Intelligent test paper composition and scoring method and system based on knowledge graph

Abstract

The invention discloses an intelligent test paper composition and scoring method and system based on a knowledge graph. The method comprises the following steps: S1, constructing a student knowledge graph state sub-graph; s2, constructing a question bank knowledge graph; s3, on the basis of a node embedding modeling method of an improved DeepWalk algorithm, performing joint training and low-dimensional representation learning on knowledge point nodes by utilizing a continuous bag-of-words model of multi-head context aggregation, and constructing a topic adaptation degree scoring index; s4, outputting an optimal test paper scheme set; s5, collecting a current answer record when the student completes the test paper; and S6, extracting a semantic feature vector and a behavior feature vector, constructing a fusion score feature vector, and generating scores of test paper subjective question answers. According to the invention, accurate test paper composition and multi-dimensional subjective question intelligent scoring for individual knowledge states of students are realized, the method is suitable for personalized learning evaluation and teaching feedback scenes in an education evaluation platform, and the method has the advantages of high personalization, fine feedback and automatic scoring.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and education technology, and in particular to a knowledge graph-based intelligent test paper generation and scoring method and system. Background Technology

[0002] In the current development of intelligent education systems, personalized teaching and intelligent assessment have become key research directions. Traditional test paper generation and scoring methods rely heavily on teachers' experience and fixed templates, lacking the dynamic perception and adaptability to students' knowledge mastery levels. This fails to effectively meet the differentiated needs of students in terms of learning ability, knowledge mastery, and behavioral strategies. Although some systems have introduced question bank management and difficulty matching mechanisms, combining questions through static tag filtering, the lack of deep modeling of students' knowledge structures and semantic reasoning capabilities results in test papers that fail to accurately reflect students' weak knowledge points and cannot achieve personalized optimization in terms of question structure, knowledge coverage, and difficulty distribution.

[0003] In recent years, knowledge graphs, as an effective tool for expressing conceptual relationships and knowledge structures, have been widely applied in the field of educational intelligence to describe the knowledge point connections between students and questions. However, existing methods mostly remain at the level of graph structure visualization and simple path querying, failing to delve into the potential semantic information between nodes and structures in the graph, especially in the semantic fusion modeling of student answer behavior and answer text information, which still has significant shortcomings. At the same time, subjective question scoring still mainly relies on manual marking or coarse-grained scoring methods based on keyword matching, which cannot comprehensively and accurately assess the depth of students' understanding of knowledge points and the rationality of their answer strategies, and lacks the ability to perform semantic integrity analysis and behavioral feature modeling.

[0004] Existing test paper generation strategies generally employ heuristic rules or single-objective optimization methods, making it difficult to simultaneously consider multiple educational objectives, such as knowledge point coverage, question type ratio, difficulty balance, and personalized adaptability. The lack of a unified multi-objective combination optimization mechanism leads to unstable test paper quality and insufficient adaptability, hindering widespread application in actual teaching. Therefore, there is an urgent need for an intelligent test paper generation and scoring method that integrates knowledge graph modeling, graph embedding learning, language model semantic analysis, and time-series behavioral modeling. This method should comprehensively characterize students' knowledge status, accurately achieve personalized question matching, and provide multi-dimensional automatic scoring and dynamic feedback on answers, thereby improving the intelligence and personalization level of the education system. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent test paper generation and scoring method and system based on knowledge graphs. This invention fully integrates graph neural network modeling, deep embedding learning, pre-trained language models, time series behavior modeling, and multi-objective genetic optimization algorithms. It describes in detail how to generate personalized test papers based on the state subgraph of the student knowledge graph and the question bank knowledge graph, and how to complete multi-dimensional automatic scoring and dynamic feedback updates by combining subjective question text and behavior sequence data. It has the advantages of strong adaptability, precise scoring, feedback loop, and high degree of automation.

[0006] The knowledge graph-based intelligent test paper generation and scoring method and system according to embodiments of the present invention includes the following steps:

[0007] S1. Collect students' historical answer records and learning behavior data to construct a state subgraph of the student knowledge graph;

[0008] S2. Construct a knowledge graph for the question bank;

[0009] S3. A node embedding modeling method based on the improved DeepWalk algorithm, which uses a continuous bag-of-words model with multi-head context aggregation to jointly train and learn low-dimensional representations of knowledge point nodes, and constructs a question fit scoring index.

[0010] S4. Set the objective function for test paper generation, use the non-dominated sorting genetic algorithm to combine and optimize the candidate test paper generation set, and output the optimal test paper scheme set;

[0011] S5. Collect the student's current answer record when completing the test paper;

[0012] S6. Based on the scoring feature extraction module and the subjective question scoring and feedback module, extract semantic feature vectors and behavioral feature vectors, construct a fused scoring feature vector, and generate the score for the subjective questions in the test paper.

[0013] Optionally, the question bank knowledge graph connects each question in the question bank knowledge graph with the covered knowledge points, and records the question type, difficulty level, and structured information of the knowledge unit to which it belongs.

[0014] Optionally, S3 specifically includes:

[0015] S31. Perform unsupervised node embedding modeling on the student knowledge graph state subgraph and the question bank knowledge graph respectively. Use the improved DeepWalk algorithm to take each node in the student knowledge graph state subgraph and the question bank knowledge graph as the starting point and perform a random walk operation of a preset number of rounds. According to the set walk length and jump rules, sample and generate a set of random walk sequences composed of node numbers.

[0016] The random walk operation refers to starting from each node in the graph and sequentially visiting adjacent nodes according to a set number of walk rounds, walk length, and jump probability, generating a set of node sequences that reflect the structural adjacency relationships and semantic relevance of the graph, which is used to construct the training corpus for the subsequent embedding model. The random walk sequence is used to characterize the structural adjacency relationships between nodes in the graph; the jump rules are probabilistically modeled based on the structural distance and semantic weights between the current node and its adjacent nodes.

