An Adaptive Question Generation Method and System Based on Retrieval Enhancement Algorithm

Abstract

The present invention relates to an adaptive question generation method and system based on a retrieval enhancement algorithm. The method includes collecting the user's answer record data through a data collection module to obtain a user state vector and a feedback label; adjusting the question generation strategy according to the user state vector through a reinforcement learning unit, and transferring the learning experience of other users to the adjustment of the question generation strategy of new users through a transfer learning unit; using the retrieval enhancement algorithm to retrieve through a knowledge graph in a knowledge base, calculating the similarity between the user state and the nodes in the knowledge graph, and selecting the node with the highest similarity as the retrieval result; generating questions using a graph-based retrieval enhancement generation algorithm and a large language model, and pushing them to the user. Compared with the prior art, the present invention has the advantages of high personalization, real-time matching, and diversification.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and in particular to an adaptive question generation method and system based on a retrieval enhancement algorithm. Background Art

[0002] With the rapid development of information technology, the digital transformation of the education industry is gradually advancing. The traditional teaching mode often relies on teachers to design test questions and set questions according to a one-size-fits-all standard, making it difficult to make personalized adjustments according to the learning progress and knowledge mastery of each student. At present, some intelligent education platforms have adopted data analysis and machine learning technologies for personalized teaching, but they are still limited to static knowledge point recommendations, question generation, or the push of review content. The existing automatic question generation technology relies on pre-set rules or templates to generate questions, lacking flexibility and intelligent control in the question generation process, and it is difficult to simulate the thinking process and question-setting strategies of teachers, resulting in the generated test questions being inaccurate and diverse enough in practical applications. In addition, the question-setting system is difficult to effectively simulate the creative thinking of teachers during the question-setting process. When setting questions, teachers not only consider the learning situation of students, but also flexibly adjust the type, difficulty, and knowledge coverage of questions according to teaching objectives, the difficulty of knowledge points, and the learning background of students. The deficiencies of the existing automated question-setting systems lead to the lack of innovation and pertinence in the generated questions. At the same time, the learning state of students is dynamically changing. A single static question bank and fixed question-setting rules only generate a fixed set of test questions based on the historical data or one-time assessment results of students, lacking the ability to adjust the test question content in real time. This static generation method cannot timely track the weak points of students' knowledge and learning progress, and has poor results in personalized teaching. In recent years, retrieval enhancement algorithms have made remarkable progress in the field of natural language processing. The retrieval enhancement algorithm model combines two technical means of information retrieval and text generation, retrieves relevant information from a large-scale external database, and combines the generation model to generate specific content. However, the application of retrieval enhancement algorithms in the education industry is still in its initial stage. There is still a lack of mature technical solutions and implementation methods for combining retrieval enhancement algorithms with the historical learning data of students to dynamically generate test questions that meet the personalized needs of students.

[0003] Chinese Patent Application CN118861240A discloses a method and device for generating subject test questions based on large language models and retrieval enhancement. The method for generating subject test questions based on large language models and retrieval enhancement converts them into test questions in json format through a format parser based on large language models, and then flattens them into text-form test questions through rules; uses the flattened test questions to retrieve relevant texts in textbooks; inputs the flattened test questions, question types, knowledge points, and retrieved relevant texts into a question generator based on large language models to generate test questions. However, this application has the deficiency of not being dynamically adjusted according to user feedback. While generating test questions with retrieval enhancement, it mainly relies on knowledge points in textbooks and lacks personalized design. Therefore, how to automatically generate test questions suitable for the individual characteristics of users according to the user's learning progress, knowledge mastery, and weak links is a technical problem to be solved. Summary of the Invention

[0004] The purpose of the present invention is to provide an adaptive question generation method based on a retrieval enhancement algorithm. By generating question generation metrics through reinforcement learning, optimizing data through transfer learning, retrieving through a retrieval enhancement algorithm, and generating test questions based on a graph-based retrieval enhancement generation algorithm and large language models, it solves the problems of insufficient accuracy in personalized question generation and weak dynamic adjustment ability in the prior art, and realizes real-time matching between test question generation and students' learning status.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] According to one aspect of the present invention, an adaptive question generation method based on a retrieval enhancement algorithm is provided. The specific steps include:

[0007] S1. The data collection module collects the user's answer record data, filters out outliers and performs data preprocessing to obtain a user status vector and feedback labels, and inputs them into the optimization and continuous learning module;

[0008] S2. The optimization and continuous learning module adjusts the test question generation strategy according to the user status vector through the reinforcement learning unit, and transfers the learning experience of other users to the adjustment of the test question generation strategy of new users through the transfer learning unit;

[0009] S3. According to the user status vector and feedback labels, a retrieval enhancement algorithm is used to retrieve through a knowledge graph in the knowledge base, calculate the similarity between the user status and the nodes in the knowledge graph, and select the node with the highest similarity as the retrieval result;

[0010] S4. The test question generation module generates test questions using a graph-based retrieval enhancement generation algorithm and large language models according to the user status vector, test question generation strategy, and retrieval result, and pushes them to the user and stores them in the knowledge base.

[0011] Furthermore, the answer record data in S1 includes user ID, answer, answer correctness, answering duration, current learning progress mark, and subjective feedback text; extract feedback keywords from the subjective feedback text and convert them into structured feedback tags, where the feedback tags include feedback question difficulty and corresponding feedback knowledge point name; preprocess the answer record data to obtain the user's answering accuracy rate, answer error type code, and quantified learning progress.

[0012] Furthermore, the state space of the reinforcement learning unit in S2 is the user state vector, including answering accuracy rate, answer error type code, answering duration, and quantified learning progress; the action space of the reinforcement learning unit includes adjusting the question difficulty weight and adjusting the knowledge point coverage.

[0013] In the reinforcement learning unit, a reward function is set according to the user state vector to drive the adjustment of the question generation strategy, and the reward function has the following expression:

[0014] ,

[0015] where , , and are reward weights; is the answering accuracy rate; is the error type code; is the total number of error types; is the answering duration; is the quantified learning progress; is the label of the current user state vector;

[0016] The adjustment of the question generation strategy is carried out through the value function, and the expression is:

[0017] ,

[0018] where is the current policy; is the next moment policy parameter; is the step size amplitude parameter used to control the adjustment of the question generation strategy; is the reward function; is the user state vector value function.

[0019] Furthermore, the transfer learning in S2 uses a long short-term memory network to predict the user's current knowledge weak points based on the user's historical answer record data, and feedbacks them to the reinforcement learning unit based on the knowledge weak points.

[0020] Cluster according to the subjects selected by the user. When a user with insufficient historical answering record data enters, calculate the cosine similarity of the user state vectors between the current user and each cluster based on the existing answering record data, and select the most similar cluster for migration; the judgment basis for insufficient historical answering record data is that the number of operations of this user is lower than the first threshold. When the number of times of using transfer learning data accumulates to the first threshold, reduce the transfer learning weight according to the preset step size and transition to pure reinforcement learning.

[0021] Further, in S3, the nodes in the knowledge graph include the knowledge point name and the difficulty of the question, the edge attributes include the relationship type and the similarity score between nodes. In the knowledge graph, the relationship type is established by comparing the similarity score between nodes with the second threshold. The relationship types include belonging to knowledge points, associated wrong questions, easily confused knowledge points, prerequisite knowledge points, and extended knowledge points.

[0022] Calculate the user similarity between the user state vector and the nodes, and expand the similarity of the nodes corresponding to the feedback labels by a preset ratio. Sort according to the obtained user similarity, and select a preset number of nodes with the largest similarity to form a retrieval set as the retrieval result, and calculate the user similarity The expression of

[0023] is:

[0024] where is the current user state vector, is the embedded representation of the nodes in the knowledge graph; is the element in the node set .

[0025] Further, the generated questions in S4 include the question surface, the answer, the difficulty level, and the corresponding knowledge points. The generated questions also need to go through difficulty verification, and the verification is that the difficulty level number needs to be lower than the answering correct rate in the user state vector multiplied by a preset multiple.

