Knowledge graph model fusing knowledge, capability and data and construction method

Abstract

The invention discloses a knowledge graph model fusing knowledge, capability and data and a construction method. According to the model, attributes of entities are expanded into three dimensions of knowledge attributes, capability attributes and data attributes, so that organic unification of knowledge storage, decision execution and data updating is realized. Wherein the capability attribute enables the entity to have an extensible intelligent decision-making capability through an integrated decision-making model and a standardized interface; data attributes support real-time data updating and model training through a hierarchical organization structure of a data complete set, a data subset and a data label. The invention further provides a corresponding construction method. The method comprises the steps of concept and relation definition, knowledge attribute extraction, data attribute collection, capability attribute training, collaborative knowledge updating and the like. According to the method, the problem that a traditional knowledge graph model lacks decision-making ability is solved, the application value of the knowledge graph in enterprise intelligent decision-making is improved, and the method has good practicability.

Description

Technical Field

[0001] The present invention relates to the field of knowledge engineering, and in particular to a knowledge graph model and a construction method that integrates knowledge, capabilities and data. Background Art

[0002] As a structured knowledge representation method, knowledge graph plays a key role in modern information processing and decision support. It provides a systematic solution for the representation, retrieval and reasoning of complex information by organizing concepts, entities and their relationships in the form of a graph structure. In the business field, knowledge graphs have been widely used in multiple scenarios such as customer profiling, competition analysis, risk control, product recommendations, etc., showing significant technical value and business potential.

[0003] However, with the deepening of digital transformation and the rapid development of artificial intelligence technology, the existing knowledge graph model has exposed obvious technical limitations in the application of the business field. The existing knowledge graph mainly adopts the traditional entity-relationship-entity model, which only focuses on the storage and representation of static knowledge. This model is difficult to effectively express and encapsulate decision logic and cannot meet the needs of intelligent decision-making systems for knowledge application. For example, in customer relationship management, it can only represent the basic attributes of customers, but it is difficult to integrate decision-making capabilities such as customer value assessment and churn warning.

[0004] The existing model also lacks a standardized capability encapsulation mechanism, which leads to the separation of decision logic and knowledge structure. In practical applications, it is often necessary to build a decision model separately outside the knowledge graph, which not only increases the complexity of the system, but also reduces the efficiency of knowledge reuse. Especially in scenarios where rapid response to market changes is required, this separation seriously restricts the adaptability of the system.

[0005] In the current information explosion environment, business data presents the characteristics of multi-source heterogeneity and real-time update. Traditional knowledge graph models lack a systematic management mechanism for dynamic data, and it is difficult to effectively handle high-frequency updated market data, user behavior data, etc. According to statistics, the average annual growth rate of enterprise data volume exceeds 40%, and the update efficiency of existing models is difficult to match this growth rate. At the same time, the existing models lack a coordination mechanism between knowledge, capabilities and data, which makes it difficult to achieve the comprehensive utilization of multi-dimensional information in complex decision-making scenarios. For example, in brand reputation management, historical knowledge, real-time public opinion data and evaluation models need to be considered at the same time, but existing technologies are difficult to support this comprehensive analysis demand.

[0006] In view of the above problems, it is urgent to develop a new knowledge graph model that can organically integrate knowledge representation, decision-making capabilities and data management functions to provide more complete knowledge service support for modern business intelligence systems. The present invention is an innovative solution proposed in this context. Summary of the invention

[0007] The present invention aims to solve the above problems and proposes a knowledge graph model and construction method that integrates knowledge, capabilities and data. By introducing capability attributes and data attributes, the knowledge graph is endowed with general decision-making capabilities.

[0008] In order to achieve the above-mentioned object of the present invention, according to the first aspect of the present invention, the present invention provides a knowledge graph model integrating knowledge, capability and data, including the following contents:

[0009] Concept module, which is used to define the basic architecture of the knowledge graph;

[0010] Entity module, used to represent entity nodes in the knowledge graph;

[0011] Relationship module, used to represent the connection relationship between entities;

[0012] Furthermore, each entity node in the entity module contains three types of attributes:

[0013] Knowledge attributes, used to store static feature information related to entities;

[0014] Capability attributes, which are used to execute decision logic associated with the entity;

[0015] Data attributes, which are used to manage dynamic data associated with an entity;

[0016] The knowledge attributes, capability attributes and data attributes are associated with each other through a unified identifier to achieve collaborative operations of knowledge storage, decision execution and new information accommodation.

