Personalized Dynamic Question Generation Method and System Based on Large Language Model

Abstract

The present invention discloses a personalized dynamic question generation method and system based on a large language model. The steps of the method are as follows: S1. Collect multi-source data such as students' historical answer records, subject scores, and learning directions, and construct a structured feature vector; S2. Evaluate the students' mastery of the complete knowledge system through a pre-trained model, and calculate the knowledge point coverage; S3. Extract a default question set from the question bank, use the attention mechanism to convert the students' knowledge point mastery scores into weights, dynamically adjust the difficulty distribution of the questions, and strengthen the training of weak links; S4. Integrate the student characteristics, mastery scores, coverage, and adjusted difficulty distribution, and generate questions and push them through a personalized question generation model fine-tuned by a large language model. The present invention utilizes the powerful semantic understanding and generation capabilities of the large language model, can analyze the students' learning trajectories, dynamically adjust the difficulty of the questions, the question type distribution, and the knowledge point coverage range, so as to achieve accurate and personalized question recommendation.

Description

Technical Field

[0001] The present invention belongs to the field of generative artificial intelligence, and particularly relates to a personalized dynamic question generation method and system based on a large language model. Background Art

[0002] With the rapid development of online education and intelligent learning systems, personalized teaching based on artificial intelligence has become a research hotspot in the education field. Traditional question generation methods mainly rely on manual compilation by teachers or fixed question banks, making it difficult to adapt to the personalized needs of students in real time, resulting in low matching degree of learning content, uncontrollable difficulty, and difficulty in effectively improving students' learning efficiency and learning effect.

[0003] Currently, some intelligent education systems attempt to use technologies such as knowledge graphs and rule matching to achieve personalized question generation, but there are still many limitations: 1) Inaccurate personalized matching: Most existing systems set questions based on static rules and fail to fully combine multi-dimensional information such as students' historical learning data, cognitive abilities, and interest preferences for dynamic adjustment, resulting in insufficient accuracy of question recommendation. 2) Limited ability to control question difficulty: Traditional methods are difficult to dynamically optimize question difficulty according to students' learning progress and answering performance, which may lead to excessive question difficulty increasing learning pressure or too low difficulty affecting the challenge, ultimately reducing the learning effect. 3) Weak feedback and optimization mechanism: Existing intelligent question generation systems mainly rely on simple accuracy statistics for adjustment, lacking advanced algorithms for in-depth optimization, and it is difficult to provide accurate personalized learning feedback, restricting the continuous improvement of learning effect.

[0004] In recent years, the rapid development of large language models (LLMs) has provided a new technical path for personalized question generation. Pre-trained language models, with their deep semantic understanding, conditional generation strategies, and attention mechanisms, can more accurately match students' learning needs and dynamically adjust question difficulty and knowledge point distribution in combination with their learning trajectories and goals. However, existing question generation methods based on large language models still face the following challenges: 1) Insufficient utilization of personalized data: Most systems only rely on students' historical answering records and fail to fully integrate multi-dimensional learning data such as learning habits, interest preferences, and goal planning, resulting in limited accuracy of personalized recommendation. 2) Imperfect dynamic question generation strategy: Existing methods lack an efficient adaptive optimization mechanism and are difficult to adjust question difficulty, question type distribution, and knowledge point coverage based on real-time feedback, affecting the personalized learning experience. 3) Insufficient multi-modal information fusion ability: Current systems mainly rely on text data for question generation, ignoring the enhancing effect of multi-modal information such as images and voices on learning effect, resulting in a single question presentation method and difficulty in meeting the needs of different learning scenarios. Summary of the Invention

[0005] The object of the present invention is to solve the problems existing in the prior art, and discloses a personalized dynamic question generation method and system based on a large language model.

