Zhi mu certification examination method and system based on large language model and ability atlas

Abstract

The application discloses a kind of based on big language model and ability atlas's wisdom pastor authentication examination method and system, belong to artificial intelligence technical field.The application first constructs the ability atlas based on knowledge, skill, behavior three-dimensional ontology model;Obtain examinee multimodal data and generate initial ability vector;Through double-path dynamic gate graph neural network, real-time update ability vector, the network includes erase gate and increase gate and is updated by graph convolution layer propagation;According to ability short board, determine aggressive factor and combine incidental factor, use big language model to dynamically generate the next logical unique test question;Repeat interaction until the end of examination;Finally, according to all updated ability vector, generate digital ability portrait report.The application realizes dynamic individualized surveying and mapping, anti-cheating and deep characterization of talent ability structure, improves the precision and security of wisdom pastor authentication examination.

Description

Technical Field

[0001] This invention relates to the field of digitalization and intelligentization technology in animal husbandry, specifically to a smart animal husbandry (hereinafter referred to as smart animal husbandry) certification examination method and system based on a large language model and ability graph. Background Technology

[0002] With the global digital transformation of agriculture, the livestock industry is undergoing a fundamental change from extensive management to precise intelligent sensing. As a core position connecting biotechnology, information technology, and automation engineering, the identification and certification of the professional capabilities of intelligent livestock engineers are becoming increasingly important. Current vocational skills certification systems generally adopt a standardized examination model based on a fixed question bank. This model has the following prominent problems: First, the update speed of the fixed question bank lags far behind the rapid iteration of digital technologies, resulting in a disconnect between exam content and actual job requirements. Second, random question selection cannot accurately detect the individualized weaknesses of candidates, often resulting in repeated testing of strengths and omission of weaknesses, failing to truly reflect the candidate's overall capabilities. Third, standardized exams focus on fragmented knowledge point assessments, neglecting the comprehensive ability structure of candidates across knowledge, skills, and behavior, especially lacking quantitative assessment of deeper behavioral traits such as professional ethics and decision-making style. Fourth, the widespread adoption of online certification increases the risk of question leaks and violations, and fixed questions cannot fundamentally prevent such violations. In addition, early deep knowledge tracing models (DKT) suffer from bottlenecks such as insufficient modeling accuracy, difficulty in capturing high-order features, and unstable training when dealing with knowledge graphs with complex structural dependencies.

[0003] Therefore, there is an urgent need in this field for a smart animal certification examination method and system that can achieve dynamic and personalized mapping, anti-cheating, and in-depth characterization of talent ability structure. Summary of the Invention

[0004] To address the problems of outdated fixed question banks, inaccurate detection of skill gaps due to random question selection, neglect of ability structure assessment, and high risk of violations in existing technologies, this invention provides a SmartMuse certification examination method and system based on a large language model and ability graph to solve the aforementioned technical deficiencies.

[0005] This invention proposes a method for the Zhimu certification examination based on a large language model and competency graph, comprising the following steps: S1. Construct a capability graph for the Zhimu certification exam. The capability graph is based on a three-dimensional ontology model of knowledge, skills, and behavior, and includes multiple capability nodes and semantic relationships between capability nodes. S2. Obtain the multimodal data of the examinee, map the multimodal data to a high-dimensional semantic space through a multimodal data mapping engine based on contrastive learning, generate an initial ability vector, and use the initial ability vector as the current ability vector; S3. Present the current test questions to the test takers and collect their interaction data, including answer results and digital behavior footprints; S4. Based on the capability graph, a dual-path dynamic gated graph neural network is used to update the current capability vector according to the interaction data, and an updated capability vector is generated. The dual-path dynamic gated graph neural network includes an erase gate and an add gate, which are used to remove outdated information and add capability increments, respectively. The update of the current capability node is propagated to the adjacent nodes through a graph convolutional layer. S5. The updated ability vector is used as the new current ability vector. The aggressive factor of the test question is determined based on the ability weakness where the ability value is lower than the preset threshold. Combined with the incident factor, the next test question is dynamically generated using the large language model. The aggressive factor determines the core difficulty and cognitive load of the test question, and the incident factor determines the narrative form of the test question. S6. Take the next question as the current question and repeat steps S3 to S5 until the exam termination condition is met. S7. Generate a digital competency profile report for the examinee based on all updated competency vectors generated during repeated execution.

[0006] Preferably, in step S2, the multimodal data of the examinee is acquired, and the multimodal data is mapped to a high-dimensional semantic space through a multimodal data mapping engine based on contrastive learning to generate an initial ability vector. This specifically includes the following sub-steps: S21. Obtain the candidate's static and dynamic data. Static data includes resume, project experience report and award certificate. Dynamic data includes historical learning log, historical test record and interactive event stream of the practical simulator. S22. Using a multimodal data mapping engine based on the Transformer architecture, the data of different modalities are aligned into a unified high-dimensional semantic space through the triplet loss function to obtain the aligned multimodal features. The triplet loss function is constructed in a way that makes the distance between the anchor sample and the positive sample less than the distance between the anchor sample and the negative sample minus the preset boundary margin. S23. The aligned multimodal features are processed by the InfoNCE loss function to enhance the mapping accuracy between the candidate's ability vector and the question difficulty vector. The InfoNCE loss function is constructed in a way that maximizes the cosine similarity between positive pairs while minimizing the cosine similarity between negative pairs. S24. Cluster the multimodal features processed in step S23 into an initial capability vector in a high-dimensional semantic space.