[0017] Furthermore, the capability attributes include the following:

[0018] Decision models, based on custom machine learning models or general large language models, are used to perform reasoning decisions;

[0019] A standardized interface is used to realize the interaction between the decision model and external systems.

[0020] Furthermore, the data attributes include the following:

[0021] The complete set of data uses a distributed storage architecture to store all data related to a specific entity;

[0022] A data subset is a data set selected from the entire data set according to specific conditions or objectives;

[0023] Data labeling, adding labels or classification information to the elements in the data subset;

[0024] The full data set, data subset and data tags are associated with each other through an indexing mechanism to support real-time updating and efficient retrieval of data.

[0025] In order to achieve the above-mentioned object of the present invention, according to the second aspect of the present invention, the present invention provides a method for constructing a knowledge graph integrating knowledge, capabilities and data, wherein the knowledge graph model described in the first aspect is used in the execution process, and the method comprises the following steps:

[0026] S1, definition phase, defining the core concepts of the domain and the main relationships between entities based on expert experience;

[0027] S2, knowledge construction stage, extracting entity names, knowledge attributes of entities and relationships between entities from open websites to form a basic knowledge graph;

[0028] S3, data construction phase, collects relevant information from open websites based on the keywords of the entity, configures the update strategy, and establishes data attributes for each entity;

[0029] S4. In the capability building phase, according to the decision-making capability requirements, a customized machine learning model is trained or a general large language model is configured to establish capability attributes for each entity.

[0030] Furthermore, the step S1 includes the following steps:

[0031] S11, domain analysis stage, through expert experience and literature research, determine the application field and core objectives of the knowledge graph;

[0032] S12, Concept design stage, the core concept system in the design field, including the hierarchical structure of concepts and the relationship between concepts;

[0033] S13. In the entity relationship design phase, define the relationship types between entities, including the directionality and constraints of the relationships.

[0034] Furthermore, the step S2 comprises the following steps:

[0035] S21, information collection stage, collecting structured and semi-structured information related to the entity from data sources such as open websites and knowledge bases;

[0036] S22, knowledge extraction stage, design hierarchical regular expressions to extract entity names, knowledge attributes of entities and relationships between entities;

[0037] S23, entity alignment stage, calculate the feature vector of each entity according to the entity name and the knowledge attribute of the entity, and merge entities with similar feature vectors;

[0038] S24, knowledge storage stage, storing the above results in a database to form a basic knowledge graph;

[0039] S25, knowledge correction stage, uses collaborative process to correct the above basic knowledge graph.

[0040] Furthermore, the step S25 includes the following steps:

[0041] S251, in the preparation phase, the reviser creates an independent branch when making updates;

[0042] S252, modification stage, the modifier performs content modification operations on the knowledge graph on its branch;

[0043] S253, Merge Application Phase: After completing the update, the reviser initiates a pull request to merge the changes into the main branch;

[0044] S254, During the review phase, auditors review the changes and discuss and modify them when necessary;

[0045] S255, merging stage. After the review is passed, the reviewer will merge the changes into the main branch to complete the revision of the knowledge graph.

[0046] Furthermore, step S3 includes the following steps:

[0047] S31, information collection stage: configure information collection strategies to collect text information related to the entity from open websites, including media reports, user feedback, industry policies, etc.

[0048] S32, information deduplication stage, using MD5 hash algorithm to dedupe text information;

[0049] S33, in the stage of establishing the complete set of data, the deduplication results are stored in the database, and a multi-dimensional index is established to form the complete set of data attributes;

[0050] S34, the data set update phase, adopts the sliding time window update strategy;

[0051] S35, data subset establishment stage, filtering information from the entire data set according to specific conditions to form a data subset;

[0052] S36: Data label establishment phase: adding labels or classification information to the elements in the data subset to form a data label corresponding to this data subset.

