text entity distribution matrix;

[0125] Call the speech recognition entity list, compare the text with the entity sets from the speech sources, and calculate the distribution of the same entities across different sources using a text matching algorithm. First, perform entity extraction on the speech text and the literature text respectively, count the occurrences of each entity in different sources, and perform normalization to ensure the comparability of the entity occurrence probabilities across different text sources. During the normalization process, set the normalized frequency for each entity, and the calculation method is as follows:

[0126] ;

[0127] where represents the normalized entity occurrence frequency, represents the number of occurrences of the entity in the text, represents the total number of words in the text. For example, in the speech text, "wind-cold common cold" appears 85 times, and the total number of words in the speech text is 1000, then its normalized frequency is calculated as follows:

[0128] ;

[0129] Similarly, in the literature text, this entity may appear 92 times, and the total number of words in the literature is 1200, then its normalized frequency is calculated as follows:

[0130] ;

[0131] Subsequently, establish the cross-text entity distribution matrix. The occurrence frequency of each entity can be used for subsequent matching calculations. During this process, to exclude interference information, first perform stop word filtering to remove common but semantically unhelpful words such as "of", "is", "can", etc., and then apply a lemmatization method to merge different variants (such as "ginger decoction" and "decoction of ginger") to ensure that the same entity has a unified representation in different texts. Finally, use the TF-IDF (Term Frequency-Inverse Document Frequency) method to calculate the importance weights of the entities and store them in the database for subsequent calculation of the cross-text entity distribution matrix.

[0132] S312: Based on the cross-text entity distribution matrix, calculate the entity co-occurrence frequencies of multiple text sources, analyze the co-occurrence times of the target entity in multiple documents, calculate the association strength of entity pairs, using the formula:

[0133] ;

[0134] Calculate the relevance of entities in multiple documents by calculating the association strength of each entity pair, representing the two entities in the entity pair with rows and columns, and generating a cross-document entity co-occurrence matrix;

[0135] Among them, represents the entity association strength, represents the entity in the standardized occurrence frequency in the speech text, represents the entity in the standardized occurrence frequency in the literature text, and are the entity means of the two text sources respectively, represents the number of entities counted, is the index of the entity;

[0136] Based on the cross-text entity distribution matrix, calculate the entity co-occurrence frequency of different text sources, count the co-occurrence times of the same entity in different documents, and calculate the association strength of the entity pair. First, conduct a co-occurrence statistics on each entity pair (such as "wind-cold cold - ginger soup"), record the number of times it appears simultaneously in the same text segment or adjacent paragraphs, and ensure the accuracy of the co-occurrence calculation. The formula is set as follows:

[0137] ;

[0138] Among them, represents the entity association strength, represents the entity in the standardized occurrence frequency in the speech text, represents the entity in the standardized occurrence frequency in the literature text, and are the entity means of the two text sources respectively, represents the number of entities counted. Taking "wind-cold cold - ginger soup" as an example, in the speech text, its standardized frequency is 0.085, while in the literature text, the standardized frequency of this entity is 0.0767. The calculation process is as follows:

[0139] ;

[0140] Calculated:

[0141] ;

[0142] ;

[0143] To ensure the accuracy of entity co-occurrence data and exclude the influence of random co-occurrence on the calculation results, a minimum threshold for co-occurrence frequency is set. For example, the number of co-occurrences of an entity pair in 1000 documents should not be less than 5 times, otherwise the entity pair will be excluded. The final calculation results are stored in the database and output to the cross-document entity co-occurrence matrix.

[0144] S313: Call the cross-document entity co-occurrence matrix, parse the syntactic structure of the statement, analyze the modification relationship and semantic dependency path between entities, evaluate the interaction of multiple entities in the sentence structure, and construct an entity relationship derivation graph;

[0145] Call the cross-document entity co-occurrence matrix, parse the syntactic structure of the statement, analyze the modification relationship and semantic dependency path between entities, and determine the interaction of different entities in the sentence structure. First, apply the dependency parsing method to decompose the statement and determine the functional roles of entities in the sentence, such as subject, object, attributive, etc. For example, after parsing the sentence "Wind-cold cold can be relieved by ginger soup", "wind-cold cold" is the subject, "ginger soup" is the object, and "relieved" is the main verb, and the dependency path is "wind-cold cold, relieved, ginger soup".