[0006] The specific technical solutions adopted by the present invention are as follows:

[0007] In the first aspect, the present invention provides a personalized dynamic question generation method based on a large language model, which includes:

[0008] S1. Obtain multi-dimensional information data of the target student user, including historical answering records, historical scores of each subject, and learning direction guidance information, and construct them into a structured student feature vector through feature engineering;

[0009] S2. Input the student feature vector into a pre-trained knowledge point mastery degree evaluation model to obtain the mastery degree score of the target student user on each knowledge point in the complete knowledge point distribution, and then calculate the knowledge point coverage based on the mastery degree scores of the target student user on all knowledge points;

[0010] S3. Extract a default question set to be recommended to the target student from the current question bank, and analyze the question difficulty distribution of the default question set on all knowledge points; convert the mastery degree scores of the target student user on all knowledge points into attention weights for each knowledge point through an attention mechanism, and then use the attention weights to dynamically adjust the question difficulty distribution, increase the question difficulty corresponding to the knowledge points with attention weights exceeding the weight threshold, and decrease the question difficulty corresponding to the knowledge points with attention weights lower than the weight threshold to obtain an adjusted question difficulty distribution;

[0011] S4. Construct the default questions of the target student, the mastery degree scores of the target student user on all knowledge points, the knowledge point coverage, and the adjusted question difficulty distribution into a prompt text in combination with a prompt template, input the prompt text into a personalized question generation model obtained by fine-tuning a large language model (LLM) on the question generation task, and output a new question set by the personalized question generation model and push it to the target student user.

[0012] As a preference of the above first aspect, the learning direction guidance information includes the subjects that the target student user needs to focus on learning and / or the knowledge points that need to be focused on learning in each subject.

[0013] As a preference of the above first aspect, the knowledge point mastery degree evaluation model is obtained by performing supervised training on a multi-layer perceptron, a random forest, a support vector machine, or a convolutional neural network on a labeled data set.

[0014] Preferably, as for the first aspect above, the knowledge point coverage is calculated based on the expected distribution of the mastery level of knowledge points predefined by the teacher user and the mastery level scores of the target student user on all knowledge points. The calculation method is as follows: for each knowledge point in the complete knowledge point distribution, compare the expected mastery level score corresponding to this knowledge point in the expected distribution of the mastery level of knowledge points with the mastery level score of the target student user on this knowledge point, and take the smaller value as the representation value. Then, calculate the ratio of the sum of the representation values of all knowledge points to the sum of the expected mastery level scores of all knowledge points, and the calculated ratio is used as the current knowledge point coverage of the target student user.

[0015] Preferably, as for the first aspect above, when fine-tuning the large language model (LLM) for the question generation task, optimize the model parameters of the large language model with the goal of maximizing the sum of the knowledge point coverage and the knowledge point difficulty matching degree to obtain a personalized question generation model.

[0016] Preferably, as for the first aspect above, the knowledge point difficulty matching degree is the mean square error between the adjusted question difficulty distribution input to the model and the question difficulty distribution of the new question set output by the model.

[0017] Preferably, as for the first aspect above, during the actual application process of the personalized question generation model, it is necessary to construct an incremental learning data set by collecting feedback data, and optimize the model parameters of the personalized question generation model regularly through the Bayesian optimization algorithm with the goal of maximizing the sum of the knowledge point coverage and the knowledge point difficulty matching degree.

[0018] In the second aspect, the present invention provides a personalized dynamic question generation system based on a large language model, which includes:

[0019] An information collection module for recording multi-dimensional information data of each student user on the online learning platform;

[0020] A target user selection module for allowing the user to select the target student user for whom personalized test questions need to be pushed;

[0021] A personalized push module for performing personalized test question pushing for the target student user selected by the user according to the multi-dimensional information data recorded in the information collection module according to the personalized dynamic question generation method based on the large language model described in any one of the above first aspect solutions.

[0022] In the third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the personalized dynamic question generation method based on the large language model described in any one of the above first aspect solutions.

