A multi-dimensional adaptive examination generation method and system based on a dynamic skill map

Abstract

The application relates to the technical field of talent evaluation, and discloses a multi-dimensional adaptive examination generation method and system based on a dynamic skill map. The method comprises the following steps: real-time analysis of recruitment website and enterprise job description data, extraction of atomized skill phrases containing skill names, levels and correlation degrees, and generation of unified semantic representation; obtaining a standardized skill map through a large language model and transmitting the standardized skill map to a central database through a blockchain; generating an evolution strategy based on skill node weights and user historical answering characteristics, and differentiating and executing the evolution strategy by edge nodes; selecting a target edge node and configuring a skill knowledge cache; analyzing a candidate natural language answering request to generate ability matching characteristics, determining a candidate test question set, and outputting a test paper; and updating an ability evaluation model and an evolution strategy in combination with an answering result and a response time. The application can improve the authority and reliability of the evaluation result to meet the scientific and accurate needs of enterprises for talent evaluation.

Description

Technical Field

[0001] This application relates to the technical field of talent assessment, and in particular to a multi-dimensional adaptive test generation method and system based on dynamic skill graphs. Background Technology

[0002] Currently, traditional talent assessment systems generally employ static skill map construction methods, relying on historical job descriptions or industry standards updated at fixed intervals. This results in severely outdated skill map data, failing to reflect the dynamic changes in market demand for skills in real time. Furthermore, the lack of standardized skill representation leads to significant differences in how different companies describe the same skill, using terms like "data analysis," "data mining," and "statistical analysis." This hinders the effective integration of skill maps across companies and industries, impacting the accuracy and comparability of assessment results.

[0003] Furthermore, existing adaptive testing systems often employ a single-dimensional difficulty control strategy in their test paper generation logic, adjusting the difficulty of questions solely based on candidates' historical answer accuracy rates. This lacks multi-dimensional comprehensive optimization of skill coverage, question type ratios, and difficulty distribution. The interpretability of the assessment results is also relatively weak, failing to provide a chain of evidence linking questions to skill points. This results in a lack of authority and credibility, making it difficult to meet enterprises' needs for scientific and precise talent assessment.

[0004] As can be seen from the above, how to improve the authority and credibility of the assessment results to meet the needs of enterprises for scientific and precise talent evaluation still needs to be addressed. Summary of the Invention

[0005] To enhance the authority and credibility of assessment results and meet enterprises' needs for scientific and precise talent evaluation, this application provides a multi-dimensional adaptive test generation method and system based on dynamic skill graphs.

[0006] Firstly, this application provides a multi-dimensional adaptive exam generation method based on a dynamic skill graph, employing the following technical solution:

[0007] A multi-dimensional adaptive test generation method based on dynamic skill graphs includes:

[0008] The data acquisition terminal performs real-time analysis on job description data from recruitment websites and corporate job descriptions, extracts atomic skill phrases from the job descriptions and generates a unified skill semantic representation. The atomic skill phrases include skill name, skill level and skill relevance.

[0009] The skill semantic representation is based on a large language model for semantic disambiguation and synonym normalization. It uses a hierarchical density clustering algorithm to achieve automatic normalization and deduplication of skills, establishes cross-enterprise and cross-industry skill semantic mapping relationships, and transmits the standardized skill graph to the central database through a blockchain distributed version management mechanism.

[0010] Based on the skill node weights in the standardized skill graph and the user's historical answering behavior characteristics, a skill graph evolution strategy is dynamically generated when skill requirements change dynamically. The skill graph evolution strategy is executed by a local processing module deployed on an edge computing node. The local processing module responds in real time to sudden changes in skill requirements and performs batch optimization for periodic changes in skill requirements.

[0011] After the skill graph evolves, the evolved skill graph content is obtained. In the computing architecture composed of edge nodes and central servers, target edge nodes are selected based on skill node weights and user geographic location information, and the skill knowledge cache set of the target edge nodes is dynamically configured.

[0012] The system obtains the candidate's natural language answer request, parses the corresponding candidate ability assessment information from the natural language answer request, generates multi-dimensional ability matching features based on the ability assessment information and the skill node data in the configured cache, determines a candidate question set based on the multi-dimensional ability matching features using a dynamic weight multi-objective optimization algorithm, outputs the corresponding adaptive test paper based on the candidate question set, obtains the candidate's answer results and response time on the adaptive test paper, and updates the ability assessment model and skill graph evolution strategy based on the answer results and response time.

[0013] Optionally, the extraction of atomized skill phrases employs a fusion model of bidirectional long short-term memory networks and conditional random fields, and the method also includes:

[0014] Convert the job description text into a standardized sequence of word vectors;

[0015] The word vector sequence is input into the BiLSTM network, and through the collaborative computation of the forward hidden layer and the backward hidden layer, the output feature vector fused with context-dependent information is obtained.

[0016] The feature vector output by the BiLSTM network is input into the CRF layer. The optimal annotation sequence is calculated based on the preset skill phrase annotation rules. The accurate atomic skill phrases are extracted based on the optimal annotation sequence. The skill phrase annotation rules include annotation labels corresponding to skill name, level, and relevance.

[0017] Optionally, the large language model achieves semantic disambiguation and synonym normalization through a Prompt Tuning fine-tuning strategy. Other methods include:

[0018] Construct a domain-specific prompt template, which includes skill domain-specific information corresponding to the skill type guidance statement, skill domain-specific information corresponding to the description of typical application scenarios, and skill domain-specific information corresponding to the skill level matching reference case.

[0019] Standardized atomic skill phrases are used as training data, input into a large language model, and fine-tuned based on the constructed prompt templates;

[0020] The atomized skill phrases to be processed are input into the fine-tuned large language model, which outputs semantically consistent normalized skill descriptions and generates a skill semantic similarity matrix. The skill semantic similarity matrix is ​​used to help determine the clustering threshold of the hierarchical density clustering algorithm.

[0021] Optionally, the generation process of the skill graph evolution strategy incorporates a skill demand heat prediction model, and the method also includes:

[0022] A skill demand heat prediction model is constructed using a temporal convolutional network. The input of the skill demand heat prediction model is defined as a historical skill demand data sequence, and the output is the trend of skill demand heat change within a future preset period.

[0023] Historical skill demand data is input into a temporal convolutional network for training. Once the temporal convolutional network converges, it can predict the future demand for skills.

[0024] The predicted skill demand intensity is used as a dynamic adjustment factor for skill node weights; a skill graph evolution strategy that adapts to changes in potential skill demand is generated by combining skill node weights and users' historical answering behavior characteristics.

[0025] Optionally, the selection of target edge nodes is achieved through a multi-objective optimization algorithm, and the method also includes:

[0026] A multi-objective optimization function is constructed, and the optimization objectives are determined to be minimizing the network transmission latency between the examinee and the edge node, maximizing the utilization rate of the edge node cache resources, and balancing the load pressure of the edge node cluster. The function constraints are the upper limit of the edge node cache capacity and the network bandwidth threshold.

[0027] A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function and output the Pareto optimal set of target edge nodes;

[0028] By combining the geographical location information of the examinees and the real-time network quality data of the examinees, the optimal target edge node is selected from the Pareto optimal set.

[0029] Optionally, the generation of multi-dimensional ability matching features can include additional dimensions such as skill transferability and answer stability. Other methods may also include:

[0030] The ability to transfer skills is quantitatively assessed by the accuracy of candidates' answers to cross-skill related questions; the stability of answers is quantitatively assessed by the variance of the response time of questions of different difficulties at the same skill level.

[0031] The newly added dimension features and the core competency dimension features extracted based on the candidates' ability assessment information are normalized to eliminate differences in units; among them, the core competency dimension features include the degree of skill mastery and the suitability of the test questions.

[0032] An attention mechanism is used to assign initial weights to each dimension, and combined with the skill node data in the configured cache, a personalized multi-dimensional ability matching feature vector is generated.

[0033] Optionally, the update of the ability assessment model and skill map evolution strategy adopts an incremental learning mechanism, and the method also includes:

[0034] For the skill nodes corresponding to newly added answer data, only the relevant parameters in the ability assessment model and skill graph evolution strategy are locally updated, without retraining the entire ability assessment model.

[0035] An exponential moving average algorithm is used to dynamically decay the weights of skill nodes that have no long-term answer data support. By combining the decayed skill node weights, local update results, and the personalized multi-dimensional ability matching feature vector, the ability assessment model and skill graph evolution strategy are optimized.

[0036] Secondly, this application provides a multi-dimensional adaptive exam generation system based on a dynamic skill graph, employing the following technical solution:

[0037] A multi-dimensional adaptive test generation system based on dynamic skill graphs includes:

[0038] The job skill analysis and semantic generation module performs real-time analysis on job description data from recruitment websites and company job descriptions at the data collection end, extracts atomic skill phrases from the job descriptions and generates a unified skill semantic representation. The atomic skill phrases include skill name, skill level and skill relevance.

[0039] The skill semantic normalization and graph transmission module performs semantic disambiguation and synonym normalization based on a large language model. It achieves automatic normalization and deduplication of skills through hierarchical density clustering algorithm, establishes cross-enterprise and cross-industry skill semantic mapping relationship, and transmits the standardized skill graph to the central database through blockchain distributed version management mechanism.

[0040] The skill graph dynamic evolution module dynamically generates a skill graph evolution strategy when skill requirements change dynamically, based on the skill node weights in the standardized skill graph and the user's historical answer behavior characteristics. The skill graph evolution strategy is executed by a local processing module deployed on an edge computing node. The local processing module responds in real time to sudden changes in skill requirements and performs batch optimization for periodic changes in skill requirements.

[0041] The edge node selection and cache configuration module obtains the evolved skill graph content after the skill graph evolves. In the computing architecture composed of edge nodes and central server, it selects target edge nodes based on skill node weights and user geographic location information, and dynamically configures the skill knowledge cache set of the target edge nodes.

