Large model driving type data classifying and grading system

Abstract

The invention relates to a large-model-driven data classification and grading system, belongs to the technical field of data processing, and solves the problems that an existing data classification and grading result is inaccurate and a large amount of time and manpower are consumed. Comprising a rule base, and an overall planning large model, a classification agent and a grading agent which are constructed based on LangChain; the overall planning large model comprises a task tree generation component, a task distribution component and a task tracking component; the task tree generation component extracts tasks from the rule base to construct a task tree; the task distribution component takes out the tasks from the task tree and distributes the tasks to the classification agents or the grading agents; the task tracking component receives the classification intelligent agent and the execution result of the grading intelligent agent and evaluates the classification intelligent agent and the execution result; the classification agent calls a data classification tool for classification and sends the classification result to the task tracking assembly; and the grading agent calls the data grading tool for grading, and sends a grading execution result to the task tracking component. The quality and efficiency of data classification and grading are improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a large model driven data classification and grading system. Background Art

[0002] With the development of data classification and grading technology and the improvement of relevant standards, complex and diverse business scenarios and multi-dimensional data classification and grading goals have become difficulties faced by data classification and grading.

[0003] Most of the existing data classification and grading technologies are based on pre-defined static rules. With the development of data classification and grading technologies and related standards, multi-dimensional data classification and grading has become the mainstream. When facing new business scenarios and multi-dimensional data classification and grading goals, data classification and grading technologies based on static rules require a lot of time and manpower to add and modify rules.

[0004] Moreover, the data classification results may vary greatly for similar data in different business scenarios. Conventional data classification methods only focus on the data itself, without considering the particularity of the business scenario or multi-dimensional indicators. Summary of the invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a large model driven data classification and grading system to solve the problem that the existing data classification and grading results are inaccurate and consume a lot of time and manpower.

[0006] The embodiment of the present invention provides a large model driven data classification and grading system, including: a rule base, and a comprehensive large model, a classification agent and a grading agent built on LangChain; The overall planning model includes a task tree generation component, a task distribution component and a task tracking component; the task tree generation component extracts tasks from the rule base according to the received task identifier and constructs a task tree; the task distribution component extracts pending tasks from the task tree and distributes them to the classification agent or the hierarchical agent according to the evaluation result of the task tracking component; the task tracking component receives the execution results of the classification agent and the hierarchical agent and evaluates them, and sends the evaluation results to the task distribution component; The classification agent calls the data classification tool to classify the relevant data according to the received to-do tasks, and sends the classification execution results to the task tracking component; The grading agent calls the data grading tool to grade the relevant data according to the received to-do tasks and the preset grading indicators, and sends the grading execution results to the task tracking component.

[0007] Based on further improvements of the above system, the rule base includes multiple classification tasks and hierarchical tasks, each task includes: task ID, parent task ID, child task ID, task name, data source, data description, task goal, task description, task requirements, task parameters, tool model and tool hyperparameters.

[0008] Based on the further improvement of the above system, the task tree generation component extracts tasks from the rule base according to the received task identifier to construct a task tree, with the task corresponding to the task identifier as the root task, and traverses the rule base layer by layer to take out all subtasks according to the subtask identifier of the root task until the subtask identifier is empty to obtain tree-structured data.

[0009] Based on the further improvement of the above system, the task distribution component takes out the to-do tasks from the task tree according to the evaluation results of the task tracking component, including: Initially, take out the root task from the task tree as the to-do task; When the received evaluation result is normal, the next task is taken from the task tree as a to-do task in a breadth-first manner; When the received evaluation result is timeout and the number of executions of the to-do task does not exceed the threshold, the to-do task is taken out again; When the received evaluation result is abnormal, or the number of executions of the to-do task exceeds the threshold, the to-do task and all subtasks with the to-do task as the parent task will no longer be executed, and the next task will be taken from the task tree as the to-do task in a breadth-first manner.

[0010] Based on the further improvement of the above system, the classification agent calls the data classification tool to classify the relevant data according to the received to-do tasks, including: According to the task identifier of the to-do task, identify whether the received to-do task is an existing task or a new task. If it is an existing task that has not been modified, call the data classification tool based on the to-do task to classify the relevant data; if it is a new task or a modified existing task, and the data source of the to-do task is a structured database table, identify whether the data description of the to-do task can be aligned with the metadata information of each field in the database table. If so, call the data classification tool based on the aligned to-do task to classify the relevant data; otherwise, send the misaligned exception information as the classification execution result to the task tracking component.