[0007] Preferably, in step S4, based on the capability map, a dual-path dynamic gated graph neural network is used to update the current capability vector according to the interaction data to generate an updated capability vector, which specifically includes the following sub-steps: S41. Obtain the input feature vector at the current time step. The input feature vector includes the question ID embedding, the ability node number corresponding to the current question in the ability graph, the correct answer identifier, and the standardized response time. S42. The information that needs to be removed from the current capability vector is calculated using the erase gate. The expression for the erase gate is: ; in, Indicates the erase gate output. This represents the input feature vector at the current time step. Represents the weights of a two-dimensional matrix. Indicates the bias term. This represents the Sigmoid activation function; S43. By increasing the incremental gate computing power, the expression for increasing the gate is: ; in, This indicates that an additional gate output has been added. Represents the weight matrix. Indicates the bias term. The sigmoid activation function is used; the dual-path dynamic gated graph neural network also utilizes the GELU activation function to generate candidate capability state vectors. ; S44. Generate the updated capability vector using the hidden state update expression. The hidden state update expression is: ; in, This represents the updated capability vector. This represents the capability vector from the previous time step. Represents the candidate capability state vector. Represents element-wise product. Represents a unit vector; S45. The graph convolutional layer is used to propagate the update information of the current capability node to the adjacent capability nodes in the capability graph that have semantic relationship with the current capability node. The graph convolutional layer uses an attention mechanism as the aggregation function for message passing.

[0008] More preferably, step S4 also includes the following sub-steps: The collected digital behavioral footprints are used as components of the input feature vector and input into a dual-path dynamic gated graph neural network. Use a pre-trained probabilistic prediction model to determine whether the collected digital behavioral footprints match the ability level represented by the current ability vector; When the judgment result is a mismatch, the unreliability score of the corresponding ability node is increased, and a related question with a higher difficulty than the current question is generated in the subsequent step S5 for cross-validation.

[0009] Preferably, in step S5, the updated ability vector is used as the new current ability vector. Based on the ability weaknesses where the ability value is below a preset threshold, the aggressive factor of the question is determined. Combined with the incident factor, the next question is dynamically generated using a large language model. Specifically, this includes the following sub-steps: S51. Extract capability dimensions whose values ​​are below a preset threshold from the updated capability vector as capability weaknesses; S52. Automatically identify the radical factor based on the ability gaps. The radical factor specifies the core knowledge points and cognitive constraints that the next test question must include. S53. Generate random factors using random number seeds or candidate preference templates. Random factors set the industry background, narrative style, and contextual description for the next question. S54. Embed the radical factor and the occasional factor into the prompt words, and call the large language model to generate the next question with a unique logical chain.

[0010] Preferably, in step S7, a digital competency profile report of the examinee is generated based on all updated competency vectors generated during repeated execution, specifically including the following sub-steps: S71. Based on all the updated capability vectors generated during the repeated execution, generate a capability structure radar chart with three dimensions: knowledge, skills, and behavior. S72. Based on the updated ability vector at each time step, generate a cognitive map evolution path showing how the ability status changes with the number of answer steps during the exam. S73. Extract the capability gaps from the final updated capability vector and call the large language model to generate personalized follow-up learning suggestions; S74. Combine the capability structure radar chart, cognitive map evolution path, and follow-up training recommendations into a digital capability profile report.

[0011] Preferably, in step S1, constructing the competency map for the Zhimu certification exam specifically includes the following sub-steps: S11. Collect unstructured text in the field of Smart Animal Husbandry. The unstructured text includes Smart Animal Husbandry industry standards, equipment manuals and expert experience reports. Identify the capability points and their hierarchical relationships related to Smart Animal Husbandry engineer certification from the unstructured text through ontology learning technology. Classify the capability points according to knowledge dimension, skill dimension and behavior dimension to generate the initial nodes and semantic edges of the capability graph. S12. Obtain structured data in the field of intelligent animal husbandry. The structured data includes the existing intelligent animal husbandry engineer certification test question bank and job skills database. The structured data is mapped to supplementary nodes and edges of the capability graph through RDB-to-Ontology technology. The supplementary nodes and edges are then merged into the initial nodes and semantic edges to form a basic capability graph. S13. Incrementally update the basic competency graph using a large language model: The large language model performs semantic checks on the basic competency graph, identifies nodes affected by changes in the technology of the Zhimu domain, generates new hierarchical connection methods, and adds the new hierarchical connection methods to the basic competency graph to obtain the competency graph for the Zhimu certification exam.

[0012] This invention also proposes a SmartMother certification examination system based on a large language model and competency graph, for implementing any of the methods described above, the system comprising: The competency graph construction module is configured to build a competency graph for the Zhimu certification exam. The competency graph is based on a three-dimensional ontology model of knowledge, skills, and behavior, and contains multiple competency nodes and semantic relationships between competency nodes. The multimodal mapping module is configured to acquire the examinee's multimodal data, map the multimodal data to a high-dimensional semantic space through a contrastive learning-based multimodal data mapping engine, generate an initial ability vector, and use the initial ability vector as the current ability vector. The interactive data collection module is configured to present the current test questions to the test takers and collect their interactive data, including answer results and digital behavioral footprints. The capability tracking module is configured to update the current capability vector based on the capability graph using a dual-path dynamic gated graph neural network according to the interaction data, and generate an updated capability vector. The dual-path dynamic gated graph neural network includes an erase gate and an add gate, which are used to remove outdated information and add capability increments, respectively, and propagate the update of the current capability node to the adjacent nodes through a graph convolutional layer. The test question generation module is configured to use the updated ability vector as the new current ability vector, determine the aggressive factor of the test question based on the ability weakness where the ability value is lower than a preset threshold, and combine it with the incident factor to dynamically generate the next test question using a large language model. The aggressive factor determines the core difficulty and cognitive load of the test question, and the incident factor determines the narrative form of the test question. The next test question is then fed back to the interactive acquisition module as the current test question to trigger the cyclical execution of the interactive acquisition module, ability tracking module and test question generation module until the test termination condition is met. The report generation module is configured to generate a digital competency profile report for the examinee based on all updated competency vectors generated during repeated execution.

[0013] The present invention also proposes a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the Zhimu certification examination method based on a large language model and competency graph as described above.

[0014] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent herdsman certification examination method based on a large language model and competency graph as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention constructs a capability map that includes a three-dimensional ontology model of knowledge, skills and behavior, and incorporates behavioral dimensions such as professional ethics, decision-making style and psychological traits into a quantifiable evaluation system. This realizes a paradigm shift from knowledge point assessment to capability structure mapping, and can comprehensively and deeply depict the candidate's capability profile.

