A medical field knowledge fusion method based on label and content hierarchical fusion

Abstract

The application relates to a medical field knowledge fusion method based on label and content hierarchical fusion, and belongs to the technical field of knowledge fusion and artificial intelligence. Medical knowledge of different knowledge graphs is fused to construct a medical field knowledge base and provide large-scale medical industry knowledge services. The method comprises three steps of defining a meta label of an entity in a medical knowledge graph, data label alignment, and label content fusion. The method has the following characteristics: firstly, in the data label alignment, a data label alignment method based on integration of string measurement and semantic measurement is designed, and data label alignment is performed by using string features and semantic features. Secondly, in the label content fusion, a joint mechanism of content aggregation verification and unstructured knowledge generation is proposed, which aims to extract common knowledge of different granularities in different knowledge graphs, verify contradictory knowledge, and convert common triple knowledge into unstructured sentences to construct a medical knowledge base.

Description

TECHNICAL FIELD

[0001] The present application relates to a medical field knowledge fusion method based on label and content hierarchical fusion, and belongs to the technical field of knowledge fusion and artificial intelligence. BACKGROUND

[0002] Massive knowledge is contained in multi-source heterogeneous network data, which usually has the characteristics of heterogeneity, diversity and fragmentation. Knowledge graph, such as YAGO and DBpedia, refers to the use of graph data structure to organize human knowledge into a structured knowledge system to describe things and their relationships in the real world. Due to the difficulty of a single knowledge graph to cover all knowledge resources, and in order to extract and maximize the use of knowledge from massive knowledge resources, knowledge fusion technology emerges as the times require.

[0003] Knowledge fusion aims to align, associate and fuse two or more related knowledge graphs into a larger knowledge graph, and then build a knowledge base to provide high-quality knowledge sharing and knowledge services. For example, entity alignment refers to identifying whether different entities in two or more knowledge graphs represent the same thing or object in the real world. By aligning and fusing knowledge from different sources, different languages and different structures, knowledge fusion technology is beneficial to enrich, supplement, update and deduplicate existing knowledge graphs and knowledge bases. According to the difference of fusion objects, knowledge fusion can be divided into mode layer fusion and data layer fusion. Mode layer fusion refers to the fusion of knowledge system, which fuses two or more heterogeneous knowledge systems and constructs mapping relationships for the same categories, attributes and relationships. Data layer fusion refers to the fusion of instances (entity instances, relationship instances) in different knowledge graphs. On the one hand, knowledge fusion technology is widely used in information recommendation, question and answer systems, information retrieval and other fields, and has important application value. On the other hand, knowledge fusion helps computers better understand human intelligence and human thinking, and can effectively promote network data search and provide knowledge support for many application tasks of artificial intelligence.

[0004] The present application studies knowledge fusion technology for medical field knowledge graph. For example, the national population and health science data sharing platform pharmacy data center and other medical comprehensive websites contain a large amount of semi-structured knowledge of drug instructions. The drug knowledge of these different websites can be mutually supplemented and fused to form a set of hierarchical drug knowledge base. Therefore, the present application aims to fuse medical field knowledge of different knowledge graphs and build a medical knowledge base for medical artificial intelligence service needs.

[0005] There are two problems in the knowledge fusion of medical knowledge graphs at present. First, the data labels of different knowledge graphs are synonymous and different in name. For example, the label "pharmacokinetics" of a drug in one knowledge graph and the label "pharmacodynamics" of a drug in another knowledge graph have the same semantics but different names. In view of the same semantics and different names of entity labels in different knowledge graphs, the alignment of entity labels needs to be solved, and the alignment relationship between the labels "pharmacokinetics" and "pharmacodynamics" can be identified.