[0011] Based on the further improvement of the above system, the data classification tool is called to classify the relevant data, including: Using the data extraction tool bound to the first LLM model, obtain the data to be classified from the data source of the to-do tasks; When the tool model and tool hyperparameters of the to-do task are not empty, the corresponding tool model is extracted from the data classification tool bound to the first LLM model, the hyperparameters are obtained from the tool hyperparameters and set to the tool model, and the first LLM model calls the tool model to complete the data classification task for the classified data according to the task objectives of the to-do task; otherwise, classification prompt words are constructed according to the to-do task, and the first LLM model completes the data classification task according to the classification prompt words.

[0012] Based on the further improvement of the above system, it is possible to identify whether the data description of the to-do task letter can be aligned with the metadata information of each field in the database table, including: Filter out the attributes in the data description that do not exist in the metadata information of each field as the attributes to be identified; Get the similarity threshold in the task parameters of the to-do task, identify whether there is a field in the metadata information whose similarity with the attribute to be identified exceeds the similarity threshold, if so, alignment can be achieved, and the field corresponding to the maximum similarity is used to replace the attribute to be identified; otherwise, alignment cannot be achieved.

[0013] Based on the further improvement of the above system, the grading agent calls the data grading tool to grade the relevant data according to the received to-do tasks and the preset grading indicators, including: Using the data extraction tool bound to the second LLM model, obtain each attribute and its data record from the data source in the to-do task; According to the task requirements of the to-do tasks and the data records in each attribute, the scenario weight vector of each attribute on each grading indicator is obtained; According to the embedding vector representation of each attribute, obtain the indicator compliance vector of each attribute on each grading indicator; According to the scenario weight vector and indicator compliance vector of each attribute, the data grading tool is called to obtain the level of each attribute.

[0014] Based on the further improvement of the above system, the scenario weight vector of each attribute on each grading indicator is obtained by randomly taking out at least one data record corresponding to the attribute from the data source, constructing a weight allocation prompt word according to the data record, task requirements and each grading indicator, and passing it to the second LLM model.

[0015] Based on the further improvement of the above system, the indicator compliance vector of each attribute on each grading indicator is obtained, which is to pass the embedded vector representation of each attribute into the trained neural network model and output the probability distribution that meets each grading indicator.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. Trigger the overall planning model to plan and distribute pending tasks from the rule base. During the iterative execution process, the classification agent and grading agent are used to complete the classification and grading tasks in the pending tasks respectively. The overall planning model infers and evaluates the execution results. This is completed through collaboration among multiple parties. While ensuring the quality of data classification and grading results, it saves time and labor costs and improves data processing efficiency.

[0017] 2. While classifying and grading, it supports the addition of new tasks and the modification of existing tasks, realizing the automatic expansion and optimization of task rules in the rule library.

[0018] 3. Based on multi-dimensional grading indicators, the scenario weight vector and indicator compliance vector are extracted from task requirements, data content and data attributes. The data grading task is completed by fusing the two vectors, making the grading results better aligned with the business scenario, enhancing data security and reducing business risks.

[0019] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components; Figure 1 Schematic diagram of the structure of a large model driven data classification and grading system in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0022] A specific embodiment of the present invention discloses a large model-driven data classification and grading system, which is the basis for enterprises to manage data assets. Figure 1 As shown, the system includes: a rule base, a comprehensive large model, a classification agent, and a hierarchical agent built on LangChain; The overall planning model includes a task tree generation component, a task distribution component and a task tracking component; the task tree generation component extracts tasks from the rule base according to the received task identifier and constructs a task tree; the task distribution component extracts pending tasks from the task tree and distributes them to the classification agent or the hierarchical agent according to the evaluation result of the task tracking component; the task tracking component receives the execution results of the classification agent and the hierarchical agent and evaluates them, and sends the evaluation results to the task distribution component; The classification agent calls the data classification tool to classify the relevant data according to the received to-do tasks, and sends the classification execution results to the task tracking component; The grading agent calls the data grading tool to grade the relevant data according to the received to-do tasks and the preset grading indicators, and sends the grading execution results to the task tracking component.