[0016] (2) This invention uses a multimodal data mapping engine based on contrastive learning to align heterogeneous data such as candidates’ resumes, project experience, and learning logs into a unified high-dimensional semantic space by using triplet loss function and InfoNCE loss function. This effectively solves the cold start problem of the exam and enables the initial ability vector to objectively reflect the candidate’s historical background, providing a scientific benchmark for personalized question setting.

[0017] (3) This invention uses a dual-path dynamic gated graph neural network DPDG-GKT for ability tracking. It dynamically adjusts the ability vector through the two-path system of erasing and adding gates, and independently adjusts each dimension of the ability vector using two-dimensional matrix weights. This can accurately capture the ability gain and forgetting of candidates in specific scenarios. At the same time, the update of the current ability node is propagated to the adjacent nodes through the graph convolutional layer, realizing a penetrating assessment that measures a whole area based on answering one question, which significantly improves the accuracy and stability of ability prediction.

[0018] (4) This invention uses a large language model to dynamically generate the next test question based on the candidate's ability weaknesses. It locks the core difficulty and cognitive load through aggressive factors and ensures the uniqueness of the test question description form through occasional factors, so that the logical chain of each candidate's test question is unique. This effectively reduces the risk of question bank leakage and illegal operation from the physical and logical levels. At the same time, combined with reverse logic trap questions, it can effectively filter candidates who practice questions in a pattern and enhance the credibility of the test.

[0019] (5) This invention collects digital behavioral footprints in real time, judges the consistency between behavior and ability level through probability prediction model, and adds unreliability score when behavior is abnormal and generates more difficult related questions for cross-validation. It realizes the evolution from single score to execution risk report, and can objectively reveal the candidate's decision-making traps and professional ethics tendencies. Attached Figure Description

[0020] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 This is a flowchart of the Zhimu certification examination method based on large language models and competency graphs; Figure 2 This is a structural diagram of the Zhimu certification examination system based on a large language model and competency graph; Figure 3 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] Figure 1 A flowchart illustrating the Smart Shepherd certification examination method based on a large language model and competency graph is shown. Figure 1 As shown, the method includes the following steps S1 to S7.

[0024] Step S1: Construct the competency map for the Smart Farm certification exam. The competency map is based on a three-dimensional ontology model of knowledge, skills, and behavior, i.e., the KSB three-dimensional competency structure. Specifically, the knowledge dimension focuses on an individual's mastery of facts, concepts, principles, and logical relationships in a specific domain. The knowledge dimension covers TCP / IP protocols, LoRaWAN (Low Power Wide Area Network Protocol) networking methods, and MQTT (Message Queuing Telemetry Transport Protocol) message brokers in network engineering; distributed databases, livestock farming platform architecture, and streaming data processing in information systems; and the theoretical foundations of sensor transmission protocols, edge gateway optimization, and RFID technology in IoT devices. Knowledge points are connected in the map through semantic relationships such as premise, inclusion, and association, forming a semantic network with a topological structure. The skills dimension is defined as the procedural ability to apply knowledge to solve specific tasks, emphasizing the ability to do things and the path to solve problems. In the Smart Farm scenario, skills are refined into fault diagnosis (e.g., sensor data drift diagnosis and signal interference troubleshooting in extreme environments), system integration (e.g., logical linkage between automated feeding systems and environmental control systems), and solution planning (e.g., designing intelligent transformation solutions for large-scale pig farms, balancing cost and power consumption). The behavioral dimension examines an individual's psychological traits, professional ethics, and decision-making style during task execution. By monitoring the candidate's Digital Behavioral Footprint (DBF) during the exam, it assesses behavioral consistency (i.e., the alignment of the problem-solving path with the expert model), emotional stability (emotional fluctuations under pressure), professional ethics (e.g., abnormal integrity, screen-switching detection), and decision-making style (rational or impulsive). The competency graph contains multiple competency nodes, which are connected through semantic relationships such as premise, inclusion, and association.

[0025] Specifically, step S1 is implemented through the following sub-steps: S11. Collect unstructured text in the SmartMule field. Unstructured text includes SmartMule industry standards, equipment manuals, and expert experience reports. Using ontology learning technology, identify competency points and their hierarchical relationships related to SmartMule engineer certification from the unstructured text. Classify the competency points according to knowledge, skill, and behavioral dimensions to generate initial nodes and semantic edges for the competency graph.

[0026] S12. Obtain structured data in the Smart Animal Husbandry (SAG) domain. This structured data includes existing SAG engineer certification question banks and job skills databases. Using RDB-to-Ontology technology (a mapping technology from relational databases to ontology), the structured data is mapped to supplementary nodes and edges in the capability graph. These supplementary nodes and edges are then merged into the initial nodes and semantic edges to form a basic capability graph.

[0027] S13. Incrementally update the basic competency graph using a large language model. When new technologies emerge in the Zhimu field, the large language model performs semantic checks on the basic competency graph, identifies affected nodes, generates new hierarchical connection methods, and adds these new hierarchical connection methods to the basic competency graph to obtain the competency graph for the Zhimu certification exam.

[0028] Step S2: Before the certification begins, the system collects the candidate's full digital footprint through a legitimate interface, i.e., acquires the candidate's multimodal data. This multimodal data includes static and dynamic data; static data includes work resumes, project experience reports, and award certificates, while dynamic data includes historical learning logs, historical test records, and interactive event streams from the practical simulator. Then, a multimodal data mapping engine based on contrastive learning maps the multimodal data to a high-dimensional semantic space, generating an initial capability vector, which is then used as the current capability vector. Specifically, step S2 includes the following sub-steps: S21. Obtain the static and dynamic data of the candidates.

[0029] S22. Utilizing a multimodal data mapping engine based on the Transformer architecture (a neural network architecture based on a self-attention mechanism), the data from different modalities are aligned into a unified high-dimensional semantic space using a triplet loss function, resulting in aligned multimodal features. Triplet loss function. The expression is:

[0030] in For the target candidates' multimodal feature anchor points, Positive examples are expert samples with similar capability structures. Negative examples are those where there are significant differences in ability. To pre-set boundary margin, This represents the distance metric. The specific parameters of the Transformer architecture for the mapping engine are: 6 encoder layers, 768 hidden dimensions, 12 attention heads, and a 3072-dimensional hidden dimension in the feedforward network. It employs learnable positional encoding and the GELU (Gaussian Linear Unit) activation function, and uses a normalization scheme for the front-end layers. Each encoder layer contains multi-head self-attention and feedforward network sublayers.