[0023] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0024] The memory is used to store computer programs;

[0025] The processor is used to, when executing the computer program, implement the personalized dynamic question generation method based on the large language model as described in any of the above-mentioned first aspect solutions.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention provides a personalized dynamic question generation method and system based on the large language model. Through a multi-dimensional data-driven dynamic question generation mechanism, the present invention combines multi-source information such as the student's historical answer records, historical academic performance in various disciplines, and learning direction guidance to construct an intelligent and personalized dynamic question generation model. This method utilizes the powerful semantic understanding and generation capabilities of the large language model to analyze the student's learning trajectory, dynamically adjust the question difficulty, question type distribution, and knowledge point coverage, so as to achieve precise and personalized question recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the steps of the personalized dynamic question generation method based on the large language model;

[0029] Figure 2 It is a schematic diagram of the module composition of the personalized dynamic question generation system based on the large language model;

[0030] Figure 3 It is a schematic diagram of the composition of the computer electronic device. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] In order to make the above-mentioned objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made with reference to the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below. The technical features in various embodiments of the present invention can be combined with each other without conflict.

[0032] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be construed as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features.

[0033] The present invention provides a personalized dynamic question generation method based on a large language model, which can be used to recommend questions to students on an online learning platform in a personalized manner to meet the students' own learning progress and learning needs, and maximize the effect of online learning.

[0034] As Figure 1 shown, in a preferred embodiment of the present invention, the personalized dynamic question generation method based on a large language model includes steps S1 to S4, and the specific implementation of each step will be described in detail below.

[0035] S1. Obtain multi-dimensional information data of the target student user, including historical answering records, historical scores of each subject, and learning direction guidance information, and construct them into a structured student feature vector through feature engineering.

[0036] It should be noted that the target student user refers to the object for whom the current personalized dynamic question generation method on the online learning platform needs to generate questions. The historical answering record refers to the answering record of the student user for the exercises on the platform, including answering time, answering accuracy rate, answering ideas, and answering duration, etc. The historical scores of each subject refer to the historical score data of the student in each subject, such as the scores of unit tests, mid-term and final exams, and mock exams. The learning direction guidance information refers to the situation where teachers or parents, etc. guide the next learning focus direction of the student, such as learning suggestions, tutoring content, learning habit guidance, etc. information, which needs to include the subjects that the target student user needs to focus on learning, and can also include the knowledge points that need to be focused on learning in each subject, and can be adjusted according to actual needs.

[0037] In an embodiment of the present invention, multi-dimensional information data can be collected for student users through steps such as collecting historical answering situations, subject score data, learning guidance situations, future goals and plans, etc. Collection of historical answering situations: Collect the answering records of students during the past learning process, including answering time, answering accuracy rate, answering ideas (extracting key information from the answering text through natural language processing technology), and answering duration, etc., and convert them into structured data for storage. Collection of subject score data: Obtain the historical score data of students in each subject from the score management system of the online learning platform, including the scores of unit tests, mid-term and final exams, and mock exams. Normalize the score data to facilitate integration with other dimensional data. Collection of learning guidance situations: Through questionnaires, interviews, or online feedback systems, collect the learning guidance situations of teachers and parents for students, including learning suggestions, tutoring content, learning habit guidance, etc. These information are stored after extracting key features through text analysis technology. Collection of future goals and plans: Collect the learning goals and future plans of students, including short-term goals (such as improving the score of a certain subject) and long-term goals (such as applying for a target college or major). Convert these goals into structured knowledge point requirements through natural language processing technology. For the data collected in the above learning guidance situations and future goals and plans, the core keywords need to be extracted, that is, the subjects that need to be focused on learning and the knowledge points that need to be focused on learning in each subject, so as to be used as key features for the recommendation task applicable to the present invention. Among them, information such as the target college or major in the long-term goal can be converted into key features through the key subjects concerned by different colleges and majors.

