System and method for generating test cases for a recruitment platform

Abstract

This application discloses a test case generation system and method for recruitment platforms. The system includes a format conversion agent, a test point generation agent, a test case generation agent, and a historical regression test case recommendation agent. During system operation, test requirement data from the recruitment platform is acquired and preprocessed by the format conversion agent to output test requirement text. The test point generation agent performs multimodal parsing on the requirement text, constructing a three-dimensional feature vector containing business entities, interaction logic, and constraints. This feature vector is then combined with recruitment business prompts to complete scenario recognition and business rule injection, forming a scenario-based three-dimensional feature vector. Based on this vector, a set of test points for recruitment scenarios is generated. The test case generation agent matches corresponding generation strategies from a strategy library and generates recruitment platform test cases by combining the multi-dimensional feature vector. The historical regression test case recommendation agent retrieves matching regression test cases. This application improves the quality and efficiency of generating test cases for recruitment platforms.

Description

Technical Field

[0001] This application relates to the fields of software testing and artificial intelligence technology, and in particular to a system and method for generating test cases for a recruitment platform. Background Technology

[0002] With the rapid development of the online recruitment industry, recruitment platforms are becoming increasingly complex in their business functions, encompassing job postings, resume screening, cross-city commuting matching, intelligent recommendations, high-concurrency access, and multi-device UI display. As business complexity continues to increase, the frequency of iterative development is accelerating. Each feature iteration or optimization requires the testing team to quickly and comprehensively design and execute corresponding test cases.

[0003] However, current testing on recruitment platforms largely relies on manually written test cases, leading to low testing efficiency, inconsistent test case quality, difficulty in guaranteeing test coverage, and an inability to adapt to the complex testing requirements of recruitment platforms, such as high-frequency iteration, high concurrency, and multi-device compatibility. For example, product requirements for recruitment platforms are typically provided in multiple formats, such as product requirement documents, Excel field constraint tables, and UI design diagrams. Testers need to manually read, understand, and integrate this multimodal information, a tedious process that is prone to overlooking key constraints, resulting in incomplete test data. Secondly, the transformation from requirement text to test points depends on the tester's depth of business understanding and test design experience. For complex recruitment business rules (such as the triggering conditions for urgent hiring tags, skill weight calculations for different educational backgrounds, etc.), inexperienced testers struggle to effectively decompose scenarios and design exception flows, easily creating test blind spots. In addition, the current use of uniform templates for writing test cases fails to adaptively match the optimal generation strategy according to different types of test points (such as functional flow, boundary values, exception handling, etc.), resulting in highly redundant and untargeted test case sets, and a large workload for writing them. Summary of the Invention

[0004] In view of this, embodiments of this application provide a system and method for generating test cases for a recruitment platform, which is used to solve at least one of the above-mentioned technical problems.

[0005] Firstly, embodiments of this application provide a test case generation system for a recruitment platform. The system includes a format conversion agent, a test point generation agent, a test case generation agent, and a historical regression test case recommendation agent. After startup, the system performs the following processes: acquiring test requirement data from the recruitment platform; preprocessing the test requirement data using the format conversion agent to obtain test requirement text in the target format; performing multimodal requirement parsing on the test requirement text using the test point generation agent to construct a three-dimensional feature vector of business entities, interaction logic, and constraints; and combining recruitment business prompts to perform scene recognition and business logic analysis on the three-dimensional feature vector. Rule injection yields a scenario-based 3D feature vector; test points are generated based on this vector to create a recruitment scenario test point set; a test case generation agent performs test type classification and feature extraction on the recruitment scenario test point set to obtain a multi-dimensional feature vector; based on the test type classification results, a corresponding test case generation strategy is matched from a pre-defined strategy library; recruitment platform test cases are generated based on the test case generation strategy and the multi-dimensional feature vector; a historical regression test case recommendation agent queries a historical test case library for historical regression test cases that match the test requirement data; and the recruitment platform test cases and historical regression test cases are output.

[0006] Secondly, embodiments of this application provide a method for generating test cases for a recruitment platform. This method is based on the recruitment platform test case generation system provided in the first aspect. The method includes: acquiring test requirement data of the recruitment platform; preprocessing the test requirement data through a format conversion agent to obtain test requirement text in a target format; performing multimodal requirement parsing on the test requirement text through a test point generation agent to construct a three-dimensional feature vector of business entities, interaction logic, and constraints; combining recruitment business prompts to perform scene recognition and business rule injection on the three-dimensional feature vector to obtain a scenario-based three-dimensional feature vector; generating test points based on the scenario-based three-dimensional feature vector to obtain a recruitment scenario test point set; classifying test types and extracting features from the recruitment scenario test point set through a test case generation agent to obtain a multi-dimensional feature vector; matching the corresponding test case generation strategy from a preset strategy library according to the test type classification result; generating recruitment platform test cases based on the test case generation strategy and the multi-dimensional feature vector; querying historical regression test cases that match the test requirement data from a historical test case library through a historical regression test case recommendation agent; and outputting the recruitment platform test cases and historical regression test cases.

[0007] The recruitment platform test case generation system and method provided in this application address several issues. Firstly, through a multi-agent collaborative architecture involving format conversion, test point generation, test case generation, and historical regression test case recommendation agents, it solves problems such as low efficiency, poor test case quality, and insufficient coverage associated with manual testing. This achieves full automation and standardization of the recruitment platform test case generation process, improving both efficiency and quality. Secondly, the test point generation agent, through multimodal requirement parsing and three-dimensional feature vector construction, accurately extracts business entities, interaction logic, and constraints from requirements. By combining recruitment business prompts for scenario recognition and business rule injection, it dynamically matches and integrates numerous business scenarios from the recruitment industry into the current requirements. This ensures that the generated recruitment scenario test point set is not merely a list of surface functions but deeply embedded with industry-specific rules and complex interaction logic, enhancing the depth of test point coverage for business risks and compensating for scenario gaps caused by insufficient human experience. Compared to traditional unified template generation methods, the test case generation agent can adaptively match the optimal generation strategy based on different test point types, achieving differentiated and precise test case generation. This reduces redundant test cases and lowers test execution costs. Meanwhile, by extracting features from multiple dimensions, the adaptability of test cases to recruitment business scenarios was further enhanced, and the defect detection capability of test cases was improved.

[0008] On the other hand, the agent recommending historical regression test cases can quickly filter out highly matching regression test cases from a massive historical test case library and output them in conjunction with newly generated test cases. This reduces the workload of repeatedly writing regression test cases, shortens the testing cycle, further reduces testing costs, and allows test cases from each iteration to be accumulated and reused, forming a virtuous cycle in testing work. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings of the embodiments of this application will be briefly described below.

[0010] Figure 1 This is a structural block diagram of a system for generating test cases for a recruitment platform provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the execution flow of a recruitment platform test case generation system provided in an embodiment of this application.

[0012] Figure 3 This is a schematic diagram of an operation of a recruitment platform test case generation system provided in an embodiment of this application.

[0013] Figure 4 This is a schematic diagram of the execution flow of a format conversion agent provided in an embodiment of this application.

[0014] Figure 5 This is a schematic diagram of another execution flow of the format conversion agent provided in the embodiments of this application.

[0015] Figure 6 This is a schematic diagram of an execution flow for generating an intelligent agent for test points provided in an embodiment of this application.

[0016] Figure 7 This is a schematic diagram of an execution flow for generating an intelligent agent for test points provided in an embodiment of this application.

[0017] Figure 8 This is a schematic diagram of an execution flow for generating an intelligent agent for test points provided in an embodiment of this application.

[0018] Figure 9 This is a schematic diagram of an execution flow for generating an intelligent agent for test cases provided in an embodiment of this application.

[0019] Figure 10 This is a schematic diagram of an execution flow for generating an intelligent agent for test cases provided in an embodiment of this application.

[0020] Figure 11 A structural block diagram of a historical regression test case recommendation agent provided in the embodiments of this application.

[0021] Figure 12 This is a schematic diagram of the execution flow of a historical regression test case recommendation agent provided in an embodiment of this application. Detailed Implementation

[0022] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0023] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0024] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0025] Various modifications and variations can be made to this application without departing from its spirit or scope, which will be apparent to those skilled in the art. Therefore, this application is intended to cover modifications and variations falling within the scope of the corresponding claims (the claimed technical solutions) and their equivalents. It should be noted that the implementation methods provided in the embodiments of this application can be combined with each other without contradiction.

[0026] Before describing the technical solutions provided in the embodiments of this application, in order to facilitate understanding of the embodiments of this application, this application first specifically explains the problems existing in the prior art: With the rapid development of the online recruitment industry, recruitment platforms are becoming increasingly complex in their business functions, encompassing job postings, resume screening, cross-city commuting matching, intelligent recommendations, high-concurrency access, and multi-device UI display. As business complexity continues to increase, the frequency of iterative development is accelerating. Each feature iteration or optimization requires the testing team to quickly and comprehensively design and execute corresponding test cases.

[0027] However, current testing on recruitment platforms largely relies on manually written test cases, leading to low testing efficiency, inconsistent test case quality, difficulty in guaranteeing test coverage, and an inability to adapt to the complex testing requirements of recruitment platforms, such as high-frequency iteration, high concurrency, and multi-device compatibility. For example, product requirements for recruitment platforms are typically provided in multiple formats, such as product requirement documents, Excel field constraint tables, and UI design diagrams. Testers need to manually read, understand, and integrate this multimodal information, a tedious process that is prone to overlooking key constraints, resulting in incomplete test data. Secondly, the transformation from requirement text to test points depends on the tester's depth of business understanding and test design experience. For complex recruitment business rules (such as the triggering conditions for urgent hiring tags, skill weight calculations for different educational backgrounds, etc.), inexperienced testers struggle to effectively decompose scenarios and design exception flows, easily creating test blind spots. In addition, the current use of uniform templates for writing test cases fails to adaptively match the optimal generation strategy according to different types of test points (such as functional flow, boundary values, exception handling, etc.), resulting in highly redundant and untargeted test case sets, and a large workload for writing them.

[0028] To address at least one of the aforementioned technical problems, this application provides a system and method for generating test cases for a recruitment platform.

[0029] The following section first introduces the test case generation system for the recruitment platform provided in this application.

[0030] Figure 1 This is a structural block diagram of a system for generating test cases for a recruitment platform provided in an embodiment of this application. For example... Figure 1 As shown in the embodiments of this application, the recruitment platform test case generation system 10 may include a format conversion agent 101, a test point generation agent 102, a test case generation agent 103, and a historical regression test case recommendation agent 104. The format conversion agent 101 can be used to preprocess the test requirement data to unify the requirement format. The test point generation agent 102 can be used to parse the test requirements and generate a set of test points adapted to the recruitment scenario. The test case generation agent 103 can be used to generate test cases for the recruitment platform based on the test points. The historical regression test case recommendation agent 104 can be used to retrieve historical regression test cases matching the test requirements from the historical test case library.

[0031] Figure 2 This is a schematic diagram of the execution flow of a recruitment platform test case generation system provided in an embodiment of this application. Figure 3 This is a schematic diagram illustrating the operation of a test case generation system for a recruitment platform provided in an embodiment of this application. (Combined with...) Figure 2 and Figure 3As shown, the recruitment platform test case generation system is used to execute the following steps S201 to S206 after startup.

[0032] S201: Obtain test requirement data from the recruitment platform.

[0033] In some examples, users can upload test requirement data from the recruitment platform through a pre-defined upload portal. Test requirement data includes, but is not limited to, text, images, or tables describing the test requirements; it can also be multi-format test requirement data containing at least two of the following: text, images, and tables.

[0034] S202: The test requirement data is preprocessed by the format conversion agent to obtain the test requirement text in the target format.

[0035] In S202, a format conversion agent can perform unified preprocessing on the test requirement data of the recruitment platform, such as parsing, four-dimensional integrity verification, and / or intelligent information completion. Finally, the test requirement data is converted into test requirement text in the target format so that the subsequent test point generation agent can process it. For example, the target format includes, but is not limited to, standardized text, JSON structured format, etc.

[0036] S203: Generate an intelligent agent through test points to perform multimodal requirement parsing on the test requirement text, construct a three-dimensional feature vector of business entities, interaction logic, and constraints; and combine recruitment business prompts to perform scene recognition and business rule injection on the three-dimensional feature vector to obtain a scenario-based three-dimensional feature vector; generate test points based on the scenario-based three-dimensional feature vector to obtain a recruitment scenario test point set.