[0031] S23. The aligned multimodal features are processed using the InfoNCE loss function to enhance the mapping accuracy between the candidate's ability vector and the question difficulty vector. The expression for the InfoNCE loss function is:

[0032] in Represents cosine similarity. This represents the feature vector of the current question. Represents the current capability vector. This represents the temperature coefficient. The loss function is constructed to maximize the cosine similarity between positive pairs while minimizing the cosine similarity between negative pairs.

[0033] S24. Cluster the multimodal features processed in step S23 into initial capability vectors in a high-dimensional semantic space. The training data sources for the mapping engine include: large-scale professional text corpora, collected from historical project documents, equipment maintenance manuals, and LoRaWAN industry standards of over 3,000 senior Zhimu engineers; multimodal interaction records, integrating hundreds of thousands of interaction event streams from the practical simulator, covering candidates' troubleshooting paths, parameter adjustment logic, and decision-making time; and unstructured archives, including resumes, professional certificates, and award reports extracted using OCR (Optical Character Recognition) and PDF parsing technologies. By minimizing the triplet loss function and the InfoNCE loss function, candidates with the same capability level are forced to cluster within a spatial neighborhood, ensuring that the initial capability vectors can serve as a scientific benchmark for evaluation. This high-dimensional architecture enables the model to effectively extract potential project practical capabilities from unstructured resumes and perform cross-modal semantic alignment with dynamic learning logs.

[0034] After acquiring the candidates' multimodal data, this embodiment also performs the following preprocessing operations: For candidates' resumes and certificates, semantic extraction is performed using a large language model. For example, when a resume mentions "participating in the intelligent transformation of a pig farm with a scale of 10,000 heads," the system presets higher initial ability scores in the network engineering and solution planning dimensions of the ability graph. For historical learning logs, interaction data such as the candidate's dwell time on technical documents and the rollback rate of video learning are extracted from the learning platform. Frequent rollback operations usually indicate a high cognitive load, and this data is used to assist in generating initial ability vectors. All unstructured text generates initial ability vectors through a mapping encoder.

[0035] Step S3: Present the current test question to the examinee and collect the examinee's interaction data. The interaction data includes the answer results and digital behavioral footprint. The current test question presented for the first time is either a system-preset initial test question or a question matched from the test question bank based on an initial ability vector. The digital behavioral footprint includes time-dimensional indicators, path and strategy indicators, editing and uncertainty indicators, and suspected cheating detection indicators.

[0036] The time dimension indicators include single-question response time and inter-action latency. This embodiment uses the Drift Diffusion Model (DDM) to deconstruct the candidate's response time and accuracy into interpretable psychological parameters: processing efficiency measures the rate at which an individual extracts effective information from the questions; a high drift rate corresponds to a rational decision-making style; reaction caution sets the evidence threshold required for decision-making; a low threshold indicates an impulsive style. Path and strategy indicators record the candidate's real-time operation sequence, skipped-question sequence, revisit frequency, and interface focus loss trajectory. For example, if a candidate browses the entire paper before answering questions, it indicates strong metacognitive planning ability; if a candidate frequently revisits already answered questions, it may reflect lower confidence or a cautious decision-making style. This embodiment uses the Dynamic Time Warping (DTW) algorithm to compare the similarity between the candidate's real-time operation sequence and the expert baseline path, using the formula:

[0037] in To align the path, For local distances, a smaller DTW distance indicates that the candidate's troubleshooting path conforms to professional engineering logic. Editing and uncertainty indicators include the frequency of answer modifications, the number of times the back button is used, and traces of time spent in the draft area, reflecting the candidate's cognitive certainty and cognitive load. Suspected cheating detection indicators include abnormal behavioral sequences such as frequent loss of page focus, abnormal text selection, and right-clicking, which are used as deductions in the credibility score within the behavioral dimension.

[0038] Step S4: During the assessment process, the candidate's ability status is constantly changing. This embodiment uses a Dual-Path Dynamic Gating Graph Neural Network (DPDG-GKT) as the core engine for ability tracking. By simulating the candidate's cognitive evolution on the ability graph, DPDG-GKT can effectively capture the highly coupled ability characteristics of the engineer in the network, device, and scenario, achieving a penetrating assessment that "draws the whole picture by answering one question." Specifically, based on the ability graph, DPDG-GKT updates the current ability vector according to the interaction data, generating an updated ability vector. DPDG-GKT includes an Erase Gate and an Add Gate, which are used to remove outdated information and add ability increments, respectively, and propagate the update of the current ability node to neighboring nodes through a Graph Convolutional Layer (GCN).

[0039] Once the exam begins, the interaction results for each question are fed into the model as input. These results include answer correctness, time consumption, and digital behavior footprint characteristics. The system identifies the core nodes corresponding to the current question based on the ability graph, such as MQTT Service Quality Level (QoS) settings. The DPDG-GKT's add gate captures the examinee's skill proficiency in specific intelligent scenarios, while the erase gate dynamically reduces the mastery probability of relevant nodes based on the examinee's error patterns and hesitation behavior.

[0040] Specifically, step S4 includes the following sub-steps: S41. Obtain the input feature vector for the current time step. The input feature vector includes the question ID embedding, the ability node number corresponding to the current question in the ability graph, the correct answer identifier, and the standardized response time. The standardized response time is a processing efficiency value or response caution value obtained by decomposing the candidate's original response time using a drift-diffusion model. The dimensions of the input feature vector are analyzed as follows: the question ID embedding is 128-dimensional, used to capture the semantic features of the question itself, its overall difficulty, and its unique identifier in the entire question bank; the ability graph node number is N-dimensional and uses one-hot encoding, where N is the total number of nodes in the ability graph, directly indicating which specific atomic ability point in knowledge, skills, or behavior the current question belongs to; the correct answer identifier is 1-dimensional, with a value of 0 or 1; the standardized response time is 1-dimensional, reflecting cognitive load. In addition, it may optionally include K-dimensional historical values ​​of relevant nodes, that is, the mean of the test point and its key predecessor nodes in the previous round of ability vector, used to provide contextual priors.

