Dynamic intelligent test paper compiling method, system and device based on multiple item types and difficulty gradient

Abstract

This application discloses a dynamic intelligent test paper generation method, system, and device based on multiple question types and difficulty gradients, relating to the field of online education technology. It first cleans and deduplicates the question bank, constructs a question profile containing multi-dimensional information, and labels the difficulty. Then, it receives the test paper generation configuration, generates hard constraints, divides the test paper into multiple difficulty gradient layers, and sets target difficulty and continuity constraints. Subsequently, it filters candidate sets according to question type and gradient layer, and generates initial test papers in layers and segments. It iterative replacement and exchange are used to repair constraint conflicts, and the final version is optimized under multi-objective criteria. Finally, it dynamically updates ability parameters and adjusts subsequent target difficulty based on answer data, outputting the test paper and verification indicators. This method can adjust subsequent test paper generation strategies in real time based on the examinee's answer performance, ensuring the quality of the question bank, ensuring that the test paper meets hard constraints and that the difficulty progresses smoothly, improving test paper quality and assessment effectiveness, enhancing the adaptability and accuracy of personalized testing, and is suitable for various personalized testing scenarios.

Description

Technical Field

[0001] This application relates to the field of online education technology, specifically to a dynamic intelligent test paper generation method, system, and device based on multiple question types and difficulty levels. Background Technology

[0002] With the rapid development of smart education, intelligent test paper generation technology serves as a core support for online examinations, personalized assessments, and skills evaluations. Its scientific rigor and adaptability directly impact assessment effectiveness and the accuracy of learning diagnosis. Currently, intelligent test paper generation has gradually replaced traditional manual test paper generation, becoming one of the key technologies in the field of educational informatization. However, it still faces many challenges in adaptability and accuracy in practical applications. Existing mainstream test paper generation methods primarily focus on "meeting fixed parameter constraints," that is, selecting test question combinations based on preset parameters such as question type ratios, total scores, and difficulty thresholds. While this improves test paper generation efficiency, it struggles to match the personalized assessment needs and differentiated characteristics of various question types in different scenarios.

[0003] Existing intelligent test paper generation technologies have significant core flaws, specifically in four dimensions: First, the difficulty evaluation dimension is too simplistic. Most methods rely solely on "answer accuracy" as the only difficulty assessment indicator, ignoring both the specific differences in question types (such as the quality of distractors in objective questions, score fluctuations in subjective questions, and the complexity of operational steps in practical questions, all of which affect the perceived difficulty) and individual user differences (such as a user's knowledge weaknesses, answering speed, types of past mistakes, and learning habits, all of which can alter their perception of the difficulty of the same question). Second, the difficulty gradient is too coarse, generally adopting a three-level division model of "easy, medium, and difficult," which fails to accurately cover the needs of users at all levels, from those with weak foundations to those with advanced abilities. This results in basic users feeling frustrated due to overly difficult questions, while high-achieving users are unable to demonstrate their true abilities due to overly easy questions, significantly reducing the differentiation and diagnostic value of the assessment. Third, the test paper assembly process is static. Once the initial parameters are determined, they remain fixed and cannot respond to the user's answering status in real time. For example, when a user answers difficult questions correctly consecutively or frequently answers basic questions incorrectly, the difficulty of the test paper cannot be dynamically adjusted, ultimately leading to insufficient assessment accuracy and difficulty in accurately depicting the user's true knowledge level. Fourth, the adaptability to multiple question types is lacking. The difficulty transmission patterns of different question types and the overall difficulty balance of the test paper are not considered. For example, simply combining different question types such as multiple choice, fill-in-the-blank, and problem-solving questions in proportion ignores the cumulative effect of difficulty between different question types, which may cause the overall difficulty of the test paper to deviate from the expected target.

[0004] The aforementioned shortcomings make it difficult for existing intelligent test paper generation technologies to meet the core needs of current personalized educational assessments and precise skills evaluations, resulting in insufficient scientific rigor and poor user adaptability of assessment results. Therefore, how to integrate features of multiple question types, achieve refined difficulty gradient division, and dynamically adapt to the user's real-time status to improve the scientific rigor and accuracy of test paper generation is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution:

[0006] In a first aspect, embodiments of this application provide a dynamic intelligent test paper generation method based on multiple question types and difficulty gradients, including:

[0007] The question bank is cleaned and deduplicated to construct a question profile containing multi-dimensional information about the questions, and then the profiles are integrated and labeled with difficulty.

[0008] Receive test paper configuration and generate hard constraints for test paper generation, divide the test paper into multiple difficulty gradient layers, and set target difficulty and continuity constraints from easy to difficult.

[0009] Calculate candidate scores and filter candidate sets according to question type and gradient level, and generate initial test papers in layers and segments;

[0010] Constraint conflicts were resolved through iterative replacement and exchange, and the final draft of the test paper was optimized under multi-objective criteria.

[0011] In dynamic scenarios, the ability parameters are updated based on the answer data, and the difficulty of subsequent targets is adjusted to output the test paper and verification indicators.

[0012] In one possible implementation, the step of cleaning and deduplicating the questions in the question bank to construct a question profile containing multi-dimensional information about the questions and then fusing and labeling the difficulty includes:

[0013] The question bank data is cleaned, deduplicated, and version-merged. A question profile is then created for each question, including question type, knowledge point coverage, difficulty, discrimination, response time, exposure risk, and quality score.