[0037] In S203, the test point generation agent can perform multimodal parsing on the test requirement text output by the format conversion agent, extracting business entities, interaction logic, and various constraints from the test requirements to construct a three-dimensional feature vector. This transforms unstructured requirements into structured features, enabling a quantitative expression of the requirement's semantics. Then, combined with preset recruitment business prompts, recruitment scenario recognition and the injection of specific business rules are completed, resulting in a scenario-based three-dimensional feature vector. This ensures that the three-dimensional feature vector accurately matches the recruitment platform's business scenarios and core rules, facilitating the generation of test points that meet recruitment business requirements and reducing the generation of invalid test points that are detached from actual business needs. Finally, test points are automatically generated based on the scenario-based three-dimensional feature vector, forming a comprehensive test point set adapted to the recruitment scenario.

[0038] S204: The test case generation agent performs test type classification and feature extraction on the test point set of the recruitment scenario to obtain a multi-dimensional feature vector; based on the test type classification results, the corresponding test case generation strategy is matched from the preset strategy library; based on the test case generation strategy and the multi-dimensional feature vector, test cases for the recruitment platform are generated.

[0039] In S204, the test case generation agent can classify the test points in the recruitment scenario test point set based on their business attributes and test objectives. For example, test types may include functional testing, UI / device compatibility testing, and high-concurrency stability testing. Then, features are extracted from each test point in the recruitment scenario test point set to obtain a multi-dimensional feature vector for each test point. This multi-dimensional feature vector can characterize key information such as business constraints, device attributes, and timing characteristics of the test points, facilitating the differentiation of core differences between different test points and generating targeted, non-redundant, and differentiated test cases. Next, based on the test case generation strategy and the multi-dimensional feature vectors, targeted recruitment platform test cases that fit the recruitment business rules are generated.

[0040] S205: Recommend the agent to query historical regression test cases that match the test requirement data from the historical test case library using historical regression test cases.

[0041] In S205, the historical regression test case recommendation agent can use the test requirement data from the recruitment platform to query historical regression test cases that match the test requirement data from the historical test case library through keyword retrieval and / or semantic retrieval.

[0042] S206: Output test cases for the recruitment platform and historical regression test cases.

[0043] The system can integrate recruitment platform test cases generated by the test case generation agent and historical regression test cases filtered by the historical regression test case recommendation agent to obtain a test case set. This test case set is then displayed to the user through an interface, and users can download and edit the test case set to meet their testing needs.

[0044] The recruitment platform test case generation system provided in this application, on the one hand, solves the problems of low efficiency, poor test case quality, and insufficient coverage of manual testing through a multi-agent collaborative architecture of format conversion agent, test point generation agent, test case generation agent, and historical regression test case recommendation agent. This achieves full automation and standardization of the recruitment platform test case generation process, improving the efficiency and quality of recruitment platform testing. On the other hand, the test point generation agent, through multimodal requirement parsing and three-dimensional feature vector construction, can accurately extract business entities, interaction logic, and constraints from requirements. By combining recruitment business prompts for scenario recognition and business rule injection, it can dynamically match and integrate numerous business scenarios in the recruitment industry into the current requirements. This makes the generated recruitment scenario test point set no longer a superficial list of functions, but deeply embedded with industry-specific rules and complex interaction logic, improving the depth of test point coverage of business risks and compensating for scenario gaps caused by insufficient human experience. Compared to traditional unified template generation methods, the test case generation agent can adaptively match the optimal generation strategy according to different types of test points, achieving differentiated and accurate generation of test cases, reducing redundant test cases and lowering test execution costs. Meanwhile, by extracting features from multiple dimensions, the adaptability of test cases to recruitment business scenarios was further enhanced, and the defect detection capability of test cases was improved.

[0045] On the other hand, the agent recommending historical regression test cases can quickly filter out highly matching regression test cases from a massive historical test case library and output them in conjunction with newly generated test cases. This reduces the workload of repeatedly writing regression test cases, shortens the testing cycle, further reduces testing costs, and allows test cases from each iteration to be accumulated and reused, forming a virtuous cycle in testing work.

[0046] To facilitate understanding, the following examples illustrate the test case generation system for recruitment platforms.

[0047] According to some embodiments of this application, optionally, the test requirement data can be multi-format test requirement data containing at least two of the following: text, images, and tables. For example, in some examples, the test requirement data may include a Word document containing requirements for a cross-city commuting filtering function, an Excel document containing commuting field constraint tables, and UI screenshots of the commuting filtering interface design. In other examples, the test requirement data may include a PDF document containing a resume upload interface document, an Excel document containing resume attachment size constraint tables, and a text document containing requirements for handling resume upload exceptions. These various data formats can be interconnected to collectively constitute the test requirement data for the recruitment platform.

[0048] The format conversion agent can be deployed with a fine-tuned text parsing model. This model can be fine-tuned using a test requirement dataset from a recruitment platform. The test requirement dataset can contain various formats of requirement data relevant to recruitment scenarios, such as Word-formatted job function requirement documents, Excel-formatted field constraint tables, PDF-formatted interface documents, and UI screenshots (e.g., job filtering / resume submission interfaces). It can cover specific business information such as job details, filtering rules, and resume interaction.

[0049] Figure 4 This is a schematic diagram illustrating the execution flow of a format conversion agent provided in an embodiment of this application. Figure 4 As shown, S202: The test requirement data is preprocessed by the format conversion agent to obtain the test requirement text in the target format, which may include the following steps S401 to S403.

[0050] S401: Extract features from test requirement data using a text parsing model to obtain layout structure features and cross-modal semantic association features.

[0051] In S401, a text parsing model can be used to extract layout structure features and cross-modal semantic association features from test requirement data. Layout structure features can include heading hierarchy features and / or table structure features from the test requirement data. For example, heading hierarchy features can include multi-level headings from the test requirement data, such as first-level headings (e.g., resume management function test requirements), second-level headings (e.g., resume attachment upload test), and third-level headings (e.g., upload format constraint test).

[0052] The table structure features involve the accurate identification of various structured tables, such as Excel field constraint tables and job screening condition tables. Specifically, it can include the table's row and column division, cell belonging relationships, and key correspondences such as field name-constraint conditions and functional module-test requirements.

[0053] Cross-modal semantic association features are primarily used to capture the semantic relationships between different modalities of content in test requirement data. These can include semantic association features and / or association weight features between at least two items from text, images, and tables. Cross-modal semantic association features aim to achieve semantic linkage between requirement data of different formats, breaking down the information fragmentation problem of single-format data. For example, the visual position of an "urgent recruitment" tag in a UI screenshot can be semantically associated with the generation rules of the "urgent recruitment" tag in the text requirement, and their association weight can be calculated. Similarly, a semantic association can be established between the 20MB size limit for resume attachments in an Excel spreadsheet and the "upload interface max_size parameter is 20971520 bytes" in a PDF interface document. The semantic association features between the two can be extracted, and the association weight can be calculated, thereby establishing information correspondence between different content formats.

[0054] S402: Perform four-dimensional integrity verification on test requirement data based on layout structure features and cross-modal semantic association features. The four-dimensional integrity verification includes essential element detection, logical completeness verification, reference integrity verification and version consistency verification. Four-dimensional integrity verification can be performed based on pre-defined business rules and document specifications in the recruitment platform's domain knowledge base, thereby accurately adapting to recruitment scenario requirements. During essential element detection, it can check whether the test requirement data is missing core elements based on the essential element list, such as function name, field constraints, business triggering conditions, interface association ID, and data format requirements, combined with information boundaries defined by layout structure features. For example, it can check whether the cross-city commuting filtering function requirement includes sub-elements such as city list, commuting mode, and time threshold. If missing, the missing type is marked, such as completely missing or partially missing.

[0055] During logical completeness verification, semantic association features can be used to detect logical contradictions in the test requirement data, such as the simultaneous existence of filters for commuting time ≤ 30 minutes and filters for commuting time ≤ 20 minutes. During reference integrity verification, layout structure features can be used to identify recruitment-specific references in the test requirement data, such as interface IDs, document numbers, and job category codes. These references can then be linked to the recruitment platform's interface document library and historical requirement document library to verify their existence and availability, avoiding issues such as referencing invalid interfaces or expired documents.

[0056] During version consistency verification, the version number can be extracted from the test requirement data and compared with historical version documents at the field level using cross-modal semantic association features to verify the consistency of functional rules. For example, if the maximum size of resume attachments in historical versions was 10MB, and the current version has changed it to 20MB without clear explanation of the change, it is determined to be a version inconsistency.

[0057] S403: Generate test requirement text based on the test requirement data after four-dimensional integrity verification.

[0058] In S403, test requirement data that passes the four-dimensional integrity check can be converted into test requirement text in the target format. If the four-dimensional integrity check fails, the user can be prompted to supplement, modify, or re-upload the data.

[0059] Thus, by deploying a text parsing model fine-tuned for recruitment scenarios, multi-format test requirement data can be accurately parsed, overcoming the shortcomings of traditional methods that can only unify formats and struggle to extract features and related information. By extracting layout structure and cross-modal semantic association features, information hierarchy can be determined and semantic linkage of multi-format data can be achieved, providing support for subsequent verification and integration. Through four-dimensional integrity verification, issues such as missing elements, logical contradictions, invalid references, and version inconsistencies can be identified, ensuring data integrity and rationality. Finally, standardized test requirement text is generated, unifying and integrating multi-format data, providing high-quality input for subsequent test point generation, reducing parsing difficulty, and adapting to the multi-format and multi-modal testing requirements of recruitment platforms.

[0060] In some specific embodiments, the text parsing model may optionally include a Visual Transformer (ViT) model and a LayoutLMv3 model. The ViT model and the LayoutLMv3 model can work together in a hybrid architecture.

[0061] Accordingly, S401: The test requirement data is parsed using the fine-tuned text parsing model to obtain layout structure features and cross-modal semantic association features, which may include the following steps one and two.

[0062] Step 1: Extract features from the test requirement data using the visual transformer model to obtain layout structure features.

[0063] Test requirement data can be converted into image data, such as converting Word or PDF documents into RGB images by page number, and converting Excel spreadsheets into structured pixel matrices. Then, the image data is input into the ViT model, which extracts features from the image data to obtain layout structure features, such as the positional information and structural features of titles and tables.

[0064] Step 2: Extract text information from the test requirement data using the layout language model and generate text word embedding vectors. Convert the layout structure features into relative position codes and concatenate the relative position codes with the text word embedding vectors to obtain multimodal feature vectors. Based on the multimodal feature vectors and cross-modal attention mechanisms, calculate the association weights between at least two items in the text, image, and table and establish semantic associations to obtain semantic association features and / or association weight features.

[0065] Specifically, text information can be extracted from the test requirement data using the LayoutLMv3 model, and the extracted text information can be encoded to generate text word embedding vectors. At the same time, the layout structure features can be converted into relative position encoding supported by the LayoutLMv3 model, and the relative position encoding can be concatenated with the text word embedding vectors to obtain a multimodal feature vector that integrates layout information and text information.

[0066] Next, using the cross-modal attention mechanism of the LayoutLMv3 model, the association weights between at least two items in the text, image, and table are calculated. If the association weight is greater than or equal to a preset threshold, it is determined to be a strong semantic association. Semantic associations are then established for data of different formats, such as the semantic linkage between UI screenshots and text requirements, and the semantic linkage between tables and interface documents, thereby obtaining cross-modal semantic association features and association weight features. The size of the preset threshold can be flexibly adjusted according to the actual situation, and this application does not limit it. For example, the preset threshold can be 0.7.

[0067] Thus, the text parsing model employing the ViT+LayoutLMv3 hybrid architecture helps address the technical limitations of a single model in simultaneously handling layout parsing and cross-modal semantic association. The ViT layer can accurately parse the layout structure of multi-format requirement data in recruitment scenarios, determine information hierarchy and table relationships, define precise boundaries for subsequent semantic extraction, reduce interference from irrelevant information, and improve the accuracy of feature extraction. The LayoutLMv3 model enables cross-modal semantic linkage between text, images, and tables, capturing the relationships between business information specific to recruitment scenarios, thereby helping to improve the accuracy of subsequent four-dimensional integrity verification.

[0068] Figure 5 This is a schematic diagram illustrating another execution flow of the format conversion agent provided in an embodiment of this application. For example... Figure 5 As shown, according to some embodiments of this application, optionally, S202: preprocessing the test requirement data through a format conversion agent to obtain test requirement text in the target format may also include the following steps S501 to S505.

[0069] S501: Analyze the missing items in the test requirement data after the four-dimensional integrity verification to determine the missing items.