[0006] Second, the knowledge of different knowledge graphs is redundant and contradictory. For example, the content of the "ingredient" label of the drug "Ujibai Fengwan" in one knowledge graph is "chicken (hair and claw intestine), antler glue, turtle shell (processed), oyster (calcined), sand worm, ginseng, astragalus, angelica, white peony root, rhizoma anemarrhenae (vinegar), licorice, rehmia, prepared rehmannia, chuanxiong, silver radix bupleuri, salvia miltiorrhiza, yam, gordon euryale (fried), antler cream. The auxiliary material is honey". The content of the "ingredient" label in another knowledge graph is "ginseng, astragalus, white peony root, angelica, prepared rehmannia, chuanxiong, chicken, antler glue, thymus, raw land". The two label contents have knowledge redundancy. When semi-structured knowledge from different data sources is fused into the same label, on the one hand, the knowledge of different knowledge graphs needs to be complementary to each other, and the consistency of the knowledge, i.e. the knowledge cannot be contradictory. On the other hand, it is necessary to avoid the repetition and redundancy of knowledge of different granularities. SUMMARY

[0007] The purpose of the present application is to propose a label and content hierarchical fusion method for multi-source data, to fuse medical knowledge of different knowledge graphs, to construct a medical knowledge base, and to provide large-scale medical industry knowledge services. The method includes three steps of defining the meta-label of the entity in the medical knowledge graph, data label alignment, and label content fusion. The characteristics of the method are as follows. First, in the data label alignment, a data label alignment method based on the integration of string measurement and semantic measurement is designed, and the label alignment is performed by using string features and semantic features. Second, in the label content fusion, a joint mechanism of content aggregation verification and unstructured knowledge generation is proposed, aiming to extract common knowledge of different granularities in different knowledge graphs, verify contradictory knowledge, and convert the verified triple knowledge into unstructured sentences to construct a medical knowledge base.

[0008] A medical field knowledge fusion method based on label and content hierarchical fusion, comprising the following steps:

[0009] Step 1: defining a meta-label for the entity in the two medical knowledge graphs to be fused;

[0010] The medical knowledge graph K m is a triple set, i.e. K m= (e, l, c), where e represents an entity, l represents a data label (attribute or relationship), and c represents the content (attribute value or entity) corresponding to the data label l, referred to as label content;

[0011] Defining the meta-label of the medical field entity as the category of the data label of different medical knowledge graphs, and using the meta-label as the data label for unified naming;

[0012] Step 2: Aligning data labels using an integrated method based on string metrics and semantic metrics, including the following steps:

[0013] Step 2.1: Constructing a semantic similarity matrix between meta-labels and data labels;

[0014] First, for any meta-label t F and any data label t of different medical knowledge graphs, convert the meta-label t F and t into feature vectors using the pre-trained language model BERT, and perform feature processing using the maximum pooling layer. After processing, generate the feature φ(t F ) of the meta-label t F and the feature φ(t) of the data label t.

[0015] Then, generate the semantic similarity matrix M sem between meta-labels and data labels based on cosine similarity, as shown in formulas (1) and (2):

[0016]

[0017]

[0018] where M sem ∈R m×n , m represents the number of meta-labels, n represents the number of data labels, a ij represents the element corresponding to the i-th row and j-th column of M sem , ||φ(t F )|| represents the norm of φ(t F ), and ||φ(t)|| represents the norm of φ(t).

[0019] Step 2.2: Identify the meta-label category to which the data label belongs;

[0020] First, for each row element in the semantic similarity matrix M sem , sort the elements in descending order according to the similarity size represented by the matrix elements, and select the data labels corresponding to the top K (K ≤ n) matrix elements with the highest similarity.

[0021] Then, for each meta-label t Fmark the class of the K data labels with the same or similar implied semantics as the meta-label t F The marking method is shown in equation (3):

[0022]

[0023] wherein sortTopK(·) represents a function of extracting the top K data labels based on similarity ranking, represents the i-th row element of the semantic similarity matrix M sem , (t1, t2,..., t K ) represents the K data labels with the class of the meta-label t F .

[0024] Step 2.3: Generate the string similarity matrix of the meta-label and the data label;

[0025] First, based on the string edit distance, generate the string similarity matrix M str between all meta-labels and the K data labels with the class of each meta-label, the calculation method is shown in equations (4) and (5):

[0026]

[0027]

[0028] wherein M str ∈R m×K , m represents the number of meta-labels, {ops} represents a set of string edits, max(·,·) represents the maximum value, len(·) represents the length of the string, represents the i-th meta-label, t ij represents the j-th data label with the class of the meta-label .