[0038] In addition, when constructing a structured student feature vector through feature engineering above, information in different dimensions needs to be vectorized according to their respective data characteristics. Numerical data can be directly used after standardization or normalization. Categorical features can be converted into feature vectors through one-hot encoding or label encoding. Text data needs to be preprocessed first, such as word segmentation, stop word removal, keyword extraction, etc., and then converted into numerical vectors through a word vector model. Therefore, for all data, vectors can be formed dimension by dimension, and then structured to form the final student feature vector. , where represents the i-th dimensional feature value, and n represents the total number of dimensions.

[0039] S2. Input the student feature vector into a pre-trained knowledge point mastery degree evaluation model to obtain the mastery degree score of the target student user on each knowledge point in the complete knowledge point distribution, and then calculate the knowledge point coverage based on the mastery degree scores of the target student user on all knowledge points.

[0040] It should be noted that the above complete knowledge point distribution includes all the knowledge points in the course. This knowledge point distribution is generally determined through the course syllabus and teaching objective evaluation in the course design stage and can be specified by the teacher user. The present invention needs to use a knowledge point mastery degree evaluation model to predict the mastery degree score of the target student user on each knowledge point based on the student feature vector. The mastery degree score can adopt a set of discrete score labels, and the specific score range can be designed according to actual needs. For example, in one embodiment, the mastery degree score range can be set from 0 to 5 points. 0 points represent completely not mastered, 5 points represent completely mastered, and the intermediate 1, 2, 3, and 4 points represent the gradually increasing mastery degree in turn.

[0041] In the embodiment of the present invention, the above knowledge point mastery degree evaluation model can be trained and obtained through a simple linear transformation model or a non-linear transformation model. For example, models such as a multi-layer perceptron, a random forest, a support vector machine, or a convolutional neural network can be used. After initialization, supervised training is carried out on the labeled data set, and the loss used for training can be the mean square error (MSE) between the predicted value and the labeled value of the mastery degree score distribution on all knowledge points.

[0042] In addition, in the embodiment of the present invention, the above knowledge point coverage is calculated based on the expected distribution of the knowledge point mastery degree predefined by the teacher user and the mastery degree scores of the target student user on all knowledge points. Among them, the expected distribution of the knowledge point mastery degree can be predefined by the teacher user, and an expected mastery degree score to be achieved needs to be set for the importance of each knowledge point. For example, assuming that the complete knowledge point distribution includes 4 knowledge points A, B, C, and D, the teacher user can set the expected mastery degree scores that the students need to achieve according to their respective importance in the course. For example, knowledge points A and B are extremely important and must be completely mastered, and the expected mastery degree scores can be set to 5 points each. Knowledge point C is of general importance, and the expected mastery degree scores of both can be set to 3 points. Knowledge point D is less important, and the expected mastery degree score can be set to 2 points. Thus, the expected distribution of the knowledge point mastery degree corresponding to the complete knowledge point distribution [A, B, C, D] is [5, 5, 3, 2].

[0043] In an embodiment of the present invention, based on a predefined expected distribution of knowledge point mastery levels and the mastery level scores of a target student user on all knowledge points given by a knowledge point mastery level evaluation model, the coverage of knowledge points currently mastered by the target student user can be calculated. The specific calculation method is as follows: For each knowledge point in the complete knowledge point distribution, compare the expected mastery level score corresponding to the knowledge point in the expected distribution of knowledge point mastery levels with the mastery level score of the target student user on this knowledge point, and take the smaller value as the representation value. Then, calculate the ratio of the sum of the representation values of all knowledge points to the sum of the expected mastery level scores of all knowledge points. The calculated ratio is used as the current knowledge point coverage of the target student user. The calculation method of this knowledge point coverage can be expressed by the formula:

[0044]

[0045] where: K is the predefined expected distribution of knowledge point mastery levels, is the distribution of mastery level scores of the target student user on all knowledge points, and m is the total number of knowledge points in the complete knowledge point distribution. is the expected mastery level score corresponding to the i-th knowledge point, is the mastery level score of the target student user on the i-th knowledge point. In the above formula, is used to calculate the overlapping part of the expected distribution of knowledge point mastery levels K and the distribution of students' knowledge mastery level scores to reflect the true mastery level of students on each knowledge point. represents the sum of the expected mastery level scores of all knowledge points, which is used as a measurement standard to ensure that the knowledge point coverage is between [0, 1]. The higher the knowledge point coverage , the more the question content can meet the learning needs of students on weak knowledge points.