[0053] Further, when a training customized machine learning model is adopted, the step S4 comprises the following steps:

[0054] S411, data establishment phase, select or establish appropriate data subsets and data labels from data attributes according to decision-making capability requirements;

[0055] S412, in a data preprocessing stage, the text information in the data subset and the data label is segmented and standardized, and the text is converted into a feature vector using an appropriate text vectorization method;

[0056] S413, in the model training stage, the data is split into a training set and a validation set, a neural network model is trained, and a decision model is obtained;

[0057] S414, model performance evaluation phase, calculate accuracy, precision, recall and F1 score, and generate a model performance evaluation report;

[0058] S415, in the standardized interface establishment phase, the above decision model is placed into the capability attributes of the corresponding entity, the capability description information and performance evaluation report are supplemented, and a standardized interface and permission strategy are established for the model.

[0059] Furthermore, when a general large language model is configured, step S4 includes the following steps:

[0060] S421, prompt word establishment stage, according to the decision-making ability requirements, construct a structured prompt template containing task description, context information and constraints to form a standardized functional prompt of the large language model;

[0061] S422, data establishment phase, selecting or establishing appropriate data subsets and data labels from data attributes according to decision-making capability requirements;

[0062] S423, model performance evaluation stage, calculate accuracy, precision, recall and F1 score, and generate a model performance evaluation report;

[0063] S424, the standardized interface establishment phase, places the above decision model into the capability attributes of the corresponding entity, supplements the capability description information and performance evaluation report, and establishes a standardized interface and permission strategy for the model.

[0064] The beneficial effects provided by the present invention are: compared with the existing knowledge graph, the knowledge graph model proposed by the present invention realizes the unification of knowledge storage, decision execution and data update. Among them, the capability attribute enables the entity to have scalable decision-making capabilities; the data attribute realizes real-time data update and model training support. The present invention also proposes a corresponding knowledge graph construction method, which is expected to significantly enhance the application value of the knowledge graph in business intelligence, decision support and knowledge management. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a structural schematic diagram of a knowledge graph model integrating knowledge, capability and data according to the present invention;

[0066] Explanation of numbers: 1-concept module, 2-entity module, 3-relationship module, 21-knowledge attribute, 22-ability attribute, 23-data attribute, 221-decision model, 222-standardized interface, 231-full data set, 232-data subset, 233-data label.

[0067] Figure 2 It is a flowchart of a method for constructing a knowledge graph integrating knowledge, capabilities and data according to the present invention;

[0068] Figure 3 It is a flowchart of the knowledge construction stage of the present invention;

[0069] Figure 4 It is a flowchart of the data construction phase of the present invention;

[0070] Figure 5 It is a flowchart of the process of using a training customized machine learning model in the capability building phase of the present invention;

[0071] Figure 6 It is a flowchart of configuring a general large language model in the capability building phase described in the present invention. DETAILED DESCRIPTION

[0072] In order to make the objectives, technical solutions 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 specific embodiments.

[0073] As attached Figure 1 As shown, the present invention provides a knowledge graph model integrating knowledge, capabilities and data, including the following contents:

[0074] Concept module (1), used to define the basic architecture of the knowledge graph;

[0075] Entity module (2), used to represent entity nodes in the knowledge graph;

[0076] Relationship module (3), used to represent the connection relationship between entities;

[0077] Furthermore, each entity node in the entity module (2) contains three types of attributes:

[0078] Knowledge attributes (21), used to store static feature information related to the entity;

[0079] Capability attributes (22), used to execute decision logic associated with the entity;

[0080] Data attributes (23), used to manage dynamic data associated with an entity;

[0081] The knowledge attributes (21), capability attributes (22) and data attributes (23) are associated with each other through a unified identifier to achieve collaborative operations of knowledge storage, decision execution and new information accommodation.

[0082] Furthermore, the capability attribute (22) includes the following:

[0083] A decision model (221), based on a custom machine learning model or a general large language model, for performing reasoning decisions;

[0084] A standardized interface (222) is used to implement the interaction between the decision model and an external system.