[0042] The adaptive exam generation and model update module obtains the natural language answer request input by the examinee, parses the corresponding examinee ability assessment information based on the natural language answer request, generates multi-dimensional ability matching features based on the ability assessment information and the skill node data in the configured cache, determines a candidate question set based on the multi-dimensional ability matching features through a dynamic weight multi-objective optimization algorithm, outputs the corresponding adaptive exam paper based on the candidate question set, obtains the examinee's answer results and response time on the adaptive exam paper, and updates the ability assessment model and skill graph evolution strategy based on the answer results and response time.

[0043] Thirdly, this application provides an electronic device that adopts the following technical solution:

[0044] An electronic device includes a processor running a program for the multi-dimensional adaptive test generation method based on dynamic skill graphs as described in any one of the preceding claims.

[0045] Fourthly, this application provides a computer storage medium, which adopts the following technical solution:

[0046] A computer storage medium storing a program for the multi-dimensional adaptive test generation method based on dynamic skill graphs as described in any one of the above.

[0047] In summary, this application includes at least one of the following beneficial technical effects:

[0048] By employing a standardized skills modeling and dynamic adaptation mechanism throughout the entire process, the authority and credibility of talent assessment are significantly enhanced, providing core support for enterprises' scientific and precise talent evaluation. The solution controls accuracy from the source of job skills, precisely extracting atomic skill phrases through a BiLSTM and CRF fusion model. Combined with a large language model fine-tuned by Prompt Tuning and a hierarchical density clustering algorithm, it achieves cross-enterprise and cross-industry skill semantic normalization. Furthermore, blockchain-based distributed version management ensures the immutability and traceability of the skill graph data, constructing an authoritative and unified skills assessment benchmark. This fundamentally solves the problems of vague skill definitions and inconsistent standards in traditional assessments. Simultaneously, a dynamic graph evolution strategy based on skill demand heat prediction ensures that assessment content matches changes in enterprise job skill requirements in real time, making the assessment dimensions closely aligned with actual enterprise hiring standards.

[0049] Building upon this foundation, the solution enhances assessment accuracy through multi-dimensional competency modeling and an adaptive exam generation mechanism. It adds dimensions of skill transferability and answer stability, incorporates an attention mechanism to achieve personalized competency feature matching, and uses a dynamic weighted multi-objective optimization algorithm to select questions, generating adaptive exam papers that closely match the candidates' actual abilities. Coupled with an incremental learning update mechanism and a skill decay strategy, the competency assessment model is continuously iterated and optimized, ensuring that the assessment results accurately reflect the candidates' skill mastery level and job suitability potential. The collaborative architecture of edge nodes and the central server further guarantees the smoothness of the examination process and the efficiency of data processing, ultimately forming a closed-loop chain of "authoritative skill benchmarks - dynamic demand adaptation - accurate competency assessment - continuous model optimization," fully meeting enterprises' core needs for scientific and precise talent assessment. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating a multi-dimensional adaptive test generation method based on a dynamic skill graph, according to an exemplary embodiment.

[0051] Figure 2 This is a structural block diagram of a multi-dimensional adaptive test generation system based on a dynamic skill graph, according to an exemplary embodiment. Detailed Implementation

[0052] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0053] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0054] This application discloses a multi-dimensional adaptive test generation method based on a dynamic skill graph, referring to... Figure 1 ,include:

[0055] S100 performs real-time analysis of job description data from recruitment websites and company job descriptions at the data acquisition end, extracts atomic skill phrases from the job descriptions and generates a unified skill semantic representation. The atomic skill phrases include skill name, skill level and skill relevance.

[0056] Specifically, S100 includes the following detailed execution:

[0057] Step 1, Real-time collection and access of job description data:

[0058] The data acquisition end first establishes a multi-source data access channel to achieve real-time connection to two types of data sources: recruitment websites and company job descriptions. For recruitment website data, a targeted web crawler is deployed (with anti-crawling mechanisms configured, such as dynamic User-Agent switching and request frequency control) to accurately capture core fields such as job description text, job responsibilities, and job requirements from the job details page. For company job descriptions, batch uploading and real-time parsing of various document formats (Word, PDF, Excel, TXT) are supported, and the text content is extracted using a document parsing component (such as Apache Tika).

[0059] Meanwhile, to ensure data real-time performance, a message queue (such as Kafka) is used to temporarily store and transfer the collected raw data, and a data collection frequency threshold is set (which can be customized according to the company's needs, such as collecting data every 5 minutes during peak recruitment season and every 30 minutes during off-season) to ensure timely acquisition of the latest job description data.

[0060] Step 2, Preprocessing of collected data:

[0061] The collected raw job description data undergoes standardized preprocessing to eliminate data noise and format differences, laying the foundation for subsequent skill phrase extraction. Specific operations include:

[0062] Denoising processing: Filter out irrelevant characters (such as special symbols, advertising watermarks, line breaks, spaces), and eliminate invalid text (such as repeated general template statements like "This position welcomes outstanding talents to join").

[0063] Format unification: Uniformly convert the text from different data sources to UTF-8 encoding, and segment long texts (split according to semantic logic, with each segment not exceeding 500 words to avoid semantic fragmentation during subsequent model processing).

[0064] Word segmentation and stop word removal: Use Chinese word segmentation tools (such as jieba, HanLP) to segment the preprocessed text to generate a sequence of words; at the same time, load a preset stop word list (including general words without actual skill pointers such as "of", "is", "responsible", etc.), filter out stop words, and retain the core words with actual semantics.

[0065] Data verification: Verify the integrity of the text through a rule engine (such as eliminating invalid job descriptions with less than 20 characters), and mark the data source (such as "Zhilian Recruitment-Product Manager Position", "A Enterprise-Back-end Development Job Description") for subsequent traceability and management.

[0066] Step 3, Precise extraction of atomic skill phrases: <S

[0067] Based on the preprocessed job description text data, use a fusion model of bidirectional long short-term memory network (BiLSTM) and conditional random field (CRF) to extract atomic skill phrases containing skill names, skill levels, and skill correlations. The core execution process is as follows:

[0068] Feature input preparation: First, convert the sequence of words obtained by word segmentation after preprocessing into a standardized fixed-dimensional word vector sequence (implemented using pre-trained Chinese word vector models such as Word2Vec, BERT), and at the same time incorporate词性 features (such as nouns, verbs, adjectives), and construct a multi-dimensional input feature matrix through feature concatenation to provide structured and standardized input data for subsequent model calculations, ensuring that skill texts with different expressions have unified feature dimensions.

[0069] Context feature capture: Input the constructed multi-dimensional input feature matrix (i.e., the standardized word vector sequence) into the BiLSTM network. Use the forward hidden layer of BiLSTM to capture the semantic information of the previous text and the backward hidden layer to capture the semantic information of the subsequent text. Through the collaborative calculation of the two hidden layers, output a feature vector that integrates the complete context-dependent information. This step can accurately identify the semantic boundaries of skill names. For example, it can distinguish whether "Python" in the context is used as a skill name rather than an ordinary noun, solving the problem of blurred boundaries caused by single-direction semantic analysis.

[0070] Labeling Optimization and Phrase Extraction: The feature vector output by the BiLSTM network is input into the CRF layer. Based on the preset skill phrase labeling rules (where skill name is labeled "SKILL", skill level is labeled "LEVEL", and skill relevance is labeled "RELATION"), the optimal labeling sequence is calculated using the sequence labeling optimization capability of the CRF layer. Atomized skill phrases are extracted based on this optimal labeling sequence, specifically including:

[0071] Skills covered include professional skills such as "Java development" and "data analysis," as well as general skills such as "project management" and "communication and coordination."

[0072] Skill level: Identified by adverbs of degree such as "proficient", "master", "understand", and "possess basic skills";

[0073] Skill relevance: Determined through semantic dependency analysis. For example, in the sentence "possesses Java development skills and can complete database design", the relevance between "Java development" and "database design" is marked as "high".

[0074] Phrase validation and filtering: The atomic skill phrases extracted from the CRF layer are manually validated according to rules to remove non-skill phrases such as "excellent" and "positive" that do not have actual skill references. Finally, a structured set of atomic skill phrases is output (format example: {Skill name: Java development, skill level: proficient, skill relevance: database design - high}), ensuring the accuracy and effectiveness of the extraction results.

[0075] By using a phased processing model that combines BiLSTM and CRF, along with word vector standardization, contextual semantic capture, annotation rule optimization, and manual verification, we have achieved accurate and complete extraction of atomic skill phrases (including name, level, and relevance) from job description texts. This solves the problems of ambiguous skill boundaries and inaccurate identification of levels and relevance in traditional methods, and provides high-quality structured basic data for subsequent unified skill semantic representation.

[0076] Step 4, Generation of unified skill semantic representation:

[0077] To address the issue of discrepancies in skill descriptions across different data sources (e.g., "Python programming" and "Python development" are essentially the same skill), a unified semantic representation of the extracted atomic skill phrases is generated. Specific steps:

[0078] Semantic feature extraction: A pre-trained large language model (such as BERT-base-chinese) is used to semantically encode each atomic skill phrase, generating a fixed-dimensional semantic vector;

[0079] Standardization of representation: Based on an industry skills dictionary (integrating standard skills terms from multiple industries such as IT, finance, and manufacturing), the semantic vector is normalized to ensure that different expressions of the same skill correspond to the same semantic vector;

[0080] Structured storage: Atomized skill phrases are bound to their corresponding unified semantic vectors to form a key-value pair structure of "skill phrase - semantic vector", which is then stored in a temporary database to provide input data for subsequent semantic disambiguation and normalization processing of S200.

[0081] By collecting and preprocessing multi-source data in real time, the standardization of job description data was achieved, eliminating format differences and noise interference from different data sources. The precise extraction by the BiLSTM-CRF fusion model ensured the integrity and accuracy of atomic skill phrases, clarifying the three core elements of skills (name, level, and relevance), and avoiding the omissions and errors in traditional skill extraction. The generation of unified skill semantic representations initially solved the pain point of inconsistent skill descriptions, laying a semantic foundation for the subsequent construction of cross-enterprise and cross-industry skill maps.

