A test paper generation method and system

Abstract

The application discloses a test paper generation method and system, belongs to the intelligent education technical field, acquires student answering record, question-knowledge point mapping, test question bank and knowledge text information, and constructs a knowledge point graph; a student-question-knowledge three-layer heterogeneous graph is constructed, and student answering performance on a test paper is predicted; the answering performance is taken as environmental feedback input of reinforcement learning, a question set of the current test paper is taken as a state, questions in the test paper are replaced or reserved as actions, a multi-target reward function is constructed to determine a comprehensive reward value of the test paper in each dimension, and the optimal test paper is generated by iteration optimization with the maximum comprehensive reward value as a target; based on the adjacent relationship in the knowledge point graph, candidate subtest sets are formed by selecting candidate test questions from the test question bank, and multiple sets of parallel test papers equivalent to the optimal test paper are generated by using an equivalent question replacement strategy. Higher efficient, reasonable and fair parallel test paper generation can be realized by the method.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and more specifically to a test paper generation method and system. Background Technology

[0002] As a powerful assessment tool, exam papers can not only objectively measure students' knowledge mastery, but also further enhance teaching effectiveness when combined with data analysis and personalized feedback in a smart education environment. Therefore, in the era of smart education, exam papers remain an important tool for assessing learning outcomes and promoting precision teaching. The task of Exam Paper Generation (EPG) relies on the efficient organization and mining of educational data and resources. Building upon EPG, the parallel paper generation (PPG) task requires designing multiple different versions of exam papers. These papers have nearly identical assessment effectiveness while containing nearly different question sets to ensure fairness in assessment and reduce the risk of cheating.

[0003] Traditionally, manually generating high-quality exam papers is a task that demands both physical and mental effort from teachers. In recent years, with the development of algorithms, research on exam paper generation methods has mainly focused on improving and applying metaheuristic algorithms, such as genetic algorithms, and using them to optimize several key attributes of the exam paper, such as difficulty, discrimination, and answering time. These methods are typically based on a global sampling framework, capable of searching for combinations of questions within a large scope to generate exam papers that meet expected standards. Over the past decade, these methods have made significant progress in the field of automatic exam paper generation due to their advantages of simple algorithm principles, ease of implementation, and relatively low computational cost, driving the development and implementation of intelligent exam paper generation technology.

[0004] However, existing methods still suffer from two prominent problems. Firstly, they rely excessively on expert experience in setting and labeling test paper attributes. This requires teachers or domain experts to manually assign values ​​to indicators such as question difficulty and discrimination. This approach is not only costly in terms of manpower and time but also inevitably carries subjectivity and bias, affecting the objectivity and universality of the generated test papers. Secondly, most existing optimization methods focus on the question bank, emphasizing the overall quality of the question combination, such as question type coverage, knowledge point distribution, or attribute balance, while often neglecting students' performance and feedback when actually completing the generated test papers. This "question-centric" optimization perspective leads to a lack of consideration for individual differences, making it difficult to meet the needs of differentiated instruction and personalized learning, and limiting the application value of EPG technology in real-world educational scenarios.

[0005] To address the limitations of traditional methods, some studies have introduced Deep Knowledge Tracking (DKT) based on LSTM to track student learning progress and predict test performance. These methods, combined with various generation strategies, enhance EPG (Extended Learning Platform) effectiveness, demonstrating superiority over metaheuristic algorithms. However, existing methods fail to fully utilize prior knowledge graphs and the overall structure, neglecting the intrinsic connections between individual students, knowledge, and questions, leading to inconsistencies between assessment effectiveness and teaching objectives during test generation. Summary of the Invention

[0006] To address the problems existing in the above-mentioned fields, this invention proposes a test paper generation method and system. Based on the adjacency relationship in the knowledge point graph, by selecting candidate questions from the question bank to form a candidate sub-question set, it can comprehensively enhance the equivalence of parallel test papers at the question, knowledge, and individual levels, and achieve more efficient, reasonable, and fair parallel test paper generation.

[0007] To address the aforementioned technical problems, this invention discloses a test paper generation method, comprising the following steps:

[0008] Acquire student answer records, question-knowledge point mapping, question bank and knowledge text information, extract student embedding representation, question embedding representation and knowledge embedding representation respectively, and construct student-question interaction graph, question-knowledge point association graph and knowledge point graph that integrates semantics and co-occurrence relationship;

[0009] The knowledge embedding representation is integrated into the student embedding representation and the question embedding representation. By constructing a three-layer heterogeneous graph of student-question-knowledge and extracting the high-order feature representation of each node in the heterogeneous graph, the student's answer performance on the test paper can be predicted.

[0010] Students' performance on the test paper is used as the environmental feedback input for reinforcement learning. The current set of questions in the test paper is used as the state, and replacing or retaining questions in the test paper is used as the action. By constructing a multi-objective reward function that integrates the difficulty, effectiveness, rationality and discrimination of the test paper, the comprehensive reward value of the test paper in each dimension of the multi-objective reward function is determined. With the goal of maximizing the comprehensive reward value, the optimal test paper is generated through iterative optimization.

[0011] Based on the adjacency relationship in the knowledge point graph, candidate questions are selected from the question bank to form a candidate sub-question set. Using the equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined. Through iterative optimization, multiple sets of parallel test papers equivalent to the optimal test paper are generated.

[0012] Preferably, the step of generating multiple parallel test papers equivalent to the optimal test paper by selecting candidate questions from the question bank based on the adjacency relationship in the knowledge point graph, forming a candidate sub-question set, and using an equivalent question replacement strategy to replace and combine questions that are different but cover the same knowledge points as the optimal test paper, specifically includes:

[0013] Randomly select a question from the candidate question set as a replacement candidate. When the replacement candidate satisfies the requirement of maximizing the comprehensive reward value, the original question in the optimal test paper is replaced. Otherwise, the search is expanded to the candidate question set of the neighboring knowledge nodes. If no replacement candidate with the maximum comprehensive reward value is found after the neighborhood expansion, the original question is retained without replacement.

[0014] By iterating continuously in the state and action space, and using an equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined to replace the questions in the optimal test paper. This optimizes the overall difficulty, effectiveness, rationality and discrimination of the test paper, with the goal of maximizing the overall reward value. Through iterative optimization, multiple sets of parallel test papers equivalent to the optimal test paper are generated.

[0015] Preferably, the step of using students' performance on the test paper as environmental feedback input for reinforcement learning, using the current set of questions in the test paper as the state, and replacing or retaining questions in the test paper as actions, involves constructing a multi-objective reward function that considers the test paper's overall difficulty, effectiveness, reasonableness, and discrimination, determining the overall reward value of the test paper in each dimension of the multi-objective reward function, and aiming to maximize the overall reward value. This is achieved through iterative optimization to generate the optimal test paper, specifically including:

[0016] A dual-deep Q-network is used to iteratively train the state and action of reinforcement learning. The dual-deep Q-network takes the student's answer to the test paper as the environmental feedback input, takes the current set of questions in the test paper as the state, and takes the equivalent question replacement strategy to replace or retain questions in the test paper as the action, with the goal of maximizing the comprehensive reward value.