[0085] Furthermore, the data attributes (23) include the following:

[0086] The complete set of data (231) uses a distributed storage architecture to store all data related to a specific entity;

[0087] A data subset (232), a data set selected from the entire data set according to a specific condition or objective;

[0088] Data labeling (233), adding labels or classification information to the elements in the data subset;

[0089] The full data set, data subset and data tags are associated with each other through an indexing mechanism to support real-time updating and efficient retrieval of data.

[0090] As attached Figure 2 As shown, the present invention provides a method for constructing a knowledge graph integrating knowledge, capabilities and data, comprising the following steps:

[0091] In the definition phase (S1), the core concepts of the domain and the main relationships between entities are defined based on expert experience.

[0092] Step S1 includes the following steps:

[0093] In the domain analysis phase (S11), the application domain and core objectives of the knowledge graph are determined through expert experience and literature research. In this embodiment, the goal is to establish a knowledge graph in the business domain, which mainly includes enterprise entities, product entities, and personnel entities to support business intelligence applications such as enterprise brand analysis, market forecasting, and strategic decision-making;

[0094] In the conceptual design stage (S12), the core concept system of the design field includes the hierarchical structure of concepts and the relationship between concepts. In this embodiment, three core concepts are defined: company, product, and personnel;

[0095] In the entity relationship design phase (S13), the relationship types between entities are defined, including the directionality and constraints of the relationship. In this embodiment, seven types of relationships between entities are defined: company entity-investment-company entity, company entity-subordinate-company entity, company entity-competition-entity company, product entity-subordinate-company entity, personnel entity-subordinate-company entity, product entity-competition-product entity, and personnel entity-subordinate-personnel entity.

[0096] In the knowledge construction phase (S2), entity names, knowledge attributes of entities, and relationships between entities are extracted from open websites to form a basic knowledge graph;

[0097] As attached Figure 3 As shown, step S2 includes the following steps:

[0098] In the information collection stage (S21), structured and semi-structured information related to the entity is collected from data sources such as open websites and knowledge bases. In this embodiment, an online company knowledge base is selected as the data source, a customized web crawler is developed, and a depth-first search strategy is used to traverse the industry, company, product and personnel pages of the target library to obtain semi-structured web page information;

[0099] In the knowledge extraction stage (S22), a hierarchical regular expression is designed to extract entity names, knowledge attributes of entities, and relationships between entities. In this embodiment, the extracted knowledge attributes include: company establishment time, industry classification, company size, legal representative, company address, product type, personnel position and other information. For example, the regular expression for extracting company establishment time is as follows:

[0100] pattern = r'Establishment time [::]\s*(\d{4}-\d{2}-\d{2})'

[0101] Furthermore, the relationships between entities extracted include: investment, affiliation, and competition. For example, the regular expression for extracting the affiliation relationship between a person and a company is as follows:

[0102] pattern = r'Shareholders[::]\s*([\u4e00-\u9fa5]+)'

[0103] In the entity alignment stage (S23), the feature vector of each entity is calculated based on the entity name and the knowledge attributes of the entity, and entities with similar feature vectors are merged. In this embodiment, the entity alignment stage firstly performs word segmentation on the entity name, converts the word into a 300-dimensional word vector using the Word2Vec algorithm, and takes the mean of these word vectors as the entity name vector E. n ; Then the knowledge attributes of the entity are converted into attribute vector E through one-hot encoding and TF-IDF method a ; Then E n With Ea The entity feature vector E is obtained by weighted fusion, and the formula is as follows:

[0104] E=0.4*E n +0.6*E a

[0105] Furthermore, the cosine similarity is used to calculate the similarity between any two entity feature vectors E1 and E2. The formula is as follows:

[0106]

[0107] Furthermore, when the similarity is greater than 0.95, the pair of entities are marked as objects to be merged. When merging, the entity created earlier is retained as the main entity, the relationships and attributes of the entity to be merged are transferred to the main entity, and finally the entity node to be merged is deleted;

[0108] In the knowledge storage stage (S24), the above results are stored in the database to form a basic knowledge graph. In this embodiment, the graph database Neo4j is used for storage. First, the entities are stored as nodes in the graph database, and each node contains the entity name and knowledge attributes; then, the relationship between entities is stored as an edge in the graph database, and each edge contains attributes such as the relationship name and relationship type; indexes are established through the Cypher query language to optimize query performance;

[0109] In the knowledge correction stage (S25), the above basic knowledge graph is corrected using a collaborative process.

