A knowledge base model training method and system based on AI big model technology

Abstract

The present invention discloses a knowledge base model training method and system based on AI big model technology, which relates to the field of computer technology. The method includes: conducting demand analysis in combination with actual business scenarios, determining knowledge base training indicators according to the analysis results; training and optimizing knowledge base models based on knowledge base training indicators and actual business data; improving and updating knowledge base models by using external data access; and adding a timeliness filtering mechanism to the knowledge base model. The knowledge base is customized based on business needs, and a new knowledge access API is provided to effectively solve the cold start problem and knowledge timeliness problem of the knowledge base model.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a knowledge base model training method and system based on AI big model technology. Background Art

[0002] The knowledge base model is a method that uses machine learning technology to extract knowledge from a large amount of data and construct a structured knowledge base. This model has been widely used in information retrieval, question-answering systems, recommendation systems, and other fields. However, the current knowledge base model still has some defects that cannot be ignored, such as the cold start problem and the knowledge timeliness problem. The cold start problem refers to the fact that the knowledge base model is often difficult to effectively handle new or uncommon entities and concepts; the knowledge timeliness problem refers to the fact that the knowledge base needs to be constantly updated to reflect changes in the real world, but the existing models face challenges in the dynamic updating of knowledge. Summary of the invention

[0003] The present invention provides a knowledge base model training method based on AI big model technology, comprising:

[0004] Step 1: Conduct demand analysis based on actual business scenarios and determine knowledge base training indicators based on the analysis results;

[0005] Step 2: Train and optimize the knowledge base model based on knowledge base training indicators and actual business data;

[0006] Step 3: Use external data access to improve and update the knowledge base model;

[0007] Step 4. Add a timeliness filtering mechanism to the knowledge base model.

[0008] The knowledge base model training method based on the AI ​​big model technology described above, in which the knowledge base model training and optimization are performed based on the knowledge base training indicators and actual business data, specifically includes the following sub-steps:

[0009] A data lake that extracts entities from actual business data and builds a knowledge base data structure;

[0010] Create a knowledge base model;

[0011] Initialize knowledge base model parameters using data in the data lake based on user needs;

[0012] The model parameters are further optimized based on the evaluation results of the knowledge base model based on the training indicators.

[0013] The knowledge base model training method based on the AI ​​big model technology described above, in which the knowledge base model parameters are initialized using the data in the data lake according to user needs, specifically includes the following sub-steps:

[0014] Initialize the statistical functions in the knowledge base model based on user needs and data in the data lake;

[0015] The historical statistical data of the predicted items are used as training data to train the knowledge base model and initialize the model's hyperparameters.

[0016] As described above, a knowledge base model training method based on AI big model technology, in which the statistical function in the knowledge base model is initialized according to user needs and data in the data lake, specifically includes the following sub-steps:

[0017] Create a table of empty parameter statistics functions and associate each empty parameter statistics function with its commonly used demand expression language;

[0018] Extract the demand expression language in the user demand and retrieve the empty parameter statistical function associated with the expression language;

[0019] Extract the statistical items in the user requirements and assemble them with the empty parameter statistical function into a complete statistical function.

[0020] The knowledge base model training method based on the AI ​​big model technology as described above, wherein the model parameters are further optimized based on the evaluation results of the knowledge base model based on the training indicators, specifically includes the following sub-steps:

[0021] Create a function to calculate the performance value of training indicators;

[0022] Evaluate the knowledge base model using the training indicator performance value calculation function;

[0023] Optimize the hyperparameters in the model based on the model evaluation results.

[0024] The knowledge base model training method based on AI big model technology as described above, in which the knowledge base model is improved and updated by using external data access, specifically includes the following sub-steps:

[0025] Pull new data regularly through the external data access interface;

[0026] Identify entities in new data and integrate them with entities in the data lake;

[0027] Incrementally train the knowledge base model using the consolidated entities in the data lake.

[0028] The present invention also provides a knowledge base model training system based on AI big model technology, including: a human-computer interaction module, a knowledge storage module, a knowledge base model training module, and an external data access module;

[0029] The human-computer interaction module is used to analyze the needs of users using computer language technology, and use the model trained by the knowledge base model training module to output results that meet user needs;

[0030] The knowledge storage module is used to store entities extracted from actual business data and trained knowledge base models;

[0031] The knowledge base model training module is used to use the business data in the knowledge storage module to produce an output model that meets user needs;

[0032] The external data access module is used to improve and update the entities in the knowledge base model and the knowledge storage module by using external data access.