[0017] During training, the agent uses an ε-greedy strategy to balance exploration and exploitation. It samples historical state-action-reward sequences through a priority experience replay mechanism and updates the parameters of the dual-deep Q-network through multiple rounds of iterative training. The agent gradually learns the optimal strategy for selecting questions in different states. By simulating students' performance on the test paper, it generates the optimal test paper that maximizes the comprehensive reward value across dimensions such as comprehensive difficulty, effectiveness, rationality, and discrimination, thus meeting multiple objective requirements.

[0018] Preferably, the integration of knowledge embedding representation into student embedding representation and question embedding representation involves constructing a three-layer heterogeneous graph of student-question-knowledge layers and extracting high-order feature representations of each node in the heterogeneous graph to predict student performance on the test paper. Specifically, this includes:

[0019] Obtain the low-dimensional vectors of student embeddings, question embeddings, and knowledge embeddings respectively;

[0020] By constructing a three-layer heterogeneous graph of student-question-knowledge, high-order feature representations of each node are extracted;

[0021] The low-dimensional vector and high-order feature representations are fused together, and the neighbor node information is iteratively aggregated. The student's performance is determined by predicting the probability of answering unanswered questions on the test paper.

[0022] Preferably, the step of acquiring student answer records, question-knowledge point mappings, question banks, and knowledge text information, and extracting student embedding representations, question embedding representations, and knowledge embedding representations respectively, to construct a student-question interaction graph, a question-knowledge point association graph, and a knowledge point graph integrating semantics and co-occurrence relationships, specifically includes:

[0023] Based on the acquired student answer records, question-knowledge point mapping, question bank and knowledge text information, the embedded representations of students, questions and knowledge are extracted respectively through graph attention network;

[0024] The graph attention network consists of an embedding layer and a fully connected layer. In the embedding layer, a multi-layer graph attention mechanism is introduced to obtain student, question, and knowledge embedding representations respectively. In the fully connected layer, the knowledge embedding representations are integrated into the student embedding representation and question embedding representation respectively to construct a student-question interaction graph, a question-knowledge point association graph, and a knowledge point graph that integrates semantics and co-occurrence relationships.

[0025] Preferably, the step of generating multiple sets of parallel test papers equivalent to the optimal test paper through iterative optimization further includes evaluating the equivalence of the generated parallel test papers at four levels: individual student, student group, knowledge and questions, based on knowledge distribution entropy, knowledge frequency quotient and connected component entropy.

[0026] Preferably, the acquisition of student answer records, question bank, question-knowledge point mapping, and knowledge text information specifically includes:

[0027] When retrieving student answer records, the student's historical answer log is exported from the online education platform. This log includes the student ID, question ID, answer result (correct or incorrect), answer timestamp, and number of answer attempts.

[0028] The question-knowledge point mapping is obtained by subject matter experts or course teachers through manual annotation based on the course syllabus and teaching objectives, clarifying one or more knowledge points tested by each question, and forming a structured question-knowledge point association matrix;

[0029] Obtain knowledge text information, including the unique identifier, name, detailed description text, and relative importance weight of each knowledge point in the course knowledge system;

[0030] The acquired student answer records, question bank, question-knowledge point mapping, and knowledge text information are cleaned and preprocessed.

[0031] Preferably, it also includes a test paper generation system, comprising:

[0032] The knowledge point graph construction module is used to obtain student answer records, question-knowledge point mapping, test question bank and knowledge text information, extract student embedding representation, question embedding representation and knowledge embedding representation respectively, and construct student-test question interaction graph, test question-knowledge point association graph and knowledge point graph that integrates semantics and co-occurrence relationship;

[0033] The heterogeneous graph construction module is used to integrate knowledge embedding representations into student embedding representations and question embedding representations. By constructing a three-layer heterogeneous graph of student-question-knowledge and extracting high-order feature representations of each node in the heterogeneous graph, it can predict students' performance on the test paper.

[0034] The optimal test paper generation module uses students' performance on the test paper as environmental feedback input for reinforcement learning. It takes the current set of questions in the test paper as the state and replaces or retains questions in the test paper as actions. By constructing a multi-objective reward function that considers the comprehensive difficulty, effectiveness, reasonableness and discrimination of the test paper, it determines the comprehensive reward value of the test paper in each dimension of the multi-objective reward function. With the goal of maximizing the comprehensive reward value, it generates the optimal test paper through iterative optimization.

[0035] The parallel test paper generation module is used to generate multiple sets of parallel test papers that are equivalent to the optimal test paper by selecting candidate questions from the question bank based on the adjacency relationship in the knowledge point graph and using the equivalent question replacement strategy to replace and combine questions that are different but cover the same knowledge points as the optimal test paper. Through iterative optimization, multiple sets of parallel test papers are generated.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The test paper generation method proposed in this invention constructs a three-layer heterogeneous graph of student-question-knowledge, which can integrate students' answer records, question-knowledge relationships, and knowledge text information based on graph cognition diagnosis to form a unified modeling framework that can represent cross-layer interaction and intra-layer structure. Through feature fusion and reasoning of the heterogeneous graph, a rich representation of the student-question-knowledge structure is obtained, which can accurately predict students' performance on the test paper and provide complete information support for subsequent test paper updates and parallel test paper generation. Multi-objective reinforcement learning can simultaneously optimize four core attributes: difficulty, validity, rationality, and discrimination, effectively improving test paper quality and evaluation fairness. Based on the adjacency relationship in the knowledge point graph, candidate questions are selected from the question bank to form a candidate subset. Using an equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined. Question replacement is achieved by using candidate subsets connected to knowledge nodes, ensuring the equivalence of multiple sets of parallel test papers at the four levels of problem, knowledge, individual, and group. This alleviates the problems of existing methods that rely on manual subset division and have insufficient equivalence, and improves the rationality, fairness, and intelligence level of test paper generation. Attached Figure Description

[0038] Figure 1 This is a flowchart of the test paper generation method proposed in this invention;

[0039] Figure 2 This is an overall framework diagram of the test paper generation network model provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of a subset of candidate questions formed by each knowledge node in the student-problem-knowledge heterogeneous graph provided in an embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram of the experimental results generated by parallel test paper according to an embodiment of the present invention;

[0042] Figure 5 This is a schematic diagram illustrating the total frequency of each knowledge point being tested in all parallel test papers provided in this embodiment of the invention.