[0041] The association between each test question and a node in the capability graph is established through a pre-labeled test question-capability point mapping table. When the system loads a test question, it automatically retrieves its corresponding capability point list. In the Zhimu scenario, a troubleshooting question often involves both IoT protocol knowledge points and extreme environment adaptability optimization skills. In this case, the node number component of the input feature vector will activate multiple corresponding indices. The test question-capability point mapping table also includes association strength parameters. For example, if 80% of a question focuses on LoRaWAN networking and 20% on the physical characteristics of an ammonia sensor, the information transmitted by the input feature vector will be propagated proportionally and weighted across the graph.

[0042] S42. Calculate the information that needs to be removed from the current capability vector using the erase gate. The expression for the erase gate is:

[0043] in Indicates the erase gate output. This represents the input feature vector at the current time step. Represents the weights of a two-dimensional matrix. Indicates the bias term. This represents the Sigmoid activation function. Two-dimensional matrix weights allow for independent adjustment of each dimension of the capability vector, thus precisely capturing fine-grained capability differences.

[0044] S43. Increment the computational capacity by adding gates. The expression for increasing the number of gates is:

[0045] in This indicates that an additional gate output has been added. Represents the weight matrix. Indicates the bias term. The sigmoid activation function is used; the dual-path dynamic gated graph neural network also utilizes the GELU activation function to generate candidate capability state vectors. ; S44. Generate the updated capability vector using the hidden state update expression. The hidden state update expression is:

[0046] in This represents the updated capability vector. Represents the capability vector of the previous time step, based on the current interaction. The generated candidate capability state vector, Represents element-wise product. This represents a unit vector.

[0047] S45. A graph convolutional layer is used to propagate the updated information of the current capability node to neighboring capability nodes in the capability graph that have semantic relationships with the current capability node. The graph convolutional layer is configured with 1 to 2 layers. The first layer handles message passing between the current test point and its direct prior or posterior knowledge points, and the second layer captures the higher-order features of second-order related nodes. Neighboring nodes are defined as having a one-hop or two-hop semantic relationship. The message passing aggregation function uses an attention mechanism, automatically identifying which related skill points are most affected by the current answering behavior by learning the semantic weights between nodes. For example, if a test taker performs well in MQTT message processing, the attention mechanism will assign a higher update weight to the network troubleshooting node. The result after graph convolutional aggregation is used as the initial capability vector for the next time step.

[0048] In addition, step S4 also includes a sub-step of integrating the digitized behavioral footprint into the dual-path dynamic gated graph neural network in real time to dynamically correct the current capability vector: The collected digital behavioral footprints are used as part of the input feature vector and input into a dual-path dynamic gated graph neural network.

[0049] A pre-trained probabilistic prediction model is used to determine whether the collected digital behavioral footprints match the ability level represented by the current ability vector.

[0050] When the judgment result is a mismatch, the unreliability score of the corresponding ability node is increased, and a related question with a higher difficulty than the current question is generated in the subsequent step S5 for cross-validation.

[0051] The training process of the dual-path dynamic gated graph neural network is as follows: Model training adopts a paradigm combining supervised learning and self-supervised learning, optimizing weights by minimizing the deviation between the predicted accuracy and the actual accuracy. The loss function used is the binary cross-entropy loss function. (Binary Cross-Entropy Loss) combined with L2 regularization (L2 norm regularization) to prevent overfitting:

[0052] in To answer the question with a true value, This represents the probability of the model's prediction being correct. The regularization coefficient is used. The optimizer uses Adam with an initial learning rate of 0.001 and a learning rate decay strategy, automatically reducing it to 10% of its original value after 5 rounds of stagnation in the validation set loss. The batch size is set to 32, and the training epochs are set to 100 with an early stopping mechanism. The training dataset comes from the historical answer data of 3,000 certified Zhimu engineers and 12,000 junior candidates, containing more than 800,000 valid interaction trajectories. The ground truth of the ability vectors adopts the surrogate annotation method, that is, the average performance of candidates in a series of subsequent high-difficulty comprehensive questions is used as the ground truth reference for their ability nodes in the early stages. The mapping between questions and ability points is labeled by 20 industry experts by constructing a Q-matrix (question-ability association matrix). The dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0053] Step S5: Use the updated ability vector as the new current ability vector. Determine the aggressive factor of the test question based on the ability weaknesses whose ability values ​​are below a preset threshold. Combined with the incident factor, dynamically generate the next test question using the large language model. The aggressive factor determines the core difficulty and cognitive load of the test question, while the incident factor determines the narrative form of the test question.

[0054] Specifically, step S5 includes the following sub-steps: S51. Extract capability dimensions with values ​​below a preset threshold from the updated capability vector as capability weaknesses.

[0055] S52. Automatically identify aggressive factors based on skill gaps. Aggressive factors specify the core knowledge points and cognitive constraints that the next test question must include. For example, if a candidate's IoT fault diagnosis skills are weak, the aggressive factor might require the question to include hard constraints such as LoRa spread spectrum factor configuration, sensor signal outlier handling, and data caching logic under specific memory limitations. In the Zhimu scenario application example, the aggressive factor specifies that a specific logical breakpoint must be set in the MQTT message encryption stage.