[0033] The knowledge base model training system based on AI big model technology as described above, wherein the knowledge base model training module includes a knowledge base model initialization submodule and a knowledge base model optimization submodule;

[0034] The knowledge base model initialization submodule is used to initialize the knowledge base model parameters using the data in the data lake according to user needs;

[0035] The knowledge base model optimization submodule is used to evaluate the knowledge base model and further optimize the model parameters based on the evaluation results.

[0036] As described above, the knowledge base model training system based on AI big model technology provides three services, namely data statistics service, problem solving service, and data prediction service, and also provides timeliness filtering function for training data.

[0037] As described above, the knowledge base model training system based on AI big model technology, wherein in the external data access module, improving and updating the entities in the knowledge storage module specifically includes the following sub-steps:

[0038] Extract entities, their attributes, and relationships from new data;

[0039] Calculate similarity based on the attributes and associations between new and old entities;

[0040] If the similarity is higher than 95%, they all point to the same comprehensive entity.

[0041] The beneficial effects achieved by the present invention are as follows: a customized training knowledge base is established based on business needs, and a new knowledge access API is provided, which effectively solves the cold start problem of the knowledge base model and the knowledge timeliness problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0043] Figure 1 This is a flow chart of a knowledge base model training method based on AI big model technology provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0044] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0045] Embodiment 1

[0046] like Figure 1 As shown, the first embodiment of the present invention provides a knowledge base model training method based on AI big model technology, including:

[0047] Step S10: Perform demand analysis based on actual business scenarios and determine knowledge base training indicators based on the analysis results;

[0048] Understand users' needs and expectations, including what kind of information and services they hope to obtain from the knowledge base. We have designed three types of services: data statistics service, problem-solving service, and data prediction service. Different services have different focuses. Data statistics service focuses on data calculation and aggregation capabilities, problem-solving service focuses on data retrieval and matching capabilities, and data prediction service focuses on data trend analysis capabilities. The quantitative form of these focuses is training indicators, which are used to evaluate the performance of knowledge base models in actual business scenarios.

[0049] Training indicators are customized by users according to their own needs and expectations. The indicator definition content includes indicator name, indicator weight, and indicator calculation formula. The data items in the indicator calculation formula come from two parts: model output results and training data fields.

[0050] Step S20: training and optimizing the knowledge base model based on the knowledge base training indicators and actual business data;

[0051] Actual business data refers to the business data generated within the enterprise. These business data are then processed into the data structure required by the knowledge base model. The processed data is then used to train the knowledge base model. Specifically:

[0052] Step S21: extracting entities from actual business data and establishing a data lake of knowledge base data structure;

[0053] For structured data, the extracted fields are entities, the field values ​​are attributes, and the associations between fields are the connections between entities; for semi-structured data, regular expressions are used to obtain field entities and field values; for unstructured data, such as long texts, rule templates can be used to extract entities and the connections between entities. The extracted data is stored in the data lake of the knowledge base in a graph structure.

[0054] Step S22: creating a knowledge base model;

[0055] The knowledge base model is expressed as:

[0056] where λ 1 ~λ n is the regression coefficient, F 1 (P)~F n (P) is a variety of statistical functions, γ1~γn are the regression orders of different statistical items, c is the variation that cannot be explained by the model, n is the total number of statistical items, P is the input set, T(P) is the entity extraction function, O(P) is the attribute extraction function, G(P) is the relationship extraction function, σ is the data processing identifier, 0 represents the numerical processing method, 1 represents the text processing method, when σ=0, then execute λ 1 F 1 (P) γ1 +λ 2 F 2 (P) γ2 +Λ+λ n F n (P) γn , when σ=1, T(P)+O(P)+G(P) is executed.

[0057] In actual use, the data statistics service will assemble F according to the statistical method mentioned in the input box. 1 (P) function. For example, if the user wants to summarize the sales of mineral water this week, the sum function is needed. The sum is the daily sales of mineral water from Monday to now, so F 1 (P) is a summation function, where P is the summation term; the data prediction service will convert the lambda in the model 1 ~λ n , γ1~γn and c are used as hyperparameters to participate in model training. When the data prediction service is not enabled, λ 1 ~λ n , the parameters γ1~γn are all set to the default value 1, and the default value of c is 0; the problem-solving service is implemented based on the three functions T(P)+O(P)+G(P).

