CN-121434412-B - Test paper generation method and system

CN121434412BCN 121434412 BCN121434412 BCN 121434412BCN-121434412-B

Abstract

The invention discloses a test paper generation method and a system, which belong to the technical field of intelligent education, and are characterized by obtaining student answer records, question-knowledge point mapping, a question library and knowledge text information, constructing a knowledge point map, constructing a student-question-knowledge three-layer abnormal composition, predicting answer performances of students on test paper, taking the answer performances as environment feedback input of reinforcement learning, taking a question set of a current test paper as a state, replacing the questions in the test paper or reserving the current test paper as an action, determining comprehensive reward values of the test paper in each dimension by constructing a multi-objective reward function, maximizing the comprehensive reward values as targets, generating an optimal test paper by iterative optimization, selecting candidate test questions from the question library based on an adjacent relation in the knowledge point map, and generating a plurality of parallel test papers equivalent to the optimal test paper by using an equivalent question replacement strategy. By the method, more efficient, reasonable and fair parallel test paper generation can be realized.

Inventors

WANG HUIYA
YAN HAOCHEN
LI KAIXUAN
Takashi Takashi
WANG BIXUAN
FENG JUN

Assignees

西北大学

Dates

Publication Date: 20260508
Application Date: 20251103

Claims (8)