[0070] After the four-dimensional integrity verification, non-core missing information in the test requirement data can be analyzed. Non-core missing information (i.e., missing items) can be concise descriptions that do not affect the core intent of the requirement, insufficient standardization, or missing related information. For example, if the test requirement data describes cross-city commuting filtering as supporting commuting time filtering, but analysis reveals a missing transportation mode association rule, then the transportation mode association rule is identified as a missing item. Similarly, if the test requirement data describes supporting resume format uploads, but does not specify the exact upload format or size limits, then the resume upload format and single file size limit are identified as missing items.

[0071] S502: Standardize and align the missing items with the preset entities in the recruitment industry knowledge graph to obtain the target entity.

[0072] The missing items are matched and aligned with the core entities (such as job titles, resumes, companies, screening criteria, and modes of transportation) pre-defined in the recruitment industry knowledge graph. For example, subway, driving, and public transportation are grouped together as modes of transportation entities, and minutes and hours are grouped together as time units entities. The core entities that match and align with the missing items are used as target entities.

[0073] S503: Based on the target entity, retrieve related entities and entity relationships associated with the target entity from the recruitment industry knowledge graph.

[0074] Based on the target entity, the system retrieves related entities and corresponding entity relationships within a recruitment industry-specific knowledge graph that have direct or indirect business connections with the target entity. Examples include job postings with filtering rules, resume upload requirements, company-urgent hiring tags, cross-city commuting calculations with transportation modes, and transportation modes with speed thresholds. Taking cross-city commuting calculations as an example, the system can retrieve relationships such as cross-city commuting calculations with transportation modes, transportation modes with speed thresholds, and cross-city commuting calculations with city distance, along with their corresponding related entities.

[0075] S504: Construct a local subgraph based on the target entity, related entities, and entity relationships, and infer the missing related information and business logic in the local subgraph through a graph neural network model.

[0076] Using the target entity as the central node, the retrieved related entities and entity relationships are used as child nodes and edges to construct a local subgraph. Taking cross-city commuting computation as an example, with cross-city commuting computation as the central node, a local subgraph is constructed that includes child nodes such as transportation mode, speed threshold, and city distance, and edges such as cross-city commuting computation-transportation mode and transportation mode-speed threshold are constructed.

[0077] Then, the local subgraph is input into a pre-defined Graph Neural Network (GNN) model. Based on the graph attention mechanism of the GNN model, the dependency weights between nodes in the local subgraph are calculated. Based on the business relationships between nodes, the missing related information and business logic in the test requirement data are inferred. For example, based on nodes and relationships such as city distance of 12km, transportation mode - subway, and subway - speed threshold of 30km / h in the local subgraph, the missing commute time = distance ÷ speed × 60 = 24 minutes is inferred. Finally, according to the business rules of the recruitment platform, the inferred related information and business logic are verified. If the commute time calculation result is ≥0 and the speed threshold is within the reasonable range of 10-120km / h in the recruitment scenario, the verification is deemed successful, thereby reducing the occurrence of inference results that do not conform to business specifications.

[0078] S505: Based on related information and business logic, complete the information for test requirement data.

[0079] Adhering to the principle of not modifying the core business intent of the test requirements and only supplementing non-core missing information, information supplementation was performed on the test requirement data based on related information and business logic. For example, the support for resume upload format was supplemented to support PDF / Word / image resume uploads, with a single file size ≤20MB. On the other hand, recruitment terminology in the test requirement data was standardized and unified. For example, terms such as salary and wages were uniformly supplemented and labeled as monthly salary to avoid ambiguity in subsequent parsing.

[0080] Accordingly, S403: Based on the test requirement data after four-dimensional integrity verification, generating test requirement text may include the following steps: The test requirement data after information completion is formatted and converted to obtain the test requirement text.

[0081] In this way, intelligent information completion enables the accurate completion of non-core missing information in the test requirement data of the recruitment platform, effectively solving the problems of insufficient simplification and standardization of information description in multimodal requirement documents, and subsequent parsing bias caused by missing related information. Combining a recruitment industry-specific knowledge graph and GNN model, the graph attention mechanism is used to achieve accurate reasoning of recruitment business-related information, and the completion results are more in line with the specific business needs of the recruitment scenario, such as job rules, resume interaction, and screening logic.

[0082] According to some embodiments of this application, optionally, S403: converting the format of the test requirement data after information completion to obtain test requirement text may include the following steps three to six.

[0083] Step 3: Convert the format of the test requirement data after information completion to obtain standardized test requirement text.

[0084] The titles, lists, tables, and business rule descriptions in the test requirement data are uniformly converted into standard Markdown syntax. For example, the core functional module title "Cross-city Commuting Filtering Function" in Word is converted into "Cross-city Commuting Filtering Function," and the resume field constraint table in Excel is converted into Markdown table format. Then, the converted Markdown data is further converted into JSON format to obtain standardized test requirement text.

[0085] Step 4: Construct the first abstract syntax tree of the test requirement data after information completion and the second abstract syntax tree of the standardized test requirement text, and calculate the node matching rate between the first abstract syntax tree and the second abstract syntax tree.

[0086] First, a first abstract syntax tree (AST) is constructed for the original test requirement data (before conversion) after information completion, and a second abstract syntax tree (AST) is constructed for the standardized test requirement text (after conversion). The AST can include core structured information from the recruitment requirement data, such as table structure, rule lists, conditional statements, heading levels, and business tags, ensuring that the AST can fully represent the syntactic logic and hierarchical relationships of the data. Then, the node structure, hierarchical relationships, and logical connections between the first and second ASTs are compared. The consistency of AST nodes for recruitment scenario-specific structured information (such as table row and column division, filtering rule conditional statements, and heading level division) is primarily verified. The ratio of the number of matching nodes to the total number of nodes is calculated to obtain the node matching rate between the first and second ASTs. The node matching rate can be used to measure the fidelity of the syntactic structure before and after format conversion.

[0087] Step 5: Calculate the semantic similarity between the completed test requirement data and the standardized test requirement text.

[0088] It can calculate the semantic similarity, such as cosine similarity, between the test requirement data after information completion and the standardized test requirement text to obtain the semantic similarity result.

[0089] Step 6: When the node matching rate is greater than or equal to the first preset threshold and the semantic similarity is greater than or equal to the second preset threshold, the standardized test requirement text is used as the test requirement text.

[0090] When the node matching rate is greater than or equal to the first preset threshold and the semantic similarity is greater than or equal to the second preset threshold, the dual-channel verification is deemed successful, and the standardized test requirement text is used as the test requirement text. The values ​​of the first and second preset thresholds can be flexibly set according to actual conditions, and this application does not limit them.

[0091] If the node matching rate is less than the first preset threshold, or the semantic similarity is less than the second preset threshold, step three is executed again.

[0092] Thus, through the dual-channel quality inspection mechanism of abstract syntax tree comparison and semantic similarity calculation, the consistency of syntactic structure and the fidelity of business semantics before and after format conversion can be verified. This effectively solves problems such as loss of structured information and deviation of core business semantics that occur during format conversion, ensuring the accuracy of test requirement text, reducing the conversion deviation of recruitment terminology and business rules, and further improving the accuracy of test requirement preprocessing. It also provides high-quality standardized input for subsequent multimodal requirement parsing and test point generation.

[0093] According to some embodiments of this application, optionally, the test point generating agent may be deployed with a fine-tuned semantic feature extraction (such as BERT) model and a fine-tuned non-textual feature recognition (such as UNet) model.

[0094] Figure 6 This is a schematic diagram of an execution flow for generating an intelligent agent for test points provided in an embodiment of this application. For example... Figure 6 As shown, S203: Generate an intelligent agent through test points to perform multimodal requirement parsing on the test requirement text and construct a three-dimensional feature vector of business entities, interaction logic and constraints, which may include the following steps S601 to S605.

[0095] S601: Extract semantic features from the test requirement text based on the semantic feature extraction model to obtain the core business rules; capture the implicit constraints in the test requirement text through the self-attention mechanism; and generate the text semantic feature vector based on the core business rules and implicit constraints.

[0096] In S601, the self-attention mechanism of the BERT model can be used to capture long-distance dependencies and core semantics in text, extracting core business rules specific to the recruitment scenario. These rules include time thresholds for cross-city commuting screening, matching rules for resume skill keywords, generation conditions for urgent hiring tags, and size limits for resume attachments. Simultaneously, leveraging the weight allocation characteristics of the self-attention mechanism, implicit constraints not explicitly stated in the text can be uncovered. For example, companies must complete qualification verification before posting urgent positions, and unlogged users can only view screening results and cannot perform screening operations. Next, the core business rules and the captured implicit constraints are fused to generate a text semantic feature vector. This text semantic feature vector can represent the core business rules, interaction relationships, and implicit constraints in the job posting.

[0097] S602: Identify non-textual information in the test requirement text based on the non-textual feature recognition model, and convert the identified non-textual information into non-textual feature data; non-textual information includes flowcharts, UI diagrams and / or data tables.

[0098] The UNet model's encoder-decoder structure identifies non-textual information within the test requirement text, performing semantic segmentation and feature extraction. For example, for flowcharts and UI diagrams, it identifies core elements such as interactive nodes, operational flow, and visual display locations, converting them into structured node feature data. For data tables, it uses fully connected layers to perform feature parsing, transforming the data into standardized table feature vectors. Next, it outputs non-textual feature data, such as node feature data and table feature vectors, that are adapted to the dimensions of the text semantic feature vectors, ensuring subsequent fusion calculations with the text semantic feature vectors.

[0099] S603: The text semantic feature vector and non-text feature data are fused to obtain the multimodal fusion feature of recruitment demand.

[0100] According to the recruitment business modules, such as cross-city commuting screening, resume skills matching, generation of urgent recruitment tags, and resume upload, the text semantic feature vectors and non-text feature data are correlated and aligned to ensure that text and non-text information within the same business module can be accurately matched. Then, the aligned text semantic feature vectors and non-text feature data are fused. For example, the text semantic feature of commuting time ≤ 30 minutes is fused with the visual features of the commuting screening entry in the UI diagram and the numerical constraint features of the commuting threshold in the data table. Similarly, the text semantic features of the urgent recruitment tag generation rules are fused with the generation process features of the urgent recruitment tags in the flowchart and the display location features of the urgent recruitment tags in the UI diagram. The fusion results in a multimodal fusion feature for recruitment needs, which includes both core business rules and implicit constraints at the text level and visual and structured features at the non-text level, comprehensively representing the multimodal information of recruitment needs.

[0101] S604: Extract recruitment business features corresponding to each dimension of business entities, interaction logic, and constraints from the multimodal fusion features of recruitment needs.

[0102] In S604, the multimodal fusion features of recruitment needs can be decomposed dimensionally, extracting recruitment business features corresponding to three dimensions: business entities, interaction logic, and constraints. This ensures that the features of each dimension align with the core business needs of recruitment. For example, when extracting features of the business entity dimension, two core entities can be extracted: "Subject Who" and "Object What." The subject can include job seekers, company HR, recruitment platform systems, etc., while the object can include job information, resumes, screening criteria, urgent hiring tags, skill keywords, commuting time, etc.

[0103] When extracting interaction logic features, the action relationships, data flow relationships, or operation sequence between business entities can be decomposed from multimodal fusion features, thereby determining the sequential logic of each interaction step and the direction of data transmission. For example, the interaction logic features of cross-city commuting screening are: job seeker selects target city → selects commuting mode → inputs commuting time threshold → platform system calculates job commuting time → filters out jobs that meet the threshold → displays job list.

[0104] When extracting constraint features, various restrictive rules in the recruitment business interaction process can be mined from multimodal fusion features, such as numerical constraints, scenario constraints, permission constraints, and format constraints. For example, the constraint features for cross-city commuting screening are: numerical constraint: commuting time threshold ≥ 0 minutes and ≤ 1440 minutes; scenario constraint: only cross-city positions support this screening; format constraint: commuting time only supports integer input; permission constraint: unlogged job seekers can only view and cannot screen.

[0105] S605: According to the preset weights, the recruitment business features corresponding to each dimension of business entities, interaction logic and constraints are vectorized and fused to obtain a three-dimensional feature vector.

[0106] In S605, the recruitment business features of three dimensions—business entity, interaction logic, and constraints—can be vectorized separately, converting the feature information of each dimension into standardized feature vectors. Then, according to the preset feature weight allocation rules for the recruitment scenario, the feature vectors of the three dimensions are weighted and fused to obtain a three-dimensional feature vector. The feature weight allocation rules can be flexibly set according to the actual situation, and this application does not limit them. For example, in some examples, the feature weight allocation rules may include a weight of 0.3 for business entities, a weight of 0.4 for interaction logic, and a weight of 0.3 for constraints.