[0023] During implementation, the overall planning and distribution of pending tasks from the rule base is carried out by triggering the overall planning model. During the iterative execution process, the classification agent and grading agent are used to complete the classification and grading tasks in the pending tasks respectively. The overall planning model infers and evaluates the execution results to ensure the quality of data classification and grading results and optimize the task rules in the rule base.

[0024] It should be noted that the rule base includes multiple classification tasks and grading tasks, each task includes: task ID, parent task ID, child task ID, task name, data source, data description, task goal, task description, task requirements, task parameters, tool model and tool hyperparameters; the task ID is automatically generated when it is saved to the rule base, and the parent task ID and child task ID are automatically associated according to the hierarchical relationship of the tasks; the data description is used to describe data information in multiple dimensions from the attribute name, attribute Chinese name, attribute type, length or precision; the task goal includes the data processing type and the data processing basis; the data processing type includes data classification and data grading; the task requirements are used to indicate the special processing requirements of the task, for example, focusing on data leakage loss, or no special requirements; the task parameters are used to indicate the parameters required by the classification agent or the grading agent during the execution of the task, such as the similarity threshold, the number of sample sampling, etc.; the tool model is used to indicate the tool model used when processing the task, and the tool model comes from the tool library. The tool types in the tool library include: general tools, data classification tools and data grading tools; general tools include but are not limited to: data extraction tools, data vectorization tools and image processing tools; data classification tools include but are not limited to: text-CNN classification model, text-SVM classification model, text-decision tree classification model and image-CNN classification model; data grading tools include but are not limited to: MLP grading model and random forest grading model; tool hyperparameters are the paths to the configuration files used to specify the hyperparameters required for the tool models.

[0025] For example, the task rules for “department employee classification” in the rule base are shown in Table 1.

[0026] Table 1 Tasks for “Department Employee Classification”

[0027] It should be noted that before data classification and grading, data information from different data sources should be collected first, including structured data and unstructured data. For example, structured data in the telecommunications industry includes master data, transaction data, and analytical data, and unstructured data includes text data and multimedia data.

[0028] Furthermore, the collected data is preprocessed, including: data attribute extraction and data cleaning. Among them, data attribute extraction is to obtain the data attributes of each data table and the relationship between data tables from structured data; data cleaning includes: default value processing, abnormal value processing, data format conversion and other processing.

[0029] Furthermore, the data attributes are vectorized and stored in a vector database. For example, the Word2Vec model is used to obtain the vector embedding representation of the data attributes and stored in the Chroma or Pinecone vector database to facilitate subsequent retrieval and analysis.

[0030] When performing data classification and grading, the task rules in the rule base are displayed on the interface, or the task rules in the rule base are retrieved by entering keywords of the task name, and the task rules determined by the operator are submitted and passed into the overall model to start executing the data classification and grading task.

[0031] It should be noted that the overall planning model is a large model application built on LangChain that combines multiple components to complete the data processing process, including: task tree generation component, task distribution component and task tracking component.

[0032] Among them, the task tree generation component extracts tasks from the rule base according to the received task identifier and constructs them into a task tree. The task corresponding to the task identifier is taken as the root task, and all subtasks are traversed layer by layer from the rule base according to the subtask identifier of the root task until the subtask identifier is empty to obtain tree-structured data.

[0033] Furthermore, the constructed task tree is displayed on the interface, and for each task node on the task tree, the corresponding task and its detailed information are obtained from the rule library according to its corresponding task identifier, and modification operations are provided; at the same time, operators can add new tasks according to actual needs, delete or modify existing tasks, and add corresponding detailed information to the new task rules.

[0034] The adjusted task tree is used as the to-be-processed task tree corresponding to the current operator. After submission, the task tree is received by the task distribution component, and the root task is initially taken out from the task tree as the to-be-processed task. According to the data processing type in the task target of the to-be-processed task, the to-be-processed task is distributed to the classification agent or the hierarchical agent.

[0035] It should be noted that the classification agent is a first LLM (Large Language Model) model application based on LangChain that includes a series of processing processes; the hierarchical agent is a second LLM model application based on LangChain that includes a series of processing processes. LLM models include but are not limited to: GPT series models, LLaMA models and Bert models.