[0082] S200's skill semantic representation is based on a large language model for semantic disambiguation and synonym normalization. It uses a hierarchical density clustering algorithm to achieve automatic normalization and deduplication of skills, establishes cross-enterprise and cross-industry skill semantic mapping relationships, and transmits the standardized skill graph to the central database through a blockchain distributed version management mechanism.

[0083] Specifically, the detailed execution of S200 includes:

[0084] Step 1, semantic disambiguation processing of skill semantic representation:

[0085] Semantic disambiguation of atomized skill phrases generated by S100 is performed using a large language model (such as BERT-base-chinese) to address the ambiguity of skill descriptions. Specific operations include: semantic context modeling, placing skill phrases within their original contexts and capturing the specific meaning of skills in specific contexts using a large language model; polysemous word identification, recognizing polysemous words in skill descriptions and determining their actual meanings based on the context; dictionary-based semantic disambiguation, using an industry-specific skill dictionary as a semantic reference to perform multi-dimensional disambiguation of skill descriptions; and outputting standardized semantic representations, mapping the disambiguated skill descriptions to a unified semantic space to ensure that different descriptions of the same skill receive the same semantic representation.

[0086] Step 2, Synonym Normalization Process:

[0087] After semantic disambiguation, the skill descriptions undergo synonym normalization to unify different descriptions of the same skill into a standard description. Specific steps include: 1) Synonym database construction: Based on industry skill dictionaries and historical data, a synonym database containing skill synonym mappings is built. 2) Synonym matching: Using word vector-based similarity calculations (such as cosine similarity), skill descriptions are matched against the synonym database. 3) Normalization decision: Based on the matching results and a confidence threshold (e.g., >0.8), successfully matched skill descriptions are normalized to the standard description. 4) Normalization verification: The accuracy of the normalization results is ensured through manual rule verification (e.g., checking whether the normalized skill descriptions conform to industry standards).

[0088] Step 3: Hierarchical density clustering algorithm performs automatic normalization and deduplication:

[0089] Hierarchical density clustering algorithm is used to automatically normalize and deduplicate normalized skills, forming a structured skill set. Specific operations include: density calculation, which calculates the local density of the semantic vector of each skill (based on Euclidean distance or cosine similarity) to identify high-density regions; hierarchical clustering, which uses a density-based hierarchical clustering algorithm to cluster similar skills into the same cluster based on the density calculation results; intra-cluster optimization, which further optimizes the skills within each cluster to ensure semantic consistency; and deduplication, which removes duplicates from the clustered skill set, retaining representative skill descriptions from each cluster and deleting duplicate skills to ensure the conciseness and accuracy of the skill map.

[0090] Step 4: Establish cross-enterprise and cross-industry skill semantic mapping relationships:

[0091] Based on the normalized and deduplicated skill sets, cross-enterprise and cross-industry skill semantic mapping relationships are established. Specific operations include: cross-enterprise mapping, mapping expressions representing the same skill across different enterprises; cross-industry mapping, mapping expressions representing the same skill across different industries; mapping relationship storage, storing the established mapping relationships as key-value pairs of skills and mapped skills for easy subsequent querying and application; and mapping relationship verification, ensuring the accuracy of the mapping relationships through expert review and data validation.

[0092] Step 5: The blockchain distributed version management mechanism transmits the data to the central database.

[0093] The standardized skill graph is transmitted to a central database via a blockchain-based distributed version control mechanism. Specific operations include: version control, creating a new blockchain block for each skill graph update, containing the updated content, update time, and updater information; distributed storage, storing skill graph update records across multiple nodes on the blockchain to ensure data decentralization and immutability; consensus mechanism, employing a consensus mechanism (such as PBFT or PoS) to ensure all nodes agree on skill graph updates; data transmission, transmitting the standardized skill graph to the central database via the blockchain network to ensure data integrity and consistency; and version backtracking, supporting the querying and backtracking of historical versions of the skill graph for convenient data management and issue tracing.

[0094] By employing semantic disambiguation and synonym normalization in skill semantic representation, the problems of polysemy and variability in skill descriptions are resolved. Hierarchical density clustering algorithm is used to achieve automatic normalization and deduplication of skills, ensuring the conciseness and accuracy of the skill graph. By establishing cross-enterprise and cross-industry skill semantic mapping relationships, skill barriers between different enterprises and industries are broken down. The introduction of blockchain distributed version management mechanism ensures the integrity, consistency, and traceability of skill graph data.

[0095] S300 dynamically generates skill graph evolution strategies when skill requirements change dynamically, based on the skill node weights in the standardized skill graph and the user's historical answer behavior characteristics. The skill graph evolution strategies are executed by the local processing module deployed on the edge computing node. The local processing module responds in real time to sudden changes in skill requirements and performs batch optimization for periodic changes in skill requirements.

[0096] Specifically, the detailed execution of S300 includes:

[0097] Step 1, Preprocessing for integrating skill node weights with user's historical answering behavior characteristics:

[0098] First, retrieve the standardized skill map generated by S200 from the central database, and extract the initial weights of all skill nodes in the map. The initial weights are calculated based on the frequency of skill occurrence in job descriptions across enterprises and industries, the average skill relevance, and industry expert ratings (e.g., the initial weight of core skills is set to 0.8, and that of general skills is set to 0.4).

[0099] Meanwhile, user historical answering behavior features are extracted from the user answering behavior database, including: the accuracy rate of questions corresponding to each skill node, the answering response time, the distribution of wrong question types, and the stability coefficient of skill mastery (the variance of the accuracy rate of questions of different difficulty for the same skill).

[0100] The preprocessing of the above two types of data includes the following operations:

[0101] Missing value imputation: For skill nodes with no answer records, the average answer characteristics of users in the same industry and the same position are used to fill in the missing values.

[0102] Data normalization maps indicators such as skill node weights, answer accuracy, and response time to the [0,1] interval to eliminate differences in units.

[0103] Feature fusion is used to construct a correlation matrix of "skill node weight - answer behavior feature". Each element in the matrix corresponds to the weight of a single skill node and the user's mastery feature of that skill, providing a data foundation for the generation of subsequent evolutionary strategies.

[0104] Step 2, Identification and classification of dynamic changes in skill requirements:

[0105] Build a skills requirement change monitoring engine to capture dynamic change signals of internal and external skills requirements in real time. The specific operation is as follows:

[0106] External demand changes are monitored by connecting to real-time job data from recruitment websites and industry skills trend reports via API interfaces. By comparing the skill differences between new job descriptions and existing skill maps, emerging skill demands (such as "generative AI applications") or changes in the popularity of skill demands (such as the demand for "Python skills" increasing from 15% to 30%) are identified.

[0107] Internal demand change monitoring: Based on internal job adjustment notices and talent assessment feedback reports, identify changes in the company's customized skill requirements (such as a department adding a "Industrial Internet Platform Operation and Maintenance" skill requirement).

[0108] The change types are classified into sudden changes (such as new skill requirements caused by policy adjustments and emergency recruitment by enterprises) and cyclical changes (such as skill demand fluctuations during peak recruitment seasons and quarterly skill updates in the industry) based on the frequency of occurrence, scope of impact, and response time requirements. Response priorities and processing time thresholds are marked for each type of change.

[0109] Step 3, Dynamic Generation of Skill Graph Evolution Strategy:

[0110] Based on the integrated "skill node weight - answer behavior feature" correlation matrix and the classified skill demand changes, a multi-objective optimization algorithm is used to generate a skill graph evolution strategy. Specific operations include:

[0111] The strategic objectives are set, and the core objectives of the evolution strategy are clearly defined as "adapting to changes in skill requirements", "improving the accuracy of user ability assessment", and "optimizing the redundancy of skill graph nodes".

[0112] Sudden Change Strategy Generation: In response to sudden changes in skill requirements, a real-time updated strategy is generated. This includes adding new skill nodes, adjusting the weight of related skill nodes (e.g., adding the skill "Generative AI Application" and increasing the weight of its related skill "Python Development"), and marking the skill as a high-priority assessment node.

[0113] Periodic change strategy generation: In response to periodic changes in skill requirements, batch optimization strategies are generated, including batch adjustment of skill node weights (such as increasing the weight of general skills such as "resume writing" and "interview skills" during peak recruitment season), merging low-frequency redundant skill nodes, and optimizing skill relationships (such as adjusting the relationship between "data analysis" and "data visualization").

[0114] Strategy validation involves verifying the effectiveness of the evolution strategy through simulation testing. If the improvement in the matching degree between the test results and the actual job requirements after the strategy is applied is less than 5%, the strategy parameters are adjusted until the preset target is met.

[0115] Step 4, Deployment and policy execution of the local processing module on the edge computing node:

[0116] The generated skill graph evolution strategy is distributed to the local processing module deployed on the edge computing node, and a differentiated processing flow is executed according to the change type. The specific operations are as follows:

[0117] The edge node module configuration preloads basic skill graph data and lightweight optimized algorithm models for the local processing module, ensuring that the module has the ability to independently handle changes in local skill requirements without relying on the real-time computing power support of the central server.

[0118] Real-time response to sudden changes: When a sudden change in skill requirements is detected, the central server will send a simplified evolution strategy instruction (such as "add skill node X") to the edge nodes. The local processing module will immediately execute the strategy, update the locally cached skill graph fragments, and synchronously adjust the question generation rules for subsequent adaptive exams. The response time is controlled within 1 minute.

[0119] Batch optimization for periodic changes: When periodic changes in skill requirements are detected, the local processing module accumulates change instructions within a certain period (such as skill weight adjustment requirements within a week) and performs batch optimization operations during non-peak assessment periods (such as nighttime). This includes batch updating skill node weights and reconstructing skill relationships. After optimization, the results are synchronized to the central database to avoid occupying computing resources during peak assessment periods.