[0058] Step S23: Initialize the knowledge base model parameters using the data in the data lake according to user needs, which specifically includes the following sub-steps:

[0059] Step S231: Initialize the statistical functions in the knowledge base model according to user needs and the data in the data lake;

[0060] First, we need to extract statistical items or prediction items from user needs. If it is a prediction requirement, we need to determine whether the prediction item can be directly retrieved from the data lake. If it can be, the statistical function is a direct return function. If not, we need to first inquire about the statistical method of the prediction item, and then assemble the statistical function according to the statistical method, and use the statistical function to obtain the prediction item. The assembly steps of the statistical function are as follows:

[0061] Step S2311: Create a table of empty parameter statistics functions and associate each empty parameter statistics function with its commonly used demand expression language;

[0062] The empty parameter statistical function table stores commonly used statistical functions, including sum, average, square difference, etc. It also supports custom statistical functions. Users can create them according to business needs and associate them with commonly used demand expressions to ensure that employees can use concise and common expressions to obtain statistical data and complete data statistics in a conversational manner.

[0063] Step S2312: extracting the demand expression language in the user demand, and retrieving the empty parameter statistical function associated with the expression language;

[0064] Step S2313: extract the statistical items in the user requirements and assemble them with the empty parameter statistical function into a complete statistical function.

[0065] Step S232: using the historical statistical data of the prediction items as training data to train the knowledge base model and initialize the hyperparameters of the model;

[0066] The relevant data of the prediction items are serialized according to the recording time, and then grouped according to the prediction period. The assembled statistical function is used to obtain the statistical value of the prediction item at each timestamp in the group as the statistical function F in the model. 1 (P)~F n The return value of (P) corresponds to the first value of the next group of data as the output value. The grouped data is used to train the knowledge base model and initialize the hyperparameters in the knowledge base model.

[0067] Step S24: Further optimize the model parameters based on the evaluation results of the knowledge base model based on the training indicators, which specifically includes the following sub-steps:

[0068] Step S241: creating a training indicator performance value calculation function;

[0069] The training indicator performance value calculation function is expressed as:

[0070] , where Q j(C) is the calculation function defined when creating the jth indicator, δ j is the target value of the jth training indicator, is the standardization parameter of the jth training indicator, μ j The weight set for the jth training indicator, j ranges from 1 to m, m is the number of training indicators, and C is the input parameter set.

[0071] Step S242: Evaluate the knowledge base model using the training indicator performance value calculation function;

[0072] First, a set of data is extracted from the training samples as the validation set, and the model output and other training indicators are passed into the training indicator performance value calculation function, and the training indicator performance value is returned. If the performance value is lower than 9, it means that the knowledge base model performs poorly.

[0073] Step S243: Optimize the hyper-parameters in the model according to the model evaluation results;

[0074] Retrain the data samples where the model performs poorly, and output the evaluation results in real time during the training process. Observe the changing trend of the evaluation results during hyperparameter iteration, and make gradual adjustments until the evaluation results meet expectations.

[0075] Step S25: the optimized knowledge base model is labeled and stored according to its purpose;

[0076] The stored knowledge base model can be used directly next time the same requirement occurs, saving time for re-analysis and generation and improving user experience.

[0077] The working process of the knowledge base model in actual use is as follows: first, the user selects the service to be used and enters his or her own requirements from the input box; the knowledge base extracts the key information in the requirements and enters the corresponding execution body of the knowledge base model according to the selected service; the knowledge base model extracts the data related to the requirements and uses the execution body to output the processing results; the language is organized according to the response template and returned to the user through a dialog box.

[0078] Step S30: improving and updating the knowledge base model by using external data access;

[0079] Because entities and knowledge in the real world are constantly changing, an external data access API is needed to update the entity and knowledge base models in the data lake to ensure the accuracy and practicality of the knowledge. The specific steps include the following:

[0080] Step S31: regularly pull new data through the external data access interface;

[0081] When pulling data, you must first verify the security of the data source, which includes but is not limited to government open data, academic research data, purchased commercial data, or shared data from partners.