[0017] S32. Using the random walk sequence as training corpus, a continuous bag-of-words model is used to construct a center node prediction task. For each center node, training samples are constructed with it as the center in the random walk sequence. A fixed window size is set, and a preset number of context nodes are selected from both sides of the center node. The embedding vectors corresponding to each group of context nodes are aggregated and encoded, and used as input to the continuous bag-of-words model to predict the center node number corresponding to the context node set. A training sample pair set with the context node set as input and the center node as output is constructed.

[0018] S33. During the training process, for each training sample pair, a multi-head parallel nonlinear transformation and weighted aggregation are performed based on the embedding vector of the context node to generate a context aggregation representation vector; the context aggregation representation is used as the input for center node prediction in the continuous bag-of-words model to predict the embedding representation of the corresponding center node in the training sample pair.

[0019] S34. Based on the vector difference between the center node embedding representation output by the continuous bag-of-words model with multi-head context aggregation structure and the true embedding representation of the corresponding center node in the training samples, a loss function is constructed. The backpropagation algorithm and gradient descent method are used to iteratively optimize the model parameters to complete the low-dimensional embedding vector learning of the knowledge point nodes of the student knowledge graph state subgraph and the question bank knowledge graph.

[0020] S35. Calculate the question embedding representation of the knowledge point nodes covered by each question in the question bank knowledge graph, and calculate the weighted average embedding representation of the knowledge point nodes related to the question in the student knowledge graph state subgraph. The weighting coefficient is set according to the mastery of the corresponding knowledge point node to obtain the student embedding representation.

[0021] S36. Use the similarity between the question embedding representation and the student embedding representation as the question fit scoring index;

[0022] S37. Sort all questions according to the question suitability scoring index, and select questions with scores higher than the preset threshold to form a candidate set for test paper compilation.

[0023] Optionally, S4 specifically includes:

[0024] The improved DeepWalk algorithm uses each knowledge node in the graph as a starting point and samples a set of random walk sequences reflecting structural adjacency relationships according to a set number of random walk rounds, walk length, and jump rules. The jump rules combine structural distance and semantic weights between nodes to construct transition probabilities, enhancing the modeling ability of local structure and semantic relevance in the graph. Subsequently, the generated random walk sequences are used as training corpus for a continuous bag-of-words model. A fixed window size is used to construct context-center node training sample pairs. Multi-head parallel nonlinear transformations and weighted aggregations are performed on the embedding vectors of context nodes to generate context aggregation representation vectors, which are used to predict the corresponding center node embedding representation. A loss function is constructed by calculating the vector difference between the predicted and true embeddings. Backpropagation and gradient descent are used to optimize the model parameters, thereby obtaining low-dimensional node embedding vectors with stronger semantic expressive power. These are further used to construct question embedding representations and student embedding representations, and the similarity between the two is used as a question fit scoring index to achieve accurate matching between the question and the student's current knowledge state.

[0025] Optionally, S4 specifically includes:

[0026] S41. Set the objective function for test paper compilation. The objective function for test paper compilation should also consider the following constraints: knowledge point coverage, difficulty matching, reasonable distribution of question types, and priority of weak point coverage as the basis for weak point priority scoring.

[0027] S42. Using the candidate test paper set generated in step S37 as the initial population, encode the question composition of each individual test paper and construct a genetic algorithm optimization problem based on question combination.

[0028] S43. Initialize the parameters of the non-dominated sorting genetic algorithm, including population size, maximum number of iterations, crossover probability, and mutation probability;

[0029] S44. In each generation, calculate the fitness vector of each individual in the current population based on the constraints in the objective function of the test paper.

[0030] S45. Perform a non-dominated sorting operation. Based on the dominance relationship of the Pareto optimal solution set, select individuals with better fitness to form the next generation population, and perform crossover and mutation operations to generate new test paper combination individuals. The non-dominated sorting operation refers to, in each generation of the genetic algorithm, according to the multiple optimization objectives set in the test paper combination objective function, performing a dominance relationship judgment on all test paper individuals in the current population based on the Pareto optimal solution set, and dividing the test paper individuals into several non-dominated levels. Among any two test paper individuals, if one test paper individual is not inferior to the other test paper individual in all objectives, and is superior to the other test paper individual in at least one objective. For each individual test paper, the former is considered to dominate the latter, thus forming a non-dominant ranking relationship. After stratifying and ranking individuals according to the non-dominant level, further ranking is performed within each non-dominant level based on crowding distance. The crowding distance is used to measure the density of the distribution of individuals in the objective function space, and test paper individuals that cover the boundary area and are evenly distributed are preferentially retained. In the population update stage, according to the non-dominant level priority and the crowding distance ranking results, a set number of high-quality individuals are selected to form the next generation population, and crossover and mutation operations are performed on the individuals in the population to generate new test paper combination individuals, thus completing the genetic evolution process of the population.

[0031] S46. Iteratively execute the genetic evolution process until the maximum number of iterations is met, and finally output a set of optimal test paper schemes that meet the set constraints and are optimally adapted.

[0032] Optionally, the genetic evolution process includes: calculating the fitness vector for each individual test paper in the current population; selecting high-quality individuals to form the next generation population based on non-dominated sorting and crowding distance; in each generation, selecting parent individuals from the current population using a roulette wheel selection strategy; generating offspring individuals by performing partial exchange operations on question combinations according to a set crossover probability; performing local question replacement operations on offspring individuals according to a preset mutation probability; recalculating the fitness vector of the updated individuals and participating in the next round of population evolution until the maximum number of iterations is reached or the population fitness change is lower than a set threshold, at which point the algorithm is considered to have converged, and the Pareto optimal test paper scheme set in the converged generation population is output.

[0033] Optionally, the current answer record consists of subjective question answer text and behavior sequence. The behavior sequence includes answer data content and answer process behavior data, including answer time, question switching order, dwell time, and modification records.