[0146] Next, calculate the syntactic dependency degree of the entity pair to evaluate the tightness of the association between entities in the syntactic structure. The following formula is used:

[0147] ;

[0148] Among them, represents the syntactic dependency degree of the entity pair, represents the number of times the entity pair appears in the dependency relationship, represents the total co-occurrence times of this entity. For example, in 100 co-occurrences, the "wind-cold cold - ginger soup" has a syntactic dependency relationship 82 times, then its syntactic dependency degree is calculated as follows:

[0149] ;

[0150] In addition, considering that some entities may have an indirect relationship (such as "wind-cold cold" and "ginger soup" may be indirectly related through "relieved"), a path length parameter is introduced to calculate the influence of the syntactic dependency path. Set the path length as the shortest syntactic distance between entities, then the syntactic corrected dependency degree is calculated as:

[0151] ;

[0152] Set that "wind-cold cold" and "ginger soup" are connected through "relieved", and the path length , then calculate:

[0153] ;

[0154] That is, after the path is corrected, the grammatical dependency decreases, indicating that the entity pair may be indirectly related. Finally, all calculation results are mapped to a directed graph to form an entity relationship derivation graph.

[0155] See also Figure 5 ,The specific process of obtaining the TCM knowledge graph is as follows:

[0156] S411: calling the entity relationship derivation graph, converting multiple medical entities into nodes of the graph, establishing edge connections according to the semantics, co-occurrence frequency, and dependency relationship between entities, calculating the connection strength of each node, and obtaining the entity network topology structure;

[0157] The entity relationship derivation graph is called to convert each medical entity into a node of the graph, and edge connections are established based on the semantics, co-occurrence frequency and dependency relationship between entities. Each edge represents the semantic association relationship between two entities. In this process, all extracted medical entities are first deduplicated and standardized to ensure that different expressions of the same concept (such as "ginger soup" and "ginger soup") can be processed uniformly. Then, the connection frequency of each entity in the corpus is calculated, and the entity pairs with high-frequency association are screened out using the association analysis method. For example, if "wind-cold cold" and "ginger soup" co-occur 200 times in 1,000 traditional Chinese medicine literature, and "wind-cold cold" and "relief" co-occur 150 times, the connection strength between the two will be different. Subsequently, the graph theory method is used to calculate the connection strength of each entity node, and the nodes are preliminarily classified, such as classifying nodes belonging to the same traditional Chinese medicine treatment category. To calculate the node influence of the entity, the following formula is set:

[0158] ;

[0159] in, Represents the influence index of the entity node, Representing Entity The connection frequency, Represents the number of connections of the node in the network. Represents the number of entities counted. If the connection frequency of "cold" is set to 200 and the number of nodes is 10, the calculation is as follows:

[0160] ;

[0161] ;

[0162] This calculation is used to measure the importance of physical nodes in the network and is applied to subsequent network optimization to ultimately form a physical network topology.

[0163] S412: Based on the entity network topology, the connection strength between multiple nodes is calculated, and synonymous entities are merged. The formula is set as follows:

[0164] ;

[0165] Calculate the fusion adjustment factors between entities and generate a synonymous entity fusion matrix;

[0166] in, represents the fusion adjustment factor, Representing Entity The structural weights in the network, Representing Entity Frequency in the knowledge base, Represents the number of entities counted, is the index of the entity;

[0167] Based on the entity network topology, the connection strength between different nodes is calculated, and synonymous entities are fused to reduce redundant information and enhance the overall stability of the network. To measure the degree of fusion of two entities, the calculation formula is as follows:

[0168] ;

[0169] in, represents the fusion adjustment factor, For Entity The structural weights in the network, For Entity Frequency in the knowledge base, is the number of entities counted. Assuming that the structural weight of a certain entity for "cold-ginger soup" is 0.85 and the knowledge base frequency is 0.90, and the structural weight of another entity for "cold-relief" is 0.78 and the knowledge base frequency is 0.82, the calculation is as follows:

[0170] ;

[0171] ;

[0172] ;

[0173] This value is used to determine the rationality of the fusion between entities. If the value exceeds the set threshold (such as 0.8), the entity pair is considered to be merged. The calculation results are stored in the synonymous entity fusion matrix.