[0014]

[0015] in: For question type identification, For knowledge point coverage vectors, For difficulty parameters, For discrimination parameter, To estimate the response time, In order to expose the risk factor, Rate the quality;

[0016] The difficulty parameter is obtained by combining the annotation difficulty and the statistical difficulty. middle:

[0017]

[0018] in: The difficulty level is labeled by humans or experts. To reduce the difficulty of statistics, The fusion coefficient controls the level of trust in the tags.

[0019] In one possible implementation, the step of receiving the test paper configuration and generating hard constraints for test paper generation, dividing the test paper into multiple difficulty gradient layers, and setting target difficulty and continuity constraints from easy to difficult, includes:

[0020] Receive test paper configuration and generate a set of hard constraints, which includes at least question type quota constraints, total time constraints and knowledge point coverage lower limit constraints.

[0021] The test paper is divided into multiple difficulty gradients. Target difficulty is set for each gradient, and gradient continuity constraints are set to ensure that the difficulty of each gradient increases smoothly from easy to difficult.

[0022] In one possible implementation, the receiving test paper configuration and generating a set of hard constraints, the set of hard constraints including at least question type quota constraints, total time constraints, and knowledge point coverage lower limit constraints, including:

[0023] For the question Set binary variable If selected for the exam paper ,otherwise ;

[0024] Input test paper configuration: Question type set Number of questions per question type The total score, total duration, and target weights for each knowledge point are used to generate a constraint set, which includes:

[0025] Question type and quantity constraints: ;

[0026] Total duration constraints: ;

[0027] Knowledge point coverage constraints: ;

[0028] in: To allow the maximum response time, To cover weights, represent the question Does it cover the knowledge points? , For knowledge points Minimum coverage.

[0029] In one possible implementation, dividing the test paper into multiple difficulty gradient layers, setting target difficulty for each gradient layer, and setting gradient continuity constraints to ensure that the difficulty of each gradient layer smoothly increases from easy to difficult includes:

[0030] Divide the exam paper into Each difficulty gradient layer generates a target difficulty center for that layer. ;

[0031]

[0032] Set gradient continuity constraints to ensure that the actual average difficulty of each gradient layer is... satisfy:

[0033] ;

[0034] in: As a difficulty gradient layer, The number of gradient layers. For the first The center of difficulty of the layer target, The linear gradient step size is... For the first The actual average difficulty of each level. The allowable deviation is defined as such, with smaller allowable deviations indicating stricter requirements. This is a set of questions assigned to a section according to the order of the exam papers. It represents the minimum gradient increment between adjacent segments.

[0035] In one possible implementation, the step of calculating candidate scores and filtering candidate sets according to question type and gradient layer, and generating the initial test paper in a hierarchical and segmented manner includes:

[0036] For each question type and each grade level, candidate scores are calculated. These candidate scores comprehensively consider the degree of match between the question difficulty and the target difficulty, discrimination, quality score, exposure risk, and penalty for duplication with historical test papers.

[0037]

[0038] in: For the title In difficulty gradient layer Candidate scores, For difficulty matching function, , Match weight coefficients to difficulty level. The weighting coefficients are the discrimination parameter. This is the weighting coefficient for the quality score. The exposure risk coefficient is a weighting factor. For repeated penalty items, Related to the similarity of historical exam papers or neighboring exam papers, The weighting coefficient for the repeated penalty term. This represents the difficulty tolerance scale within the difficulty gradient layer. The smaller the size, the more emphasis is placed on aligning with the difficulty of the target;

[0039] Candidate sets are obtained by filtering candidates based on their scores, and initial test papers are generated by performing tiered and segmented question selection on the candidate sets. The tiered and segmented question selection includes at least the following: prioritizing the coverage of scarce knowledge points and then supplementing the question type quota and gradient layer quota according to the candidate scores.

[0040] In one possible implementation, the step of repairing constraint conflicts through iterative replacement and optimization of the final exam paper under multi-objective criteria includes:

[0041] The initial test paper is subjected to constraint violation detection, and the test paper is made to satisfy the set of hard constraints and the gradient continuity constraint through an iterative repair method of question replacement and exchange.

[0042] The test paper was finalized based on multi-objective optimization criteria, which include at least reducing the difficulty gradient deviation, reducing the knowledge point coverage gap, reducing the timeout penalty, reducing the accumulation of exposure risk, and increasing the discrimination benefit.

[0043] In one possible implementation, the finalization of the test paper based on multi-objective optimization criteria includes at least reducing difficulty gradient deviation, reducing knowledge point coverage gaps, reducing timeout penalties, reducing exposure risk accumulation, and improving discrimination gains, including:

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] in: For topic selection vectors, For gradient deviation loss, For knowledge gaps and losses, As a penalty for exceeding the time limit, To accumulate risk, For differentiation revenue, and They are respectively , , , and The target weight.

[0051] In one possible implementation, the step of updating ability parameters based on answer data and adjusting subsequent target difficulty in a dynamic scenario, and outputting the test paper and verification indicators, includes:

[0052] Obtain user answer data and update user ability parameters. Based on the updated user ability parameters, adjust the target difficulty range or target difficulty sequence of subsequent test papers to generate the next test paper or the next module of test questions.

[0053] Output the final test paper and generate consistency verification results. The consistency verification results include at least the following: question type quota satisfaction, difficulty gradient deviation, knowledge point coverage satisfaction, total time satisfaction, and exposure risk and repetition rate indicators.