[0107] Thus, by deploying a BERT-UNet hybrid architecture fine-tuned for recruitment scenarios, deep fusion and parsing of textual and non-textual multimodal requirements were achieved, effectively solving the problems of unfocused understanding and unclear logical organization of long recruitment text requirements by intelligent agents. Through precise extraction, alignment, fusion, and dimensional decomposition of semantic and non-textual features, features such as business entities, interaction logic, and constraints can be aligned with core recruitment scenarios such as job screening and resume submission. Furthermore, by constructing three-dimensional feature vectors, natural language and visualized recruitment requirements can be transformed into structured feature data, providing high-quality feature input for the scenario-based generation of subsequent test points, ensuring comprehensive coverage and alignment with actual business needs.

[0108] Figure 7 This is a schematic diagram of an execution flow for generating an intelligent agent for test points provided in an embodiment of this application. For example... Figure 7As shown, according to some embodiments of this application, optionally, S203: combining recruitment business prompts to perform scene recognition and business rule injection on the three-dimensional feature vector to obtain a scene-based three-dimensional feature vector may include the following steps S701 to S704.

[0109] S701: Perform semantic analysis on recruitment business prompts to obtain test scenario and test point keywords.

[0110] For example, recruitment business prompts can be used to set test scenarios, test point keywords, etc. In S701, the recruitment business prompts can be semantically parsed to obtain test scenarios and test point keywords. The test scenario represents the business module and test type being tested, while the test point keywords represent the core focus of the test. For example, parsing the prompt "resume upload UI device compatibility test" yields a test scenario of resume upload business module + UI device compatibility test type, and test point keywords such as resume upload, UI device, and compatibility. Similarly, parsing the prompt "cross-city commuting screening boundary value test" yields a test scenario of cross-city commuting screening business module + boundary value test type, and test point keywords such as cross-city commuting screening and boundary value.

[0111] S702: Based on the test scenario and test point keywords, match the target scenario node in the pre-built graph neural network knowledge graph.

[0112] The graph neural network knowledge graph uses recruitment business scenarios as nodes and scenario associations, historical defect associations, and user behavior associations as edges, and sets business constraints and / or historical defect rules for each node.

[0113] Specifically, the pre-built Graph Neural Network (GNN) knowledge graph can use recruitment business scenarios as core nodes. Node types can include core business scenarios and test-type scenarios. For example, core business scenarios can include cross-city commuting screening, resume skill matching, urgent hiring tag generation, and resume submission. Test-type scenarios can include UI device compatibility testing, functional testing, and high-concurrency stability testing. The GNN knowledge graph can construct connection edges between nodes using scenario relevance, historical defect relevance, and user behavior triggering as weights to represent the relationships between different scenarios. Simultaneously, each scenario node is pre-configured with corresponding exclusive business constraints and historical defect rules, forming a complete scenario knowledge system. In S701, based on test scenarios and test point keywords, the corresponding target scenario nodes can be retrieved and matched in the pre-built GNN knowledge graph.

[0114] S703: Extract the target business constraints and / or target historical defect rules corresponding to the target scene nodes from the graph neural network knowledge graph, and convert them into rule feature vectors that match the dimensions of the three-dimensional feature vector.

[0115] Based on the target scenario nodes, target business constraints and / or target historical defect rules corresponding to the target scenario nodes are extracted from the graph neural network knowledge graph. For example, if the target scenario node is the UI device compatibility test for resume upload, the corresponding target business constraints are extracted as needing to cover mainstream target models, foldable / notch-shaped special screen types, and target operating systems, and the target historical defect rules are the test rules corresponding to button offsets in the resume upload interface during foldable screen adaptation in the past. Then, the target business constraints and / or target historical defect rules are vectorized and converted into standardized rule feature vectors through feature encoding.

[0116] S704: The regular feature vector is fused with the three-dimensional feature vector to obtain the scene-based three-dimensional feature vector.

[0117] According to preset weights, the rule-based feature vector and the 3D feature vector can be fused to obtain a scenario-based 3D feature vector. For example, in some examples, the preset weight allocation is: 70% for the 3D feature vector, used to retain the core business entities, interaction logic, basic constraints, and other core features of the recruitment requirements themselves; and 30% for the rule-based feature vector, used to strengthen the specific business constraints and historical defect rules corresponding to the target scenario. Then, the 70% 3D feature vector and the 30% rule-based feature vector are fused to obtain the scenario-based 3D feature vector.

[0118] In this way, by parsing prompt words, matching with the GNN knowledge graph, and injecting business rules, the problems of traditional test point generation lacking scenario guidance and being disconnected from actual business are solved. Utilizing a knowledge graph specific to the recruitment scenario can accurately match test intent, while scenario-based 3D feature vectors can effectively reduce redundant test points, laying the foundation for subsequent generation of business-aligned and comprehensive test cases.

[0119] Figure 8 This is a schematic diagram of an execution flow for generating an intelligent agent for test points provided in an embodiment of this application. For example... Figure 8 As shown, according to some embodiments of this application, optionally, S203: generating test points based on scenario-based three-dimensional feature vectors to obtain a recruitment scenario test point set may include the following steps S801 to S803.

[0120] S801: Construct a recruitment business state machine based on the interaction logic in the scenario-based 3D feature vector, traverse the state nodes and state transition paths of the recruitment business state machine, and generate normal flow test points.

[0121] First, the interaction logic in the scenario-based 3D feature vector is decomposed to extract the key operational steps of the recruitment business. Each independent operation is set as a state node of the state machine, defining the input conditions, operation actions, and output results of each node. The flow relationships between nodes are sorted out, and the state transition path is built to form a complete recruitment business state machine. For example, taking the resume submission scenario, the state machine can be: Not logged in → Log in → Enter job details → Click the submit button → Select resume → Confirm submission → Submission successful / Submission failed. The arrows represent state transitions and corresponding flow conditions. The content before and after each arrow is a state node, and the arrow represents the state transition path.

[0122] After the state machine is built, the core state nodes and state transition paths are traversed to generate normal flow test points for each core path. Each normal flow test point is labeled with three key pieces of information: the test business module, the test operation steps, and the test objective, which are consistent with the normal operating scenarios of the recruitment platform. For example, for the cross-city commuting screening business scenario, the normal flow test points generated after traversing the state machine path include: Test Point 1 (Test Business Module: Cross-city Commuting Screening; Test Operation: After logging in, job seekers select the departure city (City A), destination city (City B), commuting mode (subway), commuting time threshold (30 minutes), and click the filter button; Test Objective: Verify that the platform can correctly filter cross-city positions that meet the commuting time condition).

[0123] S802: Based on the constraints in the scenario-based 3D feature vector, an abnormal scenario of the recruitment business is generated through a pre-trained adversarial generative network, and abnormal flow test points are generated based on the abnormal scenario.

[0124] To fill the gaps in the coverage of normal flow test points, and targeting abnormal scenarios that are prone to problems in the recruitment business, various abnormal scenarios and corresponding abnormal flow test points are generated based on the constraint dimensions in the scenario-based three-dimensional feature vector, such as numerical values, scenarios, permissions, and format constraints, combined with a pre-trained recruitment scenario-specific generative adversarial network (GAN).

[0125] GAN models can use historical defect reports, abnormal operation logs, and boundary value test cases from recruitment platforms as training samples, integrating abnormal scenarios from the recruitment industry's business rule base. Through multiple rounds of iterative training, the GAN model accurately learns the characteristics of abnormal scenarios, boundary value distribution, and abnormal response logic in the recruitment business, thereby ensuring that the generated abnormal scenarios and corresponding abnormal flow test points align with the platform's testing requirements and effectively reducing the generation of invalid or business-deviation-related content.

[0126] In the abnormal scenario generation stage, the constraints in the scenario-based 3D feature vector are input into the generator of the GAN model. The generator then generates various abnormal scenarios for the recruitment business based on the learned abnormal features and constraint rules. Considering the characteristics of recruitment scenarios, abnormal scenarios are divided into four main categories: boundary value anomalies, operational anomalies, format and data anomalies, and special scenario anomalies.

[0127] Boundary value anomalies refer to scenarios that violate the numerical range limits in the constraints, such as commuting time threshold input of 0 minutes, 1440 minutes (the upper limit of the reasonable range in the constraints), -5 minutes (negative value anomaly), 1441 minutes (exceeding the upper limit of the reasonable range), resume attachment size input of 20MB (the upper limit of the constraints), 20.1MB (exceeding the upper limit), 0MB (lower limit anomaly), and skill matching weight input of 0%, 100%, -5% (negative value anomaly).

[0128] Operational anomalies refer to scenarios that violate the standard operating procedures of the recruitment business, such as job seekers who are not logged in attempting to filter cross-city positions (violating permission constraints), directly entering time thresholds without selecting a commuting method (violating operational process constraints), or selecting non-existent positions when submitting resumes (violating data validity constraints).

[0129] Format and data anomalies refer to input data formats and data types that violate constraints, such as entering non-numeric characters for commuting time, uploading resumes in unsupported formats, or entering negative numbers for company recruitment budgets. Special scenario anomalies refer to abnormal scenarios that are not adapted to special devices or special operating environments, such as opening the job filtering interface on a foldable screen phone, submitting resumes on a target model of mobile phone or computer, or generating multiple urgent job tags simultaneously in a high-concurrency scenario.

[0130] After generating the abnormal scenarios, corresponding abnormal flow test points are generated for each scenario. Each abnormal flow test point is labeled with the test business module, abnormal operation / boundary value parameters, and expected abnormal response, ensuring that each abnormal flow test point corresponds to a potential defect scenario in the recruitment business and can accurately verify the recruitment platform's ability to handle abnormal scenarios. For example, for the boundary value abnormal scenario of a 20.1MB resume attachment, the generated abnormal flow test point is as follows: Test business module: Resume upload; Abnormal operation: After logging in, the job seeker selects a 20.1MB PDF resume and clicks the upload button; Expected abnormal response: The platform pops up an upload failure message, clearly indicating that the resume attachment size exceeds the limit (maximum 20MB), and does not perform the upload operation.

[0131] S803: Based on the normal flow test points and the abnormal flow test points, obtain the test point set for the recruitment scenario.

[0132] By integrating normal flow test points and abnormal flow test points, a complete set of test points for recruitment scenarios is obtained. This set of test points covers both the regular core scenarios of recruitment business and supplements various abnormal scenarios that are prone to defects. It fits the actual testing needs of the recruitment platform, and the constructed recruitment scenario test point set is comprehensive and highly targeted, effectively solving the problem of incomplete coverage of traditional test point generation, especially insufficient coverage of abnormal scenarios.

[0133] Thus, through S801 to S803, the collaborative generation of test points for normal and abnormal flows was achieved, resulting in a comprehensive and highly targeted test point set for the recruitment scenario. A state machine built based on interaction logic ensures that normal flow test points align with the business process, eliminating invalid test points that deviate from the core business. An abnormal scenario and abnormal flow test point are generated using a recruitment scenario-specific GAN model. Combined with constraints in the scenario-based 3D feature vectors, this ensures that the generated abnormal scenarios closely match the characteristics of the recruitment business, accurately covering boundary values, operational anomalies, and other scenarios prone to defects, thereby improving defect detection capabilities.

[0134] According to some embodiments of this application, optionally, S803: obtaining a set of test points for recruitment scenarios based on normal flow test points and abnormal flow test points may include the following steps seven and eight.

[0135] Step 7: Based on reinforcement learning algorithms, perform multiple rounds of iterative optimization on normal flow test points and abnormal flow test points. In each iteration, calculate the reward value of each test point using a preset target reward function, and select normal flow test points and abnormal flow test points whose reward values ​​are greater than or equal to a preset threshold. The target reward function rewards based on recruitment business coverage and defect detection potential, and penalizes based on test point redundancy and business deviation.

[0136] In step seven, the PPO algorithm optimized for the recruitment scenario is used to address issues such as excessive redundancy, insufficient defect detection capability, and deviation from business rules in the initial test points. High-quality test points are then selected through multiple rounds of iteration.

[0137] Specifically, a target reward function for the pre-defined recruitment scenario is used. This reward function adopts a design approach that combines rewards and penalties, and its expression is R=R1+R2-R3-R4.

[0138] R1 is the recruitment business coverage reward, which measures the completeness of the test point's coverage of the recruitment business. It is assigned a value based on the proportion of the covered business modules and core interaction steps, with a range of [0,10]. The higher the coverage ratio, the higher the reward. It is mainly used to ensure that test points covering the core business are retained first during the optimization process.