[0056] S53. Generate incidental factors using random number seeds or candidate preference templates. Incidental factors set the industry background, narrative style, and scenario description for the next question, such as setting it as the maintenance of environmental control equipment in a high-altitude pig farm or the logic debugging of an automatic feeder in a laying hen house, ensuring the uniqueness of each candidate's paper in terms of textual expression. Incidental factors are used to ensure formal uniqueness, and their generation method is to set the narrative background using random number seeds or candidate preference templates. In the Smart Livestock scenario application example, the scenario can be set as a 10,000-head pig farm suffering from lightning strikes and frequent disconnections of the MQTT Broker (message broker server). In addition, a logic trap component can be set to detect candidates' thinking defects and set inducement items for candidates' decision-making styles, such as impulsive ones, such as setting interference options that appear logically correct but violate biosafety protection red lines. Specifically, the system will detect the candidate's answer patterns in several consecutive similar questions. When patterned answering behavior is detected, a trap question that appears similar but has the opposite core logic is generated by fine-tuning the radical factor, or an option that violates professional principles is set in the interference options. If a candidate violates the same professional principle or logical chain in a similar situation, the system will automatically lower their credibility score in the professional ethics dimension.

[0057] S54. Integrate aggressive and incidental factors into the prompt words, and call the large language model to generate the next question with a unique logical chain. The prompt word engineering adopts a strategy based on few-shot learning and thought chains: the system extracts detailed definitions of target ability points from the ability map, provides the large language model with 2 to 3 high-quality expert-level practical questions for reference, and requires the model to internally output a problem-solving thought chain before outputting the question. During the generation process, adaptive prompting technology dynamically adjusts the level of detail of the instructions based on the quality feedback from the previous generation by the model, thereby achieving self-correction in the question-generating process.

[0058] To further enhance anti-cheating capabilities, this invention also implements a unique question logic defense: even for two candidates who start the exam at the same time, the questions they see in the latter half of the exam will be significantly different due to differences in their ability evolution, depending on the differences in their answering process. This large-scale unique distribution reduces the possibility of cheating by taking screenshots or sharing questions to zero. Specifically, in step S54, a unique variable associated with the candidate (such as the hash value of the candidate's ID or the feature code of the initial ability vector) is injected into the prompt, thereby ensuring that the questions generated for different candidates targeting the same ability weakness have completely different logical chains and expressions.

[0059] In addition, step S5 implements a closed-loop testing strategy targeting the ability structure: when the system detects that the candidate has shown extremely high proficiency in a certain sub-domain, it automatically shifts the focus of the next test question to the higher-order ability points adjacent to that sub-domain in order to detect the actual boundaries of the candidate's ability; when the candidate makes repeated mistakes in a certain basic link, the system generates a series of low-level logic questions with subtle differences to diagnose whether the lack of ability stems from blind spots in knowledge points or lack of proficiency in skill application.

[0060] Step S6: Set the next question as the current question and repeat steps S3 to S5 until the exam termination condition is met. The exam termination condition may be reaching a preset number of questions, the change in the ability vector being less than a set threshold, or the examinee voluntarily submitting to end the exam.

[0061] Step S7: Generate a digital competency profile report for the examinee based on all updated competency vectors generated during repeated execution. At the end of the exam, the system generates a detailed digital competency profile report, rather than a single score.

[0062] Specifically, step S7 includes the following sub-steps: S71. Based on all the updated capability vectors generated during the repeated execution process, generate a capability structure radar chart with three dimensions: knowledge, skills, and behavior, to demonstrate the balance of the three dimensions of KSB.

[0063] S72. Based on the updated ability vector at each time step, generate a cognitive map evolution path of ability status changes with the number of answer steps during the exam, thereby identifying whether the examinee is explosive, stable or fatigued.

[0064] S73. Extract the capability gaps from the final updated capability vector, call the large language model to generate personalized follow-up learning suggestions, and provide targeted follow-up learning plans.

[0065] S74. Combine the capability structure radar chart, cognitive map evolution path, and follow-up training recommendations into a digital capability profile report.

[0066] To verify the technical effectiveness of this invention, a comparative experiment was conducted in a pilot program for the Zhimu Engineer certification. Using the traditional fixed question bank sampling method, the ability prediction accuracy (AUC) was 0.71, the cheating detection rate was 12%, the assessment time was fixed at 120 minutes, and the profile completeness was only a single score. Using the method combining the DPDG-GKT model of this invention with a large language model, the ability prediction accuracy reached 0.88, an improvement of 23.9%; the cheating detection rate reached 99.8%, a significant improvement; the assessment time was dynamically adjusted from 45 to 90 minutes, saving approximately 30%; and the profile completeness was improved to a 128-dimensional ability vector plus behavioral style, achieving a dimensional leap. Experimental data shows that the DPDG-GKT model, through optimization of two-dimensional matrix weights and Xavier initialization, significantly outperforms the traditional deep knowledge tracing DKT model in terms of stability and accuracy when processing long sequence operation data.

[0067] Further reference Figure 2 As an implementation of the above method, this invention also proposes a Smart Shepherd certification examination system based on a large language model and ability graph, which can be applied to various electronic devices. This Smart Shepherd certification examination system 200 based on a large language model and ability graph includes the following modules: The Capability Graph Construction Module 210 is configured to construct a capability graph for the Zhimu certification exam. The capability graph is based on a three-dimensional ontology model of knowledge, skills, and behavior, and contains multiple capability nodes and semantic relationships between capability nodes.

[0068] The multimodal mapping module 220 is configured to acquire the examinee's multimodal data, map the multimodal data to a high-dimensional semantic space through a contrastive learning-based multimodal data mapping engine, generate an initial ability vector, and use the initial ability vector as the current ability vector.

[0069] The interactive data collection module 230 is configured to present the current test questions to the test takers and collect their interactive data, including answer results and digital behavioral footprints.

[0070] The capability tracking module 240 is configured to update the current capability vector based on the capability graph using a dual-path dynamic gated graph neural network according to the interaction data, and generate an updated capability vector. The dual-path dynamic gated graph neural network includes an erase gate and an add gate, which are used to remove outdated information and add capability increments, respectively, and propagate the update of the current capability node to the adjacent nodes through a graph convolutional layer.

[0071] The test question generation module 250 is configured to use the updated ability vector as the new current ability vector, determine the aggressive factor of the test question based on the ability weakness where the ability value is lower than a preset threshold, and combine it with the incident factor to dynamically generate the next test question using a large language model. The aggressive factor determines the core difficulty and cognitive load of the test question, and the incident factor determines the narrative form of the test question. The next test question is then fed back to the interactive acquisition module as the current test question to trigger the cyclic execution of the interactive acquisition module, ability tracking module and test question generation module until the test termination condition is met.