[0029] Step 2.4: Construct the data label set with the class of each meta-label;

[0030] Compare the K elements of each row of the string similarity matrix M str with the set threshold value T. For the meta-label , construct the data label set Tag i with the class of the meta-label, the construction method is shown in equation (6):

[0031]

[0032] wherein b ij represents the element corresponding to the i-th row and the j-th column of the string similarity matrix M str , the initial value of Tag i is If b ij ≥ T, Tag i represents a set of data tags with the i-th meta-tag and the class of the meta-tag Tag i .

[0033] Further, for data tags of different knowledge graphs with the class of the meta-tag , the meta-tag is used for unified naming.

[0034] Step 3: Adopting the joint mechanism of content aggregation verification and unstructured knowledge generation to perform label content fusion, and fusing the label contents of the labels with the same class of meta-tags. For two triples (e1, l1, c1) and (e2, l2, c2) from different medical knowledge graphs, if the data tag l1 of the entity e1 and the data tag l2 of the entity e2 have the same class of meta-tags, then the contents c1 and c2 of the two data tags l1 and l2 are fused.

[0035] Comprising the following steps:

[0036] Step 3.1: Adding label content traceability;

[0037] Adding label content traceability means that first, data tags with empty label contents are removed. Second, the source website url of the data tag and its content is added after the non-empty label content, and the content and the url are separated by the symbol "##" to achieve precise traceability.

[0038] Step 3.2: Aggregating and verifying the label content;

[0039] Label content aggregation and verification means extracting different granularities of common knowledge from different label contents and verifying conflicting knowledge, i.e., inconsistent knowledge. First, extract the common knowledge of the label contents c1 and c2. The label content can be a word, phrase or sentence, and all common substrings with a length greater than 1 in the label content are extracted as common knowledge. Then, verify the conflicting triples in different knowledge graphs. Two triples are conflicting if the entities and label contents of the two triples are the same, but the two data tags have conflicting relationships. The following three groups of data tags with conflicting relationships are constructed:

[0040] (drug synergy, drug antagonism, conflicting relationship),

[0041] (drug synergy, incompatibility, conflicting relationship),

[0042] (drug antagonism, incompatibility, conflicting relationship).

[0043] That is, the data label "drug synergy" is in conflict with "drug antagonism", the data label "drug synergy" is in conflict with "compatibility contraindication", and the data label "drug antagonism" is in conflict with "compatibility contraindication".

[0044] For entities e and label contents c, if (e, drug synergy, c) and (e, drug antagonism, c) are both true, then the two triples are determined as conflicting knowledge.

[0045] Step 3.3: generating unstructured sentences;

[0046] The common triples extracted from different knowledge graphs are converted into unstructured sentences. The method of generating unstructured sentences is: first, for two triple sets from two knowledge graphs, select triples with the same entity and data label, and splice the label contents of these triples. Second, use the sentence template of the data label to generate unstructured sentences.

[0047] Step 3.4: display the label content fusion result;

[0048] In order to visually and intuitively display the label content fusion result, on the one hand, the common knowledge of each knowledge graph is displayed as a summary knowledge; on the other hand, the conflicting knowledge is highlighted, and the label content of each data label is displayed.

[0049] Advantages

[0050] The method of the application fuses knowledge of two medical knowledge graphs, adopts a medical field knowledge fusion method based on label content and hierarchical fusion, and can realize label alignment and label content fusion. Compared with the prior art, the method has the following advantages:

[0051] (1) The method can fuse two medical knowledge graphs, construct a data label and label content fusion result centered on medical field entities, and provide large-scale medical field knowledge services;

[0052] (2) For the problem of different knowledge graphs of data label synonymy and different names during data label alignment, the method designs a label alignment method based on integration of string measurement and semantic measurement. First, based on semantic similarity measurement, the categories of data labels with the same or similar latent semantics are marked as pre-defined meta-labels. Then, the string similarity measurement is used to further measure the similarity of data labels and meta-labels with similar semantics. For data labels with the same meta-label category, the corresponding meta-label is used for uniform naming. The label alignment method of the application measures the similarity of data labels and meta-labels from the character level and the semantic level, realizes the fusion of entity attributes and relationships at the concept level, and improves the performance of data label alignment.