[0043] Figure 6 This is a schematic diagram of the questions tested in a parallel test paper provided in an embodiment of the present invention. Detailed Implementation

[0044] The following will refer to the appendices in the embodiments of the present invention. Figures 1-6 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0045] Example 1

[0046] like Figure 1 As shown, this invention proposes a method for generating test papers, which includes the following steps:

[0047] S1: Obtain student answer records, question-knowledge point mapping, question bank and knowledge text information, extract student embedding representation, question embedding representation and knowledge embedding representation respectively, and construct student-question interaction graph, question-knowledge point association graph and knowledge point graph that integrates semantics and co-occurrence relationship;

[0048] S2: Integrate knowledge embedding representation into student embedding representation and question embedding representation. By constructing a three-layer heterogeneous graph of student-question-knowledge and extracting the high-order feature representation of each node in the heterogeneous graph, predict students' performance on the test paper.

[0049] S3: Use students' performance on the test paper as the environmental feedback input for reinforcement learning. Use the current set of questions in the test paper as the state, and replace or retain questions in the test paper as actions. By constructing a multi-objective reward function that considers the comprehensive difficulty, effectiveness, rationality and discrimination of the test paper, determine the comprehensive reward value of the test paper in each dimension of the multi-objective reward function. With the goal of maximizing the comprehensive reward value, generate the optimal test paper through iterative optimization.

[0050] S4: Based on the adjacency relationship in the knowledge point graph, candidate questions are selected from the question bank to form a candidate sub-question set. Using the equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined. Through iterative optimization, multiple sets of parallel test papers equivalent to the optimal test paper are generated.

[0051] Specifically, in step S1, student answer records, question bank, question-knowledge point mapping, and knowledge text information are obtained, including:

[0052] When obtaining students' answer records, the historical answer logs of students are exported from online education platforms (such as PTA, Junyi Academy, etc.). These logs include student ID, question ID, answer result (correct / incorrect), answer timestamp, and number of answer attempts.

[0053] The question-knowledge point mapping is obtained by subject matter experts or course teachers through manual annotation based on the course syllabus and teaching objectives, clarifying one or more knowledge points tested by each question, and forming a structured question-knowledge point association matrix.

[0054] Obtain knowledge text information, including the unique identifier, name, detailed description text, and relative importance weight of each knowledge point in the course knowledge system.

[0055] The acquired student answer records, question bank, question-knowledge point mapping, and knowledge text information are cleaned and preprocessed, including removing invalid or abnormal records (such as records with excessively short or long answer times), handling missing values, and performing natural language processing operations such as word segmentation and stop word removal on the knowledge text. Subsequently, the complete dataset is randomly divided into training, validation, and test sets according to a preset ratio (e.g., 7:2:1). The training set is used to construct a three-layer heterogeneous graph of student-question-knowledge and train a graph cognitive diagnostic model. The validation set is used to adjust model hyperparameters and implement early stopping strategies to avoid overfitting. The test set is used to finally evaluate the performance indicators of the generated test papers. It is crucial to ensure that all answer records for each student appear in only one set during the partitioning process to avoid data leakage.

[0056] In step S2, a student-question interaction graph is constructed based on student answer records, where edges represent interaction relationships between students and questions. A question-knowledge point mapping graph is then constructed. Finally, a knowledge graph is built based on the semantic similarity and co-occurrence relationships of knowledge texts, containing two edge types: similarity relationships and collaborative relationships.

[0057] like Figure 2 As shown, a test paper generation network model, GRL-EPG framework, is constructed. GRL (Graph Representation Learning, GRL) and EPG (Evolution Positioning Grid, EPG) are also included. The GRL-EPG framework comprises a heterogeneous graph construction module, a multi-layer information fusion module, and a cognitive diagnosis module, as detailed below:

[0058] Heterogeneous graph construction module

[0059] Specifically, the basic representation of the heterogeneous graph is constructed by initializing the embedding matrices of Students, Questions, and Knowledge. Let the set of students be... The set of questions is The knowledge points are set as follows , , , , where d is the embedding dimension of the embedding matrix, N is the number of students, M is the number of questions, and C is the number of knowledge points.

[0060] Figure 2 Figure (a) shows that the acquired student answer records (Interaction), question-knowledge point mapping (Relation), question bank, and knowledge text information are used as inputs to the GRL-EPG framework, as shown in Figure (a). Figure 2Figure (b) shows the extraction of student embeddings, question embeddings, and knowledge embeddings for each node. A graph attention network is then used to update the embedding relationships of each node, constructing a student-question interaction graph, a question-knowledge point association graph, and a knowledge graph that integrates semantics and co-occurrence relationships. Figure 2 As shown in Figure (c).

[0061] For each student ,topic and knowledge points One-hot vectors are generated through the encoder's encoding process (Embedding). , , Obtain students separately ,topic and knowledge points The corresponding initial embedding representation:

[0062]

[0063] During network training, these initial embeddings are adaptively adjusted based on the complex relationships in the heterogeneous graph. This adjustment is achieved by aggregating information from neighboring nodes through a multi-layer graph attention mechanism. The graph attention network employs a multi-head attention mechanism and uses layer normalization after each layer.

[0064] Specifically, regarding knowledge points In the Layer Embedded Representation Its update process can be represented as:

[0065]

[0066] in, , This represents the set of representations of all knowledge nodes at layer l, with weights. and It is used to learn the contributions of different subgraphs and balance the knowledge relationships and problem relevance of different subgraphs.

[0067] This represents information aggregated from adjacent knowledge point nodes:

[0068]

[0069] This represents information aggregated from related question nodes:

[0070]

[0071] in, and It is a trainable matrix used for linear transformations. Indicates and The set of adjacent and related knowledge nodes, Indicates and The set of adjacent questions used to test this knowledge. Indicates the connection to the current knowledge point Adjacent knowledge points In the embedding representation of the l-th layer, Indicates related issues Embedded representation at layer l.

[0072] Attention weight and Calculated in the following way:

[0073]

[0074]

[0075] in, and For a fully connected network, This indicates a splicing operation.

[0076] During problem fusion, the fusion vector for problem j in the (l+1)th iteration is... , from its previous embedding Relational aggregation of related knowledge Aggregation of interactive students renew:

[0077]

[0078]

[0079]

[0080] in, , This represents the set of representations of all question nodes at level l. and It is node-level attention weight. This represents the set of knowledge associated with question j. This represents the set of students who interact with question j. and These are graph-level attention weights. and It is a trainable matrix used for linear transformations.

[0081] The (l+1)th iteration fusion vector of student i is derived from its previous embedding. Aggregation of interactive questions renew:

[0082]

[0083] in, , This represents the set of representations of all student nodes at level l.

[0084] The aggregated information from the student-question interaction is represented as follows:

[0085]

[0086] Among them, Q i This represents the set of questions that student i interacts with. It is a trainable matrix used for linear transformations. The calculation method and similar.

[0087] Multi-layer information fusion module

[0088] After iteratively obtaining the final fused embedding vectors K (knowledge), Q (question), and S (student), a fully connected network is used to integrate the knowledge embeddings into the student embeddings and question embeddings, respectively, optimizing the graph structure and establishing a final three-layer heterogeneous graph of student-question-knowledge, as shown below. Figure 2 As shown in Figure (e), that is:

[0089]

[0090]

[0091] in, , This indicates a fully connected network.