[0054] In one possible implementation, the step of acquiring user response data and updating user ability parameters, and adjusting the target difficulty range or target difficulty sequence of subsequent test papers based on the updated user ability parameters to generate the next test paper or the next module of questions, includes:

[0055] by Indicating the candidate's level, incremental updates are made:

[0056]

[0057] in: For the first Wheel capacity estimation, This is the step size coefficient. This refers to the number of questions answered in this round. , indicating the first Is the answer to the question correct? To predict the probability of being correct, , For the first Question discrimination For the first Difficulty parameters;

[0058] Then set the target difficulty range for the next stage to , , , To broaden the scope around the difficulty of the ability.

[0059] Secondly, embodiments of this application provide a dynamic intelligent test paper generation system based on multiple question types and difficulty levels, including:

[0060] The question bank data processing module is used to clean and deduplicat the questions in the question bank, construct a question profile containing multi-dimensional information about the questions, and then integrate and label the difficulty.

[0061] The test paper configuration and constraint generation module is used to receive test paper configuration and generate hard constraints for test paper, divide the test paper into multiple difficulty gradient layers, and set the target difficulty and the continuity constraint from easy to difficult.

[0062] The candidate set generation and initial solution construction module is used to calculate candidate scores and filter candidate sets according to question type and gradient layer, and generate the initial test paper in layers and segments.

[0063] The final draft optimization module is used to fix constraint conflicts and optimize the final draft of the test paper under multi-objective criteria through iterative replacement and exchange.

[0064] The dynamic update and test paper output module is used to update ability parameters and adjust subsequent target difficulty based on answer data in dynamic scenarios, and output test papers and verification indicators.

[0065] Thirdly, embodiments of this application provide a dynamic intelligent test paper generation device based on multiple question types and difficulty levels, including:

[0066] processor;

[0067] Memory;

[0068] And a computer program, wherein the computer program is stored in the memory, the computer program including instructions that, when executed by the processor, cause the device to perform a method implementing any possible implementation of the first aspect to implement the intelligent component.

[0069] In this embodiment, the method ensures the high quality and accuracy of the question bank data through cleaning, deduplication, and question profiling, providing a solid foundation for subsequent test paper assembly. Secondly, through multi-level and multi-angle constraints and target modeling, the method ensures that the generated test papers not only meet the hard requirements of question types, knowledge point coverage, and time limits, but also achieve a smooth progression of difficulty, thus avoiding sudden jumps in difficulty. Thirdly, through candidate set generation and initial solution construction steps, the method can flexibly adjust the selection and allocation of questions according to the test paper objectives and question characteristics, ensuring that the quality and difficulty of the test paper are matched. Fourthly, the iterative mechanism of constraint repair and quality optimization finalization ensures that the generated test papers meet all constraints while maximizing the assessment effect, avoiding biases or overly simple test papers. Finally, the method supports dynamic feedback updates, enabling real-time adjustments to subsequent test paper assembly strategies based on student responses, improving the adaptability and accuracy of personalized testing. Attached Figure Description

[0070] Figure 1A flowchart illustrating a dynamic intelligent test paper generation method based on multiple question types and difficulty gradients provided in this application embodiment;

[0071] Figure 2 This is a schematic diagram of the title image provided in the embodiments of this application;

[0072] Figure 3 This is a schematic diagram illustrating dynamic capability updates provided in an embodiment of this application.

[0073] Figure 4 A schematic diagram of a final test paper for a high school mathematics stage test provided in an embodiment of this application;

[0074] Figure 5 Provided for the embodiments of this application Figure 4 Analysis of the structure of a math exam paper;

[0075] Figure 6 A schematic diagram of a dynamic intelligent test paper generation system based on multiple question types and difficulty levels, provided for an embodiment of this application;

[0076] Figure 7 This is a schematic diagram of a dynamic intelligent test paper generation device based on multiple question types and difficulty levels, provided as an embodiment of this application. Detailed Implementation

[0077] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0078] See Figure 1 This embodiment provides a dynamic intelligent test paper generation method based on multiple question types and difficulty levels, including:

[0079] S101 cleans and deduplicates the questions in the question bank, constructs a question profile containing multi-dimensional information about the questions, and integrates and labels the difficulty.

[0080] As the core foundation for test paper compilation, the question bank's data quality directly determines the scientific validity and rationality of the test paper. Therefore, the first step is to complete comprehensive preprocessing and multi-dimensional indicator calibration. During the data cleaning phase, it is necessary to systematically handle invalid questions (such as those with missing stems, contradictory conditions, or incorrect answers) and duplicate questions (including "near-duplicate questions" that are completely identical to the original question but have slightly different wording). Different versions of the same question (such as revised versions from different years or with different formatting) should be merged to ensure that each question in the question bank is unique and valid. Simultaneously, field standardization operations are required, including unifying question type codes (e.g., unifying "multiple choice" and "single choice" with code "1", "fill-in-the-blank" with code "2", etc.), standardizing the knowledge point classification system (divided according to the subject syllabus level, such as "mathematics - algebra - quadratic equations"), and unifying the answering time unit (e.g., all in minutes), laying the foundation for subsequent test paper compilation constraint matching. After completing data preprocessing, core indicators need to be calculated or updated for each question to construct a complete question profile.

[0081] In this embodiment, the question bank data is first cleaned, deduplicated, and version-merged. A question profile is then constructed for each question, including question type, knowledge point coverage, difficulty, discrimination, response time, exposure risk, and quality score. Figure 2 As shown:

[0082]

[0083] in: For question type identification, For knowledge point coverage vectors, For difficulty parameters, For discrimination parameter, To estimate the response time, In order to expose the risk factor, Rate the quality.