[0053] (3) In order to solve the problem of knowledge redundancy and conflict in different knowledge graphs during label content fusion, the method designs a joint mechanism of content aggregation verification and unstructured knowledge generation. The mechanism first fuses the label contents of the data labels with the same meta-label category in different knowledge graphs, and then aggregates and verifies the label contents, extracts common knowledge and verifies conflicting knowledge of different granularities in different knowledge graphs, generates structured sentences, and displays the label content fusion. The label content fusion method of the application not only extracts common knowledge of different granularities, but also identifies conflicting knowledge, realizes knowledge fusion at the level of domain entities, and improves the effect of label content fusion. In addition, the application displays the fused knowledge of the knowledge graph, the traceability of the data label and the label content, and ensures the traceability of the knowledge graph fusion.

[0054] (4) The method is tested in the medical field knowledge graph, and the experimental results show the effectiveness and superiority of the method. The method has broad application prospects in the fields of information recommendation, question and answer system, information retrieval, etc. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The flowchart of the medical field knowledge fusion based on label and content hierarchical fusion proposed by the application is shown. DETAILED DESCRIPTION

[0056] The preferred embodiments of the method of the application will be described in detail below in combination with examples.

[0057] EMBODIMENT

[0058] A medical field knowledge fusion method based on label and content hierarchical fusion, as shown in Figure 1 , comprising the following steps:

[0059] Step 1: defining the meta-label of the entity in the medical knowledge graph;

[0060] The medical knowledge graph K m is a set of triples, i.e. K m = (e, l, c), where e represents an entity, l represents a label (attribute or relationship), and c represents the content (attribute value or entity) corresponding to the label l, which is referred to as label content. Define the meta-label of the entity in the medical field as the category of the data label in different knowledge graphs, and use the meta-label as the unified name of the data label.

[0061] In this embodiment, the meta-label of medical field entity drug is defined. Through the analysis of medical and health field resources and drug information of medical website, the meta-label is defined. Part of the meta-labels are as follows: generic name, trade name, English name, drug alias, pinyin, main ingredient, chemical name, chemical structure, molecular formula, molecular weight, excipient, nature, specification, summary, drug category, functional classification, indication / function main indication, usage and dosage, adverse reactions, contraindications, precautions, drug for pregnant women and lactating women, drug for children, drug for the elderly, drug interaction, pharmacology and toxicology, pharmacokinetics, storage, packaging, validity period, product name, product trademark, enterprise name, manufacturer, production address, filing number, filing organ, filing content, filing date, registration number, registration number note, approval date, certificate date, approval number, execution standard, approval date, approval number.

[0062] For example, for two triples from different knowledge graphs (Wuji Baifeng Wan, storage method, sealed) and (Wuji Baifeng Wan, storage, sealed), Wuji Baifeng Wan is a medical field entity drug, "storage method" and "storage" are data labels, and "sealed" is label content. The category label of the data labels "storage method" and "storage" is all "storage", and the category label "storage" is used as the unified name of the data labels "storage method" and "storage".

[0063] Step 2: The data labels are aligned by using the integration method based on string measurement and semantic measurement, which includes the following steps:

[0064] Step 2.1: Construct the semantic similarity matrix between meta-labels and data labels;

[0065] First, for any meta-label t F and any data label t of different knowledge graphs, the pre-trained language model BERT is used to convert them into feature vectors, and the maximum pooling layer is used for feature processing. After processing, the feature φ(t F ) of the meta-label t F and the feature φ(t) of the data label t are generated.

[0066] Then, the semantic similarity matrix M sem between meta-labels and data labels is generated based on cosine similarity, and the calculation method is shown in formulas (1) and (2):

[0067]

[0068]

[0069] Wherein, M sem ∈R m×n, m represents the number of meta-labels, n represents the number of data labels, a ij represents M sem , the element corresponding to the ith row and jth column, ||φ(t F )|| represents the norm of φ(t F ), and ||φ(t)|| represents the norm of φ(t).