[0092] Cognitive Diagnostic Module

[0093] Based on this relation-aware representation, student i's performance (result) in answering question j is shown. This can be deduced from the following formula:

[0094]

[0095] in , for function, It is a fully connected layer.

[0096] The loss function uses binary cross-entropy:

[0097]

[0098] in This represents the actual answer.

[0099] For a test paper with n questions, the predicted total score for student i, i.e., student i's performance on the test paper, can be expressed as:

[0100]

[0101] In the formula This represents the score for question j.

[0102] In step S2, the set of test questions is used as input, and the students' answers to the test questions serve as the environmental feedback input for multi-objective reinforcement learning. A dual-deep Q-network is used to model and optimize the reinforcement learning in the state-action space to gradually generate test papers that satisfy multi-objective balance. For example... Figure 2 As shown in Figure (f), the entire process can be divided into four core modules, among which:

[0103] Reinforcement learning environment and definition modules

[0104] The test paper update problem is formalized as a decision task in reinforcement learning and defined as a Markov Decision Process (MDP), using tuples. <S, A, R, , γ, T> represents the state. This represents the structure and state of the test paper at step t. It can be represented as:

[0105]

[0106] in, This is a feature vector concatenation operation. Let represent the feature vector of the i-th question. Action A is defined as replacing a question in the current best test paper with a new, unique question from the question bank, according to the question selection strategy. Reward R is the feedback value Rt resulting from action t. t In performing action A t Then receive the next reward R t+1 .Strategy The state-action value function Q is defined. π (S, A), the next question to be replaced is determined by selecting the action with the largest Q value through an online network. The formula is: The discount factor γ is used to balance short-term returns with long-term goals, while the state transition T = (S t A t , R t+1, γ t+1 , S t+1 This describes the process of updating the test paper status after the action is executed.

[0107] DDQN Network and Policy Selection Module

[0108] A dual-deep Q-network (DDQN) is employed, comprising an online network and a target network: the online network computes the Q-value of the current state-action sequence to select an action, while the target network computes the target Q-value to stabilize reward estimation. The agent selects actions based on the Q-value output by the online network using an ε-greedy policy, following the policy with a probability of 1 − ε. The optimal action is selected by randomly choosing an action with probability ε, thereby achieving a balance between exploration and exploitation.

[0109] Experience replay and training optimization module

[0110] Introducing an Experience Replay buffer to store state transition experience (S t A t , R t+1 ,S t+1 During training, mini-batch experiences are randomly sampled from the buffer to calculate the target Q-value, which is the target network output corresponding to the maximum online Q-value of the next state after a discount to the immediate reward. Stochastic gradient descent optimization is performed by minimizing the mean squared error loss function. To ensure training stability, the target network employs a delayed update mechanism; after every m training steps, the target network parameters θ are updated. − Synchronization is the online network parameter θ.

[0111] Multi-objective reward and test paper assessment module

[0112] Four reward indicators were designed based on the teachers' teaching objectives: overall difficulty, effectiveness, reasonableness, and differentiation.

[0113] 1. Overall Difficulty Rewards (Difficulty), a measure of the average student score relative to the target difficulty. Consistency:

[0114]

[0115] in The total score for the exam is... The average score for students.

[0116] 2. Effectiveness Rewards (Validity) measures the consistency between the knowledge distribution in the exam paper and the course requirements:

[0117]

[0118] in For the first question in the exam paper The proportion of each knowledge point tested. This represents the required percentage of this knowledge point in the course.

[0119] 3. Reasonable rewards (Rationality) measures the similarity between the distribution of student scores and the target distribution.

[0120]

[0121] in The actual distribution of student scores The target distribution is defined as follows.

[0122] 4. Distinctive Rewards (Discrimination) measures the ability of a test paper to differentiate between students of different levels.

[0123]

[0124] in The average score of the top 27% of students This is the average score of the bottom 27% of students.

[0125] The final overall reward function is: This allows the agent to optimize multiple learning objectives simultaneously with each question update in order to generate the optimal test paper.

[0126] Example 2

[0127] This embodiment also includes generating and evaluating parallel test papers with new data after obtaining the optimal model for generating parallel test papers, specifically including the following steps:

[0128] Step 1: Obtain data on the new student population to be tested and the test question bank;

[0129] Specifically, the student population and test question bank data during the testing phase come from different educational platforms or different academic assessment datasets; when acquiring student data, the students' historical learning records are exported from the target education system, including information such as student ID, question ID, answer results, and timestamps; the question-knowledge point mapping relationship needs to be re-annotated by subject experts according to the new curriculum standards to clarify the knowledge points tested by each question; the knowledge text information includes the name, description, and importance weight of the new knowledge system.

[0130] Step 2: Replace and combine test questions of different difficulty but covering the same knowledge points, and generate multiple parallel test papers that meet the multiple objectives of reasonableness, comprehensive difficulty, discrimination and effectiveness through the trained test paper generation network model.

[0131] Specifically, the new student population and test question bank data are input into the trained GRL-EPG system (the trained test paper generation network model);

[0132] The data is first processed through the heterogeneous graph construction module to construct a three-layer heterogeneous graph of student, question, and knowledge. Then, a multi-layer graph attention network is used to generate high-order feature representations of the nodes. After that, the data enters the cognitive diagnosis module to obtain a prediction of the student's knowledge status.

[0133] The reinforcement learning generation module is based on a multi-objective reward function that integrates difficulty, effectiveness, rationality and discrimination. With the goal of maximizing the comprehensive reward value, it uses the predicted student answer scores as environmental feedback to quantify the performance of the test paper in each dimension.

[0134] Based on the adjacency relationship of the knowledge graph, multiple parallel test papers are generated using an equivalent question replacement strategy to ensure equivalence at all levels.

[0135] Based on the test paper composition output by the test paper generation network model, a multi-index evaluation method is used to obtain the generation results of parallel test papers. The evaluation is mainly carried out through group consistency and individual consistency indices, as well as knowledge distribution entropy, knowledge frequency quotient and connectivity component entropy.

[0136] Based on the candidate question set formed by knowledge nodes and their neighborhoods, an equivalent question replacement strategy is used to replace and combine different questions that cover the same knowledge points, generating multiple parallel test papers with a high degree of consistency in rationality, overall difficulty, discrimination, and effectiveness. The entire process can be divided into two core modules: candidate question generation and replacement strategy, and multi-dimensional evaluation of parallel test paper generation.

[0137] Candidate Question Generation and Replacement Strategies

[0138] Specifically, the parallel test paper generation module first constructs a candidate sub-question set. This candidate sub-question set is constructed based on the relationships between knowledge nodes in a knowledge point graph, such as... Figure 3 As shown, for each knowledge node its candidate sub-topic set It consists of the set of directly related questions and the set of questions associated with its adjacent knowledge nodes:

[0139]

[0140] in, This indicates a direct examination of knowledge points. A collection of questions, express The set of adjacent knowledge points.