1. The test paper generating method is characterized by comprising the following steps of: Obtaining student answer records, question-knowledge point mapping, a question library and knowledge text information, respectively extracting student embedded representation, question embedded representation and knowledge embedded representation, updating embedded relations of all nodes in a knowledge point map by adopting a graph attention network, and constructing a student-question interaction map, a question-knowledge point association map and a knowledge point map fusing semantic and co-occurrence relations; Integrating knowledge embedding representation into student embedding representation and question embedding representation by using a fully connected network, optimizing a graph structure, constructing three layers of different composition of student-question-knowledge, extracting high-order characteristic representation of each node of the different composition, and predicting answer sheet performance of students on test paper: Obtaining answer expression of student i to question j through full-connection network prediction The method comprises the following steps: ; Determining the loss function of a fully connected network uses binary cross entropy: ; For the test paper with n questions, the total score of the test paper prediction of the student i, namely the answer expression of the student i to the test paper is expressed as follows: ; Wherein, the Representing knowledge points Is defined by the number of knowledge points adjacent to each other, , For the integrated student to embed a representation, For the embedded representation of the integrated title, Is that The function of the function is that, Is a full-connection layer, and is formed by the following steps, In order to achieve a true answer result, A score representing the j-th question; the method comprises the steps of taking answering performance of students on test papers as environment feedback input of reinforcement learning, taking a question set of the current test papers as a state, replacing questions in the test papers or reserving the questions as actions, determining comprehensive rewarding values of the test papers on all dimensions of the multi-objective rewarding functions by constructing the multi-objective rewarding functions of comprehensive difficulty, effectiveness, rationality and discrimination of the test papers, maximizing the comprehensive rewarding values, and generating an optimal test paper through iterative optimization: Obtaining and measuring average and target difficulty of students Integrated difficulty rewards for consistency of (c) The method comprises the following steps: ; obtaining validity rewards measuring consistency of test paper knowledge distribution and course requirements The method comprises the following steps: ; obtaining rationality rewards measuring similarity of student score distribution and target distribution The method comprises the following steps: ; obtaining discrimination rewards measuring the ability of test papers to discriminate between different levels of students The method comprises the following steps: ; Wherein, the For the total score of the test paper, Is equally divided for students, The first test paper The proportion of investigation of the individual knowledge points, To be the required proportion of the knowledge point in the course, For the actual score distribution of the student, Is the target distribution, is the average score of the first 27 percent of students, Average score for the last 27% students; Determining a comprehensive rewarding function as follows: ; Through the comprehensive rewarding function, the intelligent agent optimizes a plurality of teaching targets simultaneously in each question updating so as to generate an optimal test paper; Based on the adjacency relation in the knowledge point map, candidate questions are selected from a question library to form a candidate sub-question set, questions which are different but cover the same knowledge points as the optimal test paper are replaced and combined by using an equivalent question replacement strategy, and a plurality of sets of parallel test paper equivalent to the optimal test paper are generated through iterative optimization: For each knowledge node Candidate subset of questions The method comprises the following steps of jointly forming a directly-related topic set and a topic set related to adjacent knowledge nodes: ; Wherein, the Representing direct survey knowledge points Is a set of topics of (1), Representation of Is a set of adjacent knowledge points; the equivalent question replacement strategy is to generate each question in the parallel test paper Corresponding knowledge node The candidate subtopic set is subjected to neighborhood expansion operation, test questions are randomly selected from the candidate subtopic sets to serve as replacement candidates, when the selected questions in the replacement candidates meet a multi-objective rewarding function, replacement is executed, when proper candidate questions are not found, the candidate subtopic sets searched to neighborhood knowledge nodes are sequentially expanded, and for candidate questions meeting the conditions are not found after neighborhood expansion, the original questions are reserved and are not replaced, so that consistency of test paper on knowledge coverage is guaranteed.
2. The test paper generation method according to claim 1, wherein the generating a plurality of parallel test papers equivalent to the optimal test paper based on the adjacency relation in the knowledge point map includes selecting candidate test papers from the test paper library to form candidate sub-question sets, and replacing and combining different questions covering the same knowledge point as the optimal test paper by using an equivalent question replacement strategy, and the generating the plurality of parallel test papers equivalent to the optimal test paper specifically includes: Randomly selecting test questions from the candidate subtopic sets as replacement candidates, and executing replacement action on the original questions in the optimal test paper when the replacement candidates meet the maximization of the comprehensive rewarding value, otherwise, sequentially expanding the candidate subtopic sets searched to the neighborhood knowledge nodes, and reserving the original questions without replacement action for the replacement candidates which are not found out after the neighborhood expansion; The questions in the optimal test paper are replaced by replacing and combining the questions which are different but cover the same knowledge points as the optimal test paper by continuously iterating in the state and action space and utilizing an equivalent question replacement strategy, the comprehensive difficulty, effectiveness, rationality and degree of distinction of the optimal test paper are optimized, the comprehensive reward value is maximized, and a plurality of parallel test paper equivalent to the optimal test paper are generated through iterative optimization.
3. The test paper generating method according to claim 1, wherein the answering performance of the student on the test paper is used as an environmental feedback input of reinforcement learning, the question set of the current test paper is used as a state to replace the questions in the test paper or remain as an action, the comprehensive reward value of the test paper on each dimension of the multi-objective reward function is determined by constructing the multi-objective reward function of the comprehensive difficulty, effectiveness, rationality and distinction degree of the test paper, the comprehensive reward value is maximized as a target, and the optimal test paper is generated by iterative optimization, and the method specifically comprises: Iterative training is carried out on the state and the action of reinforcement learning by adopting a double-depth Q network, the double-depth Q network takes the answer performance of students on test papers as environmental feedback input, takes the question set of the current test papers as the state, takes the equivalent question replacement strategy as the action to replace the questions in the test papers or keep the questions as the action, and aims at maximizing the comprehensive rewarding value; In the training process, the intelligent agent uses epsilon-greedy strategy to balance exploration and utilization, a priority experience playback mechanism is used for sampling a historical state-action-rewarding sequence, multiple rounds of iterative training are used for updating double-depth Q network parameters, the intelligent agent gradually learns the optimal strategy for selecting questions in different states, and the student is simulated to answer performance of the test paper, so that the optimal test paper which takes the comprehensive difficulty, effectiveness, rationality and discrimination into consideration is generated, wherein the comprehensive rewarding value in each dimension is the largest, and the multi-objective requirement is met.