[0034] Optionally, S6 specifically includes:

[0035] S61. Collect subjective question answer texts and behavioral sequence information generated by students during the process of completing personalized test papers. The behavioral sequence includes behavioral characteristics such as answer time, question switching order, dwell time, and modification records.

[0036] S62. Input the subjective question answer text into the pre-trained language model BERT and extract semantic feature vectors representing semantic integrity and keyword coverage. The semantic integrity is represented by calculating the average semantic similarity between the student's subjective question answer text and the set of embedded vectors corresponding to the knowledge points of the standard answer of the subjective question in the embedding space, which is used to measure semantic integrity. The keyword coverage is obtained by calculating the degree of matching between the key terms in the subjective question answer text and the set of keywords in the standard answer.

[0037] S63. Encode the behavior sequence into a time series vector and input it into the time series modeling based on gated recurrent units to extract the behavior feature vector representing the behavior strategy features;

[0038] S64. Perform a concatenation operation between the semantic feature vector extracted in step S62 and the behavioral feature vector extracted in step S63 to construct a fused scoring feature vector.

[0039] S65. Input the fused scoring feature vector into the feedforward neural network scoring model, and output the scores for the subjective questions on the test paper, including semantic rationality score, content completeness score, and behavioral strategy rationality score.

[0040] The knowledge graph-based intelligent test paper generation and scoring system according to an embodiment of the present invention includes the following modules:

[0041] The student knowledge graph construction module is used to collect students' historical answer records and learning behavior data to construct the student knowledge graph state subgraph;

[0042] The question bank knowledge graph construction module is used to construct the question bank knowledge graph;

[0043] The embedding modeling and question fitting scoring module is used to perform joint node embedding training on the student knowledge graph state subgraph and the question bank knowledge graph based on the improved DeepWalk algorithm and the continuous bag-of-words model, generate low-dimensional embedding vectors of knowledge point nodes, and calculate the question fitting score index.

[0044] The personalized test paper optimization module is used to set the objective function for test paper generation. It uses a non-dominated sorting genetic algorithm to combine and optimize the test paper candidate set whose question fit score index is higher than the preset threshold, and outputs the personalized optimal test paper scheme.

[0045] The scoring feature extraction module is used to input the subjective question answer text into the pre-trained language model BERT to extract feature vectors, input the time series vectors into the gated recurrent unit to extract behavioral feature vectors, and construct a fused scoring feature vector.

[0046] The subjective question scoring and feedback module is used to input the fused scoring feature vector into the feedforward neural network scoring model and output the scores of the subjective questions answered on the test paper.

[0047] The beneficial effects of this invention are:

[0048] By constructing a student knowledge graph state subgraph and a question bank knowledge graph, and introducing an improved DeepWalk algorithm and a continuous bag-of-words model with multi-head context aggregation, the system achieves dual embedding modeling of the structure and semantics of knowledge point nodes. This enables the system to accurately depict the student's current knowledge mastery status and the deep semantic information of the test questions, thereby achieving personalized and precise test question fit assessment.

[0049] This invention sets a multi-objective test paper generation constraint function, which comprehensively considers knowledge point coverage, difficulty matching, question type structure and weak point priority, and uses a non-dominated sorting genetic algorithm to perform evolutionary optimization of the test paper generation candidate set. This not only ensures that the content of the generated test paper is scientific and reasonable, but also significantly improves the efficiency and intelligence of test paper generation, meeting the actual needs of automatic construction of diverse test papers in educational scenarios.

[0050] During the scoring phase, this invention combines the pre-trained language model BERT with a GRU-based behavior sequence modeling method to effectively extract semantic integrity and keyword coverage features from subjective question answers, as well as strategy pattern features from answering behaviors. Multi-dimensional scoring is achieved through a feedforward neural network scoring model, and the scoring results are promptly fed back to the knowledge graph to realize dynamic updates and refined modeling of students' knowledge status, thus opening up a complete closed-loop path of "assessment-learning-testing".

[0051] This invention not only breaks through the bottleneck of insufficient coverage of knowledge points and adaptation to individual student differences in traditional test paper generation methods, but also solves the problems of low efficiency and poor consistency in manual subjective question scoring. It has high automation, strong adaptability and wide applicability, and has important promotional value in the fields of intelligent education assessment and personalized teaching recommendation. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is an overall flowchart of the knowledge graph-based intelligent test paper generation and scoring method proposed in this invention;

[0054] Figure 2 This is a schematic diagram of the structure of the knowledge graph-based intelligent test paper generation and scoring system proposed in this invention;

[0055] Figure 3This is a schematic diagram of the improved DeepWalk algorithm structure proposed in this invention, which is based on a knowledge graph-based intelligent test paper generation and scoring system. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1 The knowledge graph-based intelligent test paper generation and scoring method includes the following steps:

[0058] S1. Collect students' historical answer records and learning behavior data to construct a state subgraph of the student knowledge graph;

[0059] S2. Construct a knowledge graph for the question bank;

[0060] S3. A node embedding modeling method based on the improved DeepWalk algorithm, which uses a continuous bag-of-words model with multi-head context aggregation to jointly train and learn low-dimensional representations of knowledge point nodes, and constructs a question fit scoring index.

[0061] S4. Set the objective function for test paper generation, use the non-dominated sorting genetic algorithm to combine and optimize the candidate test paper generation set, and output the optimal test paper scheme set;

[0062] S5. Collect the student's current answer record when completing the test paper;

[0063] S6. Based on the scoring feature extraction module and the subjective question scoring and feedback module, extract semantic feature vectors and behavioral feature vectors, construct a fused scoring feature vector, and generate the score for the subjective questions in the test paper.