[0139] R2 is a defect detection potential reward, used to evaluate the ability of test points to discover business defects. It combines historical defect reports to statistically assign weights to high-frequency and high-risk scenarios, with a range of [0,8]. The higher the weight, the higher the reward. It is mainly used to strengthen the retention priority of test points with high defect detection potential.

[0140] R3 is a redundancy penalty used to suppress test point redundancy. It is assigned a value based on the semantic similarity and execution path similarity between test points, with a range of [0,6]. The higher the similarity, the higher the penalty. It is mainly used to avoid the repeated retention of similar test points.

[0141] R4 is the business deviation penalty, used to ensure that test points conform to recruitment business rules and target test scenarios. It is assigned a value based on the degree of deviation from business rules and target scenarios, with a range of [0,7]. The higher the deviation, the higher the penalty. It is mainly used to remove test points that deviate from the core business or have no practical testing significance.

[0142] Based on the PPO algorithm, multiple rounds of iterative optimization are performed on normal flow test points and abnormal flow test points. In each iteration, the reward value of each test point is calculated using a preset target reward function, and normal flow test points and abnormal flow test points with reward values ​​greater than or equal to a preset threshold are selected. The size of the preset threshold can be flexibly adjusted according to actual conditions, and this application does not limit it.

[0143] Step 8: Based on the selected normal flow test points and abnormal flow test points, obtain the test point set for the recruitment scenario.

[0144] After the PPO algorithm is iteratively optimized, all normal flow test points and abnormal flow test points with reward values ​​greater than or equal to the preset threshold are collected to obtain the recruitment scenario test point set.

[0145] In this way, through a dual reward and punishment mechanism, high-quality test points are accurately selected, strengthening the retention priority of test points with high defect detection potential and high business coverage, while eliminating redundant and business-deviation test points. This effectively solves the technical problems of excessive redundancy, uneven defect detection potential, and some test points deviating from the core business in the initially generated test point set, and achieves a four-dimensional balance of test point coverage, defect detection potential, redundancy, and business fit.

[0146] Compared to traditional static screening methods, this embodiment employs a reinforcement learning dynamic optimization strategy, which can adapt to changes in recruitment business scenarios and flexibly adjust test point screening rules to suit the iteration of core business and updates in testing requirements of the recruitment platform. The resulting recruitment scenario test point set exhibits superior overall performance, reducing the workload of subsequent testing while enhancing the relevance of testing. This provides more reliable support for the accurate generation of test cases and improved testing efficiency for the recruitment platform, further refining the entire process optimization system for recruitment scenario test point generation.

[0147] According to some embodiments of this application, optionally, S203: generating test points based on scenario-based three-dimensional feature vectors to obtain a recruitment scenario test point set may also include the following steps nine to eleven.

[0148] Step 9: Perform dual-channel deduplication on the generated test points.

[0149] Dual-channel deduplication can include semantic similarity deduplication and execution path comparison deduplication. Semantic similarity deduplication is used to remove test points that are semantically similar. That is, the semantic similarity between any two test points can be calculated; if the semantic similarity is greater than or equal to a preset threshold, one of the two test points is retained, and the other is deleted. Execution path comparison deduplication is used to remove test points with similar execution paths through execution path feature analysis. That is, for test points with similar execution paths, one is selected and retained.

[0150] Step 10: Match the deduplicated test points from the dual-channel system with the preset recruitment business rules, find and mark test points with rule conflicts, and remind manual review.

[0151] Preset recruitment rules can include numerical constraints, scenario constraints, permission constraints, and mutually exclusive operation rules. For example, commuting time thresholds cannot be negative, users who are not logged in cannot perform filtering operations, urgent hiring tags must be generated with a recruitment budget of ≥5000 yuan / month, and the same resume cannot be submitted to the same position repeatedly.

[0152] In step ten, identify test points that conflict with the recruitment business rules, detect issues such as rule conflicts, mutually exclusive operations, and contradictory conditions, mark conflicting test points, note the reason for the conflict and the corresponding rule, and push them to the testing end to remind manual review, so as to prevent conflicting test points from entering subsequent stages.

[0153] Step 11: Based on the manually reviewed test points, obtain the test point set for the recruitment scenario.

[0154] Testers can manually review each marked conflicting test point, determining a conflict resolution solution based on the actual business scenario of the recruitment platform, testing requirements, and recruitment business rules. The solutions fall into two main categories: First, if the test point itself has issues, such as unreasonable test objectives or incorrect execution conditions, the test point needs modification and adjustment. Step ten is then re-executed until there is no conflict with the rules. Second, if the recruitment business rules are outdated or poorly worded, inconsistent with actual testing requirements, the recruitment business rules are adjusted, and the test point is corrected accordingly to ensure consistency. After review, the verified normal flow test points and abnormal flow test points are integrated to obtain the recruitment scenario test point set.

[0155] Figure 9 This is a schematic diagram illustrating an execution flow for generating an intelligent agent for test cases provided in an embodiment of this application. For example... Figure 9 As shown, according to some embodiments of this application, optionally, S204: the intelligent agent generated by the test case performs test type classification and feature extraction on the test point set of the recruitment scenario to obtain a multi-dimensional feature vector, which may include the following steps S901 to S904.

[0156] S901: Based on the test scenarios and core information of each test point in the recruitment scenario test point set, determine the test type for each test point. The test types include functional testing, UI device compatibility testing, or high concurrency stability testing.

[0157] First, each test point is structured and parsed to extract its core information, such as the test business module, test operation, constraints, test scenario annotations, and risk level. The extracted core information is then converted into standardized text.

[0158] Then, based on the extracted test scenario annotations, the test type for each test point is determined. The test type can include functional testing, UI / device compatibility testing, or high-concurrency stability testing.

[0159] Functional test types correspond to test points marked as core business function verification, business rule implementation, or constraint condition verification. They mainly verify the accuracy of the core functions of the recruitment business, such as cross-city commuting time calculation, resume skill keyword matching, verification of urgent recruitment tag generation rules, and the completeness of the resume submission process.

[0160] UI device compatibility testing (referred to as compatibility testing) corresponds to test scenarios marked as UI display adaptation, device compatibility, or cross-system operation test points. It mainly verifies the display and operation adaptability of the recruitment platform interface under different devices and systems, such as resume display UI adaptation, job screening interface device compatibility, urgent recruitment tag visual effect cross-device display, foldable screen phone screening interface adaptation, and other test points.

[0161] High-concurrency stability testing (hereinafter referred to as stability testing) corresponds to test scenarios marked as high-concurrency operations, peak period operation, or batch operation processing. It mainly verifies the operational stability of the recruitment platform under high concurrency and peak load scenarios, such as test points for batch resume uploading under high concurrency, job filtering during peak periods, multiple users generating urgent recruitment tags at the same time, and response stability of batch resume filtering.

[0162] S902: Based on a hybrid model of object detection network and bidirectional long short-term memory, it extracts the basic features of test points. The basic features include the visual and device basic features of UI device compatibility test points, the semantic basic features of functional test points, and the temporal basic features of high concurrency stability test points.

[0163] A hybrid model based on the object detection network (YOLOv7) and the bidirectional long short-term memory network (BiLSTM) can extract different basic features for test points of different test types.

[0164] For example, for UI device compatibility test points, visual basic features and device basic features can be extracted through the YOLOv7 layer in the hybrid model. Visual basic features can include element position, size, display status, color, etc., while device basic features can include manufacturer, system version, screen type, resolution, etc.

[0165] For example, for functional test points, the bidirectional attention mechanism of the BiLSTM layer in a hybrid model can be used to capture the core semantic information in the test point text, such as the test business module, test operation steps, and explicit constraints. Simultaneously, implicit business relationships within the text can be preliminarily uncovered; for instance, in the resume upload test point, there is an implicit relationship between the text description and file format validation. The semantic base features can include the core semantic information and implicit business relationships of the functional test points.

[0166] For example, for high-concurrency stability test points, the basic temporal features can be extracted using the BiLSTM layer in the hybrid model. These basic temporal features can include peak operation periods, concurrency thresholds, and response time fluctuation patterns.

[0167] S903: Based on the scenario-based feature supplementation strategy corresponding to each test type, the scenario-based features of each test point are supplemented.

[0168] For functional test points, we can combine recruitment domain knowledge bases, business rule bases, and / or historical test cases to uncover implicit business constraints in the test points that are not explicitly stated but conform to the recruitment business logic, thus obtaining scenario-based features. For example, for the cross-city commuting screening function test point, scenario-based features (i.e., implicit constraint features) may include calculating subway / driving speed thresholds for different cities according to the recruitment platform's preset city traffic rules, rounding commuting time results to the nearest integer, and defaulting to all commuting methods when the commuting method is not explicitly specified. For the resume upload function test point, scenario-based features may include requiring a confirmation prompt before overwriting files with the same name, and supporting resume upload after interruption.

[0169] For UI device compatibility testing points, a device fingerprint database for recruitment scenarios can be used to supplement the four-dimensional information of manufacturer, system, resolution, and screen type, constructing a four-dimensional device fingerprint multi-dimensional feature vector of manufacturer-system-resolution-screen type. For example, for the UI adaptation compatibility testing point of resume display, the supplemented scenario-based feature vector is: Manufacturer: xx, System: yy, Resolution: 2560×1440, Screen type: foldable screen, Visual element adaptation standard: Buttons have no offset in the unfolded state of the foldable screen, and core elements are visible in the folded state.

[0170] For high-concurrency stability test points, historical failure mode libraries from recruitment platforms can be used to supplement failure-related scenario-based features, such as failure trigger thresholds, failure recovery times, high-frequency failure operations, and concurrency thresholds, to construct stability time-series scenario-based features. Simultaneously, based on the risk level of the test points, concurrency pressure-related features can be added to high-risk stability test points.

[0171] S904: Based on basic features and scenario-specific features, a multi-dimensional feature vector of the test point is obtained.

[0172] Based on preset weighting rules, the basic features and contextual features of each test point can be fused to obtain a multi-dimensional feature vector for each test point. For example, in some examples, functional test points can adopt a weighting of 40% for semantic basic features and 60% for implicit contextual features to strengthen the influence of implicit requirement features, generating a 512-dimensional feature vector for functional testing, ensuring that implicit constraints can accurately support test case design. UI device compatibility test points can adopt a weighting of 30% for visual basic features, 20% for device basic features, and 50% for device fingerprint contextual features to strengthen device adaptation-related features, generating a 256-dimensional feature vector for compatibility testing, adapting to testing strategies for different device combinations. High-concurrency stability test points can adopt a weighting of 40% for temporal basic features and 60% for historical fault contextual features to strengthen fault correlation features, generating a 128-dimensional feature vector for stability testing, supporting the generation of stress test-related test cases.

[0173] Thus, by classifying test types, test points can be categorized into three types: functional testing, UI / device compatibility testing, and high-concurrency stability testing. This aligns closely with the real business scenarios of recruitment platforms, providing a stable and consistent basis for subsequent feature extraction and reducing biases during the feature extraction process. The hybrid model, composed of YOLOv7 and BiLSTM, extracts corresponding visual, semantic, and temporal features based on the characteristics of different test types, moving away from reliance on a single feature dimension. This results in extracted basic features that are closer to real-world testing scenarios, providing more complete information and higher recognizability. Building upon this, a scenario-based feature supplementation approach, combined with recruitment business rules, device fingerprint databases, and historical fault databases, automatically uncovers previously unwritten implicit constraints, device compatibility conditions, and high-risk fault points. This effectively improves upon traditional methods' shortcomings in identifying implicit business rules and incomplete compatibility testing device information, making the features more closely reflect the actual operational logic of the recruitment business.

[0174] In addition, by weighting and fusing basic features and scenario-based features, and allocating feature weights according to the emphasis of different test types, the resulting multi-dimensional feature vector can more accurately reflect the business attributes and test priorities of each test point, making it easier to generate test cases that are more business-oriented, more comprehensive, and of more stable quality.

[0175] According to some embodiments of this application, optionally, S204: generating a strategy by matching the corresponding test cases from a preset strategy library based on the test type classification results may include the following steps: The test type classification results are input into a pre-built graph neural network policy inference engine to obtain a target test case generation strategy that matches the test type classification results; The graph neural network strategy inference engine uses recruitment test scenarios as nodes and strategy matching rules as edges, and integrates a pre-set test case generation strategy library; the target test case generation strategies include dedicated strategies for functional testing, dedicated strategies for UI device compatibility testing, or dedicated strategies for high concurrency stability testing.