[0084] This is a multi-dimensional vector describing the related knowledge points and the strength of those connections. In this embodiment, subject matter experts can manually annotate the knowledge points corresponding to the questions based on curriculum standards and examination syllabi, and can also mark the strength of the connections. Core knowledge points are marked as 1, and secondary knowledge points as 0.5, forming a structured vector. Analyzing student answer data, if the group of students who make mistakes on a certain question highly overlaps with the group of students who are weak on a certain knowledge point, then... The association strength of the markers in the data is updated.

[0085] This is a core indicator for measuring the difficulty level of a question, and can be obtained using student pass rates and expert experience estimates. The student pass rate is the ratio of the number of correct answers to the total number of participants; the closer the value is to 1, the easier the question, and the closer it is to 0, the more difficult it is. When a new question appears and there is no large-scale data available, teaching and research personnel estimate the difficulty based on their teaching experience and exam objectives, assigning initial values ​​of 0.2, 0.5, and 0.8 to three levels: easy, medium, and difficult.

[0086] To measure the effectiveness of the questions in differentiating candidates of different ability levels, for example, candidates are ranked by total score, the top 27% are taken as the high group and the bottom 27% as the low group, and the pass rate of the high group minus the pass rate of the low group is used as the discrimination parameter.

[0087] To estimate the approximate time required to complete a question, the following methods can be used: 1. Statistical method: Record the time taken by test-takers in real exams and practice sessions, calculate the average of a large sample as the estimated time, and then refine the estimate by classifying and statistically analyzing question types. 2. Feature fitting model: Extract question features, train a time prediction model using regression models, random forests, etc., and input new question features to output the estimated time. 3. Expert experience prediction: Curriculum and research staff estimate the time based on question complexity and number of steps, and then iteratively optimize the estimate using real-world data.

[0088] To assess the risk of overuse and leakage of exam questions, the following methods can be used: Frequency of use statistics: Statistics on the number of times questions are used in exams and practice sessions, and the target audience. Higher frequency and wider coverage indicate higher risk, calculated using time decay normalization. Scenario association analysis: Questions appearing in easily leaked scenarios such as public exam papers and free platforms are marked as high-risk; questions used only in internal mock exams or paid systems have lower risk, calculated using scenario weights. Content similarity detection: Algorithms such as cosine similarity are used to detect the similarity between the question and leaked questions; higher similarity indicates higher exposure risk.

[0089] To comprehensively evaluate the scientific validity, effectiveness, and suitability of the questions, a multi-dimensional weighted scoring method can be adopted: establish scoring dimensions, with each dimension scored by experts, and then weighted according to their respective weights to obtain the total score. Analyze abnormal responses; if abnormal responses are found, the score will be reduced, while consistently stable performance will be increased. Finally, combine manual review and user feedback to dynamically adjust the scoring. New questions entering the database will initially receive an initial score, and subsequent scores will be updated iteratively based on performance.

[0090] The difficulty parameter is obtained by combining the annotation difficulty and the statistical difficulty. middle:

[0091]

[0092] in: The difficulty level is labeled by humans or experts. To reduce the difficulty of statistics, The fusion coefficient controls the level of trust in the tags.

[0093] In this embodiment, Mapping through historical accuracy For extreme accuracy rates (such as...) or Smoothing processing is required first (e.g.) The time was corrected to 0.01. The time was adjusted to 0.99 to avoid meaningless logarithmic calculations. The discrimination index reflects the ability of a question to differentiate between candidates of different levels (e.g., questions that high-scoring candidates must answer correctly and low-scoring candidates must get wrong have high discrimination). The quality score can be comprehensively scored by combining dimensions such as the stability of the question's accuracy rate, the clarity of the question stem, and the uniqueness of the answer. The exposure risk is assessed based on factors such as the recent frequency of use of the question, whether it has been widely disseminated, and whether there is a risk of leakage. The final question profile will serve as the core basis for subsequent screening and test paper compilation.

[0094] S102 receives the test paper configuration and generates hard constraints for test paper assembly, divides the test paper into multiple difficulty gradient layers, and sets the target difficulty and the continuity constraint from easy to difficult.

[0095] The core of test paper creation is to meet specific exam requirements. Therefore, it is necessary to first accurately analyze the test paper creation objectives and transform abstract requirements into a set of quantifiable and executable hard constraints. The sources of test paper configurations are diverse. They may be unit test requirements proposed by teachers based on the teaching progress (e.g., focusing on knowledge points in a certain chapter, with multiple-choice and fill-in-the-blank questions as the main question types), or requirements for entrance examinations based on the exam syllabus (e.g., covering all knowledge points in the entire book, including a certain proportion of comprehensive questions), or personalized tests tailored to students' learning situations (e.g., intensive training for weak knowledge points). During the analysis process, it is necessary to first clarify the priority of core requirements (e.g., knowledge point coverage is a rigid requirement, and the total time can fluctuate by ±5%), and then break down the requirements into specific constraint parameters.

[0096] In this embodiment, the test paper configuration is received and a set of hard constraints is generated. The set of hard constraints includes at least question type quota constraints, total time constraints, and knowledge point coverage lower limit constraints.

[0097] Specifically, regarding the question Set binary variable If selected for the exam paper ,otherwise .

[0098] Input test paper configuration: Question type set Number of questions per question type The total score, total duration, and target weights for each knowledge point are used to generate a constraint set, which includes:

[0099] Question type and quantity constraints: ;

[0100] Total duration constraints: ;

[0101] Knowledge point coverage constraints: ;

[0102] in: To allow the maximum response time, To cover weights, represent the question Does it cover the knowledge points? , For knowledge points Minimum coverage.