[0064] This invention provides a method for intelligent test paper generation and scoring based on knowledge graphs. First, it collects students' historical answer records and learning behavior data to construct a student knowledge graph state subgraph. Then, combining this with a question bank knowledge graph, it employs an improved DeepWalk algorithm and a continuous bag-of-words model with multi-head context aggregation to achieve knowledge point node embedding modeling and question fit scoring. Based on this, it sets a test paper generation objective function and introduces a non-dominated sorting genetic algorithm to optimize and combine the candidate test paper sets, generating a personalized optimal test paper scheme. Subsequently, it collects students' current answer records after completing the test paper, extracts semantic and behavioral features through pre-trained language models BERT and GRU time series modeling modules, constructs a fused scoring feature vector, and inputs it into a feedforward neural network model to output multi-dimensional subjective question answer scoring results, achieving dynamic feedback and continuous updating of students' knowledge status.

[0065] In this embodiment, each node in the student knowledge graph state subgraph represents a knowledge point. The node contains attribute information such as the student's mastery level, the number of times they answered questions, and their accuracy rate. The edges represent the relationships between knowledge points.

[0066] In this embodiment, the question bank knowledge graph connects each question in the question bank knowledge graph with the covered knowledge points, and records the question type, difficulty level and structured information of the knowledge unit to which it belongs.

[0067] In this embodiment, S3 specifically includes:

[0068] S31. Perform unsupervised node embedding modeling on the student knowledge graph state subgraph and the question bank knowledge graph respectively. Use the improved DeepWalk algorithm, taking each node in the student knowledge graph state subgraph and the question bank knowledge graph as the starting point, and perform a preset number of random walk operations. According to the set walk length and jump rules, sample and generate a set of random walk sequences composed of node numbers:

[0069] ;

[0070] The diagram structure is as follows: For each starting node ,implement Each round of movement generates a length of [length]. Node sequence ,satisfy:

[0071] ;

[0072] All node sequences constitute a set of random walk sequences. The random walk sequence is used to characterize the structural adjacency relationship between nodes in the graph and serves as the input corpus for subsequent training. The jump rule is based on the structural distance and semantic weight between the current node and its neighboring nodes to perform probability modeling, which guides the random walk sequence to generate node paths in the graph that are more consistent with the learning state distribution.

[0073] Jump rules are constructed by probabilistic distribution functions. Implementation, used to control the random walk in the current node. Transition behavior between it and its neighboring nodes:

[0074] ;

[0075] in, This represents the node visited at step t in the random walk path. Indicates structural distance; Indicates semantic weight; Indicates the control coefficient; It is the set of adjacent nodes of the current node.

[0076] S32. Using the random walk sequence as training corpus, a continuous bag-of-words model is used to construct a center node prediction task. For each center node, training samples are constructed with it as the center in the random walk sequence. A fixed window size is set, and a preset number of context nodes are selected from both sides of the center node. The embedding vectors corresponding to each group of context nodes are aggregated and encoded, and used as input to the continuous bag-of-words model to predict the center node number corresponding to the context node set. A training sample pair set with the context node set as input and the center node as output is constructed.

[0077] S33. During training, for each training sample pair, multi-head parallel nonlinear transformation and weighted aggregation are performed based on the embedding vector of the context node to generate a context aggregation representation vector:

[0078] ;

[0079] in: Represents a nonlinear activation function; Indicates the first Each attention head to the context node

[0080] Attention weights, satisfying ; Indicates the first Aggregate representation under size.

[0081] Concatenate the aggregated representations of all heads into a context aggregated representation vector. ;

[0082] The context aggregation representation is used as input to the continuous bag-of-words model for center node prediction, which is used to predict the embedding representation of the corresponding center node in the training sample pair. A loss function is constructed based on the vector difference between the center node embedding representation output by the continuous bag-of-words model with the multi-head context aggregation structure and the true embedding representation of the corresponding center node in the training samples.

[0083] ;

[0084] Wherein, the embedding representation of the center node corresponding to the training sample pair The true embedding representation of the corresponding center node in the training samples ;

[0085] The backpropagation algorithm and gradient descent method are used to iteratively optimize the model parameters so that the context aggregation input can accurately predict the embedding representation of the corresponding center node, and complete the low-dimensional embedding vector learning of the knowledge point nodes of the student knowledge graph state subgraph and the question bank knowledge graph.

[0086] S34. Calculate the question embedding representation of the knowledge point nodes covered by each question in the question bank knowledge graph, and calculate the weighted average embedding representation of the knowledge point nodes related to the question in the student knowledge graph state subgraph. The weighting coefficient is set according to the mastery of the corresponding knowledge point node to obtain the student embedding representation.

[0087] S35. The similarity between the question embedding representation and the student embedding representation is used as a question fit scoring index to measure the degree of matching between the question content and the student's knowledge status.

[0088] S36. Sort all questions according to the question suitability scoring index, and select questions with scores higher than the preset threshold to form a candidate set for test paper compilation.

[0089] In the knowledge graph node modeling process, this invention first performs unsupervised node embedding modeling on the student knowledge graph state subgraph and the question bank knowledge graph based on an improved DeepWalk algorithm. Starting with each node in the graph, a set of structure- and semantically guided random walk sequences is generated according to the set number of walk rounds, walk length, and jump rules, and used as training corpus for the continuous bag-of-words model. During model training, context-center node sample pairs are constructed with a fixed window size. A multi-head attention mechanism is used to perform nonlinear transformation and weighted aggregation on the context embedding vectors to construct a context aggregation representation vector. This representation is then used to predict the embedding vector of the center node. A loss function is constructed by comparing the predicted embedding with the actual embedding vector, and parameter optimization is performed to obtain a more expressive knowledge point node embedding representation. Subsequently, the question embedding representation of the knowledge points covered by each question in the question bank and the weighted average embedding representation of the relevant knowledge points currently mastered by the student are calculated. The similarity between the two is used as a question fit scoring index. Finally, a set of questions with scores higher than a preset threshold is selected for subsequent personalized test paper optimization.

