Government affair multi-mode large model and government affair test case generation method and system

Abstract

The invention provides a government affair multi-mode large model and a government affair test case generation method and system. Relates to the technical field of software testing. Through the government affair multi-mode large model, the problem that the maintenance cost is high when a test case is iterated along with a product in agile development is solved; according to the government affair multi-modal large model, model fine tuning processing is adopted, and the multi-modal large model is trained in stages through a LoRA method, so that the model has the use case generation capability of different levels of services; the test case is generated by using a product prototype, a demand specification and project-related policy data, so that the problems of incomplete coverage and insufficient effectiveness when the test case is compiled are avoided; on this basis, a dominant entity relationship is extracted from the product knowledge graph and supplemented into business text information, and business logic information is further enriched; in addition, based on the Axure prototype, a complete business line is exported from the product prototype in the form of a word document, so that the subjectivity of manually constructing the business line is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of software testing, in particular to a government multi-modal large model and a government test case generation method and system. BACKGROUND

[0002] With the continuous development of science and technology and the popularization of digital management systems, more and more business scenarios need to use test cases for system function testing; among them, the digital management system of the government scenario has high popularization rate, large use amount and various function requirements, so the development and iteration speed of the government industry software is very rapid.

[0003] In the agile development process of the government industry software engineering, the product prototype is often output before the requirement specification; when writing test cases, testers also rely more on product prototype interaction logic and business processes; therefore, it is a key point to improve test efficiency by efficiently and accurately output high-quality test cases according to the product prototype.

[0004] The traditional test case is constructed by manual according to the product prototype and the requirement specification, which is time-consuming and inefficient, has high maintenance cost in the product iteration process of agile development, cannot meet the speed requirement of current development and iteration, and the case writing has human subjectivity, which is easy to cause business understanding deviation caused by subjective factors, and further affects the final test result.

[0005] Currently, although there are test case generation methods based on rule matching and text large language model technology, product prototype, requirement specification and other different modal product data are not fully utilized, which often leads to incomplete coverage and insufficient effectiveness of test cases.

[0006] Therefore, it is necessary to provide a government multi-modal large model and a government test case generation method and system to solve the technical problems of incomplete coverage and insufficient effectiveness of test case writing, high maintenance cost and poor timeliness of test cases following product iteration in the prior art. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art, provide a government multi-modal large model and a government test case generation method and system, and solve the technical problems of incomplete coverage and insufficient effectiveness of test case writing, high maintenance cost and poor timeliness of test cases following product iteration in the prior art.

[0008] To achieve the above purpose, the present application provides a government multi-modal large model generation system, which comprises:

[0009] Project data collection module: collects project-related data of the government industry software to be tested; wherein, the project-related data includes product prototypes and corresponding requirements specifications;

[0010] Knowledge graph construction module: builds product prototype knowledge graph based on project-related information;

[0011] Business logic combing module: constructs a business logic tree based on the product prototype and obtains demand data through the requirements specification; in which, the corresponding demand data is merged into the business of the business logic tree;

[0012] Business information merging module: extracts explicit entity relationships from the product prototype knowledge graph and adds text information to the corresponding business in the business logic tree;

[0013] Original test case processing module: cluster analysis is performed on original test cases to construct different types of test cases;

[0014] Among them, by associating various types of test cases with businesses to form a many-to-many relationship, samples of corresponding types are obtained, and a sample set is constructed based on the preset sample ratio of each type of sample;

[0015] Model fine-tuning processing module: Based on the model fine-tuning strategy of phased training, the multimodal large model is fine-tuned through the sample set to obtain a multimodal large model of government affairs.

[0016] As a further solution, the knowledge graph construction module performs data cleaning on the project-related data, and then performs named entity recognition and relationship extraction on the project-related data through the recognition model to obtain an entity relationship network and construct it into a product prototype knowledge graph.

[0017] As a further solution, the business logic combing module exports the complete business line of the product prototype in a Word document through the Axure prototype, and builds a business logic tree based on the complete business line; among them, the product prototype is decomposed into the most fine-grained business through the page tree. The business is bimodal information consisting of image information and text information. The text information contains global descriptions, and the image information contains the user interface.