4. The test paper generating method according to claim 1, wherein the integrating the knowledge embedding representation into the student embedding representation and the question embedding representation predicts the answering performance of the student to the test paper by constructing a student-question-knowledge three-layer different composition and extracting higher-order feature representations of each node of the different composition, specifically comprising: Respectively obtaining a student embedded representation, a question embedded representation and a knowledge embedded representation; Extracting high-order characteristic representation of each node by constructing three layers of different patterns of students, topics and knowledge; and fusing the low-dimensional vector and the high-order characteristic representation, iteratively aggregating neighbor node information, and determining answer expression by predicting answer probability of students on unanswered questions in the test paper.
5. The test paper generating method according to claim 1, wherein the obtaining of the student answer record, the question-knowledge point map, the question library and the knowledge text information extracts student embedded representation, question embedded representation and knowledge embedded representation respectively, and updates the embedded relationship of each node in the knowledge point map by using a graph attention network to construct a student-test question interaction map, a test question-knowledge point association map and a knowledge point map fusing semantic and co-occurrence relationships, specifically comprising: Generating one-hot vectors through the encoding process of the encoder according to the acquired student answer records, the question-knowledge point mapping, the question library and the knowledge text information, and respectively acquiring initial embedded representations corresponding to the students, the questions and the knowledge points; Aggregating adjacent node information through a graph attention network to update an initial embedded representation, wherein the graph attention network comprises an embedded layer and a full connection layer; And integrating the knowledge embedding representation into the student embedding representation and the question embedding representation respectively at the full-connection layer to construct a student-test question interaction diagram, a test question-knowledge point association diagram and a knowledge point map fusing semantic and co-occurrence relation.
6. The test paper generation method according to claim 1, wherein the generating of a plurality of parallel test papers equivalent to the optimal test paper through iterative optimization further comprises evaluating the equivalent of the generated parallel test papers on four levels of student individuals, student groups, knowledge and questions based on knowledge distribution entropy, knowledge frequency quotient and connected component entropy.
7. The test paper generating method according to claim 1, wherein the obtaining of the student answer record, the test question library, the question-knowledge point mapping and the knowledge text information specifically comprises: When a student answer record is obtained, a history answer log of the student is derived from an online education platform, wherein the log comprises a student ID, a question ID, an answer result comprising correct or incorrect, an answer time stamp and answer attempt times; Acquiring a question-knowledge point mapping, manually marking by a subject expert or a course teacher according to a course outline and a teaching target, determining one or more knowledge points inspected by each question, and forming a structured question-knowledge point association matrix; Acquiring knowledge text information, wherein the knowledge text information comprises unique identification, name, detailed description text and relative importance weight of the detailed description text in a course knowledge system of each knowledge point; And cleaning and preprocessing the acquired student answer records, test question library, question-knowledge point mapping and knowledge text information.
8. A test paper generation system, comprising: The knowledge point map construction module is used for acquiring student answer records, question-knowledge point mapping, a question library and knowledge text information, respectively extracting student embedded representation, question embedded representation and knowledge embedded representation, updating the embedded relation of each node in the knowledge point map by adopting a map attention network, and constructing a student-test question interaction map, a test question-knowledge point association map and a knowledge point map fusing semantic and co-occurrence relation; The heterogeneous diagram construction module is used for integrating knowledge embedding representation into student embedding representation and question embedding representation by using a fully connected network respectively, optimizing a diagram structure, and predicting answering performance of students on test paper by constructing three layers of different composition of students-questions-knowledge and extracting high-order characteristic representation of each node of the different composition. Obtaining answer expression of student i to question j through full-connection network prediction The method comprises the following steps: ; Determining the loss function of a fully connected network uses binary cross entropy: ; For the test paper with n questions, the total score of the test paper prediction of the student i, namely the answer expression of the student i to the test paper is expressed as follows: ; Wherein, the Representing knowledge points Is defined by the number of knowledge points adjacent to each other, , For the integrated student to embed a representation, For the embedded representation of the integrated title, Is that The function of the function is that, Is a full-connection layer, and is formed by the following steps, In order to achieve a true answer result, A score representing the j-th question; The optimal test paper generating module is used for inputting answer performances of students on test paper as environment feedback of reinforcement learning, replacing the questions in the test paper or reserving the questions as actions by taking the current question set of the test paper as a state, determining comprehensive rewarding values of the test paper on each dimension of the multi-target rewarding function by constructing the multi-target rewarding function of comprehensive difficulty, effectiveness, rationality and discrimination of the test paper, maximizing the comprehensive rewarding values as targets, and generating the optimal test paper by iterative optimization: Obtaining and measuring average and target difficulty of students Integrated difficulty rewards for consistency of (c) The method comprises the following steps: ; obtaining validity rewards measuring consistency of test paper knowledge distribution and course requirements The method comprises the following steps: ; obtaining rationality rewards measuring similarity of student score distribution and target distribution The method comprises the following steps: ; obtaining discrimination rewards measuring the ability of test papers to discriminate between different levels of students The method comprises the following steps: ; Wherein, the For the total score of the test paper, Is equally divided for students, The first test paper The proportion of investigation of the individual knowledge points, To be the required proportion of the knowledge point in the course, For the actual score distribution of the student, In order to achieve a distribution of the objects, Is the average score of the first 27% of students, Average score for the last 27% students; Determining a comprehensive rewarding function as follows: ; Through the comprehensive rewarding function, the intelligent agent optimizes a plurality of teaching targets simultaneously in each question updating so as to generate an optimal test paper; The parallel test paper generation module is used for selecting candidate test questions from the test question library to form candidate sub-question sets based on the adjacent relation in the knowledge point map, replacing and combining different questions covering the same knowledge point as the optimal test paper by using an equivalent question replacement strategy, and generating a plurality of sets of parallel test papers equivalent to the optimal test paper through iterative optimization. For each knowledge node Candidate subset of questions The method comprises the following steps of jointly forming a directly-related topic set and a topic set related to adjacent knowledge nodes: ; Wherein, the Representing direct survey knowledge points Is a set of topics of (1), Representation of Is defined by the number of knowledge points adjacent to each other, Representation of Is a set of adjacent knowledge points; the equivalent question replacement strategy is to generate each question in the parallel test paper Corresponding knowledge node The candidate subtopic set is subjected to neighborhood expansion operation, test questions are randomly selected from the candidate subtopic sets to serve as replacement candidates, when the selected questions in the replacement candidates meet a multi-objective rewarding function, replacement is executed, when proper candidate questions are not found, the candidate subtopic sets searched to neighborhood knowledge nodes are sequentially expanded, and for candidate questions meeting the conditions are not found after neighborhood expansion, the original questions are reserved and are not replaced, so that consistency of test paper on knowledge coverage is guaranteed.