[0176] First, core test case generation strategies for recruitment scenarios are extracted from a pre-defined test case generation strategy library and used as nodes in the GNN strategy knowledge graph. These nodes can be categorized into three main types based on test type. The first category comprises functional test strategy nodes, which may include core sub-nodes such as boundary value combination strategies, exception flow combination strategies, and business rule verification strategies. The second category comprises UI device compatibility test strategy nodes, which may include core sub-nodes such as cross-matrix strategies, screen type adaptation strategies, system compatibility strategies, and visual element verification strategies. The third category comprises high-concurrency stability test strategy nodes, which may include core sub-nodes such as dynamic stress curve adjustment strategies, concurrency gradient setting strategies, and fault recovery verification strategies.

[0177] Next, using feature similarity, business relevance, or historical matching success rate as core weights, we construct association edges between test type classification results, multi-dimensional feature vectors, and each strategy node, thus establishing the matching association relationships between different test types and strategy nodes. For example, the association edge between functional test type classification results, commuting time calculation vector, and boundary value combination strategy node has a business relevance weight set at 95%, indicating a very high degree of business fit between the two.

[0178] Next, based on the historical test case generation records, strategy matching logs, and the correspondence data between multi-dimensional feature vectors and strategies from the recruitment platform, a graph convolutional network (GCN) is used to iteratively train the graph neural network strategy inference engine in multiple rounds. This allows the engine to accurately learn the matching rules between test types, multi-dimensional features, and strategy nodes, ensuring that the output target test case generation strategy can accurately adapt to the test type and test point features.

[0179] During use, the test type classification results and multi-dimensional feature vectors are input into the graph neural network strategy inference engine. The graph convolution operation of the graph neural network strategy inference engine calculates the comprehensive weight of the associated edges, and the strategy node with the highest weight and that meets the criteria is selected as the target test case generation strategy that matches the test type.

[0180] For example, based on the functional test type classification results, the engine matches and outputs a boundary value combination strategy + anomaly flow combination strategy as the target test case generation strategy. For the UI device compatibility test type classification results, the engine matches and outputs a device cross-matrix strategy + screen type adaptation strategy as the target test case generation strategy. For the high-concurrency stability test type classification results, the engine matches and outputs a dynamic stress curve adjustment strategy + concurrency gradient setting strategy as the target test case generation strategy, ensuring that the target test case generation strategy is highly compatible with the test type and test point characteristics.

[0181] Thus, by constructing a graph neural network policy reasoning engine that integrates various recruitment business and testing-related data and adopts a knowledge graph structure of policy nodes and matching rules, it can more stably capture the intrinsic relationship between test types, feature vectors and generation strategies, improve the reliability of policy matching, and enable the output target test case generation strategy to better adapt to the core requirements of various tests.

[0182] Figure 10 This is a schematic diagram illustrating an execution flow for generating an intelligent agent for test cases provided in an embodiment of this application. For example... Figure 10 As shown, according to some embodiments of this application, optionally, S204: generating test cases for the recruitment platform based on the test case generation strategy and multi-dimensional feature vectors may include the following steps S1001 to S1003.

[0183] S1001: For functional testing, based on the functional testing-specific strategy and combined with the business constraints in the multi-dimensional feature vector, functional test cases combining boundary values ​​and anomaly flows are generated.

[0184] First, business constraints are extracted from the multi-dimensional feature vector of functional testing. These constraints can include explicit constraints, such as a commuting time threshold range of 0-1440 minutes and a maximum resume attachment size of 20MB, as well as implicit constraints, such as defaulting to all commuting methods when the commuting time is not explicitly specified, and requiring a prompt to overwrite uploaded resumes with the same name. Business constraints serve as the core basis for boundary values ​​and anomaly flow combinations. Then, based on the boundary value combination strategy in the functional testing-specific strategy, standardized boundary value combinations are generated for the numerical business constraints extracted from the feature vector. These combinations cover core boundary points such as minimum value, minimum +1, normal value, maximum -1, and maximum value. For example, for the commuting time threshold constraint, boundary value combinations are generated as 0 minutes, 1 minute, 30 minutes, 1439 minutes, and 1440 minutes. For the resume attachment size constraint, boundary value combinations are generated as 0MB, 1MB, 20MB, 19.9MB, and 20.1MB.

[0185] Secondly, based on the anomaly flow combination strategy in the functional testing-specific strategy, abnormal scenario combinations are generated based on the operational constraints and format constraints extracted from the feature vectors. This involves reasonably combining different types of abnormal scenarios to cover potential defects in functional testing. Examples include: anomaly flow combinations such as inputting non-numeric commuting time + not selecting a commuting method; resume upload format + filename containing special characters; and not logged in + resume submission operation. Finally, combining boundary value combinations and anomaly flow combinations, standardized functional test cases are generated. Each test case can include core elements such as preconditions, test steps, expected results, and test priority. For example, a test case could be: Test Case ID: Func-Recruit-001; Business Module: Cross-city commuting screening; Preconditions: Job seeker has logged into the recruitment platform; Test Steps: 1. Select departure city and destination city; 2. No commuting method selected; 3. Input commuting time threshold; 4. Click the filter button; Expected Result: The system displays a prompt message, and no job list is displayed; Test Priority: High.

[0186] S1002: For UI device compatibility testing, based on the UI device compatibility testing-specific strategy, and combining visual and device fingerprint feature information in multi-dimensional feature vectors, compatibility test cases corresponding to the cross matrix of device, system and resolution are generated.

[0187] First, visual features and device fingerprint features are extracted from the multi-dimensional feature vector of compatibility testing. Then, based on the device-OS-resolution cross-matrix strategy in the UI device compatibility testing strategy, a four-dimensional cross-test matrix of manufacturer-system version-screen type-resolution is constructed. Each cell in the matrix corresponds to a unique device-system-resolution combination. Finally, for each cell of the four-dimensional cross-test matrix, corresponding compatibility test cases are generated by combining the visual features from the multi-dimensional feature vector. Each test case can include core elements such as test device, test steps, visual verification standards, and expected results, used to verify the adaptability and display integrity of the core interface elements. For example, a test case could be: Test Case ID: Compat-Recruit-001; Business Module: Resume Display UI Adaptation; Test Device: xx (foldable screen), yy operating system, resolution 2496×2224; Test Steps: 1. Job seeker logs into the recruitment platform; 2. Uploads a PDF resume; 3. Opens the resume details page; 4. Switches between the unfolded and folded screen states and checks the interface display effect; Expected Results: Resume text, tables, and images are displayed completely, without layout errors, content truncation, or element overlap. The "Urgent Hiring" tag is in a fixed position, and the core operation buttons are visible and clickable in the folded state.

[0188] S1003: For high-concurrency stability testing, based on a dedicated strategy for high-concurrency stability testing, stability test cases are generated by combining timing and historical fault feature information from multi-dimensional feature vectors.

[0189] First, temporal features and historical fault features are extracted from the multi-dimensional feature vector of stability testing. Temporal features may include peak operation periods, concurrency thresholds, and response time fluctuation patterns. Historical fault features may include fault trigger thresholds, fault recovery times, and high-frequency fault operations.

[0190] Then, based on the dynamic stress curve adjustment strategy in the high-concurrency stability test-specific strategy, and combined with the extracted time-series features and historical fault features, the gradient curve parameters of the stress test are dynamically adjusted.

[0191] For example, the gradient curve parameters may include the initial concurrency value, gradient step size, peak concurrency, duration of each stage, and / or descent gradient. Finally, combining the dynamic stress curve parameters with the core verification objectives, standardized stability test cases are generated. Each test case can include core elements such as business modules, dynamic stress curve parameters, test steps, and expected results, used to verify the system's response stability, data accuracy, and fault recovery capabilities under high concurrency scenarios. For example, a test case could be: Test Case ID: Stable-Recruit-001; Business Module: High-concurrency batch resume upload; Dynamic Stress Curve: 0-60 seconds (concurrency from 100 to 1000 users / second), 60-120 seconds (concurrency from 1000 to 3000 users / second), 120-180 seconds (concurrency maintained at 3000 users / second for continuous testing), 180-240 seconds (concurrency from 3000 to 100 users / second); Test Steps: 1. 1. Simulate multiple users logging into the recruitment platform simultaneously; 2. Batch upload PDF resumes (each resume is 10MB) according to the concurrency set by the dynamic stress curve; 3. Monitor the system response time and resume upload success rate in real time; 4. Verify the system recovery status after the stress curve is completed; Expected results: No resume upload failures or duplicate uploads throughout the entire process, system response time ≤ 3 seconds, no service crashes or data loss, the system can quickly recover to normal operation during the stress curve decline phase, and the response time recovers to within 1 second.

[0192] Thus, for the three test types—functionality, UI device compatibility, and high-concurrency stability—the generation strategies and multi-dimensional feature vectors for each type of test case are tailored to their core requirements. The boundary value + exception flow combination strategy for functional testing, combined with business constraint features, covers core defects in functional testing, particularly strengthening test scenarios corresponding to implicit constraints and improving the detection rate of functional defects. The cross-matrix strategy for compatibility testing, combined with device fingerprint features, covers mainstream devices, systems, and resolutions in recruitment scenarios, while also accommodating special screen types, solving the problem of incomplete coverage of recruitment UI adaptation scenarios and improving the comprehensiveness of compatibility testing. The dynamic stress curve adjustment strategy for stability testing, combined with time-series and historical fault characteristics, accurately simulates high-concurrency scenarios in recruitment, focusing on verifying fault thresholds and system recovery capabilities, solving the problem of fixed stress parameters in traditional stability test cases that do not fit real-world scenarios.

[0193] According to some embodiments of this application, optionally, S204: generating test cases for the recruitment platform based on the test case generation strategy and multi-dimensional feature vectors may also include the following steps twelve and thirteen.

[0194] Step 12: Based on the reinforcement learning algorithm, perform multiple rounds of iterative optimization on the test case generation strategy and the test cases it generates. In each round of iteration, calculate the reward value of the test cases through a preset multi-dimensional balanced reward function, select test cases with reward values ​​greater than or equal to a preset threshold, adjust the corresponding test case generation strategy based on the reward value, and use the adjusted test case generation strategy to generate a new round of test cases.

[0195] Similar to the test point optimization mentioned above, the core of step twelve is to use reinforcement learning algorithms (such as the PPO algorithm) combined with a multi-dimensional balanced reward function for the recruitment scenario to achieve bidirectional iterative optimization of the test case generation strategy and test cases, so that the final generated test cases not only meet the coverage requirements but also have the advantages of high efficiency and low cost.

[0196] The multi-dimensional balanced reward function is designed based on the testing objectives and resource constraints of the recruitment platform. It employs a weighted fusion + penalty design logic, rewarding based on test coverage and test execution efficiency, and penalizing based on test resource consumption. Its function expression is R = α × R1 + β × R2 - γ × R3, where α, β, and γ are the weight coefficients for test coverage, execution efficiency, and resource consumption, respectively. For example, in some examples, considering the testing priority in the recruitment scenario, the preset weight configuration is α = 0.4, β = 0.3, and γ = 0.3, with a total weight of 1.0.

[0197] R1 is the test coverage reward, used to evaluate the completeness of test cases in covering the recruitment business. The specific calculation indicators include the proportion of the number of recruitment business modules covered by test cases to the number of target modules, the proportion of the number of core business constraints covered to the number of target constraints, and the proportion of the number of device combinations covered to the number of target device combinations. The arithmetic mean of the three is taken as the final R1 score. The higher the coverage, the higher the R1 reward value.

[0198] R2 is an execution efficiency reward used to evaluate the ease of execution of test cases. Specific calculation indicators include the inverse ratio of the test case execution time to the preset standard time and the inverse ratio of the number of operation steps of the test case to the preset standard number of steps. The arithmetic mean of the two is taken as the final R2 score. The shorter the execution time and the more streamlined the operation steps, the higher the R2 reward value.

[0199] R3 is a resource consumption penalty used to evaluate the reasonableness of resource usage during test case execution. Specific calculation indicators include the average utilization rate of server CPU and memory in stability testing and the ratio of the number of devices called to the number of target devices in compatibility testing. The arithmetic mean of the two is taken as the final R3 score. The higher the resource consumption, the higher the R3 penalty value.

[0200] In step 12, based on the reinforcement learning algorithm, the test case generation strategy and the test cases it generates are iteratively optimized in multiple rounds. In each round of iteration, the reward value of the test cases is calculated through a preset multi-dimensional balanced reward function, and test cases with reward values ​​greater than or equal to a preset threshold are selected.