[0018] As a further solution, the business logic combing module obtains the business description from the image information of the business through the image description model, calculates the similarity between the business description and the demand description, and adds the demand description with a similarity higher than a specified threshold to the text information of the business.

[0019] As a further solution, the original use case processing module divides the original test cases into simple test cases and complex test cases based on use case operation type, operation influence range and involved pages and business cases; wherein, the samples are constructed as:

[0020] Simple samples, samples associated with simple test cases by business;

[0021] Complex samples, samples associated with complex test cases by business;

[0022] Mixed samples, samples associated with simple test cases, complex test cases and business.

[0023] As a further solution, the model fine-tuning processing module adopts the LoRA method to train the multi-modal large model in stages; wherein, the training in stages includes:

[0024] The first training stage: the user input includes the global description, business description, user interface and requirement description of the simple sample corresponding business, and the model output is the simple business test case list structured based on simple business;

[0025] The second training stage, on the basis of the first training stage, the user input includes the global description, business description, user interface and requirement description of the complex sample corresponding business, and the model output is the complex business test case list structured based on complex business;

[0026] The third training stage, on the basis of the second training stage, the user input includes the global description, business description, user interface and requirement description of the mixed sample corresponding business, and the model output is the mixed business test case list structured based on mixed business.

[0027] As a further solution, the model fine-tuning processing module sets the loss function by adding simple business sample loss, complex business sample loss, mixed business sample loss, test case integrity loss and test case quantity loss.

[0028] On the other hand, the present application also provides a government multi-modal large model generation method, which is applied to the government multi-modal large model generation system as claimed in any one of the above, and the model generation step includes:

[0029] Collecting project related data of the current government industry software to be detected; wherein, the project related data includes product prototype and corresponding requirement specification;

[0030] Constructing product prototype knowledge graph based on project related data;

[0031] A business logic tree is constructed based on the product prototype, and requirement data is obtained through the requirement specification; wherein, the corresponding requirement data is merged into the business of the business logic tree;

[0032] Explicit entity relationships in the product prototype knowledge graph are extracted and supplemented into the corresponding business of the business logic tree in the form of text information;

[0033] The original test cases are clustered and analyzed to construct different types of test cases;

[0034] Wherein, by associating each type of test case with the business to form a many-to-many relationship, a sample of the corresponding type is obtained, and a sample set is constructed based on the preset sample proportion of each type of sample;

[0035] Based on the model fine-tuning strategy of phased training, the multi-modal large model is fine-tuned through the sample set.

[0036] In another aspect, the present application also provides a government test case generation system, comprising:

[0037] Project data collection module: collect project-related data of the current government industry software to be detected; wherein, the project-related data includes a product prototype and a corresponding requirement specification;

[0038] Knowledge graph construction module: construct a product prototype knowledge graph based on the project-related data;

[0039] Business logic analysis module: construct a business logic tree based on the product prototype, and obtain requirement data through the requirement specification; wherein, the corresponding requirement data is merged into the business of the business logic tree;

[0040] Business information merging module: extract explicit entity relationships in the product prototype knowledge graph and supplement them into the corresponding business of the business logic tree in the form of text information;

[0041] Test case generation module: generate business-related test cases based on the government multi-modal large model; wherein, the government multi-modal large model is generated based on the government multi-modal large model generation method described above;

[0042] Test case processing module: based on the serialization generation result and the test case with complete structure, calculate the semantic correlation between each test case, eliminate the test cases with correlation higher than the specified threshold, and output the complete test case.

[0043] In another aspect, the present application also provides a government test case generation method, which is applied to the government test case generation system described above, and the case generation steps include:

[0044] Collect project related information of the current to-be-detected government industry software; wherein, the project related information includes a product prototype and a corresponding requirement specification;

[0045] Construct a product prototype knowledge graph based on the project related information;

[0046] Construct a business logic tree based on the product prototype, and obtain requirement data through the requirement specification; wherein, the corresponding requirement data is merged into the business of the business logic tree;

[0047] Extract explicit entity relationships in the product prototype knowledge graph, and supplement them into the corresponding business of the business logic tree in the form of text information;

[0048] Generate business related test cases based on the government multi-modal large model;

[0049] Based on the serialization generation result and the test cases with complete structure, calculate the semantic correlation between the test cases, eliminate the test cases with a correlation higher than a specified threshold, and output the complete test cases.