[0141] Based on this, the replacement strategy is used to treat each question in the generated parallel test papers. Corresponding knowledge nodes The system operates on its candidate sub-topics and then randomly selects a question from the candidate sub-topics as a replacement candidate. If the selected question meets the multi-objective constraints (including comprehensive difficulty, effectiveness, reasonableness, and discrimination requirements), the replacement is performed; if no suitable candidate question is found, the search is expanded sequentially to the candidate sub-topics of neighboring knowledge nodes.

[0142] For questions that still do not meet the criteria after neighborhood expansion (such as isolated knowledge nodes), the original questions will be retained without replacement to ensure the consistency of knowledge coverage in the test paper.

[0143] Multi-dimensional assessment of parallel test paper generation

[0144] After parallel test papers are generated, a systematic evaluation is required to ensure group consistency in knowledge coverage, difficulty, score distribution, discrimination ability, and question composition among different parallel test papers. In addition, several new indicators are introduced and proposed to evaluate the equivalence of parallel test papers at the individual, knowledge, and question levels, specifically evaluated from the following four levels:

[0145] Group consistency assessment: This measures the consistency of all parallel test papers across overall metrics, including overall difficulty, validity, reasonableness, and discrimination. It involves calculating the overall reward for each test paper and comparing it to a target value. High group consistency indicates that the generated parallel test papers are well-equivalent at the group level.

[0146] Individual stability assessment: This evaluates the stability of a student's scores across different parallel test papers, ensuring the comparability of scores across different versions of the test. Specifically, it incorporates Cronbach's... The coefficient (Reliability) quantifies stability at the individual level.

[0147]

[0148] in, Students exist The variance of scores on the test papers, where N is the number of students. This represents the total fractional variance.

[0149] Knowledge Distribution Assessment: Ideally, the distribution of knowledge across parallel test papers should be consistent. Therefore, Knowledge Distribution Entropy (KDE) is proposed to evaluate the consistency of this distribution.

[0150]

[0151] in, Representing knowledge points The proportion of questions covered in test paper p.

[0152] Knowledge frequency consistency assessment: Ideally, for knowledge points... Parallel exam papers should contain the same number of questions, i.e., the total number of questions. It should be able to be used in parallel volumes Therefore, the Knowledge Frequency Quotient (KFQ) is proposed to measure the consistency of the number of questions testing each knowledge point in parallel test papers.

[0153]

[0154] The modulo operator calculates the remainder.

[0155] Knowledge Structure Balance Assessment: A knowledge graph contains several connected components, each forming a locally independent knowledge subgraph with internal relationships but no external connections. Parallel exam papers should contain the same number of questions in each connected component. Therefore, Connected Components Entropy (CCE) is proposed to evaluate this consistency.

[0156]

[0157] in, Let P be the set of connected components in the knowledge graph, and let P be the set of parallel test papers. Is with Related questions The proportion of those examined.

[0158] To illustrate the effectiveness of the parallel exam paper generation method provided in Embodiment 2 of the present invention, this embodiment compares the method of the present invention with existing deep learning-based parallel exam paper generation methods (including ExamGAN and MOEPG).

[0159] Specifically, when constructing the multi-layer heterogeneous graph, this invention sets the learning rate to 0.0001, the number of training epochs to 20, the batch size to 128, the hidden layer size to 64, and uses the gradient descent algorithm as the optimizer. In this embodiment, 50 students are randomly selected from each dataset to form a class, repeated 100 times, and the students' performance on the generated test papers is evaluated. Each test paper consists of 100 questions, each question is scored as 1 point. In this embodiment, the number of parallel test paper generation is set to 20, the question embedding dimension is set to 30, the test paper state dimension is set to 30*100, the hidden layer size is set to 128, the batch size is set to 128, the learning rate is set to 0.01, the target update is set to 100, and the greedy parameter gradually decays from 1 to 0.01 at a decay rate of 0.001 in each training step.

[0160] Quantitative results such as Figure 4 As shown, Figure 4 This paper provides a detailed comparison of the comprehensive performance of three different parallel test paper generation methods—GRL-EPG, ExamGAN, and MOEPG—on three real-world educational datasets (ASSISTments0910, Statics2011, and Junyi). The experimental results clearly demonstrate that the proposed GRL-EPG (GEPG) framework significantly outperforms existing methods on most evaluation metrics, showcasing its superior parallel test paper generation capabilities.

[0161] Equivalence at the group and individual levels: GEPG performs best overall across the four group-level metrics, significantly outperforming ExamGAN in Validity and Discrimination. While its overall Difficulty and Rationality are slightly lower, they remain close to optimal. Rationality analysis shows that GEPG and ExamGAN achieve near-perfect consistency at the individual level. In contrast, while MOEPG excels in overall Difficulty and Rationality, its reliability is significantly insufficient. ExamGAN, due to pattern collapse leading to high test paper repetition, while ensuring high Reliability and the fastest generation speed (Average), violates the diversity requirement for parallel test paper generation. Overall, GEPG effectively balances the four objectives of overall Difficulty, Validity, Rationality, and Discrimination, demonstrating the strongest equivalence at both the group and individual levels.

[0162] Knowledge-level equivalence: The results of knowledge distribution entropy (KDE) and knowledge frequency quotient (KFQ) show that GEPG has the most balanced distribution of knowledge points, with most frequencies being integer multiples of the number of test papers, such as... Figure 5As shown, it demonstrates good regularity and focus on core concepts; in contrast, although ExamGAN and MOEPG cover a wider range of knowledge, they exhibit large frequency fluctuations, biases, or long-tailed distributions, and lack consistency. The GRL-EPG proposed in this invention not only maintains equivalence at the knowledge level but also possesses high interpretability of cross-exam question correspondences.

[0163] Question-level equivalence: Connected component entropy (CCE) results show that GEPG has the most consistent question distribution. Figure 6 A visual comparison of parallel test questions and knowledge points is provided. As can be seen from the graph, the number of questions under each knowledge point in GEPG is almost consistent, while ExamGAN and MOEPG show significant differences within the same connected component, exhibiting a clear imbalance. The visualization further illustrates that GEPG has better cross-test consistency at the question level and well explains its validity advantage.

[0164] The parallel test paper intelligent generation method provided in this embodiment combines heterogeneous graph modeling and multi-objective reinforcement learning techniques. Experimental results show that GEPG achieves significant improvements in test paper quality, multi-level equivalence, and generation efficiency compared to state-of-the-art methods. The main contribution of this invention lies in constructing a unified student-question-knowledge three-layer heterogeneous graph framework and innovatively introducing a multi-objective reinforcement learning mechanism on the basis of traditional test paper generation methods to better capture the complex relationships in educational data. In addition, through the proposed multi-dimensional evaluation indicators such as knowledge distribution entropy, knowledge frequency quotient, and connected component entropy, the method of this invention can generate parallel test papers that maintain a high degree of consistency at the four levels of group, individual, knowledge, and question, which helps educators to conduct fair and accurate academic assessments.