[0103] In this embodiment, the target number of questions for each question type needs to be clearly defined. The time allotted for this question type should be determined by combining the total exam time with the average time allotted for that question type (e.g., 2 minutes per multiple-choice question, 3 minutes per fill-in-the-blank question; if the total time is 60 minutes, multiple-choice questions can be set at a certain time). =15, Fill in the blanks =10), while ensuring the question type set Question type codes in the question bank A perfect match is ensured to avoid situations where no matching question is available. The estimated response time for question i is... It needs to be based on historical answer data statistics (such as taking the median answer time of all candidates). The time limit needs to be determined according to the exam organization schedule (e.g., 40 minutes for classroom tests and 90 minutes for midterm tests). The constraint is set to "≤" to reserve a certain amount of time leeway and prevent candidates from being unable to complete the exam due to taking too long on individual questions. If the variable is 0 or 1, it indicates whether the question covers knowledge point m. If it is a weight, it indicates the depth of coverage (e.g., the weight of core knowledge points is 2, and the weight of secondary knowledge points is 1). The setting should be based on the proportion of each knowledge point in the exam syllabus (e.g., core knowledge points). ≥2, secondary knowledge points ≥1), to ensure that the test paper can comprehensively assess the candidates' knowledge.

[0104] The test paper is divided into multiple difficulty gradients. Target difficulty is set for each gradient, and gradient continuity constraints are set to ensure that the difficulty of each gradient increases smoothly from easy to difficult.

[0105] To ensure the test paper effectively differentiates students and adapts to their responses, a scientific difficulty gradient needs to be constructed to avoid overly concentrated or abruptly increased difficulty levels. First, the number of gradient levels, L, needs to be determined, taking into account the type of exam and the distribution of student abilities: for routine unit tests, L=3 (easy, medium, difficult) is sufficient; for selective exams (such as Olympiad math or entrance exams), L=5 (easy, relatively easy, medium, relatively difficult, difficult) can more accurately differentiate between students of different levels. If student abilities follow a normal distribution, L can be moderate; if there is significant polarization, the number of intermediate gradient levels can be appropriately increased.

[0106] In this embodiment, the test paper is divided into Each difficulty gradient layer generates a target difficulty center for that layer. ;

[0107]

[0108] in: As a difficulty gradient layer, The number of gradient layers. For the first The center of difficulty of the layer target, The step size is the linear gradient step size.

[0109] The target difficulty level for the first level (easiest level) needs to be set according to the candidates' basic level (e.g., elementary school unit tests). =0.3, College Entrance Examination =0.2), For linear gradient step size (e.g., when L=3, Δ=0.2, then...) =0.5, =0.7), suitable for regular exams where a steady pace of difficulty increase is required.

[0110] To ensure the effectiveness of the gradient, the constraints of gradient overlap and adjacent gradient smoothing must be satisfied. In this embodiment, a gradient continuity constraint is set to make the actual average difficulty of each gradient layer... satisfy:

[0111] ;

[0112] in: For the first The actual average difficulty of each level. The allowable deviation is defined as such, with smaller allowable deviations indicating stricter requirements. This is a set of questions assigned to a section according to the order of the exam papers. It represents the minimum gradient increment between adjacent segments.

[0113] It should be pointed out that, Needs to be aligned with the target difficulty center The deviation does not exceed Easy section The limit can be appropriately relaxed (e.g., 0.1), for difficult sections. The difficulty level needs to be strictly controlled (e.g., 0.05) to avoid making easy questions too difficult or difficult questions too easy. The increment in difficulty between adjacent sections should be set according to the test takers' acceptance level (e.g., ...). =0.1), ensuring that the difficulty gradually increases without any sudden changes, reducing the psychological pressure on test takers.

[0114] S103: Calculate candidate scores and select candidate sets according to question type and gradient layer, and generate the initial test paper in layers and segments.

[0115] Because question banks are typically large, directly creating questions from the entire question bank would result in extremely high solution complexity. Therefore, it is necessary to divide the question bank by question type and difficulty level to generate targeted candidate sets. This reduces the solution complexity while ensuring the matching degree between candidate questions and target requirements. The division of question types and difficulty levels needs to consider the characteristics of the question types and the difficulty gradient: for example, multiple-choice questions can cover all L-level gradients (easy, medium, and difficult), while comprehensive problem-solving questions are usually only distributed in the medium and difficult levels; each Each corresponds to an independent candidate set, such as "multiple choice questions - easy questions" or "fill in the blank questions - difficult questions".

[0116] In this embodiment, candidate scores are calculated for each question type and each gradient level. The candidate scores comprehensively consider the degree of matching between the question difficulty and the target difficulty, discrimination, quality score, exposure risk, and penalty for duplication with historical test papers.

[0117]

[0118] in: For the title In difficulty gradient layer Candidate scores, For difficulty matching function, , Match weight coefficients to difficulty level. The weighting coefficients are the discrimination parameter. This is the weighting coefficient for the quality score. The exposure risk coefficient is a weighting factor. For repeated penalty items, Related to the similarity of historical exam papers or neighboring exam papers, The weighting coefficient for the repeated penalty term. This represents the difficulty tolerance scale within the difficulty gradient layer. The smaller the size, the more emphasis is placed on aligning with the difficulty of the target.