[0050] Compared with related art, the government multi-modal large model, the generation method and system of government test cases provided by the present application have the following advantages:

[0051] 1. The present application solves the problem of high maintenance cost when test cases follow product iteration in agile development through a government multi-modal large model; wherein, the government multi-modal large model is trained in stages through the LoRA method to make the model have the ability to generate test cases for different levels of business, solving the problem of incomplete coverage when test cases are written in the prior art;

[0052] 2. The present application generates test cases using a product prototype, a requirement specification, and project related policy information, avoiding the problems of incomplete coverage and insufficient effectiveness when test cases are written; on this basis, explicit entity relationships are extracted from the product knowledge graph and supplemented into the business text information, further enriching the business logic information;

[0053] 3. The present application exports complete business lines from the product prototype in the form of a word document based on Axure prototype, and constructs a business logic tree based on the complete business lines, thereby avoiding the subjectivity of manually constructing business lines, reducing the cost of manual processing, improving the efficiency of data processing, and enabling effective maintenance and update according to product iteration. BRIEF DESCRIPTION OF DRAWINGS

[0054] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor.

[0056] Figure 1 A government multi-modal large model generation system structure schematic diagram provided by the present application;

[0057] Figure 2 A government multi-modal large model generation method step schematic diagram provided by the present application;

[0058] Figure 3 A government test case generation system structure schematic diagram provided by the present application;

[0059] Figure 4 A government test case generation method step schematic diagram provided by the present application.

[0060] The purpose of the present application, the functional characteristics and the advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0062] Embodiment 1

[0063] Please refer to Figure 1 The embodiments of the present application provide a government multi-modal large model generation system, which comprises:

[0064] Project data collection module: collect project related data of the current to be detected government industry software; wherein, the project related data comprises product prototype and corresponding requirement specification;

[0065] Knowledge graph construction module: constructing product prototype knowledge graph based on project related data;

[0066] Business logic sorting module: constructing business logic tree based on product prototype, and obtaining requirement data through requirement specification; wherein, the corresponding requirement data is merged into the business of the business logic tree;

[0067] The business information merging module extracts explicit entity relationships in the product prototype knowledge graph and supplements them with text information into the corresponding business logic tree.

[0068] The original use case processing module performs clustering analysis on the original test cases to construct different types of test cases.

[0069] The original use case processing module performs clustering analysis on the original test cases to construct different types of test cases.

[0070] The model fine-tuning processing module performs model fine-tuning processing on the multi-modal large model based on the model fine-tuning strategy of phased training to obtain the government multi-modal large model.

[0071] It should be noted that: this embodiment utilizes the government industry product prototype, requirement specification, constructs a multi-modal data set, and fine-tunes a multi-modal large model to generate structured test cases. First, this embodiment collects product prototype data, exports it as a word document and constructs a corresponding knowledge graph, combines the requirement specification, and according to the directory, the product prototype and the requirement specification are split into different businesses; Then, according to the interaction logic, the product prototype module is input as an image, the knowledge graph relationship and the requirement specification module are input as text, and together with the designed test case generation prompt word, they constitute the user input, and construct a simple business test case and a complex business test case sample set; Finally, based on the model fine-tuning strategy of phased training, the multi-modal large model is fine-tuned by the sample set to obtain the government multi-modal large model.

[0072] This embodiment makes full use of product prototypes, requirement specifications, and project-related policy data to generate test cases, avoiding the problem of incomplete coverage and insufficient effectiveness when writing test cases; Through the government multi-modal large model, the problem of high maintenance cost when test cases follow product iteration in agile development is solved.

[0073] Further, the knowledge graph construction module performs data cleaning on the project-related data, and through the recognition model, performs named entity recognition and relationship extraction on the project-related data to obtain an entity relationship network and construct a product prototype knowledge graph.

[0074] In a specific embodiment, we collect project-related policy data, perform data cleaning (denoising, data normalization, and missing value processing), and use a general large model and a trained BERT model for named entity recognition and relationship extraction to obtain an entity relationship network.