[0201] For example, if a compatibility test case covers 80% of the target device combinations, 90% of the core business modules, and 80% of the core constraints, then R1 = (0.8 + 0.9 + 0.8) / 3 = 0.8. The test case execution time is 5 minutes, the preset standard execution time is 10 minutes, the number of operation steps is 8, and the preset standard number of steps is 10, then R2 = (5 / 10 + 8 / 10) / 2 = 0.9. The test case calls 20 devices, and the target number of devices is 60, then R3 = 20 / 60 ≈ 0.3. Substituting into the function to calculate the total reward value R = 0.4 × 0.8 + 0.3 × 0.9 - 0.3 × 0.3 = 0.32 + 0.27 - 0.09 = 0.5. Meanwhile, the preset reward value threshold can be set to 0.6 or other values ​​to filter high-quality test cases. Test cases that do not reach this threshold need to be optimized to guide strategy adjustments.

[0202] In each iteration, the specific scores (R1, R2, R3) for each use case are recorded to determine the optimization direction for the core shortcomings of low-reward use cases. For example, low reward values ​​for functional test cases stem from low R1, low reward values ​​for compatibility test cases stem from high R3, and low reward values ​​for stability test cases stem from low R2. In each iteration, test cases with reward values ​​R greater than or equal to a preset threshold are selected, marked as high-quality test cases, and retained in a temporary test case set. Low-quality test cases with reward values ​​R less than the preset threshold are not retained temporarily; only their optimization directions are extracted to guide the adjustment of the test case generation strategy.

[0203] Then, based on the optimization direction of low-quality test cases, the test case generation strategy in the PPO algorithm strategy network is adjusted accordingly.

[0204] In some examples, for functional testing-specific strategies, if a low reward value stems from low R1 (i.e., insufficient coverage), the boundary value combination strategy is adjusted to increase the boundary value combination dimension and expand the test coverage. If it stems from low R2 (i.e., low execution efficiency), the test case step generation strategy is adjusted to streamline redundant operation steps, such as merging duplicate login and logout steps, to shorten execution time. If it stems from high R3 (i.e., excessive resource consumption), the test case execution logic is optimized to reduce unnecessary duplicate verification steps and lower resource consumption.

[0205] In some examples, for UI device compatibility testing strategies, if a low reward value stems from a low R1, the device cross-matrix strategy is adjusted to supplement core device combinations, such as adding devices with the latest system versions from mainstream manufacturers. If it stems from a high R3, the device cross-matrix is ​​optimized, eliminating niche and low-usage device combinations, retaining core device combinations, and reducing the number of device calls. If it stems from a low R2, the testing steps are simplified, and similar visual verification operations are merged.

[0206] In some examples, for strategies specifically designed for high-concurrency stability testing, if the low reward value stems from a low R1, the dynamic stress curve strategy is adjusted to supplement the concurrency gradient near the fault threshold, enhancing fault scenario coverage. If it stems from a high R3, the stress curve parameters are adjusted to reduce the duration of continuous testing for non-core peak values ​​and optimize server resource allocation. If it stems from a low R2, the stress curve gradient step size is adjusted, appropriately increasing the gradient to shorten the overall test duration while ensuring the core objectives of stability testing remain unaffected.

[0207] The adjusted test case generation strategy is adopted, and combined with multi-dimensional feature vectors, a new set of test cases is generated, and the next iteration is started. The above steps are repeated until the iteration termination condition is met.

[0208] Step 13: Based on the selected test cases, obtain the test cases for the recruitment platform.

[0209] The selected test cases are integrated to obtain the recruitment platform test cases. Simultaneously, the recruitment platform test cases and the optimized test case generation strategy can be stored in the recruitment platform test database for easy future retrieval and maintenance.

[0210] Thus, by employing a multi-dimensional balanced reward function, weighted scoring and penalty constraints are applied across three dimensions: test coverage, execution efficiency, and resource consumption. This approach moves beyond relying solely on a single metric to evaluate test case quality, better aligning with the testing objectives and resource constraints of the recruitment business. It provides a stable and quantifiable benchmark for iterative optimization. Building upon this foundation, the PPO reinforcement learning algorithm is introduced, enabling simultaneous iterative optimization of the generation strategy and test cases. Each round selects high-quality test cases based on reward values ​​and adjusts the generation strategy accordingly, gradually addressing test case shortcomings until convergence is achieved, ensuring the output test cases meet preset quality requirements. Furthermore, differentiated strategy adjustments based on test type precisely address the pain points of functional, compatibility, and stability testing, improving execution efficiency and reducing server and equipment resource consumption while ensuring business coverage and defect detection capabilities. Through these steps, dynamic iterative optimization of the test case generation strategy and test cases is achieved, enhancing the overall quality of test cases on the recruitment platform.

[0211] Figure 11 A structural block diagram of the agent recommending historical regression test cases provided in the embodiments of this application. For example... Figure 11 As shown, according to some embodiments of this application, optionally, the historical regression test case recommendation agent 104 can be a multi-agent collaborative architecture. The multi-agent collaborative architecture may include a requirement parsing sub-agent 1041, a retrieval enhancement sub-agent 1042, and a ranking optimization sub-agent 1043.

[0212] Figure 12 This is a schematic diagram illustrating the execution flow of an agent that recommends historical regression test cases, as provided in an embodiment of this application. Figure 12 As shown, S205: Recommending agents to query historical regression test cases that match the test requirement data from the historical test case library through historical regression test cases may include the following steps S1201 to S1203.

[0213] S1201: The core requirement information is obtained by preprocessing the test requirement data and extracting key sentences through the requirement parsing sub-agent.

[0214] Specifically, the requirement analysis sub-agent 1041 can preprocess the input test requirement data, such as removing meaningless words and unifying terms. For example, it removes meaningless words such as "de (的)" and "le (了)", retains core recruitment terms such as commuting threshold, skill matching degree, urgent recruitment label, and resume attachment, and uniformly standardizes the term expressions. For example, it unifies job seekers / applicants as job seekers, etc., and finally outputs a standardized recruitment requirement text segment.

[0215] Then, the requirement analysis sub-agent 1041 inputs the preprocessed standardized text segment into the fine-tuned BERT-wwm model, and extracts the core requirement information of the recruitment test requirements through the self-attention mechanism of the model. The core requirement information can include business modules, core operations, change content, and constraints, etc.

[0216] S1202: Through the retrieval enhancement sub-agent, based on the core requirement information, perform keyword retrieval and semantic retrieval from the historical test case library, and recall candidate historical regression test cases that match the core requirement information.

[0217] The retrieval enhancement sub-agent 1042 can be deployed with an Elasticsearch+FAISS dual retrieval engine to achieve multi-way recall of keyword retrieval and semantic retrieval for the huge historical test case library of the recruitment platform. For the Elasticsearch keyword retrieval engine, a keyword index can be constructed based on core requirement information such as business modules, core operations, and constraints, and candidate historical regression test cases that match the keyword index are retrieved from the historical test case library. For the FAISS semantic retrieval engine, the core requirement information can be used as the key sentence, and candidate historical regression test cases that match the key sentence are retrieved from the historical test case library.

[0218] Finally, the retrieval enhancement sub-agent 1042 summarizes and deduplicates the candidate cases recalled by keyword retrieval and semantic retrieval, and obtains a set of candidate historical regression test cases that match the core requirement information.

[0219] S1203: Through the sorting optimization sub-agent, based on the preset sorting weights, sort the recalled candidate historical regression test cases to obtain the sorted historical regression test cases.

[0220] In some embodiments, the sorting optimization sub-agent 1043 can sort the candidate historical regression test cases recalled by keyword retrieval and semantic retrieval based on the Reciprocal Rank Fusion (RRF) algorithm, and dynamically adjust the retrieval weights in combination with the demand change frequency λ of each business module in the recruitment scenario to solve the problem of different matching strategies for requirements with different change frequencies in the recruitment scenario, and achieve accurate sorting of candidate historical regression test cases.

[0221] First, for each candidate historical regression test case retrieved by keyword retrieval and semantic retrieval, the RRF fusion score is calculated according to the following expression: (1) Where α is the retrieval weight, β is the semantic retrieval weight, rank1 is the ranking of the candidate historical regression test case in keyword retrieval, and k1 is a constant; rank2 is the ranking of the candidate historical regression test case in semantic retrieval, and k2 is a constant. For example, with α=0.3, β=0.7, k1=k2=60, if a candidate historical regression test case ranks 5th in keyword retrieval and 2nd in semantic retrieval, then the RRF fusion score of this test case = 0.3×(1 / (60+5))+0.7×(1 / (60+2))≈0.013.

[0222] Secondly, the search weights and semantic search weights are dynamically adjusted based on the frequency of demand changes (λ). In some examples, the frequency of demand changes (λ) is defined as the average number of monthly changes to each business module of the recruitment platform. The ranking optimization sub-agent categorizes this into three scenarios and adjusts the weights accordingly. Without manual intervention, the system automatically reads the change frequency parameters of the recruitment business modules and adjusts them in real time: (1) Low-frequency change scenarios (λ≤1): such as business modules including login / registration, basic job display, etc., the weights remain unchanged in the basic configuration, i.e. α=0.3, β=0.7; (2) Mid-frequency change scenario (1<λ≤3): such as business modules including cross-city commuting screening and resume skills matching, the semantic retrieval weight β will be increased to 0.75, while the keyword retrieval weight α will remain unchanged at 0.3; (3) High-frequency change scenarios (λ>3): such as business modules including urgent recruitment tag generation and intelligent recommendation rules, the semantic retrieval weight β is increased to 0.8 and the keyword retrieval weight α is reduced to 0.2.

[0223] Taking a high-frequency change scenario as an example, if the resume skill matching rules corresponding to the core requirement information are optimized, and the average monthly change frequency of this business module is λ=4, then the RRF fusion score calculation formula for candidate historical regression test cases in this scenario is adjusted as follows: .

[0224] Finally, the candidate historical regression test cases are sorted according to their RRF fusion scores from highest to lowest to obtain the sorted historical regression test cases for testers to carry out regression testing.

[0225] This embodiment, through the collaborative work of the requirement parsing sub-agent 1041, the retrieval enhancement sub-agent 1042, and the ranking optimization sub-agent 1043, addresses the core pain points of frequent requirement iterations and a large historical test case library in recruitment platforms. It achieves accurate matching and efficient recommendation of test requirements and historical regression test cases, effectively solving the technical problems of low efficiency and poor matching accuracy in regression test case screening in recruitment scenarios.

[0226] Based on the recruitment platform test case generation system 10 provided in any of the above embodiments, this application also provides a method for generating recruitment platform test cases, which can be implemented based on the recruitment platform test case generation system 10 provided in any of the embodiments.

[0227] like Figure 2 As shown in the embodiments of this application, the method for generating test cases for a recruitment platform may include the following steps: S201: Obtain test requirement data from the recruitment platform; S202: The test requirement data is preprocessed by the format conversion agent to obtain the test requirement text in the target format; S203: Generate an intelligent agent through test points to perform multimodal requirement parsing on the test requirement text, construct a three-dimensional feature vector of business entities, interaction logic, and constraints; and combine recruitment business prompts to perform scene recognition and business rule injection on the three-dimensional feature vector to obtain a scene-based three-dimensional feature vector; generate test points based on the scene-based three-dimensional feature vector to obtain a recruitment scene test point set; S204: The test case generation agent performs test type classification and feature extraction on the test point set of the recruitment scenario to obtain a multi-dimensional feature vector; based on the test type classification results, the corresponding test case generation strategy is matched from the preset strategy library; based on the test case generation strategy and the multi-dimensional feature vector, test cases for the recruitment platform are generated. S205: Recommend the agent to query historical regression test cases that match the test requirement data from the historical test case library using historical regression test cases; S206: Output test cases for the recruitment platform and historical regression test cases.

[0228] It should be noted that the method for generating test cases for the recruitment platform provided in this application embodiment has the same or corresponding technical features as the recruitment platform test case generation system 10 provided in any of the above embodiments, and can produce the same technical effects. For the sake of brevity, it will not be described in detail here.

[0229] The functional blocks shown in the structural block diagrams of this application can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc.; when implemented in software, they are programs or code segments used to perform the required tasks. Programs or code segments can be stored in memory or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. Code segments can be downloaded via computer networks such as the Internet or intranets.

[0230] It should be noted that this application is not limited to the specific configurations and processes described above or shown in the figures. The above descriptions are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the described systems, devices, modules, or units can be referred to the corresponding processes in the method embodiments, and need not be repeated here. It should be understood that the scope of protection of this application is not limited thereto. Any person skilled in the art can conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application.