[0110] Step S25 includes the following steps:

[0111] In the preparation phase (S251), the reviser creates an independent branch when performing updates;

[0112] In the modification stage (S252), the modifier performs content modification operations on the knowledge graph on its branch;

[0113] In the merge application phase (S253), after completing the update, the reviser initiates a pull request to merge the changes into the main branch;

[0114] During the review phase (S254), the auditor reviews the changes and discusses and modifies them if necessary;

[0115] In the merging stage (S255), after the review is passed, the reviewer will merge the changes into the main branch to complete the revision of the knowledge graph.

[0116] In the data construction phase (S3), relevant information is collected from open websites based on the keywords of the entities, and update strategies are configured to establish data attributes for each entity;

[0117] As attached Figure 4 As shown, step S3 includes the following steps:

[0118] In the information collection stage (S31), the information collection strategy is configured to collect text information related to the entity from open websites, including media reports, user feedback, industry policies, etc. In this embodiment, the collection period is set to 8:00 to 10:00 a.m. every day, the collection depth is not more than 3 layers, and the request interval is 2 seconds. Then, corresponding parsing rules are configured for different types of data sources such as news websites and social media, and the text content is extracted through XPath, regular expressions, etc. The collected content includes: obtaining media reports and industry policies from news websites, and obtaining user comments and feedback from social media;

[0119] In the information deduplication stage (S32), the MD5 hash algorithm is used to deduplicate the text information. In this embodiment, each text content is first preprocessed, including removing HTML tags, blank characters and special symbols; then the MD5 algorithm is used to calculate the hash value of the processed text to generate a 32-bit string; the hash value of the current text is compared with the hash value stored in the database. If the same hash value exists, it is determined to be duplicate content and discarded. If not, the text and its hash value are stored in the database to achieve incremental deduplication;

[0120] In the data set establishment stage (S33), the deduplication results are stored in the database, and a multi-dimensional index is established to form a data set of data attributes;

[0121] In the data update phase (S34), a sliding time window update strategy is adopted. In this embodiment, a sliding time window strategy with a length of 7 days is adopted for updating. From 8:00 to 10:00 a.m. every day, the system automatically adds newly collected data to the time window and deletes unused data outside the window range. For the data in the window, it is sorted based on the data timestamp to ensure the time sequence of the data;

[0122] In the data subset establishment stage (S35), information is filtered from the entire data set according to specific conditions to form a data subset. In this embodiment, a customer complaint data subset is established for a certain company entity. First, a complaint keyword dictionary is established, which includes common complaint terms such as "complaint", "refund", "quality issue", and "after-sales"; then, the dictionary is used to scan the entire data set of the entity to extract information containing at least 2 complaint keywords; then, the sentiment analysis model is used to classify the information and select information with negative sentiment tendency; finally, the filtered information is sorted by timestamp to form a customer complaint data subset;

[0123] In the data label establishment stage (S36), labels or classification information are added to the elements in the data subset to form data labels corresponding to the data subset. In this embodiment, the labels of the customer complaint data subset are divided into two categories: customer complaint and non-customer complaint.

[0124] In the capability building stage (S4), according to the decision-making capability requirements, a customized machine learning model is trained or a general large language model is configured to establish capability attributes for each entity. In this embodiment, the established capability attributes are as follows: Determine whether the input information is a customer complaint of a specified company entity.