[0165] This invention also proposes a test paper generation system, including a data acquisition and graph construction module and a heterogeneous graph reinforcement learning generation module;

[0166] The data acquisition and graph construction module is used to acquire student answer records, question banks, and knowledge point metadata (knowledge text information);

[0167] The heterogeneous graph reinforcement learning generation module includes a heterogeneous graph construction unit, a cognitive diagnosis and prediction unit, a test paper update unit, and a parallel generation unit;

[0168] The heterogeneous graph construction unit is used to construct a student-question interaction graph, a question-knowledge point association graph, and a knowledge point graph that integrates semantics and co-occurrence relationships. The relationships between knowledge points are determined by calculating cosine similarity through text embeddings generated by the Sentence-Transformer model and combining it with co-occurrence information in the questions.

[0169] The cognitive diagnosis and prediction unit aggregates cross-layer node information in a heterogeneous graph through a multi-layer graph attention network to diagnose students’ knowledge mastery status and predict their score distribution on any newly generated test paper.

[0170] The test paper update unit is based on the Rainbow deep reinforcement learning algorithm. It optimizes the test paper generation strategy through a dual deep Q network and a priority experience replay mechanism. Its reward function integrates four objectives: difficulty, effectiveness, rationality, and discrimination.

[0171] The parallel test paper generation unit is used to select alternative test questions based on the adjacency relationship in the knowledge point graph, and generate multiple sets of equivalent parallel test papers at four levels: group, individual, knowledge point, and test question.

[0172] The parallel test paper intelligent generation system in this embodiment achieves unified representation and fusion of multi-source heterogeneous data in the educational environment by constructing a three-layer heterogeneous graph of students, test questions, and knowledge points. This network effectively captures the complex relationships between nodes through a graph attention mechanism, providing rich semantic information for subsequent cognitive diagnosis and performance prediction. Subsequently, a multi-objective reinforcement learning framework is designed. By coordinating the optimization of four objectives—difficulty, effectiveness, rationality, and discrimination—it solves the technical limitations of traditional methods that rely on expert annotation and ignore students' real performance. That is, by simulating students' answer performance to dynamically optimize the test paper composition, it can alleviate the problem of inconsistency between the assessment effect and the teaching objectives in test paper generation, improve the system's adaptability to personalized learning needs and assessment accuracy, and enhance teaching quality. Through the proposed knowledge distribution entropy, knowledge frequency quotient, and connected component entropy indices, it efficiently realizes the generation of multi-level equivalent parallel test papers.

[0173] This embodiment also provides a storage medium on which a program for implementing the parallel test paper generation method is stored.

[0174] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a computer, perform the following steps:

[0175] Data preprocessing: Obtain student answer records, test question bank and knowledge point metadata from the education platform, and clean, deduplicate and standardize the data to construct high-quality heterogeneous graph input data.

[0176] Heterogeneous graph construction: Based on the preprocessed data, construct student-question interaction graph and question-knowledge point association graph, and construct knowledge point graph by calculating semantic similarity and co-occurrence relationship.

[0177] Graph representation learning: Using a multi-layer graph attention network to learn the embedding representation of nodes in a heterogeneous graph, fusing cross-layer information to obtain deep representations (high-order feature representations) of students, test questions and knowledge points.

[0178] Cognitive diagnosis and performance prediction: Based on the learned node representations (higher-order feature representations), diagnose students' mastery of knowledge points and predict their performance distribution on any test paper.

[0179] Reinforcement learning test paper generation: Initialize the test paper state, and based on the Rainbow reinforcement learning algorithm, guide the test paper replacement strategy through a multi-objective reward function to iteratively optimize the difficulty, effectiveness, rationality and discrimination of the test paper.

[0180] Parallel test paper generation: Based on the adjacency relationship of the knowledge point graph, alternative questions are selected from the related question subsets to generate multiple equivalent parallel test papers.

[0181] Multi-dimensional equivalence assessment: Calculate the equivalence indicators of parallel test papers at four levels: group, individual, knowledge point, and test question, including knowledge distribution entropy, knowledge frequency quotient, and connected component entropy.

[0182] Results Output and Visualization: The generated test papers and assessment results are saved as structured data files for use by the education platform, and then visualized and analyzed on the teacher's interface.

[0183] The storage medium may be a hard disk, solid-state drive, optical disk, flash drive, or cloud storage device, etc., capable of storing the above instructions in electronic form for execution.

[0184] This embodiment also provides an electronic device, which includes: at least one central processing unit (CPU), a graphics processing unit (GPU), a storage unit electrically coupled to the CPU, and input / output devices and a network communication module. The storage unit stores program code that can be read and executed by the CPU for executing the processing methods related to this invention.

[0185] This invention also provides various electronic devices, including:

[0186] Computational Units: The computational units consist of a Central Processing Unit (CPU) and a Graphics Processing Unit (GPU). The CPU possesses multi-core parallel processing capabilities, used for coordinating and managing computational tasks such as data preprocessing, heterogeneous graph construction, and reinforcement learning policy scheduling. It also supports efficient thread management to meet the real-time generation requirements of educational scenarios. The GPU significantly improves the efficiency of graph neural network training and inference, as well as reinforcement learning decision-making, through hardware acceleration. It supports large-scale matrix operations and representation learning of complex heterogeneous graph structures. Its architecture is optimized for parallel computing performance, adapting to the needs of deep learning frameworks and graph computing algorithms. Working together, these two components can efficiently handle large-scale heterogeneous graph computations and multi-objective optimization problems involved in educational data, thereby meeting the requirements for high performance and low latency.

[0187] Storage Units: Storage units include Random Access Memory (RAM) and non-volatile memory (such as solid-state drives). RAM is used to store temporary data generated during operation, heterogeneous graph node embedding representations, and intermediate state data in the reinforcement learning experience replay buffer. Its capacity and access speed are crucial to system performance and should support high-concurrency access to meet the needs of large-scale graph data processing and multi-agent decision-making, thereby ensuring the smoothness and real-time performance of the test paper generation process. Non-volatile memory is used to store long-term retained data such as program files, pre-trained model parameters, student historical answer records, test question banks, and knowledge point graphs. Solid-state drives (SSDs) or other non-volatile storage media can be used to provide fast data write and read capabilities. Its capacity design needs to support the high-dimensional feature storage and fast loading of massive educational data. In addition, storage units can adopt a hierarchical storage architecture, combining cache and main memory to optimize data access; equipped with a data backup module, using mirrored storage or RAID (Redundant Array of Independent Disks) to improve the security of educational data; and support expanding storage capacity through external interfaces (such as USB, SATA, or PCIe) to meet the needs of future growth in educational data volume. Through the above design, the storage unit can efficiently support the data processing needs of electronic devices, ensuring the stability and efficiency of the system in smart education applications.