Description

Test paper generation method and system Technical Field The invention relates to the technical field of intelligent education, in particular to a test paper generation method and a test paper generation system. Background As a powerful evaluation tool, the test paper can objectively measure the knowledge mastery degree of students, and can further improve the teaching effect when being combined with data analysis and personalized feedback in an intelligent education environment. Therefore, in the age of intelligent education, test papers remain an important tool for evaluating learning results and promoting accurate teaching. The task of coupon generation (EPG, exam Paper Generation) relies on efficient organization and mining of educational data and resources. Based on the EPG, the parallel group (PARALLEL PAPER generation, PPG) task requires designing multiple different versions of test papers that have nearly identical evaluation efficacy, while containing nearly different test question sets to ensure fairness of evaluation and reduce the risk of cheating. Traditional manual generation of high quality test paper is a dual physical and mental task for teachers. In recent years, with the development of algorithms, research on test paper generation methods has been focused on improvements and applications of meta-heuristic algorithms, such as genetic algorithms, and their use in optimizing a plurality of key attributes of test paper, such as difficulty, degree of distinction, and duration of answer. Such methods are typically based on a global sampling framework that can search a wide range of question bank combining schemes in order to generate test papers meeting expected criteria. In the past ten years, the method has remarkably developed in the field of automatic generation of test paper by virtue of the advantages of simple algorithm principle, easy implementation, relatively low calculation cost and the like, and promotes the development and landing of intelligent paper assembling technology. However, the existing method still has two prominent key problems, on one hand, expert experience is excessively relied on in the process of setting and marking the test paper attribute, namely, manual assignment is required to be carried out on indexes such as the difficulty and the distinction degree of the questions by teachers or field experts, and the mode not only consumes a great deal of manpower and time cost, but also has subjectivity and deviation inevitably, so that the objectivity and universality of the winding result are affected. On the other hand, most of the existing optimization methods look at the overall quality of the question combination, such as the question coverage rate, knowledge point distribution or attribute balance, from the perspective of the question library, but often neglect the performance and feedback of students when actually completing the generation of test papers. The optimization view angle with the problem as the center causes the examination paper to lack consideration of individual difference, is difficult to meet the demands of teaching and personalized learning according to the material, and also limits the application value of the EPG technology in real education scenes. In order to solve the limitation of the traditional method, part of research introduces Depth Knowledge Tracking (DKT) based on LSTM, which is used for tracking the learning progress of students and predicting test paper performance, and combines various generation strategies to promote EPG effect, and is more superior than meta-heuristic algorithm. However, the prior knowledge graph and the whole structure are not fully utilized in the prior method, and the inherent connection of the individual students, the knowledge and the problem level is ignored, so that the evaluation effect in the test paper generation is inconsistent with the teaching target. Disclosure of Invention Aiming at the problems in the field, the invention provides a test paper generation method and a system, which are based on the adjacency relation in the knowledge point map, and the candidate test questions are selected from the test question library to form a candidate sub-question set, so that the equivalence of parallel test paper in the aspects of problems, knowledge and individuals can be comprehensively enhanced, and more efficient, reasonable and fair parallel test paper generation can be realized. In order to solve the technical problems, the invention discloses a test paper generation method, which comprises the following steps: obtaining student answer records, question-knowledge point mapping, a question library and knowledge text information, respectively extracting student embedded representation, question embedded representation and knowledge embedded representation, and constructing a student-test question interaction diagram, a test question-knowledge point association diagram and a knowledge point diagram fusing