[0036] (1) The classification agent calls the data classification tool to classify the relevant data according to the received to-do tasks, including: According to the task identifier in the to-do task, identify whether the received to-do task is an existing task or a new task. If it is an existing task that has not been modified, call the data classification tool based on the to-do task to classify the relevant data; if it is a new task or a modified existing task, and the data source of the to-do task is a structured database table, identify whether the data description in the to-do task can be aligned with the metadata information of each field in the database table. If so, call the data classification tool based on the aligned to-do task to classify the relevant data; otherwise, send the misaligned exception information as the classification execution result to the task tracking component.

[0037] It should be noted that since the new task is newly added by the operator on the interface, the filled-in data description may not be completely consistent with the data attributes in the filled-in data source. Therefore, the data description information of the new task must be aligned before performing data classification. For the modified existing tasks whose data source is a structured database table, an additional identifier is added to facilitate the classification agent to align the data description information of the existing tasks.

[0038] Specifically, we need to identify whether the data description of the to-do task can be aligned with the metadata information of each field in the database table, including: Attributes that do not exist in the metadata information of each field are screened out from the attributes in the data description as attributes to be identified; illustratively, a preliminary screening is performed based on the attribute name.

[0039] Get the similarity threshold in the task parameters of the to-do task, identify whether there is a field in the metadata information whose similarity with the attribute to be identified exceeds the similarity threshold, if so, alignment can be achieved, and the field corresponding to the maximum similarity is used to replace the attribute to be identified; otherwise, alignment cannot be achieved.

[0040] Specifically, the semantic vectors of the attribute to be identified and each field are obtained respectively. For example, the semantic vector of the attribute to be identified is obtained according to the attribute name and the Chinese description of the attribute, and the semantic vector of the field is obtained according to the field name and the field description; the cosine similarity between the semantic vector of the attribute to be identified and the semantic vector of each field is calculated. When the cosine similarity exceeds the set similarity threshold, it means that they can be aligned, and the field corresponding to the maximum cosine similarity is taken to replace the attribute to be identified; when the cosine similarity does not exceed the similarity threshold, it means that there is a serious error in the new task, and the misaligned exception information is sent as the execution result of the classification to the task tracking component of the overall large model.

[0041] Preferably, when the cosine similarities do not exceed the similarity threshold, the similarity threshold is gradually reduced in proportion and is not less than the minimum similarity threshold until the maximum cosine similarity is greater than the reduced similarity threshold, the field corresponding to the maximum cosine similarity is taken out to replace the attribute to be identified, and the reduced similarity threshold is recorded at the same time; when the task can be correctly executed, the reduced similarity threshold is used to update the similarity threshold in the task parameters, and the tasks in the rule base are optimized.

[0042] For existing tasks that have not been modified, new tasks that have completed information alignment, and existing tasks, call data classification tools to classify relevant data, including: The data to be classified is obtained from the data source of the to-do task using a data extraction tool bound to the first LLM model; wherein the data extraction tool is a general tool in the tool library and can be bound to multiple LLM models, and the data extraction tool includes but is not limited to: connecting to a relational database and obtaining database table data; accessing a file system and obtaining file data; When the tool model and tool hyperparameters of the to-do task are not empty, the corresponding tool model is extracted from the data classification tool bound to the first LLM model, the hyperparameters are obtained from the tool hyperparameters and set to the tool model, and the first LLM model calls the tool model to complete the data classification task for the classified data according to the task objectives of the to-do task; otherwise, classification prompt words are constructed according to the to-do task, and the first LLM model completes the data classification task according to the classification prompt words.

[0043] It should be noted that the role of the first LLM model is defined as a data classification expert in the classification prompt words, and the data description of the tasks to be done is used as the context of the data to be classified; the task objectives of the tasks to be done are used as the objectives of the first LLM model; the task requirements of the tasks to be done are used as the classification constraints of the first LLM model, and the first LLM model is required to automatically select a suitable tool model from the bound data classification tools for classification by analyzing the data to be classified and the task objectives.

[0044] Furthermore, the first LLM model of the classification agent organizes the execution results and classification results according to the set format to form a classification result report, and sends it to the task tracking component of the overall large model.