[0075] BERT (Bidirectional Encoder Representations from Transformers) is a pre-training language model proposed by Google in 2018, whose core goal is to learn general language representations through bidirectional context modeling, thereby improving the performance of various natural language processing (NLP) tasks, and can perform named entity recognition (NER) and relation extraction tasks:

[0076] Further, the business logic sorting module exports the product prototype to a word document as a complete business line through Axure prototyping, and constructs a business logic tree based on the complete business line; wherein the product prototype is disassembled to the finest granularity of business through a page tree, the business is a dual-modal information composed of image information and text information, the text information contains global instructions, and the image information contains a user interface.

[0077] Specifically, after exporting the product prototype to a word document, the complete business line of the project can be obtained. According to the page tree, the prototype is disassembled to the finest granularity of business, and a business logic tree is constructed, and the logic code can be set as:

[0078] After exporting the product prototype to a word document, the complete business line of the project can be obtained. According to the page tree, the prototype is disassembled to the finest granularity of business, and a business logic tree is constructed, and the construction idea is:

[0079] function createTR(in, pre, k):

[0080] if k <= 0: / / recursive termination: no sub-business

[0081] return NULL

[0082] else:

[0083] / / create a new business

[0084] node.data = pre[0] / / data of the current sub-business

[0085] i = 0

[0086] when in[i]!= pre[0]:

[0087] i = i + 1

[0088] / / create all sub-businesses of the business

[0089] node.fir = createTR(in, pre + 1, i) / / left subtree: direct sub-business of the current business

[0090] node.sib = createTR(in+i+1, pre+i+1, k-i-1) / / right sub-tree: sibling successor of current service

[0091] return node

[0092] Input prototype Word document: parse page tree into two-dimensional structure array (module / page / component);

[0093] Generate traversal sequence:

[0094] Logic 1: execute flow ordering by service (such as the whole process from login to payment of the user).

[0095] Logic 2: order by service level (such as sub-functions contained in the payment module).

[0096] Call createTR to build logical tree: take logic 1 and logic 2 as input, and output multi-level service tree.

[0097] Each finest-grained service contains global description (post permission, component description, file rule, etc.), user interface, constitutes image and text bimodal information.

[0098] The business logic sorting module obtains the interpreted business description of the image information of the business through the image description model, and calculates the similarity between the business description and the requirement description, and supplements the requirement description with a similarity higher than a specified threshold to the text information of the business.

[0099] On this basis, the requirement data of the requirement specification is added, and each business is merged with the requirement. When merging, the image information of the business is first obtained through the image description model to obtain the interpreted business description, and then the similarity between the business description and the requirement description is calculated.

[0100] The similarity calculation can be performed by the following formula:

[0101]

[0102] Wherein, A and B are business description and requirement description respectively. The requirement content with a similarity higher than a specified threshold is supplemented to the text information of the business.

[0103] At the same time, the explicit entity relationship is extracted from the product knowledge graph and supplemented to the business text information, enriching the business logic information

[0104] Further, the original use case processing module divides the original test case into simple test case and complex test case based on the use case operation type, operation influence range and involved page and business; wherein, the sample is constructed as:

[0105] The original test cases are clustered and divided into simple and complex test case training data,

[0106] Simple test cases refer to test cases involving add, delete, modify and query operations or affecting only the current page.

[0107] Complex test cases refer to test cases involving interaction logic between multiple pages or being closely related to business.

[0108] The test cases are associated with the business to form a many-to-many relationship, and the associated data is divided into simple, complex and mixed samples, with a ratio of 3:3:4.

[0109] Simple samples are samples associated with simple test cases,

[0110] Complex samples are samples associated with complex test cases,

[0111] Mixed samples are samples associated with business, simple test cases and complex test cases.

[0112] Further, the model fine-tuning processing module adopts the LoRA method to perform stage-by-stage training on the multi-modal large model; wherein the stage-by-stage training comprises:

[0113] The first training stage: the user input includes the global description of the simple sample corresponding business, business description, user interface and requirement description, and the model output is the simple business test case list based on simple business structure;

[0114] The second training stage, based on the first training stage, the user input includes the global description of the complex sample corresponding business, business description, user interface and requirement description, and the model output is the complex business test case list based on complex business structure;

[0115] The third training stage, based on the second training stage, the user input includes the global description of the mixed sample corresponding business, business description, user interface and requirement description, and the model output is the mixed business test case list based on mixed business structure.