[0046] S3. Extract a default question set to be recommended to the target student from the current question bank, and analyze the question difficulty distribution of the default question set on all knowledge points; convert the mastery level scores of the target student user on all knowledge points into attention weights for each knowledge point through an attention mechanism, and then use the attention weights to dynamically adjust the question difficulty distribution. Increase the question difficulty corresponding to the knowledge points where the attention weights exceed the weight threshold, and decrease the question difficulty corresponding to the knowledge points where the attention weights are lower than the weight threshold to obtain an adjusted question difficulty distribution.

[0047] It should be noted that when extracting the default question set to be recommended to the target student from the current question bank, the extraction can be random or selected according to the original recommendation rules of the platform, and no specific limitation is made.

[0048] In addition, it should be noted that the current question bank in the present invention refers to the question bank existing on the online learning platform that can be recommended to students. The question bank contains a series of test questions, and each test question has corresponding associated knowledge points and question difficulties. Moreover, all the test questions in the question bank have been pre-created or assembled into a question set by teacher users or administrators. In the traditional recommendation mechanism of online learning platforms, the question sets pushed to each student user at a certain stage are the same. However, in the present invention, it is necessary to adjust the question difficulty and the distribution of knowledge points according to the current personalized status of the student, and then form a new question set.

[0049] In an embodiment of the present invention, assuming that the student feature vector of the target student user is X, through the knowledge point mastery degree evaluation model ) map the student feature vector to the knowledge point mastery degree of the target student on each knowledge point After that, for i = 1, 2, ……, m, the attention weight of each knowledge point can be further calculated through the attention mechanism , to reflect the relative importance of this knowledge point to the student. The calculation formula of the attention mechanism is as follows:

[0050]

[0051] Then, the above attention weights (i = 1, 2, ……, m) can be used to dynamically adjust the question difficulty distribution of the default question set on all knowledge points. Since the question difficulty of the test questions is generally discrete labels, the present invention can preset a single difficulty adjustment amount to adjust the original question difficulty of each test question . The question difficulty on the i-th knowledge point can be dynamically adjusted to :

[0052]

[0053] Wherein, is the original question difficulty on the i-th knowledge point, is the adjusted question difficulty on the i-th knowledge point, is the attention weight of the student on the i-th knowledge point. Since the attention weight It has been normalized, so its relative size can reflect the students' mastery of knowledge points. When it is relatively high, it indicates that the students have a relatively high level of mastery of this knowledge point, and the difficulty can be appropriately increased on the basis of the original difficulty to facilitate consolidation and advanced learning. When it is relatively low, it indicates that the students' mastery of this knowledge point is relatively weak. Therefore, the difficulty can be appropriately reduced on the basis of the original difficulty to avoid the test questions being too difficult, enabling the students to study on relatively simple test questions and gradually master this knowledge point.

[0054] The above specific value can be adjusted according to the actual situation. For example, if the highest difficulty score for the test questions on a knowledge point is set to 5 points and the lowest is 1 point, then can be set to 1 point, that is, each time the difficulty level is adjusted by 1. Of course, the can also be set in a non-fixed form, and is set to be adjusted according to and the difficulty level adjusted each time is determined by

[0055] Thus, for the difficulty distribution of the test questions in the default test question set on all knowledge points , after obtaining the adjusted test question difficulty for each knowledge point in the complete knowledge point distribution , the adjusted test question difficulty distribution can be constructed, and this distribution will be used as the input for the subsequent large language model.