[0072] The report generation module 260 is configured to generate a digital competency profile report for the examinee based on all updated competency vectors generated during repeated execution.

[0073] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the steps of the intelligent herdsman certification examination method based on a large language model and competency graph as described above.

[0074] The present invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the intelligent herdsman certification examination method based on a large language model and competency graph as described above.

[0075] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system 300 suitable for implementing terminal devices or servers of the present invention. Figure 3 The terminal device or server shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0076] like Figure 3 As shown, the computer system 300 includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the computer system 300. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0077] The following components are connected to I / O interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a liquid crystal display (LCD) and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card and a modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to I / O interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 310 as needed so that computer programs read from it can be installed into storage section 308 as needed.

[0078] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this invention. It should be noted that the computer-readable medium described in this invention can be a computer-readable signal medium or a computer-readable medium or any combination thereof. The computer-readable medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0079] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0080] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0081] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for Zhimu certification examination based on large language models and competency graphs, characterized in that, Includes the following steps: S1. Construct a capability graph for the Zhimu certification exam. The capability graph is based on a three-dimensional ontology model of knowledge, skills, and behavior, and includes multiple capability nodes and semantic relationships between capability nodes. S2. Obtain the candidate's multimodal data, map the multimodal data to a high-dimensional semantic space through a multimodal data mapping engine based on contrastive learning, generate an initial ability vector, and use the initial ability vector as the current ability vector; S3. Present the current test questions to the test taker and collect the test taker's interaction data, including the answer results and digital behavioral footprint; S4. Based on the capability graph, the current capability vector is updated using a dual-path dynamic gated graph neural network according to the interaction data, generating an updated capability vector. The dual-path dynamic gated graph neural network includes an erase gate and an add gate, used to remove stale information and add capability increments, respectively, and propagates the update of the current capability node to neighboring nodes through a graph convolutional layer. Step S4 also includes the following sub-steps: Obtain the input feature vector at the current time step. The input feature vector includes the question ID embedding, the ability node number corresponding to the current question in the ability graph, the correct answer identifier, and the standardized response time. The collected digital behavioral footprints are used as components of the input feature vector and input into a dual-path dynamic gated graph neural network. A pre-trained probabilistic prediction model is used to determine whether the collected digital behavioral footprint matches the ability level represented by the current ability vector. The digital behavioral footprint includes time-dimensional indicators, path and strategy indicators, editing and uncertainty indicators, and suspected cheating identification indicators. Among them, the time-dimensional indicators include the response time of a single question and the latency between actions; the path and strategy indicators record the candidate's real-time operation sequence, question skipping sequence, review frequency, and interface focus loss trajectory; the editing and uncertainty indicators include the frequency of answer modification, the number of times the back button is used, and traces of time spent in the draft area; the suspected cheating identification indicators include abnormal behavior sequences such as frequent loss of page focus, abnormal text selection, and right-clicking; when the judgment result is a mismatch, the unreliability score of the corresponding ability node is increased, and a related question with a higher difficulty than the current question is generated in the subsequent step S5 for cross-validation. S5. The updated ability vector is used as the new current ability vector. The aggressive factor of the test question is determined based on the ability weakness where the ability value is lower than the preset threshold. Combined with the incident factor, the next test question is dynamically generated using the large language model. The aggressive factor determines the core difficulty and cognitive load of the test question, and the incident factor determines the narrative form of the test question. S6. Take the next question as the current question and repeat steps S3 to S5 until the exam termination condition is met. S7. Generate a digital competency profile report for the examinee based on all updated competency vectors generated during repeated execution.

2. The intelligent herdsman certification examination method based on a large language model and competency graph as described in claim 1, characterized in that, In step S2, multimodal data of the examinee is acquired, and the multimodal data is mapped to a high-dimensional semantic space through a multimodal data mapping engine based on contrastive learning to generate an initial capability vector. This specifically includes the following sub-steps: S21. Obtain the candidate's static and dynamic data. The static data includes a resume, project experience report, and award certificates. The dynamic data includes historical learning logs, historical test records, and interactive event streams from the practical simulator. S22. Using a multimodal data mapping engine based on the Transformer architecture, the data of different modalities are aligned to a unified high-dimensional semantic space through the triplet loss function to obtain the aligned multimodal features. The triplet loss function is constructed in such a way that the distance between the anchor sample and the positive sample is less than the distance between the anchor sample and the negative sample minus the preset boundary margin. S23. The aligned multimodal features are processed by the InfoNCE loss function to enhance the mapping accuracy between the candidate's ability vector and the question difficulty vector. The InfoNCE loss function is constructed in a way that maximizes the cosine similarity between positive pairs while minimizing the cosine similarity between negative pairs. S24. Cluster the multimodal features processed in step S23 into an initial capability vector in a high-dimensional semantic space.

3. The intelligent herdsman certification examination method based on a large language model and competency graph as described in claim 1, characterized in that, In step S4, based on the capability map, the current capability vector is updated using a dual-path dynamic gated graph neural network according to the interaction data to generate an updated capability vector. This specifically includes the following sub-steps: The information that needs to be removed from the current capability vector is calculated using an erasure gate, the expression of which is: ； in, Indicates the erase gate output. This represents the input feature vector at the current time step. Represents a two-dimensional weight matrix. Indicates the bias term. This represents the Sigmoid activation function; By increasing the incremental gate computing power, the expression for increasing the gate is: ； in, This indicates that an additional gate output has been added. Represents the weight matrix. Indicates the bias term. The sigmoid activation function is represented; the dual-path dynamic gated graph neural network also utilizes the GELU activation function to generate candidate capability state vectors. ; The updated capability vector is generated by the hidden state update expression, which is: ； in, This represents the updated capability vector. This represents the capability vector from the previous time step. Represents the candidate capability state vector. Represents element-wise product. Represents a unit vector; The graph convolutional layer is used to propagate the update information of the current capability node to neighboring capability nodes in the capability graph that have semantic relationships with the current capability node. The graph convolutional layer uses an attention mechanism as the aggregation function for message passing.