[0125] As attached Figure 5 As shown, when a training customized machine learning model is adopted, step S4 includes the following steps:

[0126] In the data establishment phase (S411), according to the decision-making capability requirements, a suitable data subset and data label are selected or established from the data attributes. In this embodiment, the above-established data subset (S35) and data label (S36) are selected;

[0127] In the data preprocessing stage (S412), the text information in the data subset and data label is segmented and standardized, and the text is converted into a feature vector using an appropriate text vectorization method. In this embodiment, the text information is first cleaned, including removing HTML tags, special characters and extra spaces; then the cleaned text is segmented using the jieba segmentation tool, and a custom dictionary is loaded to improve the accuracy of the segmentation of professional terms; the segmentation results are standardized, including removing stop words, unifying the case, and processing synonyms and near-synonyms; finally, the pre-trained Word2Vec model is used to convert the processed word sequence into a word vector, each word is mapped to a 300-dimensional vector space, and the average value of the word vector is used as the vector representation of the text. Among them, the training parameters of the Word2Vec model are set as follows: window size is 5, minimum word frequency is 3, number of iterations is 100, and the CBOW model architecture is adopted;

[0128] In the model training stage (S413), the above data is split into a training set and a validation set, and a neural network model is trained to obtain a decision model. In this embodiment, an LSTM model with Multi-Head-Attention is used, and the input layer of the model receives a 300-dimensional word vector sequence x = {x1, x2, ..., x n}, where n represents the sequence length; the LSTM layer is configured with 128 hidden units, and its state update at time step t can be expressed as follows:

[0129] h t =LSTM ( x t ,h t-1 )

[0130] Among them, h t Represents the hidden state output at time t;

[0131] Furthermore, the Multi-Head-Attention layer sets 8 attention heads, and the calculation formula for each attention head is as follows:

[0132]

[0133] Among them, Q, K, and V are query matrix, key matrix, and value matrix respectively. k is the dimension of the key vector. The final output of the multi-head attention is obtained by concatenating the outputs of each attention head through a concat operation and then performing a linear transformation.

[0134] Furthermore, the fully connected layer contains 64 neurons and uses ReLU(x)=max(0,x) as the activation function;

[0135] Furthermore, the output layer contains 2 neurons, and the Softmax function is used for normalization to obtain the category probability distribution, and the formula is as follows:

[0136]

[0137] Among them, z i represents the output value of the i-th neuron in the output layer, N represents the total number of categories (here N = 2), and P(y = i|x represents the probability that the input x belongs to the i-th category;

[0138] Furthermore, in the model training process, the cross entropy loss function is used as the optimization target, and the formula is as follows:

[0139]

[0140] Among them, N is the number of sample categories, y i is the true label, p i is the predicted probability;

[0141] Furthermore, the optimizer uses the Adam algorithm and sets the initial learning rate to 0.001, where the momentum decay rate β1 = 0.9 and the second-order moment estimation decay rate β2 = 0.999. The training process is carried out in mini-batch mode, with each batch containing 64 samples and a total of 50 epochs. At the same time, the early stopping strategy is introduced. When the loss value on the validation set does not improve within 5 consecutive epochs, the training process is terminated to prevent the occurrence of model overfitting.

[0142] In the model performance evaluation stage (S414), the accuracy, precision, recall and F1 score are calculated to generate a model performance evaluation report. In this embodiment, the classification accuracy is first calculated, that is, the number of correctly identified samples (including correctly identified customer complaints and non-customer complaints) divided by the total number of samples; then the precision is calculated, that is, the number of correctly identified customer complaints divided by the number of all samples identified as customer complaints; then the recall is calculated, that is, the number of correctly identified customer complaints divided by the number of all actual customer complaint samples; finally, the F1 score is calculated, that is, the harmonic mean of the precision and recall. The evaluation uses 5-fold cross-validation and is tested on 1,000 annotated samples, of which customer complaint samples account for 30%;

[0143] In the standardized interface establishment phase (S415), the above decision model is placed in the capability attributes of the corresponding entity, the capability description information and performance evaluation report are supplemented, and a standardized interface and permission policy are established for the model. In this embodiment, the customer complaint identification model is first encapsulated as a REST API interface, and the request is received using the POST method. The request body contains parameters such as the text to be analyzed and the target entity ID, and the classification result and confidence level in JSON format are returned. Then, the description information of the model is added to the capability attributes of the entity, including the purpose of the model, input and output formats, performance indicators, etc.