[0045] The task tracking component obtains the evaluation result by identifying whether there are exceptions in the classification result report. Specifically, if there are no exceptions, the evaluation result is normal; if there are exceptions and the exception type is failure to obtain data or running errors, the evaluation result is abnormal and the relevant information is recorded in the log; if there are exceptions and the exception type is no response, the evaluation result is timeout and the execution count increases by 1.

[0046] Preferably, considering that when aligning information between new tasks and modified existing tasks, in some cases the similarities of multiple attributes are greater than the similarity threshold and the difference in similarity is relatively small, at this time, the attribute with the largest similarity is first selected to replace the attribute to be identified, and the data extraction tool samples data from the data source according to the sample sampling number in the task parameters for data classification, and sends the classification execution result to the task tracking component. If the evaluation result of the task tracking component is normal, the classification agent extracts all data for data classification and sends the new classification execution result to the task tracking component; if the evaluation result of the feedback component is abnormal, the classification agent selects the attribute with the second largest similarity to replace the attribute to be identified, and repeats the above process. If the evaluation result of the task tracking component is still not abnormal, the evaluation result is sent to the task distribution component.

[0047] (2) The grading agent calls the data grading tool to grade the relevant data according to the received to-do tasks and the preset grading indicators, including: Using the data extraction tool bound to the second LLM model, obtain each attribute and its data record from the data source of the to-do task; According to the task requirements of the to-do tasks and the data records in each attribute, the scenario weight vector of each attribute on each grading indicator is obtained; According to the embedding vector representation of each attribute, obtain the indicator compliance vector of each attribute on each grading indicator; According to the scenario weight vector and indicator compliance vector of each attribute, the data grading tool is called to obtain the level of each attribute.

[0048] It should be noted that the grading indicators of this embodiment include: confidentiality, availability, timeliness, integrity, data value and leakage loss; among them, confidentiality means that the data has confidentiality requirements and the data must be encrypted; availability means that the grading and desensitizing processing of the data cannot have a significant impact on the data availability; timeliness means that the data classification changes over time, such as strong timeliness within a certain period of time; data value means that the grading and desensitizing processing of the data cannot have a significant impact on the value of the data; leakage loss refers to the need to consider the economic losses caused to the enterprise by the leaked data.

[0049] Specifically, the scenario weight vector of each attribute on each grading indicator is obtained by randomly taking out at least one data record corresponding to the attribute from the data source, constructing a weight allocation prompt word according to the data record, task requirements and each grading indicator, and passing it to the second LLM model.

[0050] In the weight allocation prompt, the role of the second LLM model is defined as an indicator weight allocation expert. The description of each hierarchical indicator is used as the background, each attribute and at least one data record thereof is used as the context, the task requirements and the value range of each hierarchical indicator weight are used as the allocation constraints, and the appropriate weight value is assigned to each indicator according to the task requirements and the data records in each attribute as the allocation target, and the weight is output according to the order of the hierarchical indicators.

[0051] The scene weight vector output by the second LLM model , where n represents the total number of grading indicators, Represents the weight of the i-th grading indicator, i=1,2,...,n, which is used to indicate the importance of the i-th grading indicator in the current business scenario. The larger the value, the more important the grading indicator.

[0052] For example, the value range of each grading indicator is (0, 2). If the data record of the "telephone number" attribute is "187****0010", then the value of the grading indicator "leakage loss" may be set to 0.5, indicating that data leakage will not cause significant losses.

[0053] Furthermore, to obtain the index conformity vector of each attribute on each grading index, the embedded vector representation of each attribute is passed into the trained neural network model and the output attribute conforms to the probability distribution of each grading index, thereby realizing the mapping of the attribute vector to each grading index. Exemplarily, the neural network model adopts a multi-layer perceptron neural network.

[0054] Index Conformity Vector , which has the same dimension as the scene weight vector, It indicates the degree of conformity of the attribute vector to the i-th grading index. The larger the value, the more it satisfies the grading index.

[0055] Furthermore, the scenario weight vector of each attribute is fused on the basis of the indicator compliance vector, and the scenario compliance vector is obtained by the following formula: : .

[0056] Finally, the scenario conformity vector is passed to a data classification tool to obtain the level of each attribute. Exemplarily, the data classification tool is a trained random forest model that outputs a predicted level.

[0057] It should be noted that, in this embodiment, there are 4 levels of data classification, which respectively represent 4 levels of data importance: high, relatively high, medium and low.