[0174] S413: calling the synonymous entity fusion matrix, and updating the connection relationship between nodes in real time in combination with physician feedback to generate a TCM knowledge graph;

[0175] The synonymous entity fusion matrix is ​​called, and the connection relationship between nodes is updated in real time in combination with physician feedback, the correlation between entities is optimized, and the category attribution and relationship direction of the entities are adjusted. In this process, the number of times different physicians confirm the same entity relationship during the diagnosis and treatment process is first counted, and the physician feedback correction coefficient is set. For example, among the feedback of 50 physicians, 40 believe that "cold-ginger soup" has a direct treatment relationship, while 10 believe that they are indirectly related. The correction coefficient of the relationship is calculated as follows:

[0176] ;

[0177] in, represents the physician feedback correction factor, represents the number of physicians who confirmed the relationship, Represents the total number of participating physicians. Assuming that a certain relationship is confirmed by 40 physicians and 50 physicians participate, the calculation is as follows:

[0178] ;

[0179] Then, the correction coefficient is combined with the fusion adjustment factor to adjust the connection weight of the entity pair. Finally, based on the updated entity network structure, a TCM knowledge graph is formed, which contains the final relationship status of each entity pair and the physician feedback correction value.

[0180] See also Figure 6 , the specific process of obtaining the drug and disease relationship set is:

[0181] S511: Calling the TCM knowledge graph, extracting a set of drug entities related to multiple diseases, and obtaining a set of disease-related drugs;

[0182] The TCM knowledge graph is called to extract a set of drug entities related to a specific disease from the database. First, all drugs that may be related to the target disease are identified, and a subset of drugs that are directly used to treat the disease is screened out. During the extraction process, it is necessary to classify the disease based on the TCM syndrome differentiation, compare the characteristics of the disease entity, such as symptoms, causes, physical tendencies, etc., and find the corresponding drug entity. Preliminary screening is performed based on the relationship connection strength of the knowledge graph. To ensure the accuracy of the screening, it is necessary to traverse the connection path of each disease node in the knowledge graph to determine whether there is a drug node directly connected to the target disease. If a connection path exists, further analyze the indications, efficacy, and ingredient ratio of the drug, and screen out drugs that may be applicable to multiple diseases but have no obvious effect on the target disease. Finally, a set of candidate drugs is obtained. During the screening process, an entity relationship matching threshold needs to be set to measure the direct relationship strength between the disease and the drug. If the connection strength of a drug entity is lower than the threshold, it is removed from the candidate list.

[0183] S512: Based on the set of disease-related drugs, the attribute data of each drug entity, including efficacy, meridians, and contraindications, and the characteristic data of the disease entity, including causes, symptoms, and physiological effects, are matched and analyzed using the formula:

[0184] ;

[0185] Calculate and obtain the drug matching value;

[0186] in, Represents drug matching degree, is the attribute vector of the disease characteristics, is the mean vector corresponding to the drug, is the covariance matrix of the attributes, is the inverse matrix of the covariance matrix, represents the transpose operation;

[0187] Based on the set of disease-associated drugs, the attribute data of each drug entity (such as efficacy, meridians, and contraindications) and the characteristic data of the disease entity (such as cause, symptoms, and physiological effects) are matched and analyzed. In this process, the importance of different drug attributes needs to be weighted to ensure that the degree of influence of different attributes in the calculation process meets the clinical use standards of traditional Chinese medicine. The Mahalanobis distance is used to evaluate the matching degree between a single drug and a single disease. This distance is used to measure the normalized distance between a sample point and the overall mean. The calculation formula is as follows:

[0188] ;

[0189] in, Indicates the matching degree between the disease and the drug. is the attribute vector of the drug entity, is the mean vector of the disease entity, is the attribute covariance matrix, is the inverse matrix of the covariance matrix. This formula can fully consider the correlation between different variables and avoid matching based on absolute differences alone, which affects the matching accuracy.

[0190] Calculation Example

[0191] Suppose the characteristic vector of a disease , the mean vector of a drug , the covariance matrix , then the matching degree is calculated as follows:

[0192] calculate :

[0193] ;

[0194] calculate :

[0195] ;

[0196] calculate :

[0197] ;

[0198] After matrix calculation, we get:

[0199] ;

[0200] The matching degree is used to measure the degree of suitability of the drug to the disease. The lower the value, the higher the matching degree.