[0188] Input Devices: Input devices support the collection of educational data, including but not limited to student answer records, test bank metadata, and text descriptions of knowledge points, enabling the acquisition of both structured and unstructured data. Input devices must be compatible with standard educational data formats (such as CSV, JSON, XML, etc.) to ensure consistency in data interaction and processing with various learning management systems (LMS), while supporting multiple data input interfaces to adapt to the connection requirements of different educational platforms. Through input devices, educational data can be efficiently collected and transmitted, providing a foundation for subsequent heterogeneous graph construction and test paper generation.

[0189] Output devices include high-resolution displays for intuitively displaying the generated test paper content, multi-objective optimization curves, and multi-level equivalence evaluation results of parallel test papers, ensuring clear visibility of the output content. In addition, output devices may also include interactive visualization devices to support in-depth analysis and interactive operations on knowledge graph structures, student ability distributions, etc. These devices support multiple display modes to meet the needs of teachers' teaching evaluation or educational analysis, and provide users with more comprehensive and intuitive data visualization effects.

[0190] Network communication module: including a high-speed Ethernet adapter and an optional 5G communication module, used for data synchronization with the cloud education platform, receiving real-time student data, and distributing generated test papers to the online examination system, supporting intelligent education scenarios that combine edge computing and cloud computing.

[0191] With the configuration of the aforementioned electronic devices, the efficient educational data processing and real-time analysis and application of parallel test paper generation results described in this invention can be realized.

[0192] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0193] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for generating test papers, characterized in that, Includes the following steps: We acquire student answer records, question-knowledge point mappings, question banks, and knowledge text information. We extract student embedding representations, question embedding representations, and knowledge embedding representations respectively. We then use a graph attention network to update the embedding relationships of each node in the knowledge point graph, and construct a student-question interaction graph, a question-knowledge point association graph, and a knowledge point graph that integrates semantics and co-occurrence relationships. A fully connected network is used to integrate knowledge embeddings into student and question embeddings respectively, optimizing the graph structure and constructing a three-layer heterogeneous graph of student-question-knowledge. High-order feature representations of each node in the heterogeneous graph are then extracted to predict student performance on the test. The performance of student i on question j is predicted using a fully connected network. for: ； The loss function for a fully connected network is determined using binary cross-entropy: ； For a test paper with n questions, the predicted total score for student i, i.e., student i's performance on the test paper, is represented as: ； in, Representing knowledge points Adjacent knowledge points, , For the integrated student embedding representation, The integrated question embedding representation, for function, It is a fully connected layer. To ensure accurate and accurate answers, This represents the score for question j. Using students' performance on the test as environmental feedback input for reinforcement learning, and the current set of questions in the test as the state, replacing or retaining questions in the test is the action. By constructing a multi-objective reward function that considers the test's overall difficulty, effectiveness, reasonableness, and discrimination, the comprehensive reward value of the test in each dimension of the multi-objective reward function is determined. With the goal of maximizing the comprehensive reward value, the optimal test is generated through iterative optimization. Obtaining a measure of student average scores and target difficulty Consistent overall difficulty reward for: ； To obtain an effective reward for measuring the consistency between the knowledge distribution of the exam paper and the course requirements. for: ； Obtain a reasonable reward for measuring the similarity between the distribution of student scores and the target distribution. for: ； Obtain a discrimination reward that measures the test paper's ability to differentiate between students of different levels. for: ； in, The total score for the exam is... The average score of students For the first question in the exam paper The proportion of each knowledge point tested. This represents the required percentage of this knowledge point in the course. The actual distribution of student scores Let be the target distribution, and be the average score of the top 27% of students. This is the average score of the bottom 27% of students; The comprehensive reward function is determined as follows: ； By using a comprehensive reward function, the agent can simultaneously optimize multiple teaching objectives in each question update to generate the optimal test paper. Based on the adjacency relationships in the knowledge point graph, candidate questions are selected from the question bank to form a candidate subset. Using an equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined. Through iterative optimization, multiple parallel test papers equivalent to the optimal test paper are generated. For each knowledge node its candidate sub-topic set It consists of the set of directly related questions and the set of questions associated with its adjacent knowledge nodes: ； in, This indicates a direct examination of knowledge points. A collection of questions, express The set of adjacent knowledge points; The equivalent question replacement strategy is to replace each question in the generated parallel test paper. Corresponding knowledge nodes The system performs a neighborhood expansion operation on the candidate sub-question set and randomly selects a question as a replacement candidate from the candidate sub-question set. If the selected question in the replacement candidate satisfies the multi-objective reward function, the replacement is performed. If no suitable candidate question is found, the search is expanded sequentially to the candidate sub-question set of the neighboring knowledge nodes. If no candidate question that meets the conditions is found after the neighborhood expansion, the original question is retained without replacement to ensure the consistency of the test paper in terms of knowledge coverage.

2. The test paper generation method according to claim 1, characterized in that, The method, based on the adjacency relationships in the knowledge point graph, selects candidate questions from the question bank to form a candidate subset. Using an equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined to generate multiple parallel test papers equivalent to the optimal test paper. Specifically, this includes: Randomly select a question from the candidate question set as a replacement candidate. When the replacement candidate satisfies the requirement of maximizing the comprehensive reward value, the original question in the optimal test paper is replaced. Otherwise, the search is expanded to the candidate question set of the neighboring knowledge nodes. If no replacement candidate with the maximum comprehensive reward value is found after the neighborhood expansion, the original question is retained without replacement. By iterating continuously in the state and action space, and using an equivalent question replacement strategy, questions that are different but cover the same knowledge points as the optimal test paper are replaced and combined to replace the questions in the optimal test paper. This optimizes the overall difficulty, effectiveness, rationality and discrimination of the test paper, with the goal of maximizing the overall reward value. Through iterative optimization, multiple sets of parallel test papers equivalent to the optimal test paper are generated.

3. The test paper generation method according to claim 1, characterized in that, The process involves using students' performance on the test as environmental feedback input for reinforcement learning, taking the current set of questions on the test as the state, and replacing or retaining questions as actions. By constructing a multi-objective reward function that considers the test's overall difficulty, effectiveness, reasonableness, and discrimination, the process determines the test's overall reward value across each dimension of the multi-objective reward function. With the goal of maximizing the overall reward value, the optimal test is generated through iterative optimization. Specifically, this includes: A dual-deep Q-network is used to iteratively train the state and action of reinforcement learning. The dual-deep Q-network takes the student's answer to the test paper as the environmental feedback input, takes the current set of questions in the test paper as the state, and takes the equivalent question replacement strategy to replace or retain questions in the test paper as the action, with the goal of maximizing the comprehensive reward value. During training, the agent uses an ε-greedy strategy to balance exploration and exploitation. It samples historical state-action-reward sequences through a priority experience replay mechanism and updates the parameters of the dual-deep Q-network through multiple rounds of iterative training. The agent gradually learns the optimal strategy for selecting questions in different states. By simulating students' performance on the test paper, it generates the optimal test paper that maximizes the comprehensive reward value across dimensions such as comprehensive difficulty, effectiveness, rationality, and discrimination, thus meeting multiple objective requirements.