[0056] S4. Combine the default test questions of the target student, the mastery degree scores of the target student user on all knowledge points, the knowledge point coverage, and the adjusted test question difficulty distribution, and construct them into a prompt text according to the prompt template. Then input the prompt text into the personalized question generation model obtained by fine-tuning the large language model (LLM) on the question generation task, and the personalized question generation model outputs a new test question set and pushes it to the target student user.

[0057] It should be noted that in the process of the personalized question generation model outputting a new test question set, it is preferred that the model re-selects questions with different difficulties and related knowledge points from the existing question bank to replace the existing questions in the default test question set. Of course, based on the generative output ability of the large language model itself, the large language model can also automatically generate questions that meet the requirements to replace the existing questions in the default test question set, but it is necessary to ensure that the large language model has the ability to generate correct and reasonable questions, and it is best for the large language model to be further fine-tuned on the task of generating test questions.

[0058] ​It should be noted that the specific method of setting the prompt template above belongs to the conventional method in Prompt engineering. The prompt template can be designed according to the task requirements, and placeholders can be preset in the prompt template for each personalized input content, so that the program can automatically replace the placeholders and generate the prompt text for the input by the large language model (LLM). An exemplary prompt template is as follows: "Please perform a task of personalized recommendation of test questions for students. The current existing default test questions are [placeholder 1], the current mastery level scores of the current student on all knowledge points are [placeholder 2], the knowledge point coverage of the current student is [placeholder 3], and the adjusted question difficulty distribution for the current student should be [placeholder 4]. Based on the default test questions of the student, and considering the mastery level scores of the student on all knowledge points, the knowledge point coverage, and the adjusted question difficulty distribution, extract questions from the existing question bank to adjust the question composition in the default test questions, and obtain new test questions that meet the personalized learning needs of the student."

[0059] The [placeholder 1], [placeholder 2], [placeholder 3], and [placeholder 4] in the above prompt template respectively correspond to the default test questions of the target student, the mastery level scores of the target student on all knowledge points, the knowledge point coverage, and the adjusted question difficulty distribution. The replacement of the placeholders in the prompt template can be automated through a script program to facilitate the automation of the present invention.

[0060] In addition, it should be noted that the above S1 - S4 describe the practices of this personalized dynamic question generation method in actual applications. However, those skilled in the art should be aware that the above large language model (LLM) needs to be pre - fine - tuned to be used as a personalized question generation model. In the embodiments of the present invention, when fine - tuning the large language model (LLM) for the question generation task, the model parameters of the large language model can be optimized with the goal of maximizing the sum of the knowledge point coverage and the knowledge point difficulty matching degree to obtain a personalized question generation model. In the concept of this embodiment, the knowledge point difficulty matching degree can be set as the mean square error between the adjusted question difficulty distribution input to the model and the question difficulty distribution of the new test question set output by the model.

[0061] This optimization goal can be expressed by the formula:

[0062]

[0063]

[0064] Among them, Match represents calculating the matching degree between two distributions; T is the learning cycle, and at each moment within the cycle, test question recommendations need to be made for students. and respectively represent the corresponding ones of the student user at the t-th moment and . and are K and . are the learnable parameters of the large language model, are the learnable parameters after the large language model is optimized.

[0065] In the above optimization objective, needs to be as large as possible. However, based on the general loss form of model optimization, the loss value usually needs to be as small as possible. Therefore, in actual fine-tuning, the above optimization objective can be replaced by minimizing the following loss function:

[0066]

[0067]

[0068] Thus, the questions generated by the present invention not only cover the weak knowledge points of students, but also expand the knowledge scope in combination with future goals, improving the learning effect of students.

[0069] In addition, in the actual application process of the above personalized question generation model, it is necessary to construct an incremental learning dataset by collecting feedback data, with the maximization of the sum of the knowledge point coverage and the knowledge point difficulty matching degree as the optimization objective, and regularly update and optimize the model parameters of the personalized question generation model through the Bayesian optimization algorithm.