[0082] Step S32: Identify entities in the new data and integrate them with entities in the data lake;

[0083] The entity extraction rules for new data are consistent with the entity extraction rules for business data in step S21. After the extraction is completed, the attributes and relationships of the same entities in the data lake are integrated. However, due to the update and iteration of knowledge, some entities between the new and old data are essentially the same but have different names. Therefore, a dynamic rule for entity pointing needs to be added during integration, that is, the similarity between entities is calculated. If the similarity between entities in the new and old data exceeds 95%, they point to the same comprehensive entity. The comprehensive entity shares all the attributes and relationships of the new and old entities. In this way, no matter whether the user searches with the new entity name or the old entity name, all the data can be obtained. The formula for calculating the similarity between entities is expressed as:

[0084] where a t is the tth attribute value of data lake entity a, t ranges from 1 to w, w is the number of data lake entity attributes, a e is the vector of the e-th entity associated with data lake entity a, where e ranges from 1 to z, and z is the number of entities associated with data lake entity a. d is the dth attribute value of the new entity b, d ranges from 1 to r, r is the number of attributes of the new entity, b f is the vector of the fth entity associated with the new entity b, f ranges from 1 to p, and p is the number of entities associated with the new entity b.

[0085] Step S33: Use the integrated entities in the data lake to incrementally train the knowledge base model.

[0086] Step S40: adding a timeliness filtering mechanism to the knowledge base model;

[0087] The basis of the timeliness filtering mechanism is that the data in the data lake are marked with timestamps, and users can filter the entities in the data lake according to timeliness when using the knowledge base.

[0088] Embodiment 2

[0089] Embodiment 2 of the present invention provides a knowledge base model training system based on AI big model technology, including: a human-computer interaction module, a knowledge storage module, a knowledge base model training module, and an external data access module;

[0090] (1) Human-computer interaction module, which is used to analyze the needs proposed by users using computer language technology and output results that meet the user's needs using the model trained by the knowledge base model training module;

[0091] In the human-computer interaction module, users can choose the services they want to use, including data statistics service, problem solving service, and data prediction service. They can also filter the usage data according to timeliness, that is, filter the training data of the knowledge model library in the form of a date selection box according to timeliness requirements. The knowledge base model trained with this data can only display the data status within the selected time. The human-computer interaction module determines whether there is a trained model that can meet the user's needs. If it can, it will be used directly. If not, the demand will be transmitted to the knowledge base model training module for training of a new model, and the trained model will be used to output the results that meet the user's needs. In the human-computer interaction module, users can also set training indicators for knowledge base model training to evaluate the quality of the model.

[0092] (2) Knowledge storage module, used to store entities extracted from actual business data and trained knowledge base models;

[0093] The data lake extracts data entities from actual business data and builds the knowledge base data structure. For structured data, the extracted fields are entities, field values ​​are attributes, and the associations between fields are the connections between entities. For semi-structured data, regular expressions are used to obtain field entities and field values. For unstructured data, such as long texts, rule templates can be used to extract entities and connections between entities. The extracted data is stored in the knowledge base data lake in a graph structure. The trained knowledge base model is also stored in the knowledge storage module according to the purpose identification. The stored knowledge base model can be used directly the next time the same demand is met, saving the time of re-analysis and generation, and improving the user experience.

[0094] (3) a knowledge base model training module, which is used to use the business data in the knowledge storage module to produce an output model that meets user needs, including a knowledge base model initialization submodule and a knowledge base model optimization submodule;

[0095] 1. The knowledge base model initialization submodule is used to initialize the knowledge base model parameters using the data in the data lake according to user needs;

[0096] The knowledge base model is expressed as:

[0097] where λ 1 ~λ n is the regression coefficient, F 1 (P)~F n(P) is a variety of statistical functions, γ1~γn are the regression orders of different statistical items, c is the variation that cannot be explained by the model, n is the total number of statistical items, P is the input set, T(P) is the entity extraction function, O(P) is the attribute extraction function, G(P) is the relationship extraction function, σ is the data processing identifier, 0 represents the numerical processing method, 1 represents the text processing method, when σ=0, then execute λ 1 F 1 (P) γ1 +λ 2 F 2 (P) γ2 +Λ+λ n F n (P) γn , when σ=1, T(P)+O(P)+G(P) is executed.

[0098] Initialize the knowledge base model parameters, which includes the following sub-steps:

[0099] I. Initialize the statistical functions in the knowledge base model based on user needs and data in the data lake;

[0100] First, we need to extract statistical items or prediction items from user needs. If it is a prediction requirement, we need to determine whether the prediction item can be directly retrieved from the data lake. If it can be, the statistical function is a direct return function. If not, we need to first inquire about the statistical method of the prediction item, and then assemble the statistical function according to the statistical method, and use the statistical function to obtain the prediction item. The assembly steps of the statistical function are as follows:

[0101] i. Create a table of empty parameter statistics functions and associate each empty parameter statistics function with its commonly used demand expression language;

[0102] The empty parameter statistical function table stores commonly used statistical functions, including sum, average, square difference, etc. It also supports custom statistical functions. Users can create them according to business needs and associate them with commonly used demand expressions to ensure that employees can use concise and common expressions to obtain statistical data and complete data statistics in a conversational manner.