[0120] The edge nodes will report the strategy execution status and skill graph update logs to the central server in real time, forming a closed loop of "strategy generation-execution-feedback".

[0121] Based on the execution steps of S300 described above, firstly, by integrating skill node weights with user answering behavior characteristics, a precise correlation between "skill requirements and user capabilities" is achieved, solving the pain point that the "static" nature of traditional skill graphs cannot adapt to changes in user capabilities; secondly, by classifying and identifying changes in skill requirements and generating differentiated strategies, the skill graph ensures rapid response and efficient adaptation to sudden and periodic changes in requirements, improving the real-time nature and relevance of assessment content; and thirdly, through localized execution of edge computing nodes, lightweight and low-latency processing of strategies is achieved, ensuring real-time response to sudden requirements while avoiding the occupation of central server computing power by batch optimization, thus improving the overall system operating efficiency.

[0122] After the skill graph evolves, S400 acquires the evolved skill graph content. In the computing architecture composed of edge nodes and central servers, it selects target edge nodes based on skill node weights and user geographic location information, and dynamically configures the skill knowledge cache set of the target edge nodes.

[0123] The detailed execution of S400 includes:

[0124] Step 1, Obtaining and Verifying the Content of the Evolved Skill Map:

[0125] First, through the communication link between the central server and the edge computing nodes, the latest skill map content after the local processing module of the edge node in the S300 executes the evolution strategy is obtained. Specific operations include:

[0126] Data synchronization request: The central server sends a graph synchronization instruction to all edge nodes that have executed the evolution strategy. The instruction includes identification information such as the batch of evolution strategy execution and timestamps to ensure that the corresponding version of the evolved graph is obtained.

[0127] Extract the graph content, respond to the synchronization request at the edge nodes, and upload the evolved complete skill graph fragment (including the updated skill nodes, adjusted node weights, and optimized skill relationships).

[0128] For integrity and consistency verification, the central server uses a hash verification algorithm (such as MD5) to verify the integrity of the received graph content to avoid loss or tampering during data transmission. At the same time, it compares the same type of skill graph fragments uploaded by different edge nodes to verify the consistency of the evolution results. If there are differences, a resynchronization mechanism is triggered to correct the deviation based on the evolution strategy benchmark file stored on the central server.

[0129] Structured integration integrates all evolved graph fragments synchronized from edge nodes into a complete evolved skill graph, which is then stored in a temporary cache area of ​​the central database, providing complete data support for subsequent node selection and cache configuration.

[0130] Step 2: The initialization of the collaborative architecture between edge nodes and the central server is complete.

[0131] Activate the collaborative computing architecture between edge nodes and the central server to complete resource scheduling and communication link preprocessing. Specific operations include:

[0132] Architecture status detection: The central server uses a heartbeat detection mechanism (such as sending a detection packet every 10 seconds) to confirm the running status of all edge nodes (such as CPU utilization, memory usage, network bandwidth), marks the list of available edge nodes, and removes nodes that are offline or whose resource usage exceeds the threshold (such as CPU utilization > 80%).

[0133] Communication link optimization: a dedicated communication link between the central server and available edge nodes is built using the TCP / IP protocol, data compression algorithms (such as LZ4) are configured to reduce transmission latency, and encrypted transmission mechanisms (such as SSL / TLS) are enabled to ensure data transmission security.

[0134] The collaborative scheduling rules are loaded, and the central server loads the preset collaborative scheduling strategy, which clarifies the division of labor mechanism between the central server being responsible for global decision-making (target node selection, cache configuration rule formulation) and edge nodes being responsible for local execution (cache storage, data interaction), ensuring the orderly operation of the architecture.

[0135] Step 3, Target edge node selection based on skill node weights and user geographic location:

[0136] Once the collaborative architecture is in place, the target edge nodes are precisely selected based on core screening metrics. Specific operations include:

[0137] The core indicators are collected and preprocessed. The weight data of each skill node is extracted from the integrated evolved skill graph and sorted by weight to generate a "high-priority skill node list" (such as core skills with a weight ≥ 0.7). At the same time, the user's geographical location information is obtained through the user terminal's positioning module (such as GPS, base station positioning), converted into standardized latitude and longitude coordinates, and abnormal positioning data (such as invalid data with a positioning deviation of more than 1 kilometer) is filtered out.

[0138] The edge node candidate set is filtered by the central server based on the user's latitude and longitude coordinates, and edge nodes within a preset threshold (such as 50 kilometers) of the user's geographical range are selected to form a candidate node set; if the candidate node set is empty, the filtering range is expanded to 100 kilometers to ensure that there are matching edge nodes.

[0139] Multi-dimensional scoring and optimal node determination: A node selection scoring model is constructed, with "skill node weight matching degree" and "geographical distance" as the core scoring dimensions: Skill node weight matching degree = proportion of high-priority skills pre-cached in candidate nodes × 0.6 + node expandable cache space × 0.4, Geographical distance score = 1 / (1 + straight-line distance between user and node), Comprehensive score = weight matching degree × 0.7 + geographical distance score × 0.3; The comprehensive score of all nodes in the candidate node set is calculated, and the node with the highest score is selected as the target edge node;

[0140] Upon receiving the selection result feedback, the central server sends the IP address, communication port, and other connection information of the target edge node to the user terminal, and simultaneously sends a command to the target edge node to prepare to receive cache configuration.

[0141] Step 4, Dynamic configuration of the target edge node skill knowledge cache set:

[0142] Based on the selected target edge nodes and the evolved skill graph, a personalized dynamic configuration of the skill knowledge cache set is completed. Specific operations include:

[0143] For cache content filtering, the central server filters the skill knowledge content to be cached based on the user group characteristics covered by the target edge nodes (such as industry and common job positions) and the list of high-priority skill nodes in the evolved skill graph. This includes detailed analysis of high-weight skills, corresponding test question resource indexes, skill association data, etc. At the same time, cache priority is set, with core skill knowledge (weight ≥ 0.7) set as first-level cache and general skill knowledge (weight 0.4-0.7) set as second-level cache.

[0144] The cache capacity is dynamically allocated. The target edge node reports the current remaining cache space to the central server. The central server allocates cache resources based on the priority and size of the cached content using a greedy algorithm: it prioritizes allocating sufficient space for the first-level cache content, and allocates the remaining space proportionally to the second-level cache; if the remaining space is insufficient, it compresses non-core data in the second-level cache (such as simplified parsing of skill knowledge) to ensure that the core skill knowledge is completely cached.

[0145] For data transmission and writing of cached information, the central server transmits the filtered skill knowledge cached content in batches to the target edge nodes through an optimized communication link. During the transmission process, a fragmented transmission mechanism (such as 10MB per fragment) is used to avoid excessive bandwidth consumption in a single transmission. After receiving the data, the target edge nodes write it to the corresponding storage area according to the cache priority (such as writing the first-level cache to high-speed memory and the second-level cache to solid-state drives), and establish a cache index table to record the storage location of the cached content and the associated skill nodes.

[0146] Cache configuration verification and update: After the target edge node completes the cache writing, it sends a configuration completion signal to the central server. The central server randomly selects a portion of the cached content for verification (such as checking the weight of skill nodes and the accuracy of the question index); at the same time, it sets cache update trigger conditions (such as the skill graph evolving again, the access frequency of cached content being lower than a threshold) to ensure that the cache set always matches the latest skill requirements.

[0147] By acquiring and verifying the evolved skill graph, it is ensured that subsequent node selection and caching configuration are based on the latest and most accurate skill data, guaranteeing data consistency throughout the process. Target edge node selection based on skill node weights and user geographic location achieves dual optimization of "skill requirement adaptation" and "geographic proximity," ensuring that target nodes can provide users with the core skill-related resources they need while shortening data transmission distances. Dynamic caching configuration stores high-priority skill knowledge in advance at target edge nodes, avoiding frequent data requests to the central server during exam generation, significantly reducing answer latency and improving user experience.

[0148] S500 acquires the natural language answer request input by the examinee, parses the corresponding examinee ability assessment information based on the natural language answer request, generates multi-dimensional ability matching features based on the ability assessment information and the skill node data in the configured cache, determines the candidate question set based on the multi-dimensional ability matching features through a dynamic weight multi-objective optimization algorithm, and outputs the corresponding adaptive test paper based on the candidate question set, acquires the examinee's answer results and response time on the adaptive test paper, and updates the ability assessment model and skill map evolution strategy based on the answer results and response time.

[0149] The detailed implementation of S500 includes:

[0150] Step 1: Acquiring and preprocessing the candidate's natural language response request:

[0151] First, the system receives natural language answer requests from test takers via their terminal devices, including multimodal input such as text answers and speech-to-text content. Specific operations include: input interface adaptation, supporting multiple input formats (such as text box input, voice input, and handwriting recognition), and receiving test taker answers through a unified API interface; request content standardization, performing standardized preprocessing on natural language answer requests, including removing special characters, standardizing punctuation formats, and converting to lowercase, to ensure consistency in subsequent processing; and request content validation, verifying the validity of request content through a rule engine (e.g., removing invalid answers that are too short or requests containing illegal characters), and generating a unique identifier (such as a UUID) for each request to facilitate subsequent tracking and analysis.

[0152] Step 2, Analysis of Candidate Ability Assessment Information:

[0153] Based on the preprocessed natural language answer requests, key assessment information reflecting the candidate's abilities is parsed. Specific operations include: semantic feature extraction, using a pre-trained Chinese NLP model to perform deep semantic analysis on the answer content and extract key semantic features (such as keywords, semantic relationships, and logical structures); ability dimension mapping, mapping the extracted semantic features to skill nodes in the skill graph to identify the candidate's performance in each skill dimension; ability level quantification, based on skill node weights and the matching degree between the candidate's answer content and the standard answer, quantifying the candidate's ability level in each skill dimension (e.g., using continuous values ​​of 0-1 to represent the degree of skill mastery); and ability assessment information structuring, organizing the parsed ability assessment information into structured data, including skill node IDs, ability level values, and the matching degree between the answer content and the skill node, providing a foundation for subsequent multi-dimensional feature generation.