[0070] Step 2.2: Identify the meta-label category to which the data label belongs;

[0071] First, for each element in the semantic similarity matrix M sem , the elements are sorted in descending order according to the similarity size represented by the matrix elements, and the top K K ≤n) matrix elements corresponding to the data labels with the highest similarity are selected.

[0072] Then, for each meta-label t F , the K data labels with the same or similar implicit semantics are marked as the meta-label t F , and the marking method is shown in formula (3):

[0073]

[0074] where sortTopK(·) represents a function for extracting the top K data labels based on similarity sorting, represents the ith row element of the semantic similarity matrix M sem , and (t1, t2,..., t K ) represents the K data labels with the category of meta-label t F .

[0075] Step 2.3: Generate a string similarity matrix between meta-labels and data labels;

[0076] First, based on the string edit distance, generate a string similarity matrix M str between all meta-labels and the K data labels with the category of each meta-label, and the calculation method is shown in formulas (4) and (5):

[0077]

[0078]

[0079] where M str ∈R m×K , m represents the number of meta-labels, {ops} represents a set of string edits, max(·,·) represents the maximum value, len(·) represents the length of the string, represents the ith meta-label, t ij represents the K data labels with the category of meta-label The j-th data label.

[0080] Step 2.4: Construct a set of data labels for each meta label;

[0081] The string similarity matrix M str Each row's K elements are compared to a set threshold T. For meta tags... Construct a collection of data tags whose category is this meta tag. i The construction method is shown in formula (6):

[0082]

[0083] Among them, b ij M represents the string similarity matrix str The element corresponding to the i-th row and j-th column, Tag i The initial value is If b ij ≥T, Tag i This represents the combination of the i-th meta tag and the meta tag of category Tag. i A collection of data tags.

[0084] Furthermore, for category t Fi Data labels from different knowledge graphs, utilizing meta-labels t Fi Use a unified naming convention.

[0085] Step 3: Use a joint mechanism of content aggregation verification and unstructured knowledge generation to fuse tag content. Fuse tag content with the same meta tag category. For two triples (e1, l1, c1) and (e2, l2, c2) from different medical knowledge graphs, if the meta tag categories of data tag l1 of entity e1 and data tag l2 of entity e2 are the same, then fuse the content c1 and c2 of these two data tags l1 and l2.

[0086] Step 3.1: Add tag content tracing;

[0087] Adding tag content tracing involves, first, removing data tags with empty content. Second, after non-empty tag content, adding the data tag and its source URL, separated by the symbol "##", to achieve precise tracing.

[0088] Step 3.2: Aggregate and validate the tag content;

[0089] Tag content aggregation and validation refers to extracting common knowledge at different granularities from different tag contents and validating contradictory or inconsistent knowledge.

[0090] Firstly, common knowledge of label contents c1 and c2 is extracted. The label content can be a word, a phrase or a sentence, and all common substrings with a length greater than 1 in the label content are extracted as common knowledge.

[0091] For example, the content under the "ingredients" label of the entity "Wuji Baifeng Pill" in a knowledge graph is "black chicken (with hair and claw intestine), deer horn glue, turtle shell (processed), oyster (calcined), sand worm, ginseng, astragalus, angelica", and the content under the "ingredients" label of the entity "Wuji Baifeng Pill" in another knowledge graph is "this medicine is composed of ginseng, astragalus, white peony root, angelica, prepared rehmannia, chuanxiong, black chicken, deer horn glue, and winter melon". Therefore, the common knowledge of the two knowledge graphs is "ginseng, astragalus, angelica, black chicken, deer horn glue".

[0092] Then, the contradictory triples in different knowledge graphs are verified. Two triples are contradictory if the entities and label contents of the two triples are the same, but the two labels have contradictory relationships. The following three groups of labels with contradictory relationships are constructed:

[0093] (drug synergy, drug antagonism, contradictory relationship),

[0094] (drug synergy, compatibility contraindication, contradictory relationship),

[0095] (drug antagonism, compatibility contraindication, contradictory relationship).

[0096] That is, the data label "drug synergy" and "drug antagonism" are in a contradictory relationship, the data label "drug synergy" and "compatibility contraindication" are in a contradictory relationship, and the data label "drug antagonism" and "compatibility contraindication" are in a contradictory relationship.