4. The test paper generation method according to claim 1, characterized in that, The integration of knowledge embedding representations into student embedding representations and question embedding representations, by constructing a three-layer heterogeneous graph of student-question-knowledge and extracting high-order feature representations of each node in the heterogeneous graph, predicts students' performance on the test paper. Specifically, this includes: Obtain the low-dimensional vectors of student embeddings, question embeddings, and knowledge embeddings respectively; By constructing a three-layer heterogeneous graph of student-question-knowledge, high-order feature representations of each node are extracted; The low-dimensional vector and high-order feature representations are fused together, and the neighbor node information is iteratively aggregated. The student's performance is determined by predicting the probability of answering unanswered questions on the test paper.

5. The test paper generation method according to claim 1, characterized in that, The process involves acquiring student answer records, question-knowledge point mappings, a question bank, and knowledge text information. Student embeddings, question embeddings, and knowledge embeddings are extracted respectively. A graph attention network is then used to update the embedding relationships of each node in the knowledge point graph, constructing a student-question interaction graph, a question-knowledge point association graph, and a knowledge point graph that integrates semantics and co-occurrence relationships. Specifically, this includes: Based on the acquired student answer records, question-knowledge point mapping, question bank and knowledge text information, one-hot vectors are generated through the encoder encoding process to obtain the initial embedding representations corresponding to students, questions and knowledge points respectively. The graph attention network aggregates neighbor node information to update the initial embedding representation; wherein the graph attention network includes an embedding layer and a fully connected layer. In the embedding layer, a multi-layer graph attention mechanism is introduced to update the initial embedding representations corresponding to students, questions, and knowledge points respectively, and obtain the updated embedding representations corresponding to students, questions, and knowledge points. In the fully connected layer, the knowledge embedding representation is integrated into the student embedding representation and the question embedding representation respectively to construct the student-question interaction graph, the question-knowledge point association graph, and the knowledge point graph that integrates semantics and co-occurrence relationships.

6. The test paper generation method according to claim 1, characterized in that, The process of generating multiple parallel test papers equivalent to the optimal test paper through iterative optimization also includes evaluating the equivalence of the generated parallel test papers at four levels: individual students, student groups, knowledge, and questions, based on knowledge distribution entropy, knowledge frequency quotient, and connected component entropy.

7. The test paper generation method according to claim 1, characterized in that, The acquisition of student answer records, question bank, question-knowledge point mapping, and knowledge text information specifically includes: When retrieving student answer records, the student's historical answer log is exported from the online education platform. This log includes the student ID, question ID, answer result (correct or incorrect), answer timestamp, and number of answer attempts. The question-knowledge point mapping is obtained by subject matter experts or course teachers through manual annotation based on the course syllabus and teaching objectives, clarifying one or more knowledge points tested by each question, and forming a structured question-knowledge point association matrix; Obtain knowledge text information, including the unique identifier, name, detailed description text, and relative importance weight of each knowledge point in the course knowledge system; The acquired student answer records, question bank, question-knowledge point mapping, and knowledge text information are cleaned and preprocessed.

8. A test paper generation system, characterized in that, include: The knowledge point graph construction module is used to acquire student answer records, question-knowledge point mapping, test question bank and knowledge text information, extract student embedding representation, question embedding representation and knowledge embedding representation respectively, and use graph attention network to update the embedding relationship of each node in the knowledge point graph, and construct student-question interaction graph, question-knowledge point association graph and knowledge point graph that integrates semantics and co-occurrence relationship; The heterogeneous graph construction module integrates knowledge embeddings into student and question embeddings using fully connected networks, optimizes the graph structure, and predicts student performance on exams by constructing a three-layer heterogeneous graph (student-question-knowledge) and extracting high-order feature representations from each node. The performance of student i on question j is predicted using a fully connected network. for: ； The loss function for a fully connected network is determined using binary cross-entropy: ； For a test paper with n questions, the predicted total score for student i, i.e., student i's performance on the test paper, is represented as: ； in, Representing knowledge points Adjacent knowledge points, , For the integrated student embedding representation, The integrated question embedding representation, for function, It is a fully connected layer. To ensure accurate and accurate answers, This represents the score for question j. The optimal test paper generation module uses students' performance on the test paper as environmental feedback input for reinforcement learning. It uses the current set of questions in the test paper as the state and replaces or retains questions as actions. By constructing a multi-objective reward function that considers the test paper's overall difficulty, effectiveness, reasonableness, and discrimination, it determines the overall reward value of the test paper across each dimension of the multi-objective reward function. With the goal of maximizing the overall reward value, it iteratively optimizes and generates the optimal test paper. Obtaining a measure of student average scores and target difficulty Consistent overall difficulty reward for: ； To obtain an effective reward for measuring the consistency between the knowledge distribution of the exam paper and the course requirements. for: ； Obtain a reasonable reward for measuring the similarity between the distribution of student scores and the target distribution. for: ； Obtain a discrimination reward that measures the test paper's ability to differentiate between students of different levels. for: ； in, The total score for the exam is... The average score of students For the first question in the exam paper The proportion of each knowledge point tested. This represents the required percentage of this knowledge point in the course. The actual distribution of student scores For the target distribution, The average score of the top 27% of students This is the average score of the bottom 27% of students; The comprehensive reward function is determined as follows: ； By using a comprehensive reward function, the agent can simultaneously optimize multiple teaching objectives in each question update to generate the optimal test paper. The parallel test paper generation module is used to generate multiple sets of parallel test papers equivalent to the optimal test paper by selecting candidate questions from the question bank based on the adjacency relationship in the knowledge point graph, forming a candidate subset of questions, and using an equivalent question replacement strategy to replace and combine questions that are different but cover the same knowledge points as the optimal test paper. For each knowledge node its candidate sub-topic set It consists of the set of directly related questions and the set of questions associated with its adjacent knowledge nodes: ； in, This indicates a direct examination of knowledge points. A collection of questions, express Adjacent knowledge points, express The set of adjacent knowledge points; The equivalent question replacement strategy is to replace each question in the generated parallel test paper. Corresponding knowledge nodes The system performs a neighborhood expansion operation on the candidate sub-question set and randomly selects a question as a replacement candidate from the candidate sub-question set. If the selected question in the replacement candidate satisfies the multi-objective reward function, the replacement is performed. If no suitable candidate question is found, the search is expanded sequentially to the candidate sub-question set of the neighboring knowledge nodes. If no candidate question that meets the conditions is found after the neighborhood expansion, the original question is retained without replacement to ensure the consistency of the test paper in terms of knowledge coverage.

Images