[0058] It should be noted that the scenario weight vector reflects the importance of each indicator in the business scenario, and the indicator compliance vector reflects the degree of satisfaction of the data attributes for each indicator. The scenario weight vector and the indicator compliance vector work together to obtain the scenario compliance vector combined with the business scenario. Based on the scenario compliance vector, the data classification task based on the multi-dimensional classification indicator is completed, which improves the accuracy of data classification; and each attribute is graded to facilitate the control of data access rights according to the attribute level. For example, if the attribute is at a low level of importance, the data of the attribute is completely public; if the attribute is at a high level of importance, the data of the attribute is only accessible or used by internal personnel with specific permissions.

[0059] The grading agent forms a grading result report with the embedding vector of each attribute in the grading task, the data records extracted from each attribute, the scene weight vector, the indicator compliance vector and the grading result, and sends it to the task tracking component of the overall large model.

[0060] The task tracking component analyzes and evaluates the classification result report according to the above evaluation method for the classification result report, and sends the obtained evaluation result to the task distribution component.

[0061] Preferably, in the task tracking component, for result reports that do not contain exceptions, secondary reasoning is performed by constructing prompt words to set a series of rules and constraints. If it is inferred that the result report is unreasonable, the evaluation result is set to be abnormal.

[0062] It should be noted that the task distribution component takes out tasks from the task tree according to the evaluation results of the task tracking component, including: When the received evaluation result is normal, the next task is taken from the task tree as a to-do task in a breadth-first manner; When the received evaluation result is timeout and the number of executions of the to-do task does not exceed the threshold, the to-do task is taken out again; When the received evaluation result is abnormal, or the number of executions of the to-do task exceeds the threshold, the to-do task and all subtasks with the to-do task as the parent task will no longer be executed, and the next task will be taken from the task tree as the to-do task in a breadth-first manner.

[0063] When there are no pending tasks to be taken out by the task distribution component, the data classification and grading task of the current operating user is completed. New tasks that are executed normally are added to the rule library, and the existing tasks that are modified are updated according to the detailed information of normal execution.

[0064] It should be noted that the results of classification and grading are part of the data assets. Like the original collected data, they are vectorized and stored to facilitate the retrieval and management of data assets, forming an important data list and a sensitive data asset map. Combining the data grading results not only facilitates the configuration of permissions for each part, each role and each administrator, but also facilitates data desensitization and encryption.

[0065] Preferably, in order to improve the accuracy of data assets, the task rules in the overall large model execution rule base are periodically used to timely update the levels of data and data attributes in each category.

[0066] Compared with the prior art, the large model-driven data classification and grading system provided in this embodiment plans and distributes pending tasks from the rule base by triggering a coordinated large model. In the iterative execution process, the classification agent and the grading agent are used to complete the classification and grading tasks in the pending tasks respectively. The coordinated large model infers and evaluates the execution results, which saves time and labor costs while ensuring the quality of the data classification and grading results, and improves data processing efficiency. It supports the addition of new tasks and the modification of existing tasks while performing classification and grading, and realizes the automatic expansion and optimization of task rules in the rule base. Based on multi-dimensional grading indicators, the scenario weight vector and indicator compliance vector are extracted from the task requirements, data content and data attributes, and the data grading task is completed by the fusion of the two vectors, so that the grading results are better aligned with the business scenario, enhancing data security and reducing business risks.

[0067] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0068] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A large model driven data classification and grading system, characterized in that: include: Rule base, as well as the overall model, classification agent and hierarchical agent built on LangChain; The overall planning model includes a task tree generation component, a task distribution component and a task tracking component; the task tree generation component extracts tasks from the rule base according to the received task identifier and constructs a task tree; the task distribution component takes out pending tasks from the task tree and distributes them to the classification agent or the hierarchical agent according to the evaluation result of the task tracking component; the task tracking component receives the execution results of the classification agent and the hierarchical agent and evaluates them, and sends the evaluation results to the task distribution component; The classification agent calls a data classification tool to classify relevant data according to the received pending tasks, and sends the classification execution results to the task tracking component; The grading agent calls the data grading tool to grade the relevant data according to the received pending tasks and the preset grading index, and sends the grading execution result to the task tracking component.