[0103] ii. Extract the demand expression language in the user demand and retrieve the empty parameter statistical function associated with the expression language;

[0104] iii. Extract the statistical items in the user requirements and assemble them with the empty parameter statistical function into a complete statistical function.

[0105] II. Use the historical statistical data of the predicted items as training data to train the knowledge base model and initialize the hyperparameters of the model;

[0106] The relevant data of the prediction items are serialized according to the recording time, and then grouped according to the prediction period. The assembled statistical function is used to obtain the statistical value of the prediction item at each timestamp in the group as the statistical function F in the model. 1 (P)~F n The return value of (P) corresponds to the first value of the next group of data as the output value. The grouped data is used to train the knowledge base model and initialize the hyperparameters in the knowledge base model.

[0107] 2. The knowledge base model optimization submodule is used to evaluate the knowledge base model and further optimize the model parameters based on the evaluation results. It specifically includes the following substeps:

[0108] I. Create a function to calculate the performance value of the training indicator;

[0109] The training indicator performance value calculation function is expressed as:

[0110] Where Q j (C) is the calculation function defined when creating the jth indicator, δ j is the target value of the jth training indicator, is the standardization parameter of the jth training indicator, μ j The weight set for the jth training indicator, j ranges from 1 to m, m is the number of training indicators, and C is the input parameter set.

[0111] II. Evaluate the knowledge base model using the training indicator performance value calculation function;

[0112] First, a set of data is extracted from the training samples as the validation set, and the model output and other training indicators are passed into the training indicator performance value calculation function, and the training indicator performance value is returned. If the performance value is lower than 9, it means that the knowledge base model performs poorly.

[0113] III. Optimize the hyperparameters in the model based on the model evaluation results;

[0114] Retrain the data samples where the model performs poorly, and output the evaluation results in real time during the training process. Observe the changing trend of the evaluation results during hyperparameter iteration, and make gradual adjustments until the evaluation results meet expectations.

[0115] (4) External data access module, used to improve and update entities in the knowledge base model and the data lake by using external data access;

[0116] Because entities and knowledge in the real world are constantly changing, an external data access API is needed to update the entity and knowledge base models in the data lake to ensure the accuracy and practicality of the knowledge. The specific steps include the following:

[0117] I. Regularly pull new data through the external data access interface;

[0118] When pulling data, you must first verify the security of the data source, which includes but is not limited to government open data, academic research data, purchased commercial data, or shared data from partners.

[0119] II. Identify entities in the new data and integrate them with entities in the data lake;

[0120] Extract entities and their attributes and relationships from new data. After the extraction is completed, integrate the attributes and relationships of the same entities in the data lake. However, due to the update and iteration of knowledge, some entities between the new and old data are essentially the same but have different names. Therefore, a dynamic rule for entity pointing needs to be added during integration, that is, the similarity between entities is calculated. If the similarity between entities in the new and old data exceeds 95%, they point to the same comprehensive entity. The comprehensive entity shares all the attributes and relationships of the new and old entities. In this way, no matter whether the user searches with the new entity name or the old entity name, all data can be obtained.

[0121] The formula for calculating the similarity between entities is expressed as:

[0122] where a t is the tth attribute value of data lake entity a, t ranges from 1 to w, w is the number of data lake entity attributes, a e is the vector of the e-th entity associated with data lake entity a, where e ranges from 1 to z, and z is the number of entities associated with data lake entity a. d is the dth attribute value of the new entity b, d ranges from 1 to r, r is the number of attributes of the new entity, b f is the vector of the fth entity associated with the new entity b, f ranges from 1 to p, and p is the number of entities associated with the new entity b.

[0123] III. Incrementally train the knowledge base model using the integrated entities in the data lake.