[0144] As attached Figure 6 As shown, when a general large language model is configured, step S4 includes the following steps:

[0145] In the prompt word establishment stage (S421), according to the decision-making ability requirements, a structured prompt template containing task description, context information and constraints is constructed to form a standardized functional prompt for the large language model. In this embodiment, the role of the model is first clarified as a customer complaint identification expert, and then it is explained that it is necessary to determine whether the input text contains customer complaints about a specific company; then the judgment criteria are provided, including that the text must clearly mention the target company, express negative feedback, the complainant is a customer, and the complaint content is directly related to the company; at the same time, the output format is set to JSON, including fields such as complaint judgment results, confidence, judgment basis, mentioned entities and complaint types; in addition, the background information of the target company (such as the industry to which it belongs, main business, etc.) needs to be supplemented, and the constraints (such as only focusing on directly related complaints, excluding competitor complaints, etc.) need to be clarified; finally, specific example inputs are given to help the model better understand the task requirements.

[0146] In the data establishment phase (S422), according to the decision-making capability requirements, a suitable data subset and data label are selected or established from the data attributes. In this embodiment, the above-established data subset (S35) and data label (S36) are selected;

[0147] In the model performance evaluation stage (S423), the accuracy, precision, recall and F1 scores are calculated to generate a model performance evaluation report. In this embodiment, the classification accuracy is first calculated, that is, the number of correctly identified samples (including correctly identified customer complaints and non-customer complaints) divided by the total number of samples; then the precision is calculated, that is, the number of correctly identified customer complaints divided by the number of all samples identified as customer complaints; then the recall is calculated, that is, the number of correctly identified customer complaints divided by the number of all actual customer complaint samples; finally, the F1 score is calculated, that is, the harmonic mean of the precision and recall. The evaluation uses 5-fold cross validation and is tested on 1,000 annotated samples, of which customer complaint samples account for 30%;

[0148] In the standardized interface establishment phase (S424), the above decision model is placed in the capability attributes of the corresponding entity, the capability description information and performance evaluation report are supplemented, and a standardized interface and permission policy are established for the model. In this embodiment, the customer complaint identification model is first encapsulated as a REST API interface, and the request is received using the POST method. The request body contains parameters such as the text to be analyzed and the target entity ID, and the classification result and confidence level in JSON format are returned. Then, the description information of the model is added to the capability attributes of the entity, including the purpose of the model, input and output formats, performance indicators, etc.

[0149] In summary, the beneficial effects of the present invention are: on the basis of the traditional knowledge graph, the unification of knowledge storage, decision execution and data update is achieved. The decision support capability of the knowledge graph is significantly enhanced, which can provide effective support for the intelligent decision-making of enterprises.

[0150] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A knowledge graph model integrating knowledge, capabilities and data, characterized in that: include: Concept module, which is used to define the basic architecture of the knowledge graph; Entity module, used to represent entity nodes in the knowledge graph; Relationship module, used to represent the connection relationship between entities; Each entity node in the entity module contains three types of attributes: Knowledge attributes, used to store static feature information related to entities; Capability attributes, which are used to execute decision logic associated with the entity; Data attributes, which are used to manage dynamic data associated with an entity; The knowledge attributes, capability attributes and data attributes are associated with each other through a unified identifier to achieve collaborative operations of knowledge storage, decision execution and new information accommodation.

2. The knowledge graph model according to claim 1, characterized in that: The capability attributes include: Decision models, based on custom machine learning models or general large language models, are used to perform reasoning decisions; A standardized interface is used to realize the interaction between the decision model and external systems.

3. The knowledge graph model according to claim 1, characterized in that: The data attributes include: The complete set of data uses a distributed storage architecture to store all data related to a specific entity; A data subset is a data set selected from the entire data set according to specific conditions or objectives; Data labeling, adding labels or classification information to the elements in the data subset; The full data set, data subset and data tags are associated with each other through an indexing mechanism to support real-time updating and efficient retrieval of data.