[0154] Step 3, Generation of multi-dimensional capability matching features:

[0155] By combining candidate competency assessment information with skill node data in the configured cache, multi-dimensional features reflecting the matching degree between candidates' competencies and skill requirements are generated. Specific operations include: skill node data retrieval, extracting skill node data related to the candidate's competency assessment information from the skill knowledge cache of the target edge node, including skill node weights, associated skills, and sample questions; feature dimension construction, constructing multi-dimensional feature vectors, including skill matching degree (the degree of matching between the current skill and the candidate's competency), competency gap degree (the gap between the candidate's competency and the skill requirements of the target position), skill correlation degree (the strength of the correlation between this skill and other skills), and competency stability (the consistency of the candidate's performance on this skill); feature fusion and normalization, using a weighted fusion method to integrate the features of each dimension into a unified matching feature vector, while normalizing each dimension feature to the [0,1] interval to eliminate dimensional differences; and feature vector output, storing the generated multi-dimensional competency matching feature vector as structured data to provide input for subsequent question selection.

[0156] Step 4, Determining the candidate question set:

[0157] Based on multi-dimensional ability matching features, a dynamic weighted multi-objective optimization algorithm is used to determine the most suitable candidate question set. The specific operations include: defining the objective function and constructing a multi-objective optimization model, with objective functions including "maximizing skill matching degree," "question difficulty suitability," and "ability improvement potential"; dynamic weight allocation, dynamically adjusting the weights of each optimization objective according to the candidate's current ability level and skill requirements (e.g., setting the weight of weak areas to 0.4 and the weight of strong areas to 0.2); executing the optimization algorithm, using a genetic algorithm or particle swarm optimization algorithm to solve the multi-objective optimization problem and generate a candidate question set; question set screening and sorting, selecting questions that meet the threshold requirements from the optimization results and sorting them from high to low matching degree to ensure the diversity and relevance of the candidate question set; and set verification, verifying the matching degree between the candidate question set and the candidate's abilities through simulated testing. If the matching degree is lower than a preset threshold (e.g., 0.7), the optimization is returned for re-optimization.

[0158] Step 5, Output of the adaptive exam paper:

[0159] Based on a defined set of candidate questions, a personalized adaptive exam paper is generated and output. Specific operations include: question combination optimization, which optimizes the question combination structure according to the examinee's ability characteristics and the difficulty distribution of the question set (e.g., from easy to difficult, balanced skill coverage); exam paper format generation, which converts the optimized question combination into a standard exam paper format, including question content, options, and scoring criteria; exam paper integrity verification, which checks whether the exam paper contains sufficient skill coverage, has a reasonable difficulty distribution, and an appropriate number of questions (e.g., 15-20 questions); and exam paper output, which transmits the generated adaptive exam paper to the examinee's terminal in real time via an encrypted channel, ensuring the integrity and security of the exam paper content.

[0160] Step 6, Obtaining the answer results and response time:

[0161] After a candidate completes the adaptive testing paper, the system obtains their answer results and response time. Specific operations include: answer data collection, which collects interactive data such as the candidate's answer content, answer time, and skipped questions in real time through the terminal device; data preprocessing, which standardizes the collected answer data (e.g., converting answer time to seconds and removing invalid characters from answer content); result and time separation, which separates the answer results (correct / incorrect) from the response time for subsequent analysis; and data integrity verification, which checks the integrity of the answer data (e.g., whether all questions have answer records), removes invalid data, and ensures the accuracy of subsequent analysis.

[0162] Step 7, updating the competency assessment model and skill map evolution strategy:

[0163] Based on the acquired answer results and response time, the ability assessment model and skill graph evolution strategy are dynamically updated. Specific operations include: updating the ability assessment model using online learning algorithms (such as incremental learning and online SVM), incorporating newly acquired answer data into model training to improve the model's prediction accuracy of examinees' abilities; adjusting the skill graph evolution strategy based on the matching degree between new data and the skill graph (such as adjusting skill node weights and optimizing skill relationships); strategy verification and feedback, verifying the effectiveness of the updated ability assessment model and skill graph evolution strategy through simulation testing, and deploying the updated model and strategy to the system if verification is successful; and strategy deployment, synchronizing the updated ability assessment model and skill graph evolution strategy to the central server and edge nodes to ensure continuous system optimization.

[0164] Through the execution steps of the S500 described above, firstly, by accurately parsing natural language answer requests, a deep assessment of candidates' abilities is achieved, solving the problem of a single dimension of ability assessment in traditional examinations; secondly, by generating multi-dimensional ability matching features and dynamically optimizing weights, the accuracy and personalization of test question selection are ensured, improving the relevance of the examination; and thirdly, by providing real-time feedback on answer results and response time, a dynamic update mechanism for the ability assessment model and skill map is established, enabling the system to be continuously optimized, constantly improving the adaptability and accuracy of the examination, and providing candidates with a more accurate and efficient assessment experience.

[0165] Based on the overall technical solution in this embodiment, taking the talent assessment scenario of a large IT company recruiting backend development engineers as an example, this technical solution ensures the authority of the assessment from the source of skill standard construction. After the company's HR initiates the talent assessment request, the system uses the multi-source data collection mechanism of the S100 steps to simultaneously capture backend development job descriptions from platforms such as Zhaopin.com and BOSS Zhipin, and parses the company's internal job descriptions. With the help of the BiLSTM-CRF fusion model, it accurately extracts atomic skill phrases such as "Java development", "microservice architecture", and "database optimization", and clarifies the level requirements of each skill (such as "proficient in Spring Boot") and their correlation (such as the high correlation between "microservice architecture" and "distributed transactions"). Subsequently, through the S200 steps of semantic disambiguation and synonym normalization of the large language model, "Java programming" and "Java development" are unified into standard skill descriptions. Then, hierarchical density clustering is used to remove duplicates and form a standardized skill set. Finally, the data is transmitted to the central database through a blockchain distributed version management mechanism to ensure the immutability and traceability of the skill standards. This lays the foundation for a cross-platform, standardized, and authoritative skill benchmark for assessment, fundamentally solving the problems of vague skill definitions and inconsistent standards in traditional assessments.

[0166] Through dynamic adaptation and precise matching mechanisms, the assessment process is made more scientific and accurate, further enhancing the credibility of the results. In response to the recent sudden change in the company's "cloud-native application" skill requirements, the system quickly identified the type of change using the skill requirement monitoring engine in step S300, dynamically generated a skill graph evolution strategy, added a "cloud-native application" skill node and increased its weight, with real-time updates from the local module of the edge computing node. In step S400, the system combines the candidate's geographical location (e.g., in Hangzhou) with high-priority skill nodes, selects the nearest Hangzhou edge node, and dynamically configures a knowledge cache set containing core skills such as "cloud-native application" and "Java development." Once the assessment phase begins, candidates input a natural language request to "participate in the adaptive assessment for backend development engineers." The S500 process uses a finely tuned large language model to parse the candidate's job intentions and basic skill requirements. It then combines this with cached skill node data to generate feature vectors that include dimensions such as skill matching degree and ability gap degree. Through a dynamic weighted multi-objective optimization algorithm, candidate test questions that cover core skills and are appropriately difficult are selected to generate a personalized adaptive test paper, ensuring that the assessment content is highly consistent with the actual job skill requirements of the enterprise.

[0167] The system continuously enhances the credibility of assessment results through a closed-loop optimization mechanism, helping companies accurately select talent. After candidates complete the test, the system collects their answers (e.g., 85% accuracy rate for microservice architecture questions, 60% accuracy rate for cloud-native application questions) and response times (e.g., an average response time of 45 seconds for database optimization questions) in real time via the S500 steps. It then uses an incremental learning algorithm to update the ability assessment model, accurately identifying "cloud-native application" as a weak area for candidates. Simultaneously, the system adjusts the skill graph evolution strategy based on the answer data, increasing the assessment weight of the "cloud-native application" skill node. For subsequent assessments of similar positions, the system can directly call the optimized model and skill graph to generate more targeted assessment content. This closed-loop mechanism ensures that the assessment results not only accurately reflect the candidate's true skill level but also dynamically adapt to changes in the company's job skill requirements, helping companies quickly identify high-quality candidates with "core Java development capabilities + cloud-native application potential," fully meeting the core needs of companies for scientific and precise talent assessment.

[0168] In this embodiment of the application, the large language model achieves semantic disambiguation and synonym normalization through a Prompt Tuning fine-tuning strategy. The specific method includes:

[0169] Step 1, Domain Adaptation Hint Template Construction:

[0170] Based on the characteristics of the skill domain and assessment requirements, a customized prompt template was designed. The core of the template includes three types of skill domain-specific information:

[0171] Skill type guidance statements (such as "Please determine whether the following statement belongs to technical skills, general skills, or industry-specific skills: {skill phrases}").

[0172] Typical application scenario description (e.g., "This skill is commonly used in software development, data analysis, project management, etc. Please standardize its expression: {skill phrase}").

[0173] Skill level matching reference cases (e.g., "Example: 'Proficient in Python' and 'Master of Python Programming' are normalized to 'Python Development - Proficient', please process: {Skill Phrases}"). The template adopts a "guide + example + task" structure to ensure that the large language model accurately understands the normalization rules of the skill domain.

[0174] Step 2, Prompt Tuning of the Large Language Model:

[0175] Atomized skill phrases extracted and standardized by step S100 (e.g., {Skill Name: Java Development, Skill Level: Proficient, Skill Relevance: Database Design - High}) are used as training data. The training data is combined with the constructed domain-adaptive prompt template to form batch fine-tuning samples (format example: "Please unify skill descriptions, example: 'Python Programming - Proficient' normalized to 'Python Development - Proficient', processed to 'Java Programming - Proficient'"). The samples are input into the basic large language model, and fine-tuning is performed through Prompt Tuning (training only the adaptation layer parameters related to the prompt template, freezing the main parameters of the model). This allows the model to learn the semantic disambiguation and normalization rules of the skill domain, avoiding the resource consumption and overfitting problems caused by full fine-tuning.