[0124] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A knowledge base model training method based on AI large model technology, characterized in that: include: Step 1: Conduct demand analysis based on actual business scenarios and determine knowledge base training indicators based on the analysis results; Step 2: Train and optimize the knowledge base model based on the knowledge base training indicators and actual business data, which includes the following sub-steps: A data lake that extracts entities from actual business data and builds a knowledge base data structure; Create a knowledge base model; The knowledge base model is expressed as: ,in is the regression coefficient, For different statistical functions, are the regression orders of different statistical items, c is the variation that cannot be explained by the model, n is the total number of statistical items, P is the input set, T(P) is the entity extraction function, O(P) is the attribute extraction function, G(P) is the relationship extraction function, is the data processing identifier, 0 indicates the numerical processing method, 1 indicates the text processing method, when =0, then execute ,when =1, execute T(P)+O(P)+G(P); Initialize the knowledge base model parameters using the data in the data lake according to user needs, specifically: Initialize the statistical functions in the knowledge base model based on user needs and data in the data lake. First, create a table of empty parameter statistical functions and associate each empty parameter statistical function with its commonly used demand expression language. Then extract the demand expression language in the user needs and retrieve the empty parameter statistical function associated with the expression language. Finally, extract the statistical items in the user needs and assemble them with the empty parameter statistical function into a complete statistical function. Use the historical statistical data of the predicted items as training data to train the knowledge base model and initialize the model's hyperparameters; Further optimize the model parameters based on the evaluation results of the knowledge base model based on the training indicators; Step 3: Use external data access to improve and update the knowledge base model; Step 4. Add a timeliness filtering mechanism to the knowledge base model.

2. The knowledge base model training method based on AI large model technology according to claim 1 is characterized in that: The model parameters are further optimized based on the evaluation results of the knowledge base model based on the training indicators, which specifically includes the following sub-steps: Create a function to calculate the performance value of training indicators; Evaluate the knowledge base model using the training indicator performance value calculation function; Optimize the hyperparameters in the model based on the model evaluation results.

3. The knowledge base model training method based on AI large model technology according to claim 1 is characterized in that: Use external data access to improve and update the knowledge base model, which includes the following sub-steps: Pull new data regularly through the external data access interface; Identify entities in new data and integrate them with entities in the data lake; Incrementally train the knowledge base model using the consolidated entities in the data lake.

4. A knowledge base model training system based on AI large model technology, characterized in that: include: Human-computer interaction module, knowledge storage module, knowledge base model training module, external data access module; The human-computer interaction module is used to analyze the needs proposed by users using computer language technology, and use the model trained by the knowledge base model training module to output results that meet user needs; The knowledge storage module is used to store entities extracted from actual business data and trained knowledge base models; The knowledge base model training module is used to use the business data in the knowledge storage module to produce an output model that meets user needs, including: a knowledge base model initialization submodule and a knowledge base model optimization submodule; The knowledge base model initialization submodule is used to initialize the knowledge base model parameters using the data in the data lake according to user needs; The knowledge base model is expressed as: ,in is the regression coefficient, For various statistical functions, are the regression orders of different statistical items, c is the variation that cannot be explained by the model, n is the total number of statistical items, P is the input set, T(P) is the entity extraction function, O(P) is the attribute extraction function, G(P) is the relationship extraction function, is the data processing identifier, 0 indicates the numerical processing method, 1 indicates the text processing method, when =0, then execute ,when =1, execute T(P)+O(P)+G(P); Initializing the knowledge base model parameters specifically includes: Initialize the statistical functions in the knowledge base model based on user needs and data in the data lake. First, create a table of empty parameter statistical functions and associate each empty parameter statistical function with its commonly used demand expression language. Then extract the demand expression language in the user needs and retrieve the empty parameter statistical function associated with the expression language. Finally, extract the statistical items in the user needs and assemble them with the empty parameter statistical function into a complete statistical function. Use the historical statistical data of the predicted items as training data to train the knowledge base model and initialize the model's hyperparameters; The knowledge base model optimization submodule is used to evaluate the knowledge base model and further optimize the model parameters based on the evaluation results; The external data access module is used to improve and update the entities in the knowledge base model and the knowledge storage module by using external data access.

5. The knowledge base model training system based on AI large model technology according to claim 4 is characterized in that: The system provides three services in total, namely, data statistics service, problem solving service, and data prediction service, and also provides a timeliness filtering function for training data.

6. The knowledge base model training system based on AI large model technology according to claim 4 is characterized in that: In the external data access module, improving and updating the entities in the knowledge storage module specifically includes the following sub-steps: Extract entities, their attributes, and relationships from new data; Calculate similarity based on the attributes and associations between new and old entities; If the similarity is higher than 95%, they all point to the same comprehensive entity.

Images