[0201] S513: calling the drug matching value, identifying and recording contraindicated drugs related to multiple diseases, and forming a drug and disease relationship set;

[0202] Call the drug matching value, set the contraindication threshold, screen the drugs with matching degrees greater than the set value or with contraindications, and mark these drugs as contraindications to ensure the safety of treatment. The contraindication screening is based on the Chinese medicine pharmacopoeia and clinical feedback data. By querying the contraindication records of the drugs and comparing the properties of the symptoms, the drugs that may cause adverse reactions to the patients are excluded. In the contraindication screening process, the drug interactions that may be caused by the symptoms are statistically analyzed, and drug combinations that have had adverse reactions in historical case data are extracted. Combined with the main chemical components in the drug ingredients, it is detected whether there are specific interaction risks. For example, some drugs may enhance or inhibit the efficacy of other drugs, resulting in unstable treatment effects. If a drug is found to have a potential contraindication relationship with the target disease, the drug is excluded from the final matching set, and finally a set of drug and disease relationships is formed.

[0203] Table 1 Example of drug matching calculation

[0204]

[0205] As shown in Table 1 , drug B and drug D were marked as contraindicated drugs because their matching degree was higher than the set threshold (0.4) or they had contraindications, and finally drug A and drug C were screened out as available drugs.

[0206] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for constructing a digital human knowledge graph for the living inheritance of the experience of famous traditional Chinese medicine practitioners, characterized in that: The following steps are involved: S1: Obtain TCM literature data, disassemble the literature sentence by sentence, cut continuous sentences into independent vocabulary units, mark the grammatical role of each word in the sentence, identify multiple entities in the text by analyzing the combination rules between adjacent words, and generate a literature entity set; S2: Based on the document entity set, the oral speech of the famous traditional Chinese medicine practitioner is obtained, the speech signal is divided into multiple segments and converted into a text symbol sequence, and multiple entities in the speech are extracted by comparing the collocation rules of adjacent words in the text sequence and comparing multiple entities in the document, and a speech recognition entity list is output; S3: calling the speech recognition entity list, performing spatial mapping on the entity set of the text and speech sources, analyzing the co-occurrence frequency of entities across documents, identifying the mutual relationship between multiple entities by analyzing the grammatical structure at the sentence level, and generating an entity relationship derivation graph; The specific process of obtaining the entity relationship derivation diagram is as follows: S311: calling the speech recognition entity list, comparing the entity sets of the text and speech sources, calculating the distribution of the same entity in multiple sources, analyzing the occurrence frequency of multiple entities in the speech text and the literature text, and obtaining a cross-text entity distribution matrix; S312: Based on the cross-text entity distribution matrix, the co-occurrence frequencies of entities from multiple text sources are calculated, the co-occurrence times of the target entity in multiple documents are analyzed, and the association strength of entity pairs is calculated using the formula: ； Calculate the relevance of entities in multiple documents by calculating the association strength of each entity pair, using rows and columns to represent the two entities in the entity pair, and generate a cross-document entity co-occurrence matrix; in, Represents the entity association strength, Representing Entity Normalized frequency of occurrence in the spoken text, Representing Entity The normalized frequency of occurrence in the document text, and are the entity means of the two text sources, Represents the number of entities counted, is the index of the entity; S313: calling the cross-document entity co-occurrence matrix, parsing the grammatical structure of the sentence, analyzing the modification relationship and semantic dependency path between entities, evaluating the interaction between multiple entities in the sentence structure, and constructing an entity relationship derivation graph; S4: using the entity relationship derivation graph, converting entities into nodes and relationships into edges, calculating the association strength according to the connection frequency between nodes, performing attribute fusion on synonymous entity nodes, combining physician feedback, updating the connection relationship in real time, and outputting a TCM knowledge graph; S5: Call the TCM knowledge graph to extract a set of drug entities associated with multiple disease entities, calculate the matching degree between multiple drugs and diseases by comparing the drug entity attribute data with the disease entity feature data, identify the contraindicated drugs associated with the target disease, and generate a set of drug and disease relationship.

2. The method for constructing a digital human knowledge graph for the living inheritance of the experience of famous traditional Chinese medicine practitioners according to claim 1 is characterized in that: The document entity set specifically includes sentence decomposition results, vocabulary unit extraction results, and medical entity data sets; the speech recognition entity list specifically includes oral speech collection results, text conversion records, and speech entity sets; the entity relationship derivation graph specifically includes entity co-occurrence relationships, front and back modification relationships, and relationships between entities; the traditional Chinese medicine knowledge graph includes entity node conversion results, relationship edge conversion records, and synonymous entity attribute fusion.