[0090] In this embodiment, S4 specifically includes:

[0091] S41. Set the objective function for test paper compilation. The objective function for test paper compilation should also consider the following constraints: knowledge point coverage, difficulty matching, reasonable distribution of question types, and priority of weak point coverage as the basis for weak point priority scoring.

[0092] S42. Using the candidate test paper set generated in step S37 as the initial population, encode the question composition of each individual test paper and construct a genetic algorithm optimization problem based on question combination.

[0093] S43. Initialize the parameters of the non-dominated sorting genetic algorithm, including population size, maximum number of iterations, crossover probability, and mutation probability;

[0094] S44. In each generation, calculate the fitness vector of each individual in the current population based on the constraints in the objective function:

[0095] ;

[0096] in Indicates the first Individual test paper in the first Fitness scores on each constraint (knowledge point coverage, difficulty matching, reasonableness of question type distribution, and priority of weak point coverage) The dimension of the objective function is denoted as ; the specific representations of each objective function are as follows:

[0097] Knowledge point coverage objective function:

[0098] ;

[0099] in, Indicates the first The set of knowledge points covered by an individual This refers to the set of knowledge points that are expected to be covered.

[0100] Difficulty matching objective function:

[0101] ;

[0102] in, For the first The total number of questions contained in an individual. For the first The difficulty level of the question. The average difficulty value set for the target.

[0103] Objective function for question type structure matching:

[0104] ;

[0105] in, Indicates the first Individual Question Types Quantity, For question type The expected proportion This represents the set of all question types.

[0106] Vulnerability Priority Coverage Objective Function:

[0107] ;

[0108] in, Indicates students' understanding of knowledge points The degree of mastery, For the first The set of knowledge points contained within an individual. This indicates the level of weakness in this knowledge point.

[0109] S45. Perform non-dominated sorting operation, select individuals with better fitness based on the dominance relationship of the Pareto optimal solution set to form the next generation population, and perform crossover and mutation operations to generate new test paper combination individuals.

[0110] Among them, any two individual test papers and The fitness vectors are respectively and If the following conditions are met: Then it is considered that the individual Dominant Individual Based on this, a fast non-dominated sorting algorithm can be used to divide the population into multiple non-dominated levels; for individuals within each level, the crowding distance is calculated for sorting, and individuals with more uniform coverage space are retained first.

[0111] During the population renewal phase, the top-ranked population is selected based on non-dominant rank and crowding distance. Individuals form the next generation population; then, crossover and mutation operations are performed on them. Crossover is constructed based on partial exchanges of question combinations, while mutation is fine-tuned by replacing questions with similar fit in the question bank, thus generating new test paper combinations. The final output is a personalized test paper scheme with a reasonable structure and optimal fit.

[0112] S46. Iteratively execute the genetic evolution process until the maximum number of iterations is met, and finally output a set of personalized test paper schemes that meet the set constraints and are optimally adapted. The genetic evolution process includes: calculating the fitness vector for each test paper individual in the current population, selecting high-quality individuals to form the next generation population based on non-dominated sorting and crowding distance; in each generation, using a roulette wheel selection strategy to select parent individuals from the current population, performing partial exchange operations on question combinations according to the set crossover probability to generate offspring individuals, and performing local question replacement operations on offspring individuals according to the preset mutation probability; the updated individuals recalculate the fitness vector and participate in the next round of population evolution until the maximum number of iterations is reached or the population fitness change is lower than the set threshold, at which point the algorithm is considered to have converged, and the Pareto optimal test paper scheme set in the converged generation population is output.

[0113] The roulette wheel selection strategy refers to constructing an individual selection probability distribution based on the fitness value of each individual in the current population. Let the i-th individual in the current population... The fitness value of an individual is Then the probability of it being selected. Calculate as follows:

[0114] ;

[0115] in, Given the current population size, This represents the sum of the fitness values ​​of all individuals. Based on this probability distribution, by generating... A uniformly random number within the interval is used to select individuals within the corresponding probability interval as the parent individuals for the genetic operation, thus completing the individual selection process.

[0116] This invention optimizes personalized test paper combinations by constructing a candidate set of test papers based on question fit scores and setting a multi-objective test paper combination objective function that includes knowledge point coverage, difficulty matching, question type structure rationality, and priority coverage of weak points. A non-dominated sorting genetic algorithm is used to perform the optimization solution. Specifically, the candidate set of test papers is encoded as an initial population, and parameters such as population size, number of iterations, crossover and mutation probabilities are initialized. In each generation, the fitness vector of individual test paper samples is calculated according to the objective function. Non-dominated sorting and crowding distance are used for individual selection and retention. Simultaneously, a roulette wheel selection strategy is used to select parents, and partial crossover and local mutation of question combinations are performed to generate new individuals. In each round of evolution, the fitness of new individuals is re-evaluated and the population is updated. This iteration continues until the maximum number of iterations is reached or the change in population fitness is less than a set threshold. Finally, a Pareto-optimal personalized test paper scheme with a reasonable structure and high fit with students' knowledge status is output.

[0117] In this embodiment, the genetic evolution process includes: calculating the fitness vector for each individual test paper in the current population; selecting high-quality individuals based on non-dominated sorting and crowding distance to form the next generation population; in each generation, using a roulette wheel selection strategy to select parent individuals from the current population; performing partial exchange operations on question combinations according to a set crossover probability to generate offspring individuals; performing local question replacement operations on offspring individuals according to a preset mutation probability; recalculating the fitness vector of the updated individuals and participating in the next round of population evolution until the maximum number of iterations is reached or the population fitness change is lower than a set threshold, at which point the algorithm is considered to have converged, and the Pareto optimal test paper scheme set in the converged generation population is output.

[0118] In this embodiment, the current answer record consists of subjective question answer text and behavior sequence. The behavior sequence includes answer data content and answer process behavior data, including answer time, question switching order, dwell time, and modification records.