[0116] In one specific embodiment, the model fine-tuning processing module fine-tunes the Qwen2.5-VL model using the LoRA method.

[0117] The model fine-tuning process is divided into three stages:

[0118] The first stage trains simple samples, the user input includes business global description, business description, user interface, requirement description, the model output is the structured simple business test case list, and the training data amount accounts for 3 / 10, a total of 3000 data.

[0119] The purpose of this stage is to generate test cases for simple business scenarios based on product prototypes and requirement descriptions.

[0120] The second stage trains complex samples based on the first stage, the complete user input includes business global description, business description, requirement description, user interface, and the model output is a structured complex business test case list. The training data accounts for 3 / 10, a total of 3000 data.

[0121] The purpose of this stage is to optimize the model's ability to generate complex test cases in business scenarios.

[0122] The third stage trains mixed samples based on the second stage, supplements the user interface in the user input, and the model output is a mixed test case list. The training data accounts for 4 / 10, a total of 4000 data.

[0123] The purpose of this stage is to train the model's ability to output simple test cases and complex test cases in business scenarios.

[0124] Further, the model fine-tuning processing module sets the loss function by adding simple business sample loss, complex business sample loss, mixed business sample loss, test case integrity loss, and test case quantity loss.

[0125] In a specific embodiment, the loss function can be set as:

[0126]

[0127] Where θ is the model parameter, x i , x j , x k , x l are the samples in the simple sample set, complex sample set, mixed sample set, and all sample set, respectively.

[0128] f(x i ,θ), f(x j ,θ), f(x k ,θ), f(x l ,θ) are the model prediction results in the simple business sample set, complex business sample set, mixed business sample set, and all sample set, respectively.

[0129] y i , y j , y k , y l are the sample labels in the simple business sample set, complex business sample set, mixed business sample set, and all sample set, respectively, and L easy , L complex , L mixLoss functions in simple service samples, complex service samples, and mixed service samples, respectively, each loss function is a cross-entropy loss function;

[0130] L completion , L number are integrity loss functions and quantity loss functions, the integrity loss function represents that the test case does not meet the structured standard, and the quantity loss represents that the number of test cases generated by the model is too small, the repetition is too high, the number is redundant, and the gap between the test case generated by serialization calculation and the sample label is taken as the value of the loss function;

[0131] α easy , α complex , α mix are weights of simple service samples, complex service samples, and mixed service samples; β and γ are weight parameters.

[0132] Embodiment 2

[0133] Please refer to Figure 2 , the embodiment further provides a government multi-modal large model generation method based on the embodiment 1, the model generation step includes:

[0134] Collecting project-related data of the current government industry software to be detected; wherein the project-related data includes a product prototype and a corresponding requirement specification;

[0135] Building a product prototype knowledge graph based on the project-related data;

[0136] Constructing a business logic tree based on the product prototype, and obtaining requirement data through the requirement specification; wherein the corresponding requirement data is merged into the business of the business logic tree;

[0137] Extracting explicit entity relationships in the product prototype knowledge graph and supplementing them into the corresponding business of the business logic tree in the form of text information;

[0138] Clustering and analyzing the original test cases to construct test cases of different types;

[0139] Wherein, by associating each type of test case with the business to form a many-to-many relationship, a corresponding type of sample is obtained, and a sample set is constructed based on a preset sample proportion of each type of sample;

[0140] Based on the model fine-tuning strategy of phased training, the multi-modal large model is fine-tuned through the sample set.

[0141] It should be noted that: the embodiment provided in the embodiment 1 is executed in the order of the corresponding operation of the government multi-modal large model generation system, and the role and effect are the same as those of each module in the embodiment 1, which will not be described in detail here.