4. The intelligent herdsman certification examination method based on a large language model and competency graph as described in claim 1, characterized in that, In step S5, the updated ability vector is used as the new current ability vector. Based on the ability weaknesses where the ability value is below a preset threshold, the aggressive factor of the question is determined. Combined with the incident factor, the next question is dynamically generated using a large language model. This specifically includes the following sub-steps: S51. Extract capability dimensions whose values ​​are below a preset threshold from the updated capability vector as capability weaknesses; S52. Automatically identify the radical factor based on the aforementioned ability shortcomings. The radical factor specifies the core knowledge points and cognitive constraints that the next test question must include. S53. Generate a random factor using a random number seed or a candidate preference template. The random factor sets the industry background, narrative style, and contextual description for the next test question. S54. Embed the radical factor and the occasional factor into the prompt word, and call the large language model to generate the next question with a unique logical chain.

5. The intelligent herdsman certification examination method based on a large language model and competency graph as described in claim 1, characterized in that, In step S7, a digital competency profile report of the examinee is generated based on all updated competency vectors generated during repeated execution. This includes the following sub-steps: S71. Based on all the updated capability vectors generated during the repeated execution, generate a capability structure radar chart with three dimensions: knowledge, skills, and behavior. S72. Based on the updated ability vector at each time step, generate a cognitive map evolution path showing how the ability status changes with the number of answer steps during the exam. S73. Extract the capability gaps from the final updated capability vector and call the large language model to generate personalized follow-up learning suggestions; S74. Combine the capability structure radar chart, the cognitive map evolution path, and the subsequent training suggestions into the digital capability profile report.

6. The intelligent herdsman certification examination method based on a large language model and competency graph as described in claim 1, characterized in that, Step S1 involves constructing the competency map for the Zhimu certification exam, which includes the following sub-steps: S11. Collect unstructured text in the field of Smart Animal Husbandry, including Smart Animal Husbandry industry standards, equipment manuals and expert experience reports. Identify the capability points and their hierarchical relationships related to Smart Animal Husbandry engineer certification from the unstructured text using ontology learning technology. Classify the capability points according to knowledge dimension, skill dimension and behavior dimension to generate the initial nodes and semantic edges of the capability graph. S12. Obtain structured data in the field of intelligent animal husbandry. The structured data includes the existing intelligent animal husbandry engineer certification question bank and job skills database. The structured data is mapped to supplementary nodes and edges of the capability graph through RDB-to-Ontology technology. The supplementary nodes and edges are then merged into the initial nodes and semantic edges to form a basic capability graph. S13. Incrementally update the basic capability graph using a large language model: The large language model performs semantic checks on the basic capability graph, identifies nodes affected by changes in the technology of the Zhimu domain, generates new hierarchical connection methods, and adds the new hierarchical connection methods to the basic capability graph to obtain the capability graph of the Zhimu certification exam.

7. A SmartMother certification examination system based on a large language model and competency graph, used to implement the method as described in any one of claims 1 to 6, characterized in that, The system includes: The capability graph construction module is configured to construct a capability graph for the Zhimu certification exam. The capability graph is based on a three-dimensional ontology model of knowledge, skills, and behavior, and contains multiple capability nodes and semantic relationships between capability nodes. The multimodal mapping module is configured to acquire the examinee's multimodal data, map the multimodal data to a high-dimensional semantic space through a contrastive learning-based multimodal data mapping engine, generate an initial ability vector, and use the initial ability vector as the current ability vector. The interactive data collection module is configured to present the current test questions to the test taker and collect the test taker's interactive data, which includes answer results and digital behavioral footprints. The capability tracking module is configured to update the current capability vector based on the capability graph using a dual-path dynamic gated graph neural network according to the interaction data, generating an updated capability vector. The dual-path dynamic gated graph neural network includes an erase gate and an add gate, used to remove stale information and add capability increments, respectively, and propagates the update of the current capability node to neighboring nodes through a graph convolutional layer. The capability tracking module also includes the following sub-steps: Obtain the input feature vector at the current time step. The input feature vector includes the question ID embedding, the ability node number corresponding to the current question in the ability graph, the correct answer identifier, and the standardized response time. The collected digital behavioral footprints are used as components of the input feature vector and input into a dual-path dynamic gated graph neural network. A pre-trained probabilistic prediction model is used to determine whether the collected digital behavioral footprint matches the ability level represented by the current ability vector. The digital behavioral footprint includes time-dimensional indicators, path and strategy indicators, editing and uncertainty indicators, and suspected cheating detection indicators. Among them, the time-dimensional indicators include the response time of a single question and the latency between actions; the path and strategy indicators record the candidate's real-time operation sequence, question skipping sequence, review frequency, and interface focus loss trajectory; the editing and uncertainty indicators include the frequency of answer modification, the number of times the back button is used, and traces of time spent in the draft area; the suspected cheating detection indicators include abnormal behavior sequences such as frequent loss of page focus, abnormal text selection, and right-clicking. When the judgment result is a mismatch, the unreliability score of the corresponding ability node is increased, and a related question with a higher difficulty than the current question is generated in the subsequent question generation module for cross-validation. The question generation module is configured to take the updated ability vector as the new current ability vector, determine the aggressive factor of the question based on the ability weakness where the ability value is lower than a preset threshold, and combine it with the incident factor to dynamically generate the next question using a large language model. The aggressive factor determines the core difficulty and cognitive load of the question, and the incident factor determines the narrative form of the question. The next question is then fed back to the interactive acquisition module as the current question to trigger the cyclic execution of the interactive acquisition module, the ability tracking module, and the question generation module until the examination termination condition is met. The report generation module is configured to generate a digital competency profile report for the examinee based on all updated competency vectors generated during repeated execution.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the SmartMuse Certification Examination Method based on a large language model and competency graph as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the SmartMuse Certification Examination Method based on a large language model and competency graph as described in any one of claims 1 to 6.

Images