[0097] For entity e and label content c, if (e, drug synergy, c) and (e, drug antagonism, c) are both true, then the two triples are determined to be contradictory knowledge.

[0098] For example, one knowledge graph contains the triple (Sanhe Calcium Chewable Tablets, drug synergy, tetracycline), and another knowledge graph contains the triple (Sanhe Calcium Chewable Tablets, drug antagonism, tetracycline). These two triples belong to a group of contradictory knowledge, because the data label "drug synergy" and "drug antagonism" are in a contradictory relationship.

[0099] Step 3.3: Generate unstructured sentences;

[0100] The common triples extracted from different knowledge graphs are converted back into unstructured sentences. The method of generating unstructured sentences is as follows: first, for two triple sets from two knowledge graphs, select triples with the same entity and data label, and concatenate the label contents of these triples. Second, use the sentence template of the data label to generate unstructured sentences.

[0101] For example, the two knowledge graphs have the relationship triplets about the drug "Yataina" in common, which are "(Yataina, indication, depression), (Yataina, side effect, nausea), (Yataina, side effect, dizziness), (Yataina, side effect, sleepiness), and (Yataina, incompatibility, propylpylene phenylalanine hydrochloride)", and the generated unstructured sentences are "Yataina can be used to treat depression, Yataina can cause adverse reactions such as nausea, dizziness, and sleepiness, and Yataina cannot be used in combination with propylpylene phenylalanine hydrochloride".

[0102] Step 3.4: display the label content fusion result;

[0103] In order to intuitively and visually display the label content fusion result, on the one hand, the common knowledge of each knowledge graph is displayed as summary knowledge; on the other hand, the conflicting knowledge is highlighted, and the label content of each data label is displayed.

[0104] In order to illustrate the medical knowledge fusion effect of the present application, the same data set is used to compare the data label alignment effect under the same conditions by using two methods. The first method is based on the pre-trained model Mirror-BERT. The second method is the medical field knowledge fusion method based on label and content hierarchical fusion of the present application.

[0105] The evaluation indexes used are accuracy, recall rate and F1 value. The present application regards data label alignment as a binary classification problem, and judges whether the meta-label category to which the data label belongs is correct. (1) The accuracy, recall rate and F1 value of the first method are 63.2%, 92.1% and 72.8% respectively. (2) The accuracy, recall rate and F1 value of the label alignment method based on the integration of string measurement and semantic measurement proposed by the present application are 75.4%, 86.8% and 77.9% respectively.

[0106] The above is only a preferred embodiment of the present application, and the present application should not be limited to the content disclosed in the embodiment and the drawings. Any equivalent or modification made without departing from the disclosed spirit falls within the scope of protection of the present application.

Claims

1. A medical field knowledge fusion method based on label and content hierarchical fusion, characterized in that Comprising the following steps: Step 1: Define meta-labels for entities in the two medical knowledge graphs to be fused; Step 2: Align data labels using an integrated method based on string metrics and semantic metrics, including the following steps: Step 2.1: Construct a semantic similarity matrix between meta-labels and data labels; Step 2.2: Identify the meta-label category to which the data label belongs; Step 2.3: Generate a string similarity matrix between meta-labels and data labels; Step 2.4: Build a data label set for each meta-label category; Step 3: Use a joint mechanism of content aggregation verification and unstructured knowledge generation to fuse label content. For two triples (e1, l1, c1) and (e2, l2, c2) from different medical knowledge graphs, if the meta-label categories of the data labels l1 for entity e1 and l2 for entity e2 are the same, then the contents c1 and c2 of these two data labels l1 and l2 are fused, including the following steps: Step 3.1: Increase label content traceability; First, remove data labels with empty label content. Second, add the source URL of the data label and its content after the non-empty label content, with the symbol "##" separating the content and the URL to achieve precise traceability; Step 3.2: Aggregate and verify label content; First, extract common knowledge from label content c1 and c2, which can be words, phrases, or sentences, and extract all common substrings with a length greater than 1 as common knowledge. Then, verify conflicting triples in different knowledge graphs; Step 3.3: Generate unstructured sentences by converting the common triples extracted from different knowledge graphs into unstructured sentences; Step 3.4: Display the label content fusion results; Display the common knowledge of each knowledge graph as the summary knowledge; highlight the conflicting knowledge and display the label content of each data label.

2. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 1, wherein: In step 1, the method for defining meta-labels for entities in the two medical knowledge graphs to be fused is: Medical knowledge graph K m is a set of triples, i.e., K m = (e, l, c), where e represents an entity, l represents a data label, and c is the content corresponding to the data label l, referred to as label content. Define the meta-labels of medical field entities as the categories of data labels in different medical knowledge graphs, and use meta-labels as the unified naming of data labels.

3. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 1, wherein: In step 2.1, the method for constructing a semantic similarity matrix between meta-labels and data labels is: First, for any one meta-label t F and any one data label t of different knowledge graphs, use the pre-trained language model BERT to convert them into feature vectors, and use the maximum pooling layer for feature processing; After processing, the features φ(t F ) of the meta-label t F and the features φ(t) of the data label t are generated. Then, a semantic similarity matrix M between the meta-labels and the data labels is generated based on cosine similarity sem The calculation method is shown in formulas (1) and (2): wherein M sem ∈ R m×n , m represents the number of meta-labels, n represents the number of data labels, a ij represents the element corresponding to the ith row and jth column of M sem , ||φ(t F )|| represents the norm of φ(t F ), and ||φ(t)|| represents the norm of φ(t).

4. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 1, wherein: In step 2.2, the method for identifying the meta-label category to which the data label belongs is: First, for each row element in the semantic similarity matrix M sem , the data labels corresponding to the top K matrix elements with the highest similarity are selected according to the descending order of the similarity size represented by the matrix elements. Then, for each meta-tag t F , the class of the K data tags with the same or similar latent semantics is marked as the meta-tag t F , and the marking method is shown in formula (3): where m represents the number of meta-tags, sortTopK(·) represents a function of extracting top K data tags based on similarity ranking, represents a semantic similarity matrix M sem the i-th row element of M K represents K data tags of which the category is the meta-tag t F .

5. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 1, wherein: In step 2.3, the method for generating a string similarity matrix between meta-labels and data labels is: First, based on the string edit distance, a string similarity matrix M between all meta-tags and K data labels for each meta-tag is generated str The calculation method is shown in equations (4) and (5): where M str ∈ R m×K , m denotes the number of meta-labels, {ops} denotes a set of edits on a string, max(·, ·) denotes a maximum value, len(·) denotes a length of a string, denotes an i-th meta-label, t ij denotes a j-th data label of a class of an i-th meta-label .

6. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 1, characterized in that: In step 2.4, the method for constructing the category of the data label set of each meta-label is: The string similarity matrix M str The K elements of each row are compared with the set threshold T, and the meta-label The data label set Tag of the category is constructed for the meta-label i The construction method is shown in formula (6): where b ij represents the string similarity matrix M str the element in the i-th row and j-th column of matrix M i is initialized to if b ij ≥ T, Tag i represents the set consisting of the data tags whose i-th meta-tag is Tag i and whose category is the meta-tag Tag 7. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 6, characterized in that: Two triples are contradictory if the entity and label content of the two triples are the same, but the two data labels have a contradictory relationship. The following three groups of data labels with a contradictory relationship are constructed: (drug synergy, drug antagonism, contradictory relationship), (drug synergy, incompatibility, contradictory relationship), (drug antagonism, incompatibility, contradictory relationship), That is, the data label "drug synergy" and "drug antagonism" are in a contradictory relationship, the data label "drug synergy" and "incompatibility" are in a contradictory relationship, and the data label "drug antagonism" and "incompatibility" are in a contradictory relationship; For entity e and label content c, if (e, drug synergy, c) and (e, drug antagonism, c) are both true, then the two triples are determined to be contradictory knowledge.

8. The medical field knowledge fusion method based on label and content hierarchical fusion according to claim 7, characterized in that: In step 3.3, the method for generating unstructured sentences is: First, for two sets of triples from two knowledge graphs, select triples with the same entity and data label, and concatenate the label content of these triples; then, use the sentence template of the data label to generate unstructured sentences.

Images