3. The method for constructing a digital human knowledge graph for the living inheritance of the experience of famous veteran Chinese medicine practitioners according to claim 1 is characterized in that: The specific process of obtaining the document entity set is as follows: S111: Obtain TCM literature data, disassemble the literature content sentence by sentence, analyze the sentence structure, extract vocabulary units, perform part-of-speech tagging and syntactic role assignment for each vocabulary, identify various types of vocabulary, including nouns, verbs, and adjectives, and establish syntactic structure analysis results; S112: Based on the syntactic structure analysis result, the collocation pattern between adjacent words is analyzed, and the stability of the word combination in the sentence is calculated using the formula: ； Calculate the association strength of word pairs, and obtain a set of high-frequency word combinations by counting the co-occurrence frequencies between multiple words; in, represents the association strength of the word combination, Representative words The co-occurrence frequency of Representative words The weight in the syntactic structure, Represents the number of words involved in the calculation, An index of vocabulary; S113: calling the high-frequency word combination set, comparing with the existing medical terminology library, identifying medical entities in the text, including symptoms, drugs, treatment methods, causes, counting the distribution of multiple types of entities in the document, and establishing a document entity set.

4. The method for constructing a digital human knowledge graph for the living inheritance of the experience of famous veteran Chinese medicine practitioners according to claim 1 is characterized in that: The specific process of obtaining the speech recognition entity list is as follows: S211: Based on the document entity set, the oral speech of the famous traditional Chinese medicine practitioner is obtained, and the audio signal is preprocessed, including sampling rate adjustment, noise removal and speech enhancement, and the processed audio signal is divided into multiple segments. The speech is converted into time and frequency using short-time Fourier transform, and time domain and frequency domain features are extracted. The speech is denoised to obtain a speech segmentation result; S212: Based on the speech segmentation result, convert the audio data into a text symbol sequence to obtain speech text data; S213: calling the speech text data, combining multiple entities of the document, comparing the collocation rules of adjacent words in the text, and calculating the similarity between the candidate words and the known medical entities, using the formula: ； Calculate the matching credibility of each word in the speech text, evaluate the meaning consistency of multiple medical entity words in the speech text of famous traditional Chinese medicine practitioners and literature, detect and extract medical terms in the speech text, and obtain the speech recognition entity list; in, Represents the matching confidence of text words, Representing text words The context weight of Represents the semantic matching degree between the word and the medical entity. represents the phoneme matching probability of the vocabulary, represents the contextual deformation error of the word, Represents the number of valid words in the text, The index of the vocabulary in the speech text.

5. The method for constructing a digital human knowledge graph for the living inheritance of the experience of famous veteran Chinese medicine practitioners according to claim 1 is characterized in that: The specific process of obtaining the TCM knowledge graph is as follows: S411: calling the entity relationship derivation graph, converting multiple medical entities into nodes of the graph, establishing edge connections according to the semantics, co-occurrence frequency, and dependency relationship between entities, calculating the connection strength of each node, and obtaining an entity network topology structure; S412: Based on the entity network topology, the connection strength between multiple nodes is calculated, and synonymous entities are merged, and the formula is set as follows: ； Calculate the fusion adjustment factors between entities and generate a synonymous entity fusion matrix; in, represents the fusion adjustment factor, Representing Entity The structural weights in the network, Representing Entity Frequency in the knowledge base, Represents the number of entities counted, is the index of the entity; S413: Call the synonymous entity fusion matrix, and update the connection relationship between nodes in real time based on physician feedback to generate a traditional Chinese medicine knowledge graph.

6. The method for constructing a digital human knowledge graph for the living inheritance of the experience of famous veteran Chinese medicine practitioners according to claim 1 is characterized in that: The drug and symptom relationship set includes a set of drug entities associated with symptoms, a drug and symptom matching degree, and contraindicated drug identification results.

7. The method for constructing a digital human knowledge graph for the living inheritance of the experience of famous veteran Chinese medicine practitioners according to claim 1 is characterized in that: The specific process of obtaining the drug and disease relationship set is as follows: S511: calling the TCM knowledge graph, extracting a set of drug entities related to multiple diseases, and obtaining a set of disease-related drugs; S512: Based on the set of disease-associated drugs, the attribute data of each drug entity, including efficacy, meridians, and contraindications, and the characteristic data of the disease entity, including cause, symptoms, and physiological effects, are matched and analyzed using the formula: ； Calculate and obtain the drug matching value; in, Represents drug matching degree, is the attribute vector of the disease characteristics, is the mean vector corresponding to the drug, is the covariance matrix of the attributes, is the inverse matrix of the covariance matrix, represents the transpose operation; S513: Call the drug matching degree value, identify and record contraindicated drugs related to multiple diseases, and form a drug and disease relationship set.

Images