2. The large model driven data classification and grading system according to claim 1, characterized in that: The rule base includes multiple classification tasks and hierarchical tasks, each task includes: task identification, parent task identification, child task identification, task name, data source, data description, task goal, task description, task requirements, task parameters, tool model and tool hyperparameters.

3. The large model driven data classification and grading system according to claim 2, characterized in that: The task tree generation component extracts tasks from the rule base according to the received task identifier to construct a task tree, with the task corresponding to the task identifier as the root task, and traverses the rule base layer by layer to extract all subtasks according to the subtask identifier of the root task until the subtask identifier is empty to obtain tree-structured data.

4. The large model driven data classification and grading system according to claim 2, characterized in that: The task distribution component extracts the to-do tasks from the task tree according to the evaluation result of the task tracking component, including: Initially, take out the root task from the task tree as the to-do task; When the received evaluation result is normal, the next task is taken from the task tree as a to-do task in a breadth-first manner; When the received evaluation result is timeout and the number of executions of the to-do task does not exceed the threshold, the to-do task is taken out again; When the received evaluation result is abnormal, or the number of executions of the to-do task exceeds the threshold, the to-do task and all subtasks with the to-do task as the parent task will no longer be executed, and the next task will be taken from the task tree as the to-do task in a breadth-first manner.

5. The large model driven data classification and grading system according to claim 2, characterized in that: The classification agent calls a data classification tool to classify relevant data according to the received pending tasks, including: According to the task identifier of the to-do task, it is identified whether the received to-do task is an existing task or a new task. If it is an existing task that has not been modified, a data classification tool is called based on the to-do task to classify the relevant data; if it is a new task or a modified existing task, and the data source of the to-do task is a structured database table, it is identified whether the data description of the to-do task can be aligned with the metadata information of each field in the database table. If so, the data classification tool is called based on the aligned to-do task to classify the relevant data; otherwise, the misaligned exception information is sent to the task tracking component as the execution result of the classification.

6. The large model driven data classification and grading system according to claim 2 or 5, characterized in that: The calling of the data classification tool to classify the relevant data includes: Using the data extraction tool bound to the first LLM model, obtain the data to be classified from the data source of the to-do tasks; When the tool model and tool hyperparameters of the to-be-done task are not empty, the corresponding tool model is extracted from the data classification tool bound to the first LLM model, the hyperparameters are obtained from the tool hyperparameters and set to the tool model, and the first LLM model calls the tool model to complete the data classification task for the data to be classified according to the task objectives of the to-be-done task; otherwise, classification prompt words are constructed according to the to-be-done task, and the first LLM model completes the data classification task according to the classification prompt words.

7. The large model driven data classification and grading system according to claim 5, characterized in that: Whether the data description of the identified to-do task can be aligned with the metadata information of each field in the database table includes: Filtering out attributes that do not exist in the metadata information of each field from the attributes in the data description as attributes to be identified; Get the similarity threshold in the task parameters of the to-do task, identify whether there is a field in the metadata information whose similarity with the attribute to be identified exceeds the similarity threshold, if so, alignment can be achieved, and the field corresponding to the maximum similarity is used to replace the attribute to be identified; otherwise, alignment cannot be achieved.

8. The large model driven data classification and grading system according to claim 2, characterized in that: The grading agent calls the data grading tool to grade the relevant data according to the received pending tasks and the preset grading indicators, including: Using the data extraction tool bound to the second LLM model, obtain each attribute and its data record from the data source of the to-do task; According to the task requirements of the to-do tasks and the data records in each attribute, the scenario weight vector of each attribute on each grading indicator is obtained; According to the embedding vector representation of each attribute, obtain the indicator compliance vector of each attribute on each grading indicator; According to the scenario weight vector and indicator compliance vector of each attribute, the data grading tool is called to obtain the level of each attribute.

9. The large model driven data classification and grading system according to claim 8, characterized in that: The scene weight vector of each attribute on each grading index is obtained by randomly taking out at least one data record corresponding to the attribute from the data source, constructing a weight allocation prompt word according to the data record, task requirements and each grading index, and passing it to the second LLM model.

10. The large model driven data classification and grading system according to claim 8, characterized in that: The method of obtaining the index conformity vector of each attribute on each grading index is to pass the embedded vector representation of each attribute into the trained neural network model and output the probability distribution conforming to each grading index.

Images