[0176] Step 3, semantic disambiguation and normalization:

[0177] The atomic skill phrases generated by S100 are assembled according to the prompt template format and input into the fine-tuned large language model. Based on the domain knowledge learned by the fine-tuning, the model identifies the ambiguity of skill expressions (such as "interface" being "technical interface development" in a technical scenario and "cross-departmental communication interface" in a communication scenario, and disambiguates in combination with the job description context), and outputs standardized skill expressions with consistent semantics (such as unifying "Java programming", "Java development" and "Java technology application" into "Java development"). At the same time, the model calculates the semantic similarity between each skill expression and generates a skill semantic similarity matrix (the matrix elements are the semantic similarity values ​​of any two skill phrases, ranging from [0,1]).

[0178] Step 4, similarity matrix-assisted clustering threshold determination:

[0179] The generated skill semantic similarity matrix is ​​used as auxiliary input data for the hierarchical density clustering algorithm. The clustering threshold is dynamically determined by the distribution characteristics of similarity values ​​in the matrix (such as the statistical similarity mean, median, and 95th percentile). For example, when the 95th percentile of the similarity matrix is ​​0.85, the clustering threshold is set to 0.8. Compared with a fixed threshold, this method can adapt to the differences in skill semantic distribution in different industries and positions, ensuring that the normalization and deduplication effects of the hierarchical density clustering algorithm are more accurate.

[0180] By optimizing the large language model through domain-adaptive Prompt Tuning, accurate semantic disambiguation and synonym normalization of skill expressions are achieved. At the same time, the generated skill semantic similarity matrix provides a dynamic threshold basis for hierarchical density clustering, which not only solves the problem of skill expression differences and ambiguity across data sources, but also improves the adaptability and accuracy of subsequent skill clustering deduplication.

[0181] In this embodiment of the application, the generation process of the skill graph evolution strategy incorporates a skill demand heat prediction model, and the method specifically includes:

[0182] Step 1: Construct a skill demand heat prediction model based on Temporal Convolutional Network (TCN):

[0183] First, the input and output rules of the model are clearly defined. The input is a historical skill demand data sequence (covering multi-dimensional time-series data from the past 6-12 months, such as the frequency of each skill appearing in job descriptions, the demand share in industry skill trend reports, and the number of internal skill assessment requests within companies, standardized into equal time intervals by week / month). The output is the trend of skill demand popularity within a future preset period (e.g., 1 month / 3 months) (including absolute value of popularity and direction of change: rising / stable / falling). The model architecture is built based on a temporal convolutional network (TCN), utilizing the "causal convolution + dilated convolution" characteristics of TCN to accurately capture the long-term dependencies of historical skill demand data (e.g., the demand fluctuation pattern of a certain skill over 3 consecutive months), avoiding the gradient vanishing problem of traditional recurrent neural networks (LSTM / RNN), and adapting to the temporal characteristics of skill demand.

[0184] Step 2, Model Training and Future Skill Demand Prediction: Preprocess the collected historical skill demand data and divide it into training and validation sets in a 7:3 ratio; input the preprocessed historical data sequence into the TCN model for iterative training, adjusting hyperparameters such as convolution kernel size, dilation coefficient, and learning rate until the model's prediction error on the validation set (e.g., mean absolute error MAE ≤ 0.05) converges to a preset threshold; after the model training is complete, input the latest historical skill demand data and output the predicted demand for each skill node within a preset future period.

[0185] Step 3: Set dynamic adjustment factors for skill node weights: Convert the skill demand heat predicted by the TCN model into weight adjustment factors and formulate quantitative adjustment rules: If the predicted skill heat is "rising", set the adjustment factor to 1.2 to 1.5; if the heat is "stable", set it to 1.0; if the heat is "falling", set it to 0.7 to 0.9. Calculate the adjusted weights using the formula: Adjusted skill node weight = Original weight × Heat adjustment factor. For example, "Cloud Native Development" has an original weight of 0.7, an adjustment factor of 1.4, and an adjusted weight of 0.98, allowing the weights to reflect future demand changes in advance.

[0186] Step 4: Generate an evolutionary strategy to adapt to potential changes in skill requirements. This involves combining adjusted skill node weights (incorporating future popularity predictions), user historical answering behavior characteristics (e.g., a 50% accuracy rate on questions related to "cloud-native development"), and the existing skill requirement change classifications in S300 (sudden / periodic). For example, for "cloud-native development," which is predicted to see increased popularity, the strategy adds a high-priority assessment rule for this skill node and optimizes its association with "containerized deployment." Simultaneously, considering users' weak answering characteristics, the strategy strengthens the proportion of questions related to this skill, ensuring that the evolutionary strategy not only adapts to current needs but also proactively responds to potential future changes in skill requirements.

[0187] The skill demand heat prediction model constructed by the temporal convolutional network incorporates future skill demand changes as weight adjustment factors into the evolution strategy generation process, upgrading the skill map evolution strategy from "responding to existing demand changes" to "adapting to potential future demand changes", which greatly improves the foresight and adaptability of the skill map to the talent assessment needs of enterprises.

[0188] In this embodiment, the selection of target edge nodes is achieved through a multi-objective optimization algorithm, and the method further includes:

[0189] Step 1, Construction of multi-objective optimization function and setting of constraints:

[0190] First, we define three core optimization objectives, formulate corresponding quantitative evaluation rules, and construct a multi-objective optimization function:

[0191] Optimization Objective 1: Minimize network transmission latency between examinees and edge nodes. This is quantified by a weighted calculation of a geographical distance coefficient and the reciprocal of network bandwidth. The geographical distance coefficient is calculated by converting the latitude and longitude of the examinee and the deployment location of the edge node, while the reciprocal of network bandwidth is determined based on the bandwidth data reported by the edge node in real time. Both are assigned weights of 0.6 and 0.4, respectively. The smaller the final result, the lower the transmission latency.

[0192] Optimization Objective 2: Maximize the utilization of edge node cache resources. This is evaluated by the ratio of cached core skill data to the total cache capacity of edge nodes. The core skill data is determined based on the list of high-priority skill nodes identified in S400, and the total cache capacity is the upper limit of the edge node's hardware configuration. The closer the ratio is to 1, the more fully the cache resources are utilized.

[0193] Optimization Objective 3: Balance the load pressure of the edge node cluster. The load rate is evaluated by the ratio of the target node load rate to the average cluster load rate. The load rate comprehensively considers CPU utilization, memory usage, and network I / O utilization (each of which accounts for one-third of the weight). The balance score is obtained by subtracting the ratio from 1. The closer the score is to 1, the more balanced the load of the target node is with the overall cluster load.

[0194] Two constraints are set: first, the amount of skill data to be cached must not exceed the upper limit of the remaining cache capacity of the edge node; second, the real-time available network bandwidth of the edge node must not be less than 10Mbps (which can be flexibly adjusted according to the assessment business needs) to ensure that the selected node has the hardware resource foundation to support the transmission of examination data.

[0195] Step 2, Solving the Pareto Optimal Set using a Non-Dominated Sorting Genetic Algorithm: The non-dominated sorting genetic algorithm is used to solve the above multi-objective optimization function. The core steps are as follows:

[0196] Population initialization: The "available edge nodes within the geographical range threshold" selected in step 3 of S400 are used as the initial population. The population size is set to 20 to 50 based on the total number of edge node clusters.

[0197] Fitness assessment: Calculate the evaluation values ​​of three optimization objectives for each individual in the population (i.e., a single marginal node), and use them as the core basis for fitness evaluation;

[0198] Non-dominated sorting: According to the rule that "no other node is better than this node in all optimization objectives", all nodes in the population are sorted hierarchically, and nodes in the non-dominated layer (i.e. Pareto optimal layer) are retained. These nodes can achieve the initial balance of multiple objectives.

[0199] Crowding degree calculation and selection: Calculate the crowding degree (measure the distribution density of nodes in the target space) for non-dominated layer nodes, and prioritize nodes with high crowding degree to avoid excessive concentration of optimal solutions and ensure the diversity of solutions;

[0200] Crossover and mutation: By crossover and random mutation of node resource parameters (such as cache capacity, bandwidth, load rate), the next generation of population is generated. This process is repeated 50 to 100 times until the optimal solution of the population no longer changes for 10 consecutive generations (i.e., population convergence).

[0201] Output results: After the population converges, the Pareto optimal set is output. All nodes in the set satisfy the condition that "improving the performance of any one optimization objective will inevitably lead to a decrease in the performance of another objective", thus achieving a dynamic balance of multiple objectives.

[0202] Step 3: Filter the optimal target edge node based on real-time candidate data:

[0203] The nodes in the Pareto optimal set are further filtered to better suit the actual network conditions of the test takers, ensuring the practicality of the selection results:

[0204] Supplementary data collection: Real-time acquisition of candidates' network quality data, including core indicators such as network packet loss rate (≤5% is considered passing), latency fluctuation value (≤20ms is considered stable), and uplink / downlink speed;

[0205] Comprehensive Adaptability Scoring: A secondary screening scoring model was constructed, with the following scoring dimensions and weights: Pareto optimal ranking (0.4), real-time network quality adaptability (0.3), and geographical proximity (0.3). Network quality adaptability is calculated by subtracting the sum of packet loss rate and latency fluctuation from 1; a value closer to 1 indicates better network quality adaptability. Geographical proximity is calculated by dividing 1 by (1 + the straight-line distance between the examinee and the node); a value closer to 1 indicates closer geographical proximity.