[0070] In the embodiment of the present invention, the learning progress of students can be monitored in real time through the background management system on the online learning platform, including the number of completed questions, learning time, knowledge point mastery, etc., and the learning progress data is fed back to the dynamic adjustment module. At the same time, collect the feedback of students on the questions, including the answering accuracy rate, answering time, subjective satisfaction, etc. These data can all be used to construct an incremental learning dataset, and the parameters of the personalized question generation model are incrementally trained (Incremental Training) through the Bayesian Optimization algorithm to maximize the learning effect of students . According to the optimized model parameters , the question generation strategy can be dynamically adjusted, and the question difficulty and knowledge point distribution can be updated to ensure that the question generation model can adapt to the learning changes of students in real time and improve the accuracy of personalized question generation.

[0071] Similarly, based on the same inventive concept, as Figure 2 shown, in another preferred embodiment of the present invention, a personalized dynamic question generation system based on a large language model is further provided, which includes:

[0072] An information collection module for recording multi-dimensional information data of each student user on the online learning platform;

[0073] A target user selection module for allowing users to select target student users for whom personalized test questions need to be pushed;

[0074] A personalized push module for performing personalized test question pushing for the target student users selected by the user according to the multi-dimensional information data recorded in the information collection module, in accordance with the personalized dynamic question generation method based on the large language model as described in the above embodiments.

[0075] It should be noted that the above information collection module can be built in the background server of the online learning platform to record the historical learning records of student users in real time. The target user selection module can be located on the front-end interface of the online learning platform, providing specified functions through buttons or other means. Of course, the target user can also be default-set in the target user selection module. If the management user does not modify the default settings, the target student users will be selected according to the default settings. Each student user on the platform can be used as an object for personalized question generation. The push method in the personalized push module can be implemented through fixed modules or display logics in the user interface, or in the form of system messages, pop-ups, or emails, text messages, etc.

[0076] In addition, it should be noted that the personalized dynamic question generation method based on the large language model described in S1~S4 in the above embodiments can essentially be executed by a computer program or module.

[0077] Therefore, similarly, based on the same inventive concept, in another preferred embodiment of the present invention, there is also provided a computer program product corresponding to the personalized dynamic question generation method based on the large language model provided in the above embodiments, including a computer program / instructions, which when executed by a processor, can implement the personalized dynamic question generation method based on the large language model in the above embodiments.

[0078] Similarly, based on the same inventive concept, as Figure 3 shown, in another preferred embodiment of the present invention, there is also provided a computer electronic device corresponding to the personalized dynamic question generation method based on the large language model provided in the above embodiments, which includes a memory and a processor;

[0079] The memory is used for storing a computer program;

[0080] The processor is used for implementing the personalized dynamic question generation method based on the large language model in the above embodiments when executing the computer program.

[0081] In addition, when the logic instructions in the above-mentioned memory are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention.

[0082] Therefore, based on the same inventive concept, in another preferred embodiment of the present invention, there is also provided a computer-readable storage medium corresponding to the personalized dynamic question generation method based on a large language model provided in the above embodiment. A computer program is stored on the storage medium, and when the computer program is executed by a processor, it can implement the personalized dynamic question generation method based on a large language model in the above embodiment.

[0083] It can be understood that the above storage medium may include a random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory. At the same time, the storage medium may also be various media that can store program codes, such as a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc.