[0119] Using a Gaussian function ensures that the closer to the center of the target difficulty, the better. The higher the score on the questions, the easier the level. It can be large, a difficult problem. It can be small. , , , and It needs to be adjusted according to the exam objectives. The settings are based on the repetition rate of questions among test takers (e.g., questions recently answered by students in the same class). Take 0.8, otherwise take 0.1).

[0120] Candidate sets are obtained by filtering candidates based on their scores, and initial test papers are generated by performing tiered and segmented question selection on the candidate sets. The tiered and segmented question selection includes at least the following: prioritizing the coverage of scarce knowledge points and then supplementing the question type quota and gradient layer quota according to the candidate scores.

[0121] For example: each candidate set by score Sort the questions in descending order and select the Top-K questions as candidates for this layer. The value of K is usually 1.5-2 times the target number of questions for this question type-gradient layer (e.g., if 10 questions need to be selected for the "multiple choice - easy question layer", then K=15), so as to reserve sufficient selection space for subsequent initial test paper assembly and constraint repair.

[0122] The core of initial test paper assembly is to quickly construct feasible initial solutions that satisfy basic constraints. A layered and segmented greedy strategy is adopted to ensure efficiency while prioritizing key needs. First, the target number of questions in each question type-gradient layer needs to be determined, and then questions are selected independently within each candidate set.

[0123] The core of generating the initial test paper in this embodiment is "prioritizing the coverage of scarce knowledge points," and the scarcity weight of scarce knowledge point m. for: ,in: Let be the coverage of knowledge point m for question i. The fewer questions in the question bank related to this knowledge point, the better. The larger the number of questions, the better, ensuring the exam covers a smaller but more important range of knowledge points in the question bank. The specific construction rules are as follows: First, iterate through all candidate questions and select those with high coverage. Questions on knowledge points are scored according to candidates. Sort in descending order, prioritizing those selected until the knowledge point satisfies the condition. Requirements. For question types that already meet the knowledge point coverage constraint, according to... The remaining quota is filled in descending order to ensure the overall quality of the selected questions is optimal. The third step is cross-question type and degree-level verification to avoid over-concentration of the same knowledge point in one level and to ensure a balanced distribution of knowledge points.

[0124] S104, through iterative replacement and exchange, resolves constraint conflicts and optimizes the final draft of the test paper under multi-objective criteria.

[0125] Initial volume assembly scheme There may be violations of hard constraints, such as insufficient coverage of knowledge points, exceeding the total time limit, or abrupt changes in difficulty gradient. Therefore, it is necessary to repair this by projecting the feasible region, adjusting the solution to a feasible region that satisfies all hard constraints, while simultaneously improving the quality of the test paper. The core idea of ​​this repair is to transform the test paper assembly into a constrained multi-objective optimization problem.

[0126] This embodiment performs constraint violation detection on the initial test paper and uses an iterative repair method of question replacement to ensure that the test paper meets the set of hard constraints and the gradient continuity constraint. The test paper is finalized based on multi-objective optimization criteria, which include at least reducing difficulty gradient deviation, reducing knowledge point coverage gaps, reducing timeout penalties, reducing exposure risk accumulation, and increasing discrimination gains.

[0127] The multi-objective optimization function is:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134] in: For topic selection vectors, For gradient deviation loss, For knowledge gaps and losses, As a penalty for exceeding the time limit, To accumulate risk, For differentiation revenue, and They are respectively , , , and The target weight.

[0135] This reflects the degree of deviation from the difficulty gradient; a larger value indicates that it deviates more from the preset gradient. This reflects gaps in knowledge coverage; a larger value indicates more knowledge points that are not covered. This reflects the timeout situation; a larger value indicates a more severe timeout. This reflects the total exposure risk; a higher value indicates a higher risk. It reflects the overall discrimination ability; the larger the value, the stronger the discrimination ability.

[0136] In this embodiment, the repair strategy consists of three steps: First, locate the constraint with the most severe violation; second, identify high-penalty contribution issues. For example, questions that cause gaps in knowledge coverage, questions that deviate most from the gradient center, and questions that contribute the most to timeouts are calculated to determine their negative impact on the overall objective. The third step is to select questions from the candidate set corresponding to the question type and degree layer that can minimize the impact on the overall objective. Such as covering knowledge gaps and having similar difficulty levels The response time is short and the risk is low, so the "removal" procedure can be implemented. Introduction Exchange the symbols. Repeat the above steps until all hard constraints are satisfied or the maximum number of iterations is reached. If there are still minor violations after iterations, the constraint parameters can be slightly adjusted to ensure the feasibility of the solution.

[0137] S105 updates ability parameters and adjusts subsequent target difficulty based on answer data in dynamic scenarios, and outputs test papers and verification indicators.

[0138] Dynamic intelligent test paper generation is suitable for continuous tests or modular exams. Its core is to dynamically adjust the test paper generation strategy for subsequent modules by using real-time answer data from test takers, thereby achieving personalized test paper generation and improving the accuracy and adaptability of the exam.

[0139] See Figure 3 First, it is necessary to obtain user answer data and update user ability parameters. Based on the updated user ability parameters, the target difficulty range or target difficulty sequence of subsequent test papers is adjusted to generate the next test paper or the next module of test questions.

[0140] Specifically, with Indicating the candidate's level, incremental updates are made:

[0141]

[0142] in: For the first Wheel capacity estimation, This is the step size coefficient. This refers to the number of questions answered in this round. , indicating the first Is the answer to the question correct? To predict the probability of being correct, , For the first Question discrimination For the first Difficulty parameter of the question.