Claims

1. A test case generation system for a recruitment platform, the system comprising a format conversion agent, a test point generation agent, a test case generation agent, and a historical regression test case recommendation agent, wherein the system, upon startup, performs the following processing: Obtain testing requirement data from recruitment platforms; The format conversion agent preprocesses the test requirement data to obtain the test requirement text in the target format. The intelligent agent generates a multi-modal demand analysis on the test demand text through the test point, and constructs a three-dimensional feature vector of business entities, interaction logic and constraint conditions; Furthermore, by combining recruitment business prompts with scene recognition and business rule injection into the 3D feature vector, a scene-based 3D feature vector is obtained; Test points are generated based on scenario-based 3D feature vectors to obtain a test point set for recruitment scenarios; The test case generation agent performs test type classification and feature extraction on the test point set of the recruitment scenario to obtain a multi-dimensional feature vector. Based on the test type classification results, the corresponding test cases are matched from the preset strategy library to generate strategies; Based on test case generation strategies and multi-dimensional feature vectors, test cases for the recruitment platform are generated. The recommended agent uses the historical regression test cases to query historical regression test cases that match the test requirements data from the historical test case library. Output test cases for the recruitment platform and historical regression test cases.

2. The system of claim 1, wherein, The format conversion agent is equipped with a fine-tuned text parsing model, and the test requirement data is multi-format test requirement data containing at least two of the following: text, images, and tables. The format conversion agent preprocesses the test requirement data to obtain the test requirement text in the target format, including: The test requirement data is subjected to feature extraction by the text parsing model to obtain layout structure features and cross-modal semantic association features. The layout structure features include title level features and / or table structure features in the test requirement data. The cross-modal semantic association features include semantic association features and / or association weight features of at least two items in the text, images and tables in the test requirement data. The test requirement data is subjected to four-dimensional integrity verification based on layout structure features and cross-modal semantic association features. The four-dimensional integrity verification includes essential element detection, logical completeness verification, reference integrity verification and version consistency verification. The test requirement text is generated based on the test requirement data after four-dimensional integrity verification.

3. The system of claim 2, wherein, The text parsing model includes a visual transformer model and a layout language model; The test requirement data was analyzed using a fine-tuned text parsing model to obtain layout structure features and cross-modal semantic association features, including: The visual transformer model is used to extract features from the test requirement data to obtain layout structure features. The layout language model is used to extract text information from the test requirement data and generate text word embedding vectors. The layout structure features are converted into relative position codes, and the relative position codes are concatenated with the text word embedding vectors to obtain multimodal feature vectors. Based on the multimodal feature vectors and cross-modal attention mechanism, the association weights between at least two items in the text, image and table are calculated and semantic associations are established to obtain semantic association features and / or association weight features.

4. The system of claim 2, wherein, The test requirement data is preprocessed by the format conversion agent to obtain the test requirement text in the target format, and the process also includes: Analyze the missing items in the test requirement data after the four-dimensional integrity verification to determine the missing items; The missing items are standardized and aligned with the preset entities in the recruitment industry knowledge graph to obtain the target entity; Based on the target entity, retrieve associated entities and entity relationships related to the target entity from the recruitment industry knowledge graph; A local subgraph is constructed based on the target entity, associated entities, and entity relationships, and the missing association information and business logic in the local subgraph are inferred through a graph neural network model. Based on related information and business logic, information is supplemented in the test requirement data; Based on the test requirement data after four-dimensional integrity verification, the test requirement text is generated, including: The test requirement data after information completion is formatted and converted to obtain the test requirement text.

5. The system according to claim 4, characterized in that, The test requirement data after information completion is format-converted to obtain the test requirement text, including: The format of the test requirement data after information completion is converted to obtain standardized test requirement text; Construct a first abstract syntax tree of the test requirement data after information completion and a second abstract syntax tree of the standardized test requirement text, and calculate the node matching rate between the first abstract syntax tree and the second abstract syntax tree; The semantic similarity between the completed test requirement data and the standardized test requirement text is calculated. When the node matching rate is greater than or equal to the first preset threshold and the semantic similarity is greater than or equal to the second preset threshold, the standardized test requirement text will be used as the test requirement text.

6. The system according to claim 1, characterized in that, The test point generating agent is equipped with a fine-tuned semantic feature extraction model and a fine-tuned non-text feature recognition model. The intelligent agent generated through the test points performs multimodal requirement parsing on the test requirement text, constructing a three-dimensional feature vector of business entities, interaction logic, and constraints, including: Based on the semantic feature extraction model, semantic features are extracted from the test requirement text to obtain core business rules; implicit constraints in the test requirement text are captured through a self-attention mechanism; and text semantic feature vectors are generated based on the core business rules and implicit constraints. The non-textual feature recognition model is used to identify non-textual information in the test requirement text and convert the identified non-textual information into non-textual feature data; the non-textual information includes flowcharts, UI diagrams and / or data tables. The text semantic feature vectors and non-text feature data are fused to obtain multimodal fusion features of recruitment needs; From the multimodal fusion features of recruitment needs, we extract the recruitment business features corresponding to each dimension of business entities, interaction logic, and constraints. According to preset weights, the recruitment business features corresponding to each dimension of business entities, interaction logic, and constraints are vectorized and fused to obtain a three-dimensional feature vector.

7. The system according to claim 1, characterized in that, By combining recruitment business prompts with scene recognition and business rule injection into the 3D feature vector, a scene-based 3D feature vector is obtained, including: Semantic analysis of recruitment prompts yields keywords for test scenarios and test points. Based on the test scenario and test point keywords, target scenario nodes are matched in a pre-constructed graph neural network knowledge graph; wherein, the graph neural network knowledge graph takes recruitment business scenarios as nodes, and scenario association, historical defect association and user behavior association as edges, and sets business constraints and / or historical defect rules for each node. Extract the target business constraints and / or target historical defect rules corresponding to the target scene nodes from the graph neural network knowledge graph, and convert them into rule feature vectors that match the dimensions of the three-dimensional feature vector; The regular feature vector is fused with the three-dimensional feature vector to obtain a scene-based three-dimensional feature vector.

8. The system according to claim 7, characterized in that, Test points are generated based on scenario-based 3D feature vectors to obtain a test point set for recruitment scenarios, including: Based on the interaction logic in the scenario-based 3D feature vector, a recruitment business state machine is constructed. The state nodes and state transition paths of the recruitment business state machine are traversed to generate normal flow test points. Based on the constraints in the scenario-based 3D feature vector, abnormal scenarios of the recruitment business are generated through a pre-trained adversarial generative network, and abnormal flow test points are generated based on the abnormal scenarios. The recruitment scenario test point set is obtained based on the normal flow test points and the abnormal flow test points.

9. The system according to claim 8, characterized in that, Based on the normal flow test points and the abnormal flow test points, the recruitment scenario test point set is obtained, including: Based on reinforcement learning algorithms, the normal flow test points and the abnormal flow test points are iteratively optimized in multiple rounds. In each round of iteration, the reward value of each test point is calculated by a preset target reward function, and normal flow test points and abnormal flow test points with reward values ​​greater than or equal to a preset threshold are selected. The target reward function rewards based on recruitment business coverage and defect detection potential, and penalizes based on test point redundancy and business deviation. The recruitment scenario test point set is obtained based on the selected normal flow test points and abnormal flow test points.

10. The system according to any one of claims 7-9, characterized in that, Test points are generated based on scenario-based 3D feature vectors to obtain a recruitment scenario test point set, which also includes: The generated test points are deduplicated using a dual-channel method; wherein, the dual-channel deduplication includes semantic similarity deduplication and execution path comparison deduplication, the semantic similarity deduplication is used to remove semantically similar test points by calculating semantic similarity, and the execution path comparison deduplication is used to remove test points with similar execution paths by analyzing execution path features; The test points after dual-channel deduplication are matched with the preset recruitment business rules, and test points with rule conflicts are identified and marked, prompting manual review. The test point set for the recruitment scenario is obtained based on the test points after manual review.

11. The system according to claim 1, characterized in that, The test case-generated intelligent agent performs test type classification and feature extraction on the test point set of the recruitment scenario, resulting in a multi-dimensional feature vector, including: Based on the test scenarios and core information of each test point in the recruitment scenario test point set, the test type of each test point is determined. The test type includes functional testing, UI device compatibility testing, or high concurrency stability testing. Based on the object detection network and bidirectional long short-term memory hybrid model, the basic features of the test points are extracted. The basic features include the visual and device basic features of UI device compatibility test points, the semantic basic features of functional test points, and the temporal basic features of high concurrency stability test points. Based on the scenario-based feature supplementation strategy corresponding to each test type, the scenario-based features of each test point are supplemented. Based on the basic features and the scenario-specific features, a multi-dimensional feature vector of the test point is obtained.

12. The system according to claim 1, characterized in that, Based on the test type classification results, strategies are generated by matching corresponding test cases from a pre-defined strategy library, including: The test type classification results are input into a pre-built graph neural network policy inference engine to obtain a target test case generation strategy that matches the test type classification results; The graph neural network strategy inference engine uses recruitment test scenarios as nodes and strategy matching rules as edges, and integrates a preset test case generation strategy library; the target test case generation strategy includes a functional test-specific strategy, a UI device compatibility test-specific strategy, or a high-concurrency stability test-specific strategy.

13. The system according to claim 12, characterized in that, Based on test case generation strategies and multi-dimensional feature vectors, test cases for the recruitment platform are generated, including: For functional testing, functional test cases are generated by combining boundary values ​​and anomaly flow combinations based on functional testing-specific strategies and business constraints in multi-dimensional feature vectors. For UI device compatibility testing, based on a dedicated UI device compatibility testing strategy, and combining visual and device fingerprint feature information from multi-dimensional feature vectors, compatibility test cases are generated corresponding to the cross matrix of device, system, and resolution. For high-concurrency stability testing, stability test cases are generated based on a dedicated strategy for high-concurrency stability testing, combining timing and historical fault feature information from multi-dimensional feature vectors.

14. The system according to claim 1, characterized in that, Based on test case generation strategies and multi-dimensional feature vectors, test cases for the recruitment platform are generated, including: Based on reinforcement learning algorithms, the test case generation strategy and the test cases it generates are iteratively optimized in multiple rounds. In each iteration, the reward value of the test cases is calculated using a preset multi-dimensional balanced reward function. Test cases with reward values ​​greater than or equal to a preset threshold are selected. Simultaneously, the corresponding test case generation strategy is adjusted based on the reward value, and a new round of test cases is generated using the adjusted test case generation strategy. The multi-dimensional balanced reward function rewards based on test coverage and test execution efficiency, and penalizes based on test resource consumption. Based on the selected test cases, the test cases for the recruitment platform are obtained.

15. The system according to claim 1, characterized in that, The recommended agent for the historical regression test cases is a multi-agent collaborative architecture, which includes a demand parsing sub-agent, a retrieval enhancement sub-agent, and a ranking optimization sub-agent. The historical regression test case recommendation agent retrieves historical regression test cases from the historical test case library that match the test requirement data, including: The core requirement information is obtained by preprocessing the test requirement data and extracting key sentences through the requirement parsing sub-agent. The retrieval enhancement sub-agent, based on the core requirement information, performs keyword and semantic searches from the historical test case library to recall candidate historical regression test cases that match the core requirement information. The sorting optimization sub-agent sorts the recalled candidate historical regression test cases based on preset sorting weights, thus obtaining the sorted historical regression test cases.

16. A method for generating test cases for a recruitment platform, the method being implemented based on the recruitment platform test case generation system as described in any one of claims 1-15, the method comprising: Obtain testing requirement data from recruitment platforms; The test requirement data is preprocessed by a format conversion agent to obtain the test requirement text in the target format. By generating intelligent agents from test points, multimodal requirement parsing is performed on the test requirement text, and a three-dimensional feature vector of business entities, interaction logic, and constraints is constructed. Furthermore, by combining recruitment business prompts with scene recognition and business rule injection into the 3D feature vector, a scene-based 3D feature vector is obtained; Test points are generated based on scenario-based 3D feature vectors to obtain a test point set for recruitment scenarios; The test case generation agent performs test type classification and feature extraction on the test point set of recruitment scenarios to obtain multi-dimensional feature vectors; based on the test type classification results, the corresponding test case generation strategy is matched from the preset strategy library. Based on test case generation strategies and multi-dimensional feature vectors, test cases for the recruitment platform are generated. The agent recommends historical regression test cases by querying the historical test case library for historical regression test cases that match the test requirements data. Output test cases for the recruitment platform and historical regression test cases.

Images