[0119] In this embodiment, S6 specifically includes:

[0120] S61. Collect subjective question answer texts and behavioral sequence information generated by students during the process of completing personalized test papers. The behavioral sequence includes behavioral characteristics such as answer time, question switching order, dwell time, and modification records.

[0121] S62. Input the subjective question answer text into the pre-trained language model BERT and extract semantic feature vectors representing semantic integrity and keyword coverage. The semantic integrity is represented by the set of embedding vectors of the knowledge points corresponding to the standard answer of the subjective question. The average semantic similarity between the student's subjective question answer text and the set of embedding vectors in the embedding space is calculated to measure semantic integrity. The keyword coverage is calculated by the degree of matching between the key terms in the subjective question answer text and the set of keywords in the standard answer.

[0122] S63. Encode the behavior sequence into a time series vector and input it into the time series modeling module based on gated loop units to extract the behavior feature vector representing the behavior strategy characteristics; the behavior feature vector is used to characterize the student's focus, answering rhythm and strategy pattern during the answering process.

[0123] S64. Perform a concatenation operation between the semantic feature vector extracted in step S62 and the behavioral feature vector extracted in step S63 to construct a fused scoring feature vector, which is used as the input to the feedforward neural network scoring model.

[0124] S65. Input the fused scoring feature vector into the feedforward neural network scoring model, and output the scores for the subjective questions on the test paper, including semantic rationality score, content completeness score, and behavioral strategy rationality score.

[0125] S66. The scores of the subjective questions in the test paper are used as feedback information to update the mastery attribute value of the corresponding knowledge point node in the student's knowledge graph state subgraph, thereby enhancing the knowledge graph's ability to dynamically depict the student's current knowledge state.

[0126] In the subjective question scoring process, this invention first collects the subjective question answer text and behavioral sequence information generated by students during the completion of personalized test papers, extracting behavioral features including answer time, question switching order, dwell time, and modification records. Then, the subjective question answer text is input into a pre-trained language model BERT to extract semantic feature vectors representing semantic completeness and keyword coverage. These are then quantitatively evaluated by calculating the average semantic similarity and terminology matching degree using the knowledge point embedding vector set corresponding to the standard answer. Next, the behavioral sequence is encoded into a time series vector and input into a time series modeling module based on gated recurrent units (GRU) to extract behavioral feature vectors representing focus, rhythm, and strategy patterns. The semantic feature vector and behavioral feature vector are then concatenated to construct a fused scoring feature vector, which is input into a feedforward neural network scoring model. This model outputs multi-dimensional scoring results, including semantic rationality, content completeness, and behavioral strategy rationality. Finally, the scoring results are used as feedback information to update the mastery attribute values ​​of relevant knowledge point nodes in the student's knowledge graph state subgraph, achieving dynamic modeling and refined characterization of the student's current knowledge state.

[0127] refer to Figure 2 The knowledge graph-based intelligent test paper generation and scoring system includes the following modules:

[0128] The student knowledge graph construction module is used to collect students' historical answer records and learning behavior data to construct the student knowledge graph state subgraph;

[0129] The question bank knowledge graph construction module is used to construct the question bank knowledge graph;

[0130] The embedding modeling and question fitting scoring module is used to perform joint node embedding training on the student knowledge graph state subgraph and the question bank knowledge graph based on the improved DeepWalk algorithm and the continuous bag-of-words model, generate low-dimensional embedding vectors of knowledge point nodes, and calculate the question fitting score index.

[0131] The personalized test paper optimization module is used to set the objective function for test paper generation. It uses a non-dominated sorting genetic algorithm to combine and optimize the test paper candidate set whose question fit score index is higher than the preset threshold, and outputs the personalized optimal test paper scheme.

[0132] The scoring feature extraction module is used to input the subjective question answer text into the pre-trained language model BERT to extract feature vectors, input the time series vectors into the gated recurrent unit to extract behavioral feature vectors, and construct a fused scoring feature vector.

[0133] The subjective question scoring and feedback module is used to input the fused scoring feature vector into the feedforward neural network scoring model and output the scores of the subjective questions answered on the test paper.

[0134] The intelligent test paper generation and scoring system based on knowledge graphs provided by this invention includes: a question bank knowledge graph construction module, used to construct a question bank knowledge graph covering question types, difficulty, and knowledge point association information; an embedding modeling and question adaptation scoring module, which performs joint node embedding training on the student knowledge graph state subgraph and the question bank knowledge graph based on an improved DeepWalk algorithm and a continuous bag-of-words model, generates low-dimensional embedding vectors, and calculates the question adaptation scoring index; a personalized test paper generation optimization module, used to set the test paper generation objective function, and use a non-dominated sorting genetic algorithm to optimize the combination of question sets with scores higher than a threshold, generating a personalized optimal test paper scheme; a scoring feature extraction module, used to input the subjective question answer text into a pre-trained language model BERT to extract semantic feature vectors, input the behavior sequence into a gated recurrent unit (GRU) to extract behavior feature vectors, and construct a fused scoring feature vector; and a subjective question scoring and feedback module, which inputs the fused scoring feature vector into a feedforward neural network scoring model, outputs a multi-dimensional score for the test paper subjective question answers, and uses it to update the mastery information of relevant knowledge points in the student knowledge graph state subgraph.

[0135] Example 1:

[0136] To verify the feasibility of this invention in practice, it was applied to the mathematics teaching scenario of the ninth grade in a middle school in a certain region, and a two-week experiment was conducted on personalized test paper generation and intelligent scoring of subjective questions. This grade has four classes with a total of 198 students. The school's original teaching and assessment model generally suffers from problems such as fixed test content, low matching degree between question difficulty and students' mastery, strong subjectivity in subjective question scoring, and delayed feedback, making it impossible to accurately model and dynamically intervene in the individual differences in students' learning status.