4. A method for constructing a knowledge graph integrating knowledge, capabilities and data, characterized in that: The following steps are involved: S1, definition phase, defining the core concepts of the domain and the main relationships between entities based on expert experience; S2, knowledge construction stage, extracting entity names, knowledge attributes of entities and relationships between entities from open websites to form a basic knowledge graph; S3, data construction phase, collects relevant information from open websites based on the keywords of the entity, configures the update strategy, and establishes data attributes for each entity; S4. In the capability building phase, according to the decision-making capability requirements, a customized machine learning model is trained or a general large language model is configured to establish capability attributes for each entity.

5. The knowledge graph construction method according to claim 4, characterized in that: The step S1 specifically includes: S11, domain analysis stage, through expert experience and literature research, determine the application field and core objectives of the knowledge graph; S12, Concept design stage, the core concept system in the design field, including the hierarchical structure of concepts and the relationship between concepts; S13. In the entity relationship design phase, define the relationship types between entities, including the directionality and constraints of the relationships.

6. The knowledge graph construction method according to claim 4, characterized in that: The step S2 specifically includes: S21, information collection stage, collecting structured and semi-structured information related to the entity from data sources such as open websites and knowledge bases; S22, knowledge extraction stage, design hierarchical regular expressions to extract entity names, knowledge attributes of entities and relationships between entities; S23, entity alignment stage, calculate the feature vector of each entity according to the entity name and the knowledge attribute of the entity, and merge entities with similar feature vectors; S24, knowledge storage stage, storing the above results in a database to form a basic knowledge graph; S25, knowledge correction stage, uses collaborative process to correct the above basic knowledge graph.

7. The knowledge graph construction method according to claim 6, characterized in that: The step S25 specifically includes: S251, in the preparation phase, the reviser creates an independent branch when making updates; S252, modification stage, the modifier performs content modification operations on the knowledge graph on its branch; S253, Merge Application Phase: After completing the update, the reviser initiates a pull request to merge the changes into the main branch; S254, During the review phase, auditors review the changes and discuss and modify them when necessary; S255, merging stage. After the review is passed, the reviewer will merge the changes into the main branch to complete the revision of the knowledge graph.

8. The knowledge graph construction method according to claim 4, characterized in that: The step S3 specifically includes: S31, information collection stage: configure information collection strategies to collect text information related to the entity from open websites, including media reports, user feedback, industry policies, etc. S32, information deduplication stage, using MD5 hash algorithm to dedupe text information; S33, in the stage of establishing the complete set of data, the deduplication results are stored in the database, and a multi-dimensional index is established to form the complete set of data attributes; S34, the data set update phase, adopts the sliding time window update strategy; S35, data subset establishment stage, filtering information from the entire data set according to specific conditions to form a data subset; S36: Data label establishment phase: adding labels or classification information to the elements in the data subset to form a data label corresponding to this data subset.

9. The knowledge graph construction method according to claim 4, characterized in that: When a training customized machine learning model is adopted, the step S4 specifically includes: S411, data establishment phase, select or establish appropriate data subsets and data labels from data attributes according to decision-making capability requirements; S412, in a data preprocessing stage, the text information in the data subset and the data label is segmented and standardized, and the text is converted into a feature vector using an appropriate text vectorization method; S413, in the model training stage, the data is split into a training set and a validation set, a neural network model is trained, and a decision model is obtained; S414, model performance evaluation phase, calculate accuracy, precision, recall and F1 score, and generate a model performance evaluation report; S415, in the standardized interface establishment phase, the above decision model is placed into the capability attributes of the corresponding entity, the capability description information and performance evaluation report are supplemented, and a standardized interface and permission strategy are established for the model.

10. The knowledge graph construction method according to claim 4, characterized in that: When a general large language model is configured, step S4 specifically includes: S421, prompt word establishment stage, according to the decision-making ability requirements, construct a structured prompt template containing task description, context information and constraints to form a standardized functional prompt of the large language model; S422, data establishment phase, selecting or establishing appropriate data subsets and data labels from data attributes according to decision-making capability requirements; S423, model performance evaluation stage, calculate accuracy, precision, recall and F1 score, and generate a model performance evaluation report; S424, in the standardized interface establishment phase, the above decision model is placed into the capability attributes of the corresponding entity, the capability description information and performance evaluation report are supplemented, and a standardized interface and permission strategy are established for the model.

Images