[0206] Optimal node determination: Calculate the comprehensive score of all nodes in the Pareto optimal set, and select the node with the highest score as the final target edge node; if multiple nodes have the same score, prioritize the node with the lower load rate in the edge node cluster to further ensure the stability of the subsequent examination process.

[0207] By constructing a multi-objective optimization function and solving the Pareto optimal set using a non-dominated sorting genetic algorithm, and combining the real-time network quality of the candidates for secondary screening, a multi-objective balance of "low transmission latency, high resource utilization, and load balancing" is achieved. Compared with the traditional linear scoring model, this significantly improves the scientificity and adaptability of the selection of target edge nodes.

[0208] In this embodiment of the application, the generation of multi-dimensional ability matching features adds the dimensions of skill transferability and answer stability. The specific method includes:

[0209] Step 1, Quantitative assessment of the new dimension (skill transferability + answer stability):

[0210] Quantification of skill transferability: First, from the configured cached skill node data, select skill pairs with related relationships (such as "Java development" and "database design", "data analysis" and "data visualization"), and extract cross-skill related questions for the corresponding skill pairs (i.e., questions that require the simultaneous use of two related skills to answer); count the accuracy rate of candidates in answering these cross-skill related questions, and use the accuracy rate as a quantitative indicator of skill transferability—the higher the accuracy rate, the stronger the candidate's ability to transfer knowledge of one skill to related skill scenarios.

[0211] Quantification of answer stability: For test questions of different difficulty levels within the same skill level (such as basic, intermediate, and advanced questions corresponding to the "Python Development - Beginner" skill), extract the response time data of candidates answering these types of questions; calculate the variance of these response times (variance reflects the degree of data fluctuation), and use variance as a quantitative indicator of answer stability—the smaller the variance, the more stable the candidate's answering state is when facing test questions of different difficulty levels within the same skill level (for example, if the candidate's response time to different difficulty levels of "Python Development - Beginner" test questions is 30 seconds, 32 seconds, and 29 seconds respectively, the variance is extremely small, indicating high answer stability).

[0212] Step 2, normalization of all-dimensional features:

[0213] First, we need to clarify the full set of features to be processed: including the newly added two dimensions, "skill transfer ability" and "answer stability", as well as the core ability dimensions mentioned above (skill mastery level: the overall accuracy rate of candidates on target skill test questions; test question difficulty fit: the degree to which candidates' accuracy rate in answering test questions of corresponding difficulty matches the preset passing score for that difficulty).

[0214] Normalization is performed on all dimensional features: The Min-Max normalization method is used to uniformly map feature indicators with different units and value ranges to the [0,1] interval (for example, the accuracy of skill transfer ability ranges from 0% to 100%, and the variance of answer stability ranges from 5 to 50, which are converted to values ​​of 0-1 after normalization). This eliminates the problem of feature weight imbalance caused by differences in indicator units and value ranges, and ensures that each dimensional feature has equal weight priority in subsequent calculations.

[0215] Step 3: The attention mechanism assigns initial weights and generates feature vectors.

[0216] Initial weight allocation: An attention mechanism is introduced, which combines the skill node data in the configured cache (such as skill node weight and skill priority in job requirements) to allocate initial weights to each ability dimension. The dimensions corresponding to core skills have higher weights (for example, "Java development" is a core skill for the target position, and the corresponding "skill transferability" and "skill mastery" dimensions have weights of 0.3 and 0.2, respectively; the dimension corresponding to the general skill "communication and coordination" has a weight of 0.1). At the same time, the weights are dynamically adjusted according to the candidate's job preferences (for example, when recruiting backend development engineers, the weight of "technical skill transferability" is increased and the weight of general skill-related dimensions is decreased).

[0217] Personalized feature vector generation: The normalized feature values ​​of each dimension are weighted and fused with the initial weights assigned by the attention mechanism, and then combined with the skill node data in the cache that matches the candidate's job (such as skill association and skill priority) to finally generate a personalized multi-dimensional ability matching feature vector (format example: [skill mastery level: 0.85 (weight 0.2), skill transfer ability: 0.78 (weight 0.3), answer stability: 0.92 (weight 0.2), test difficulty suitability: 0.80 (weight 0.3)]). The vector dimensions correspond one-to-one with the ability dimensions, accurately depicting the candidate's personalized ability characteristics.

[0218] By adding dimensions of skill transferability and answer stability, and dynamically allocating weights using an attention mechanism, the multi-dimensional ability matching features more comprehensively cover the candidate's overall abilities (including not only basic skill mastery but also deeper abilities such as skill transferability and answer stability). At the same time, the feature vectors are made more aligned with the candidate's job requirements and personal ability characteristics, providing more accurate feature support for subsequent dynamic weight multi-objective optimization algorithms to select test questions and generate personalized adaptive test papers, thereby further improving the comprehensiveness and accuracy of talent assessment.

[0219] In this embodiment, the update of the capability assessment model and the skill map evolution strategy adopts an incremental learning mechanism, the specific method of which includes:

[0220] Step 1: Locate the newly added data-related skill nodes and update local parameters. First, analyze the newly submitted answer data of the candidates to accurately locate the skill nodes corresponding to these data (for example, if the candidate answered questions related to "cloud-native development", then the associated skill node is "cloud-native development" and its associated skill "containerized deployment").

[0221] For these associated skill nodes, only relevant parameters in the capability assessment model and skill graph evolution strategy are locally updated. In the capability assessment model, only local parameters such as the capability prediction weights and error correction coefficients of the corresponding skill nodes are adjusted (without retraining the overall network structure or all parameters of the model); in the skill graph evolution strategy, only relevant rules such as the priority configuration and association weights of the corresponding skill nodes are optimized (without reconstructing the entire evolution strategy framework). This local update method significantly reduces computational resource consumption and improves update efficiency.

[0222] Step 2: An exponential moving average algorithm is used to dynamically decay the weights of skill nodes with no long-term data support. This algorithm dynamically decays the weights of skill nodes in the skill graph that have "no new answer data for a long time" (e.g., the skill "Traditional SSM Framework Development" with no candidates answering the corresponding questions for three consecutive months). This algorithm smooths the change in skill node weights by assigning higher weights to recent data and lower weights to older data, avoiding abrupt weight changes. Specifically, the weight of skill nodes with no long-term data support will gradually decrease over time (e.g., a skill with an initial weight of 0.7 decays to 0.4 after three consecutive months without data), thus adapting to the natural iteration of skill requirements (a long-term lack of assessment requirements indicates a decline in the skill's job suitability).

[0223] Step 3: Integrate multiple factors to complete the final optimization of the model and strategy. Integrate the three core data aspects to complete the optimization:

[0224] First, the parameters of the capability assessment model and the parameters of the skill map evolution strategy, which were partially updated in the previous section;

[0225] Second, the skill node weights after the exponential moving average algorithm decays;

[0226] Thirdly, the previously generated personalized multi-dimensional ability matching feature vectors (including dimensions such as skill transferability and answer stability).

[0227] The specific optimization logic is as follows: Based on the results of local updates, the evolution priority of the skill graph is adjusted by using the weakened skill node weights (the assessment proportion of skills with weakened weights is reduced), and then the prediction bias of the ability assessment model is calibrated by combining personalized ability feature vectors (for example, for the characteristic of candidates having "weak skill transfer ability", the ability prediction logic of related skills in the model is optimized), and finally the precise optimization of the ability assessment model and the skill graph evolution strategy is completed.

[0228] By implementing an incremental learning mechanism to update local parameters, the high resource consumption of retraining the entire model is avoided, thus improving update efficiency. The exponential moving average algorithm is used to decay the weights of skill nodes that have not had data for a long time, enabling the skill graph to adapt to the natural iteration of job skill requirements. Multi-factor fusion optimization makes the updated model and strategy more in line with the candidate's personalized ability characteristics and real-time job requirements, further ensuring the accuracy and dynamic adaptability of talent assessment.

[0229] This application discloses a multi-dimensional adaptive test generation system based on a dynamic skill graph, referring to... Figure 2 ,include:

[0230] The Job Skill Analysis and Semantic Generation Module 001 performs real-time analysis of job description data from recruitment websites and company job descriptions at the data collection end, extracts atomic skill phrases from the job descriptions and generates a unified skill semantic representation. The atomic skill phrases include skill name, skill level and skill relevance.

[0231] The skill semantic normalization and graph transmission module 002 represents the skill semantics based on a large language model for semantic disambiguation and synonym normalization. It achieves automatic normalization and deduplication of skills through hierarchical density clustering algorithm, establishes cross-enterprise and cross-industry skill semantic mapping relationship, and transmits the standardized skill graph to the central database through blockchain distributed version management mechanism.

[0232] The skill graph dynamic evolution module 003 dynamically generates a skill graph evolution strategy when skill requirements change dynamically, based on the skill node weights in the standardized skill graph and the user's historical answer behavior characteristics. The skill graph evolution strategy is executed by the local processing module deployed on the edge computing node. The local processing module responds in real time to sudden changes in skill requirements and performs batch optimization for periodic changes in skill requirements.

[0233] The edge node selection and cache configuration module 004, after the skill graph evolves, obtains the evolved skill graph content, selects target edge nodes based on skill node weights and user geographic location information in the computing architecture formed by the collaboration of edge nodes and central server, and dynamically configures the skill knowledge cache set of the target edge nodes.

[0234] The adaptive exam generation and model update module 005 obtains the natural language answer request input by the examinee, parses the corresponding examinee's ability assessment information based on the natural language answer request, generates multi-dimensional ability matching features based on the ability assessment information and the skill node data in the configured cache, determines the candidate question set based on the multi-dimensional ability matching features through a dynamic weight multi-objective optimization algorithm, and outputs the corresponding adaptive exam paper based on the candidate question set. It obtains the examinee's answer results and response time on the adaptive exam paper, and updates the ability assessment model and skill graph evolution strategy based on the answer results and response time.

[0235] This application also discloses an electronic device, including a processor, wherein the processor runs a program of the multi-dimensional adaptive test generation method based on dynamic skill graph as described in any one of the above embodiments.