[0143] After the ability update, the difficulty range of the next stage target will be set to... , , , The difficulty level will be adjusted to match the actual abilities of the test takers. If a test taker's ability is high, the difficulty range will be increased in the next stage, with a higher proportion of difficult questions. If a test taker's ability is low, the difficulty range will be decreased, with a higher proportion of easy questions, to avoid questions that are too difficult and discourage test takers, or questions that are too easy and fail to assess their level.

[0144] Output the final exam paper and generate a consistency verification result. The final exam paper output includes: a list of questions, the order of question types, their scores, and answers with explanations. The generated verification report includes: whether the question type quotas are met, and the average score of each graded section. With the goal Check if the deviation and knowledge gap items are zero, if the total time exceeds the limit, and if the repetition rate / exposure risk is below the threshold. See also Figure 4 This is the final test paper generated from a high school mathematics stage test, with the answers and explanations attached at the end of the paper. Figure 5 The generated math test paper structure is analyzed, and teachers can perform a preliminary check on the generated test paper. If it meets the requirements, the teacher can download and use the test paper; otherwise, it can be regenerated or temporarily saved.

[0145] Corresponding to the dynamic intelligent test paper generation method based on multiple question types and difficulty gradients provided in the above embodiments, this application also provides an embodiment of a dynamic intelligent test paper generation system based on multiple question types and difficulty gradients.

[0146] join Figure 6 A dynamic intelligent test paper generation system based on multiple question types and difficulty levels 20, including:

[0147] The question bank data processing module 201 is used to clean and deduplicate the questions in the question bank, construct a question profile containing multi-dimensional information about the questions, and then integrate and label the difficulty.

[0148] The test paper configuration and constraint generation module 202 is used to receive the test paper configuration and generate hard constraints for test paper, divide the test paper into multiple difficulty gradient layers, and set the target difficulty and the continuity constraint from easy to difficult.

[0149] The candidate set generation and initial solution construction module 203 is used to calculate candidate scores and filter candidate sets according to question type and gradient layer, and generate the initial test paper in layers and segments.

[0150] The finalization module 204 is used to fix constraint conflicts and optimize the finalized test paper under multi-objective criteria through iterative replacement and exchange.

[0151] The dynamic update and test paper output module 205 is used to update ability parameters and adjust subsequent target difficulty based on answer data in dynamic scenarios, and output test papers and verification indicators.

[0152] Corresponding to the above embodiments, this application also provides a dynamic intelligent test paper generation device based on multiple question types and difficulty gradients.

[0153] See Figure 7 The dynamic intelligent test paper generation device 300 based on multiple question types and difficulty levels may include a processor 301, a memory 302, and a communication unit 303. These components communicate through one or more buses. Those skilled in the art will understand that the structure of the dynamic intelligent test paper generation device based on multiple question types and difficulty levels shown in the figure does not constitute a limitation on the embodiments of this application. It can be a bus-shaped structure or a star-shaped structure, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0154] The communication unit 303 is used to establish a communication channel, so that the dynamic intelligent test paper generation device based on multiple question types and difficulty gradients can communicate with the test paper generation device of the school or teacher.

[0155] The processor 301 serves as the control center of the dynamic intelligent test paper generation device based on multiple question types and difficulty levels. It connects various parts of the device via various interfaces and lines, and executes software programs and / or modules stored in the memory 302, as well as calling data stored in the memory, to perform various functions and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 301 may only include a central processing unit (CPU). In this embodiment, the CPU may have a single processing core or include multiple processing cores.

[0156] Memory 302 is used to store the execution instructions of processor 301. Memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0157] When the execution instructions in memory 302 are executed by processor 301, the dynamic intelligent test paper generation device 300 based on multiple question types and difficulty levels can perform some or all of the steps in the above method embodiments. For details, please refer to the method embodiments of this application, which will not be repeated here.

[0158] Corresponding to the above embodiments, this application also provides a computer-readable storage medium, wherein the computer-readable storage medium may store a program, wherein when the program runs, it can control the device where the computer-readable storage medium is located to execute some or all of the steps in the above method embodiments. In specific implementation, the computer-readable storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0159] Corresponding to the above embodiments, this application also provides a computer program product containing executable instructions that, when executed on a computer, cause the computer to perform some or all of the steps in the above method embodiments.