[0142] Embodiment 3

[0143] Please refer to Figure 3 The embodiment also provides a government affair test case generation system, comprising:

[0144] Project data collection module: collect project related data of the current government affair industry software to be detected; wherein the project related data comprises a product prototype and a corresponding requirement specification;

[0145] Knowledge graph construction module: construct a product prototype knowledge graph based on the project related data;

[0146] Business logic sorting module: construct a business logic tree based on the product prototype, and acquire requirement data through the requirement specification; wherein the corresponding requirement data is merged into the business of the business logic tree;

[0147] Business information merging module: extract explicit entity relationships in the product prototype knowledge graph, and supplement text information into the corresponding business of the business logic tree;

[0148] Test case generation module: generate business related test cases based on a government affair multi-modal large model; wherein the government affair multi-modal large model is generated based on the government affair multi-modal large model generation method in embodiment 2;

[0149] Test case processing module: generate test cases based on serialization and keep the structure of the test cases complete, calculate the semantic correlation between the test cases, eliminate test cases with a correlation higher than a specified threshold, and output complete test cases.

[0150] It should be noted that: the embodiment uses the fine-tuned model to generate test cases, effectively ensuring the structural integrity and reliability of the test cases; the test cases are generated in a workflow manner, and through data cleaning, serialization, deduplication and other preprocessing and post-processing methods, the timeliness of the test cases is ensured, and the problem of repeated generation of test cases is prevented.

[0151] The test case processing module uses the josn_repair method to serialize the generation result, keeps the structure of the test cases complete, uses an embedded model to calculate the semantic correlation between the test cases, and eliminates test cases with a correlation higher than a specified threshold to obtain the final result.

[0152] Embodiment 4

[0153] Please refer to Figure 4 The embodiment further provides a government affair test case generation method based on embodiment 3, and the case generation steps comprise:

[0154] Collect project-related information of the current to-be-detected government industry software; wherein, the project-related information includes a product prototype and a corresponding requirement specification;

[0155] Construct a product prototype knowledge graph based on the project-related information;

[0156] Construct a business logic tree based on the product prototype, and obtain requirement data through the requirement specification; wherein, the corresponding requirement data is merged into the business of the business logic tree;

[0157] Extract explicit entity relationships in the product prototype knowledge graph, and supplement them into the corresponding business of the business logic tree in the form of text information;

[0158] Generate business-related test cases based on a government multi-modal large model;

[0159] Based on the serialization generation result and the test case with complete structure, calculate the semantic correlation between each test case, eliminate the test cases with a correlation higher than a specified threshold, and output the complete test case.

[0160] It should be noted that: the embodiment is executed according to the order of the government test case generation system provided in embodiment 3, and has the same effects and advantages as the modules in embodiment 3, which will not be described in detail here.

[0161] The above is only part of the embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation, direct / indirect application in other related technical fields based on the technical concept of the present application, and the contents of the present application specification and drawings are included in the patent protection scope of the present application.

Claims

1. A system for generating a large multimodal model of government affairs, characterized by: include: Project data collection module: collects project-related data of the government industry software to be tested; wherein, the project-related data includes product prototypes and corresponding requirements specifications; Knowledge graph construction module: builds product prototype knowledge graph based on project-related information; Business logic combing module: constructs a business logic tree based on the product prototype and obtains demand data through the requirements specification; in which, the corresponding demand data is merged into the business of the business logic tree; Business information merging module: extracts explicit entity relationships from the product prototype knowledge graph and adds text information to the corresponding business in the business logic tree; Original test case processing module: Cluster analysis is performed on original test cases to construct different types of test cases. By associating each type of test case with the business to form a many-to-many relationship, samples of the corresponding type are obtained, and a sample set is constructed based on the preset sample ratio of each type of sample. Model fine-tuning processing module: Based on the model fine-tuning strategy of phased training, the multimodal large model is fine-tuned through the sample set to obtain a multimodal large model of government affairs.

2. A government affairs multimodal large model generation system according to claim 1, characterized in that: After the knowledge graph construction module performs data cleaning on the project-related data, it performs named entity recognition and relationship extraction on the project-related data through the recognition model to obtain an entity relationship network and construct it into a product prototype knowledge graph.

3. A government affairs multimodal large model generation system according to claim 1, characterized in that: The business logic combing module uses Axure prototype to export the complete business line of the product prototype in a Word document, and builds a business logic tree based on the complete business line; wherein, the product prototype is decomposed into the most fine-grained business through the page tree. The business is bimodal information consisting of image information and text information. The text information contains global descriptions, and the image information contains the user interface.