[0084] It can be understood that the above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0085] In addition, it should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the foregoing method embodiments, and will not be elaborated herein. In the embodiments provided in the present application, the division of steps or modules in the system and method is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0086] The above-described embodiments are only some preferred implementation solutions of the present invention, but are not intended to limit the present invention. Those of ordinary skill in the relevant art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by means of equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A personalized dynamic question-setting method based on a large language model, characterized in that: include: S1. Obtain multi-dimensional information data of target student users, including historical answer records, historical scores of various subjects, and learning direction guidance information, and construct them into structured student feature vectors through feature engineering; S2. Input the student feature vector into the pre-trained knowledge point mastery evaluation model to obtain the mastery score of the target student user on each knowledge point in the complete knowledge point distribution, and then calculate the knowledge point coverage based on the mastery score of the target student user on all knowledge points; The knowledge point coverage is calculated based on the expected distribution of knowledge point mastery predefined by the teacher user and the mastery scores of the target student users on all knowledge points. The calculation method is: for each knowledge point in the complete knowledge point distribution, the expected mastery score corresponding to the knowledge point in the expected distribution of knowledge point mastery is compared with the mastery score of the target student user on the knowledge point, and the smaller one is taken as the representation value. Then, the ratio of the sum of the representation values ​​of all knowledge points and the sum of the expected mastery scores of all knowledge points is calculated, and the calculated ratio is used as the current knowledge point coverage of the target student user; S3. Extract the default test question set to be recommended to the target student from the current question bank, and analyze the difficulty distribution of the default test question set on all knowledge points; convert the mastery scores of the target student users on all knowledge points into attention weights for each knowledge point through the attention mechanism, and then dynamically adjust the difficulty distribution of the questions using the attention weights, increase the difficulty of the questions corresponding to the knowledge points whose attention weights exceed the weight threshold, and decrease the difficulty of the questions corresponding to the knowledge points whose attention weights are lower than the weight threshold, to obtain the adjusted difficulty distribution of the questions; S4. The default test questions of the target students, the mastery scores of the target student users on all knowledge points, the knowledge point coverage, and the adjusted question difficulty distribution are combined with the prompt word template to construct a prompt word text. The prompt word text is input into a personalized question setting model obtained by fine-tuning the large language model on the question setting task. The personalized question setting model outputs a new test question set and pushes it to the target student users.

2. The personalized dynamic question-setting method based on a large language model as claimed in claim 1, characterized in that: The learning direction guidance information includes the subjects that need to be studied emphatically and / or the knowledge points that need to be studied emphatically in each subject given to the target student user.

3. The personalized dynamic question-setting method based on a large language model as claimed in claim 1, characterized in that: The knowledge point mastery degree assessment model is obtained by performing supervised training of a multi-layer perceptron, a random forest, a support vector machine or a convolutional neural network on a labeled data set.

4. The personalized dynamic question-setting method based on a large language model as claimed in claim 1, characterized in that: When fine-tuning the large language model on the question-setting task, the model parameters of the large language model are optimized with the maximization of the sum of knowledge point coverage and knowledge point difficulty matching as the optimization goal to obtain a personalized question-setting model; the knowledge point difficulty matching is the mean square error between the adjusted question difficulty distribution of the input model and the question difficulty distribution of the new test question set output by the model.

5. The personalized dynamic question-setting method based on a large language model as claimed in claim 4, characterized in that: In actual application, the personalized question-setting model needs to construct an incremental learning data set by collecting feedback data, take maximizing the sum of knowledge point coverage and knowledge point difficulty matching as the optimization goal, and regularly update and optimize the model parameters of the personalized question-setting model through the Bayesian optimization algorithm.

6. A personalized dynamic question setting system based on a large language model, characterized in that: include: Information collection module, used to record multi-dimensional information data of each student user on the online learning platform; The target user selection module is used for users to select target student users who need to receive personalized test questions. A personalized push module is used to push personalized test questions for target student users selected by the user according to the personalized dynamic question-setting method based on a large language model as described in any one of claims 1 to 5 based on the multi-dimensional information data recorded in the information collection module.

7. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the personalized dynamic question-setting method based on a large language model as described in any one of claims 1 to 5 is implemented.

8. A computer electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the personalized dynamic question-setting method based on a large language model as described in any one of claims 1 to 5 when executing the computer program.

Images