[0160] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0161] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A dynamic intelligent test paper setting method based on multi-topic and multi-difficulty gradient, characterized in that, include: The question bank is cleaned and deduplicated to construct a question profile containing multi-dimensional information about the questions, and then the profiles are integrated and labeled with difficulty. Receive test paper configuration and generate hard constraints for test paper generation, divide the test paper into multiple difficulty gradient layers, and set target difficulty and continuity constraints from easy to difficult, including: Receive test paper configuration and generate a set of hard constraints, which includes at least question type quota constraints, total time constraints and knowledge point coverage lower limit constraints. The test paper is divided into multiple difficulty gradients. Target difficulty is set for each gradient, and gradient continuity constraints are set so that the difficulty of each gradient increases smoothly from easy to difficult. The process of dividing the exam paper into multiple difficulty gradients, setting target difficulty for each gradient and imposing gradient continuity constraints to ensure a smooth increase in difficulty from easy to hard across all gradients includes: The test paper is divided into difficulty gradient layers, and target difficulty centers of each gradient layer are generated ; The gradient continuity constraint is set to make the actual average difficulty of each gradient layer satisfies: ； in: As a difficulty gradient layer, The number of gradient layers. For the first The center of difficulty of the layer target, The linear gradient step size is... For the first The actual average difficulty of each level. The allowable deviation is defined as such; the smaller the allowable deviation, the more stringent the requirements. This is a set of questions assigned to a section according to the order of the exam papers. The minimum gradient increment between adjacent segments. For difficulty parameters, To answer the question The set binary variable, when selected into the exam paper ,otherwise ; Calculate candidate scores and filter candidate sets according to question type and gradient level, and generate initial test papers in layers and segments; Constraint conflicts were resolved through iterative replacement and exchange, and the final draft of the test paper was optimized under multi-objective criteria. In dynamic scenarios, ability parameters are updated based on answer data, and subsequent target difficulty is adjusted. The test paper and verification indicators are output, including: Obtain user answer data and update user ability parameters. Based on the updated user ability parameters, adjust the target difficulty range or target difficulty sequence of subsequent test papers to generate the next test paper or the next module of test questions. Output the final test paper and generate consistency verification results. The consistency verification results include at least: question type quota satisfaction, difficulty gradient deviation, knowledge point coverage satisfaction, total time satisfaction, and exposure risk and repetition rate indicators. The process of acquiring user response data and updating user ability parameters, and adjusting the target difficulty range or target difficulty sequence of subsequent test papers based on the updated user ability parameters to generate the next test paper or the next module of questions, includes: To Indicate examinee level, do delta update: in: For the first Wheel capacity estimation, This is the step size coefficient. The number of questions answered in this round. , indicating the first Is the answer to the question correct? To predict the probability of being correct, , For the first Question discrimination For the first Difficulty parameters; The next stage target difficulty range is then set as , , , is the difficulty spread around the ability.

2. The dynamic intelligent paper compiling method based on multi-topic and multi-difficulty gradient according to claim 1, characterized in that, The process of cleaning and deduplicating questions in the question bank to construct a question profile containing multi-dimensional information and then integrating and labeling the difficulty includes: The question bank data is cleaned, deduplicated, and version-merged. A question profile is then created for each question, including question type, knowledge point coverage, difficulty, discrimination, response time, exposure risk, and quality score. wherein: is a question type identifier, is a knowledge point coverage vector, is a difficulty parameter, is a discrimination parameter, is an estimated answer duration, is an exposure risk coefficient, is a quality score; and fusing the labeling difficulty with the statistical difficulty to obtain the difficulty parameter In the middle: wherein: is a human or expert annotation difficulty, is a statistical difficulty, is a fusion coefficient, controlling the trust in the label. 3.The dynamic intelligent paper compiling method based on multi-topic and multi-difficulty gradient according to claim 2, characterized in that, The receiving test paper configuration and generation of a hard constraint set, the hard constraint set including at least question type quota constraints, total time constraints, and knowledge point coverage lower limit constraints, including: if (question == "What is the capital of France?") set binary variable , selected then , else ; Input group paper configuration: question type set , number of each question type , total score, total duration, knowledge point target weight, generate constraint set, the constraint set includes: Number of question types constraint: ; Total duration constraint and: ; Knowledge point coverage constraint: ; wherein: is the maximum response duration, is the coverage weight, indicating the question whether the question covers the knowledge point , is the knowledge point minimum coverage amount.

4. The dynamic intelligent paper compiling method based on multi-topic and multi-difficulty gradient according to claim 1, characterized in that, The process of calculating candidate scores and filtering candidate sets according to question type and gradient layer, and generating initial test papers in layers and segments, includes: For each question type and each grade level, candidate scores are calculated. These candidate scores comprehensively consider the degree of match between the question difficulty and the target difficulty, discrimination, quality score, exposure risk, and penalty for duplication with historical test papers. in: For the title In difficulty gradient layer Candidate scores, For difficulty matching function, , Match weight coefficients to difficulty level. The weighting coefficients are the discrimination parameter. This is the weighting coefficient for the quality score. The exposure risk coefficient is a weighting factor. For repeated penalty items, Related to the similarity of historical exam papers or neighboring exam papers, The weighting coefficient for the repeated penalty term. This represents the difficulty tolerance scale within the difficulty gradient layer. The smaller the size, the more emphasis is placed on aligning with the difficulty of the target; Candidate sets are obtained by filtering candidates based on their scores, and initial test papers are generated by performing tiered and segmented question selection on the candidate sets. The tiered and segmented question selection includes at least the following: prioritizing the coverage of scarce knowledge points and then supplementing the question type quota and gradient layer quota according to the candidate scores.

5. The dynamic intelligent paper compiling method based on multi-topic and multi-difficulty gradient according to claim 4, characterized in that, The method of repairing constraint conflicts through iterative replacement and optimization of the final exam paper under multi-objective criteria includes: The initial test paper is subjected to constraint violation detection, and the test paper is made to satisfy the set of hard constraints and the gradient continuity constraint through an iterative repair method of question replacement and exchange. The test paper was finalized based on multi-objective optimization criteria, which include at least reducing the difficulty gradient deviation, reducing the knowledge point coverage gap, reducing the timeout penalty, reducing the accumulation of exposure risk, and increasing the discrimination benefit.

6. The dynamic intelligent paper compiling method based on multi-topic and multi-difficulty gradient according to claim 5, characterized in that, The finalization of the test paper based on multi-objective optimization principles includes at least reducing difficulty gradient deviation, reducing knowledge point coverage gaps, reducing timeout penalties, reducing exposure risk accumulation, and improving discrimination gains, including: in: For topic selection vectors, For gradient deviation loss, For knowledge gaps and losses, As a penalty for exceeding the time limit, To accumulate risk, For differentiation revenue, and They are respectively , , , and The target weight.

Images