4. A government affairs multimodal large model generation system according to claim 3, characterized in that: The business logic combing module obtains the business description from the image information of the business through the image description model, calculates the similarity between the business description and the demand description, and adds the demand description with a similarity higher than a specified threshold to the text information of the business.

5. A government affairs multimodal large model generation system according to claim 1, characterized in that: The original case processing module divides the original test cases into simple test cases and complex test cases based on the case operation type, operation impact range, and involved pages and business conditions; wherein the sample is constructed as follows: Simple samples: samples that associate business with simple test cases; Complex samples: samples that are associated with business and complex test cases; Mixed samples are samples that are related to business, simple test cases, and complex test cases.

6. A system for generating a multimodal large-scale government affairs model according to claim 5, characterized in that: The model fine-tuning processing module uses the LoRA method to perform phased training on the multimodal large model; wherein the phased training includes: In the first training phase, user input includes the global description, business description, user interface, and requirement description of the business corresponding to a simple sample. The model output is a list of simple business test cases based on the simple business structure. In the second training phase, based on the first training phase, user input includes the global description, business description, user interface and requirement description of the business corresponding to the complex sample, and the model output is a list of complex business test cases based on the structure of the complex business; In the third training phase, based on the second training phase, user input includes the global description, business description, user interface and requirement description of the business corresponding to the hybrid sample, and the model output is a list of hybrid business test cases based on the hybrid business structure.

7. A system for generating a multimodal large-scale government affairs model according to claim 6, characterized in that: The model fine-tuning processing module sets a loss function by accumulating simple business sample loss, complex business sample loss, mixed business sample loss, test case integrity loss and test case quantity loss.

8. A method for generating a large multimodal model of government affairs, applied to a system for generating a large multimodal model of government affairs as claimed in any one of claims 1 to 7, characterized in that: The model generation steps include: Collect project-related information about the government software to be tested, including product prototypes and corresponding requirements specifications; Build a product prototype knowledge graph based on project-related information; Construct a business logic tree based on the product prototype and obtain demand data through the requirements specification; merge the corresponding demand data into the business of the business logic tree; Extract explicit entity relationships from the product prototype knowledge graph and add text information to the corresponding business in the business logic tree; Perform cluster analysis on the original test cases to construct different types of test cases. By associating each type of test case with the business to form a many-to-many relationship, samples of the corresponding type are obtained, and a sample set is constructed based on the preset sample ratio of each type of sample. Based on the model fine-tuning strategy of phased training, the multimodal large model is fine-tuned through the sample set.

9. A government affairs test case generation system, characterized in that: include: Project data collection module: collects project-related data of the government industry software to be tested; wherein, the project-related data includes product prototypes and corresponding requirements specifications; Knowledge graph construction module: builds product prototype knowledge graph based on project-related information; Business logic combing module: constructs a business logic tree based on the product prototype and obtains demand data through the requirements specification; in which, the corresponding demand data is merged into the business of the business logic tree; Business information merging module: extracts explicit entity relationships from the product prototype knowledge graph and adds text information to the corresponding business in the business logic tree; A test case generation module: generates business-related test cases based on the government affairs multimodal big model; wherein the government affairs multimodal big model is generated based on a government affairs multimodal big model generation method according to claim 8; Test case processing module: Generates test cases based on serialization results and retains structural integrity, calculates the semantic relevance between test cases, removes test cases with relevance higher than a specified threshold, and outputs complete test cases.

10. A method for generating government test cases, applied to a government test case generation system as claimed in claim 9, characterized in that: The use case generation steps include: Collect project-related information about the government software to be tested, including product prototypes and corresponding requirements specifications; Build a product prototype knowledge graph based on project-related information; Construct a business logic tree based on the product prototype and obtain demand data through the requirements specification; merge the corresponding demand data into the business of the business logic tree; Extract explicit entity relationships from the product prototype knowledge graph and add text information to the corresponding business in the business logic tree; Generate business-related test cases based on the multimodal big model of government affairs; Generate test cases based on serialization results and retain complete structures, calculate the semantic relevance between test cases, eliminate test cases with relevance higher than a specified threshold, and output complete test cases.

Images