[0137] During the experimental phase, we first collected 14,823 records of math homework and tests from each student since the start of the semester. These records covered multi-dimensional behavioral data, including student scores, answering time, question switching behavior, and number of modifications. Specific data is shown in Table 1. Based on this data, we constructed an individualized knowledge graph state subgraph for each student. Nodes represent knowledge points and include attributes such as student mastery, number of attempts, and accuracy rate. Edges represent dependencies between knowledge points. Simultaneously, we constructed a question bank knowledge graph based on existing question bank resources, labeling each question with information such as the knowledge points it covers, its question type, and difficulty level.

[0138] Table 1 Comparison of the Implementation Effects of Personalized Student Test Paper Generation and Grading System

[0139] Student ID Average mastery of original knowledge points Average score of personalized test papers Difference between system rating and teacher rating Average response time (minutes) Time (seconds) for scoring subjective questions Knowledge graph mastery improved by S001 0.61 82.5 2.8 36.7 0.65 +12.3% S042 0.45 76.4 3.1 39.4 0.58 +15.9% S087 0.72 88.1 4.2 34.6 0.61 +8.6% S103 0.38 70.9 3.6 42.2 0.63 +18.2% S159 0.67 85.2 2.7 35.1 0.59 +9.7% S191 0.50 79.6 3.0 38.9 0.60 +13.5%

[0140] An improved DeepWalk algorithm and a continuous bag-of-words model with multi-head context aggregation are used to jointly embed and model the question bank knowledge graph and the student knowledge graph state subgraph, completing the low-dimensional representation learning of knowledge point nodes. Based on this, the question fit score is calculated by combining the knowledge point vectors covered by each question with the weighted embedding vectors of the corresponding knowledge points in the student state graph, and questions with a fit score higher than 0.75 are selected as candidate test papers. Furthermore, a test paper assembly objective function is set, including knowledge point coverage, difficulty matching, reasonable question type distribution, and priority coverage of weak points. A non-dominated sorting genetic algorithm is used to optimize the test paper assembly, generating personalized test papers for each student.

[0141] After students complete their personalized test papers, the subjective question scoring process is automated. The students' subjective question answers are input into a pre-trained language model BERT, which extracts semantic completeness and keyword coverage features. Simultaneously, behavioral sequences generated during the student's answering process, such as answer duration, switching order, and modification frequency, are converted into time series vectors and input into a GRU model to extract behavioral strategy features. These two types of features are concatenated and input into a feedforward neural network scoring model, ultimately outputting a three-dimensional score based on semantic reasonableness, content completeness, and behavioral strategy reasonableness. Comparative analysis of the system's scoring results with those of two experienced teachers revealed an average deviation of 3.4 points and a Pearson correlation coefficient of 0.91, indicating a high degree of consistency and reliability in the scoring results.

[0142] During the experimental period, the control group used the traditional, standardized test paper format, while the experimental group used the method of this invention. Statistical results showed that the average score of students in the experimental group improved by 7.6 points, and the standard deviation decreased by 1.8, indicating that personalized test paper format better suited students' knowledge levels and achieved precise teaching. Simultaneously, the system's scoring time averaged 0.6 seconds per question, significantly lower than manual grading time, and it could update students' knowledge graph status in real time, enabling continuous teaching intervention.

[0143] As can be seen from the above data, the method of the present invention has significant advantages in terms of personalized test paper accuracy, scoring efficiency and feedback effect. It can effectively improve the intelligence level and pertinence of the teaching process, and provides a powerful supplement to the current inefficient human-based assessment process. It has good practical value and promotion potential.

[0144] 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.

Claims

1. A knowledge graph-based intelligent test paper generation and scoring method, characterized in that, Includes the following steps: S1. Collect students' historical answer records and learning behavior data to construct a state subgraph of the student knowledge graph; S2. Construct a knowledge graph for the question bank; S3. A node embedding modeling method based on the improved DeepWalk algorithm, which uses a continuous bag-of-words model with multi-head context aggregation to jointly train and learn low-dimensional representations of knowledge point nodes, and constructs a question fit scoring index. S4. Set the objective function for test paper generation, use the non-dominated sorting genetic algorithm to combine and optimize the candidate test paper generation set, and output the optimal test paper scheme set; S5. Collect the student's current answer record when completing the test paper; S6. Based on the scoring feature extraction module and the subjective question scoring and feedback module, extract semantic feature vectors and behavioral feature vectors, construct a fused scoring feature vector, and generate the score for the subjective questions in the test paper.

2. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, Each node in the student knowledge graph state subgraph represents a knowledge point. The node contains attribute information such as the student's mastery level, the number of times they answered questions, and their accuracy rate. The edges represent the relationships between knowledge points.

3. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, The question bank knowledge graph connects each question with the knowledge points covered by the question bank and records the question type, difficulty level, and the knowledge unit to which it belongs.

4. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, S3 specifically includes: S31. Unsupervised node embedding modeling is performed on the student knowledge graph state subgraph and the question bank knowledge graph respectively. An improved DeepWalk algorithm is used, taking each node in the student knowledge graph state subgraph and the question bank knowledge graph as the starting point, and performing a random walk operation for a preset number of rounds. According to the set walk length and jump rules, a set of random walk sequences composed of node numbers is sampled and generated. The random walk sequence is used to characterize the structural adjacency relationship between nodes in the graph. The jump rules are probabilistically modeled based on the structural distance and semantic weight between the current node and its neighboring nodes. S32. Using the random walk sequence as training corpus, a continuous bag-of-words model is used to construct training samples centered on each central node in the random walk sequence. A fixed window size is set, and a preset number of context nodes are selected from both sides of the central node. The embedding vectors corresponding to each group of context nodes are aggregated and encoded as input to the continuous bag-of-words model. The central node number corresponding to the context node is predicted, and a training sample pair set with the context node set as input and the central node as output is constructed. S33. During the training process, for each training sample pair, a multi-head parallel nonlinear transformation and weighted aggregation are performed based on the embedding vector of the context node to generate a context aggregation representation vector; the context aggregation representation vector is used as the input for center node prediction in the continuous bag-of-words model to predict the embedding representation of the corresponding center node in the training sample pair. S34. Based on the vector difference between the embedding representation of the central node and the true embedding representation of the corresponding central node in the training samples, construct a loss function, and use the backpropagation algorithm and gradient descent method to iteratively optimize the parameters of the continuous bag-of-words model to complete the low-dimensional embedding vector learning of the knowledge point nodes of the student knowledge graph state subgraph and the question bank knowledge graph. S35. Calculate the question embedding representation of the knowledge point nodes covered by each question in the question bank knowledge graph, and calculate the weighted average embedding representation of the knowledge point nodes related to the question in the student knowledge graph state subgraph. The weighting coefficient is set according to the mastery of the corresponding knowledge point node to obtain the student embedding representation. S36. Use the similarity between the question embedding representation and the student embedding representation as the question fit scoring index; S37. Sort all questions according to the question suitability scoring index, and select questions with scores higher than the preset threshold to form a candidate set for test paper compilation.

5. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, The specific steps of the improved DeepWalk algorithm are as follows: Starting from each knowledge point node in the student knowledge graph state subgraph and the question bank knowledge graph, a set of random walk sequences reflecting structural adjacency relationships is generated by sampling according to the set number of random walk rounds, walk length and jump rules. The generated random walk sequences are used as the training corpus for the continuous bag-of-words model, and training sample pairs are constructed with a fixed window size. Perform multi-head parallel nonlinear transformation and weighted aggregation on the embedding vector of the context node to generate a context aggregation representation vector; The loss function is constructed by calculating the vector difference between the predicted embedding and the true embedding. Backpropagation and gradient descent are used to optimize the parameters of the continuous bag-of-words model to obtain low-dimensional node embedding vectors with stronger semantic expressive power. Question embedding representation and student embedding representation are then constructed.

6. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, S4 specifically includes: S41. Set the objective function for test paper compilation. The objective function for test paper compilation should also consider the following constraints: knowledge point coverage, difficulty matching, reasonable distribution of question types, and priority of weak point coverage as the basis for weak point priority scoring. S42. Using the candidate set of test papers generated in step S37 as the initial population, encode the composition of questions for each individual test paper. S43. Initialize the parameters of the non-dominated sorting genetic algorithm, including population size, maximum number of iterations, crossover probability, and mutation probability; S44. In each generation, calculate the fitness vector of each individual in the current population based on the constraints in the objective function of the test paper. S45. Perform non-dominated sorting operation, select individuals with better fitness based on the dominance relationship of the Pareto optimal solution set to form the next generation population, and perform crossover and mutation operations to generate new test paper combination individuals. S46. Iteratively execute the genetic evolution process until the maximum number of iterations is met, and finally output a set of optimal test paper schemes that meet the set constraints and are optimally adapted.

7. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, The genetic evolution process includes: calculating the fitness vector for each individual test paper in the current population; selecting high-quality individuals to form the next generation population based on non-dominated sorting and crowding distance; in each generation, using a roulette wheel selection strategy to select parent individuals from the current population; performing partial exchange operations on question combinations according to a set crossover probability to generate offspring individuals; performing local question replacement operations on offspring individuals according to a preset mutation probability; recalculating the fitness vector of the updated individuals and participating in the next round of population evolution until the maximum number of iterations is reached or the population fitness change is lower than a set threshold, at which point the algorithm is considered to have converged, and the set of optimal test paper schemes in the converged generation population is output.

8. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, The current answer record consists of subjective question answer text and behavior sequence. The behavior sequence includes answer data content and answer process behavior data, including answer time, question switching order, dwell time, and modification records.

9. The knowledge graph-based intelligent test paper generation and scoring method according to claim 1, characterized in that, S6 specifically includes: S61. Input the subjective question answer text into the pre-trained language model BERT, and extract semantic feature vectors representing semantic integrity and keyword coverage; the semantic integrity is represented by the embedding vector set of the knowledge points corresponding to the standard answer of the subjective question, and the average semantic similarity between the student's subjective question answer text and the embedding vector set in the embedding space is calculated; the keyword coverage is obtained by calculating the degree of matching between the key terms in the subjective question answer text and the keyword set of the standard answer. S62. Encode the behavioral sequence into a time series vector and input it into time series modeling based on gated cyclic units to extract behavioral feature vectors; S63. Perform a concatenation operation on the semantic feature vector and the behavioral feature vector to construct a fused scoring feature vector; S64. Input the fused scoring feature vector into the feedforward neural network, and output the scores for the subjective questions on the test paper, including semantic reasonableness score, content completeness score, and behavioral strategy reasonableness score.

10. A knowledge graph-based intelligent test paper generation and scoring system, implementing the knowledge graph-based intelligent test paper generation and scoring method according to any one of claims 1 to 9, characterized in that, Includes the following modules: The student knowledge graph construction module is used to collect students' historical answer records and learning behavior data to construct the student knowledge graph state subgraph; The question bank knowledge graph construction module is used to construct the question bank knowledge graph; The embedding modeling and question fitting scoring module is used to perform joint node embedding training on the student knowledge graph state subgraph and the question bank knowledge graph based on the improved DeepWalk algorithm and the continuous bag-of-words model, generate low-dimensional embedding vectors of knowledge point nodes, and calculate the question fitting score index. The personalized test paper optimization module is used to set the objective function for test paper generation. It uses a non-dominated sorting genetic algorithm to combine and optimize the test paper candidate set whose question fit score index is higher than the preset threshold, and outputs the personalized optimal test paper scheme. The scoring feature extraction module is used to input the subjective question answer text into the pre-trained language model BERT to extract feature vectors, input the time series vectors into the gated recurrent unit to extract behavioral feature vectors, and construct a fused scoring feature vector. The subjective question scoring and feedback module is used to input the fused scoring feature vector into the feedforward neural network and output the score of the subjective questions in the test paper.

Images