[0236] This application also discloses a computer storage medium storing a program for the multi-dimensional adaptive exam generation method based on dynamic skill graphs as described in any one of the above embodiments.

[0237] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A multi-dimensional adaptive test generation method based on dynamic skill graphs, characterized in that, include: The data acquisition terminal performs real-time analysis on job description data from recruitment websites and corporate job descriptions, extracts atomic skill phrases from the job descriptions and generates a unified skill semantic representation. The atomic skill phrases include skill name, skill level and skill relevance. Based on the large language model, the semantic representation of the skills is subjected to semantic disambiguation and synonym normalization. The processed semantic representation of skills is automatically normalized and deduplicated through hierarchical density clustering algorithm. Based on the normalized and deduplicated semantic representation of skills, cross-enterprise and cross-industry skill semantic mapping relationship is established to generate a standardized skill graph. The standardized skill graph is then transmitted to the central database through a blockchain distributed version management mechanism. Based on the skill node weights in the standardized skill graph and the user's historical answering behavior characteristics, a skill graph evolution strategy is dynamically generated when skill requirements change dynamically. The skill graph evolution strategy is executed by a local processing module deployed on an edge computing node. The local processing module responds in real time to sudden changes in skill requirements and performs batch optimization for periodic changes in skill requirements. The generation process of the skill graph evolution strategy incorporates a skill demand popularity prediction model. The method further includes: constructing a skill demand popularity prediction model using a temporal convolutional network; defining the input of the skill demand popularity prediction model as a sequence of historical skill demand data, and the output as the trend of skill demand popularity changes within a preset future period; inputting historical skill demand data into the temporal convolutional network for training, and completing the prediction of future skill demand popularity after the temporal convolutional network converges; using the predicted skill demand popularity as a dynamic adjustment factor for skill node weights; and combining skill node weights and users' historical answering behavior characteristics to generate a skill graph evolution strategy adapted to potential skill demand changes. After the skill graph evolves, the evolved skill graph content is acquired. In a computing architecture comprised of edge nodes and a central server, a multi-objective optimization function is constructed. The optimization objectives are to minimize network transmission latency between examinees and edge nodes, maximize the utilization of edge node cache resources matching skill node weights, and balance the load pressure on the edge node cluster. The function constraints are the upper limit of edge node cache capacity and network bandwidth threshold. A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function, outputting a Pareto-optimal set of target edge nodes. Combining skill node weight matching degree, examinee geographical location information, and examinee real-time network quality data, the optimal target edge nodes are selected from the Pareto-optimal set. The skill knowledge cache set of the target edge nodes is then dynamically configured. The system acquires a natural language answer request from a test taker, parses the corresponding test taker ability assessment information from the natural language answer request, generates multi-dimensional ability matching features based on the ability assessment information and skill node data in a configured cached skill knowledge set, determines a candidate question set based on the multi-dimensional ability matching features using a dynamic weighted multi-objective optimization algorithm, outputs a corresponding adaptive exam paper based on the candidate question set, acquires the test taker's answer results and response time on the adaptive exam paper, and updates the skill graph evolution strategy based on the answer results and response time. The generation of multi-dimensional ability matching features adds the dimensions of skill transferability and answer stability. The method also includes: quantitatively assessing skill transferability by the examinee's correct answer rate on cross-skill related questions; quantitatively assessing answer stability by the variance of answer response time for questions of different difficulties at the same skill level; normalizing the features of skill transferability and answer stability, as well as the core ability dimension features extracted based on the examinee's ability assessment information, to eliminate differences in units; among which, the core ability dimension features include the degree of skill mastery and the suitability of the question difficulty; and assigning initial weights to each dimension through an attention mechanism, combined with the skill node data in the configured cache, to generate a personalized multi-dimensional ability matching feature vector.

2. The multi-dimensional adaptive exam generation method based on dynamic skill graphs according to claim 1, characterized in that, The extraction of atomized skill phrases employs a fusion model of bidirectional long short-term memory networks and conditional random fields. The method also includes: Convert the job description text into a standardized sequence of word vectors; The word vector sequence is input into the BiLSTM network, and through the collaborative computation of the forward hidden layer and the backward hidden layer, the output feature vector fused with context-dependent information is obtained. The feature vector output by the BiLSTM network is input into the CRF layer. The optimal annotation sequence is calculated based on the preset skill phrase annotation rules. The accurate atomic skill phrases are extracted based on the optimal annotation sequence. The skill phrase annotation rules include annotation labels corresponding to skill name, level, and relevance.

3. The multi-dimensional adaptive exam generation method based on dynamic skill graphs according to claim 2, characterized in that, Large language models achieve semantic disambiguation and synonym normalization through a Prompt Tuning fine-tuning strategy. Other methods include: Build domain-adaptive prompt templates, which include various skill domain-specific information such as skill type guidance statements, descriptions of typical application scenarios, and skill level matching reference cases; Standardized atomic skill phrases are used as training data, input into a large language model, and fine-tuned based on the constructed prompt templates; The atomized skill phrases to be processed are input into the fine-tuned large language model, which outputs semantically consistent normalized skill descriptions and generates a skill semantic similarity matrix. The skill semantic similarity matrix is ​​used to help determine the clustering threshold of the hierarchical density clustering algorithm.

4. The multi-dimensional adaptive exam generation method based on dynamic skill graphs according to claim 1, characterized in that, The update of the competency assessment model and skill map evolution strategy adopts an incremental learning mechanism, and the method also includes: For the skill nodes corresponding to newly added answer data, only the relevant parameters in the ability assessment model and skill graph evolution strategy are locally updated, without retraining the entire ability assessment model. An exponential moving average algorithm is used to dynamically decay the weights of skill nodes that have no long-term answer data support. By combining the decayed skill node weights, local update results, and the personalized multi-dimensional ability matching feature vector, the ability assessment model and skill graph evolution strategy are optimized.

5. A multi-dimensional adaptive exam generation system based on a dynamic skill graph, characterized in that, include: The job skill analysis and semantic generation module performs real-time analysis on job description data from recruitment websites and company job descriptions at the data collection end, extracts atomic skill phrases from the job descriptions and generates a unified skill semantic representation. The atomic skill phrases include skill name, skill level and skill relevance. The skill semantic normalization and graph transmission module performs semantic disambiguation and synonym normalization on the skill semantic representation based on a large language model. The processed skill semantic representation is automatically normalized and deduplicated using a hierarchical density clustering algorithm. Based on the normalized and deduplicated skill semantic representation, a cross-enterprise and cross-industry skill semantic mapping relationship is established to generate a standardized skill graph. The standardized skill graph is then transmitted to the central database through a blockchain distributed version management mechanism. The skill graph dynamic evolution module dynamically generates a skill graph evolution strategy when skill requirements change dynamically, based on the skill node weights in the standardized skill graph and the user's historical answer behavior characteristics. The skill graph evolution strategy is executed by a local processing module deployed on an edge computing node. The local processing module responds in real time to sudden changes in skill requirements and performs batch optimization for periodic changes in skill requirements. The generation process of the skill graph evolution strategy incorporates a skill demand popularity prediction model. The method further includes: constructing a skill demand popularity prediction model using a temporal convolutional network; defining the input of the skill demand popularity prediction model as a sequence of historical skill demand data, and the output as the trend of skill demand popularity changes within a preset future period; inputting historical skill demand data into the temporal convolutional network for training, and completing the prediction of future skill demand popularity after the temporal convolutional network converges; using the predicted skill demand popularity as a dynamic adjustment factor for skill node weights; and combining skill node weights and users' historical answering behavior characteristics to generate a skill graph evolution strategy adapted to potential skill demand changes. The edge node selection and cache configuration module, after the skill graph evolves, acquires the evolved skill graph content. Within the computing architecture formed by the collaborative efforts of edge nodes and the central server, it constructs a multi-objective optimization function. The optimization objectives are to minimize the network transmission latency between examinees and edge nodes, maximize the utilization rate of edge node cache resources matching skill node weights, and balance the load pressure on the edge node cluster. The function constraints are the upper limit of edge node cache capacity and network bandwidth threshold. A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function, outputting a Pareto-optimal set of target edge nodes. Combining the skill node weight matching degree, examinee geographical location information, and examinee real-time network quality data, the optimal target edge nodes are selected from the Pareto-optimal set. The module then dynamically configures the skill knowledge cache set of the target edge nodes. The adaptive exam generation and model update module acquires the natural language answer request input by the examinee, parses the corresponding examinee ability assessment information based on the natural language answer request, generates multi-dimensional ability matching features based on the ability assessment information and skill node data in the configured cached skill knowledge set, determines a candidate question set based on the multi-dimensional ability matching features through a dynamic weighted multi-objective optimization algorithm, outputs the corresponding adaptive exam paper based on the candidate question set, acquires the examinee's answer results and response time on the adaptive exam paper, and updates the skill graph evolution strategy based on the answer results and response time. The generation of multi-dimensional ability matching features adds the dimensions of skill transferability and answer stability. The method also includes: quantitatively assessing skill transferability by the examinee's correct answer rate on cross-skill related questions; quantitatively assessing answer stability by the variance of answer response time for questions of different difficulties at the same skill level; normalizing the features of skill transferability and answer stability, as well as the core ability dimension features extracted based on the examinee's ability assessment information, to eliminate differences in units; among which, the core ability dimension features include the degree of skill mastery and the suitability of the question difficulty; and assigning initial weights to each dimension through an attention mechanism, combined with the skill node data in the configured cache, to generate a personalized multi-dimensional ability matching feature vector.

6. An electronic device, characterized in that, Includes a processor, wherein the processor runs a program for the multi-dimensional adaptive test generation method based on dynamic skill graphs as described in any one of claims 1-4.

7. A computer storage medium, characterized in that, The program stores the multi-dimensional adaptive test generation method based on dynamic skill graphs as described in any one of claims 1-4.

Images