[0026] According to another aspect of the present invention, a system for adaptive question generation based on a retrieval enhancement algorithm is provided. The system includes a data collection module, an optimization and continuous learning module, a retrieval enhancement module, and a question generation module;

[0027] The data collection module collects the answering record data of the user, filters out outliers and performs data preprocessing to obtain the user state vector and the feedback label and inputs them into the optimization and continuous learning module;

[0028] The optimization and continuous learning module adjusts the question generation strategy according to the user state vector through the reinforcement learning unit, and transfers the learning experience of other users to the adjustment of the question generation strategy for new users through the transfer learning unit;

[0029] The retrieval enhancement algorithm module uses the retrieval enhancement algorithm to retrieve through the knowledge graph in the knowledge base according to the user state vector and feedback label, calculates the similarity between the user state and the nodes in the knowledge graph, and selects the node with the highest similarity as the retrieval result;

[0030] The question generation module uses the graph-based retrieval enhancement generation algorithm and large language model to generate candidate questions based on the user state vector, question generation strategy and retrieval result, and pushes them to the user and stores them in the knowledge base.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) Improve the accuracy of personalized question setting: By integrating reinforcement learning and transfer learning technologies, a dynamically adjustable personalized question setting mechanism is constructed. The reinforcement learning module takes the user's real-time answering data as the core input, and drives policy iteration optimization through the reward function to ensure that the question difficulty and knowledge point distribution are accurately matched with the user's current ability. The transfer learning module uses a pre-trained long short-term memory network to quickly inherit the learning experience of similar groups when the new user data is insufficient, solving the cold start problem. This technical collaboration enables the system to more accurately capture the user's real-time learning state and dynamically generate personalized and adaptable questions.

[0033] (2) Achieve efficient and diverse question generation: The system combines the knowledge graph and retrieval enhancement algorithm with the generation ability of the large language model to significantly improve the question generation efficiency and diversity. The knowledge graph is constructed based on the knowledge base and user wrong questions. The nodes cover knowledge points, questions and reference materials, and the edge relationship supports semantic depth retrieval. In the retrieval stage, the graph-based retrieval enhancement generation algorithm and large language model are used to preferentially return questions highly relevant to the user's weak points, covering the learning needs of different cognitive levels.

[0034] (3)Support the closed-loop optimization of dynamic learning tracking and question generation strategies: The system can implement a closed-loop process of data collection, strategy update, question generation, and feedback verification, achieving continuous tracking and adaptive optimization of the user's learning progress. The data collection module synchronizes the user's answering behavior in real time, filters outliers, and converts unstructured feedback into quantitative metrics. The optimization and continuous learning module dynamically adjusts the question generation strategy based on real-time data. In the retrieval enhancement algorithm module, the knowledge graph retrieves dynamically according to the user's current situation, and the subsequent retrieval can be made more targeted through different edge attribute relationship types. In addition, the difficulty verification mechanism requires that the difficulty level of the generated questions does not exceed a preset multiple of the user's correct rate. The entire system's question production process forms a closed-loop optimization, which is conducive to generating high-quality questions that better meet the user's personalized needs and improves the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of an adaptive question generation method based on a retrieval enhancement algorithm;

[0036] Figure 2 It is a structural diagram of an adaptive question generation system based on a retrieval enhancement algorithm. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.

[0038] Embodiment 1

[0039] This embodiment is an adaptive question generation method based on a retrieval enhancement algorithm, and the specific steps include:

[0040] S1. The data collection module collects the user's answering record data, filters outliers, and performs data preprocessing to obtain the user state vector and feedback label, and inputs them into the optimization and continuous learning module;

[0041] S2. The optimization and continuous learning module adjusts the question generation strategy according to the user state vector through the reinforcement learning unit, and transfers the learning experience of other users to the adjustment of the question generation strategy of new users through the transfer learning unit;

[0042] S3. According to the user state vector and feedback label, the retrieval enhancement algorithm is used to retrieve through the knowledge graph in the knowledge base, calculate the similarity between the user state and the nodes in the knowledge graph, and select the node with the highest similarity as the retrieval result;

[0043] S4. The question generation module generates questions according to the user status vector, question generation strategy, and retrieval results, using a graph-based retrieval enhanced generation algorithm and a large language model, and pushes them to the user and stores them in the knowledge base.

[0044] The answer record data includes the user ID, answer, answer correctness, answering duration, current learning progress mark, and subjective feedback text. Extract feedback keywords from the subjective feedback text and convert them into structured feedback tags, including the feedback question difficulty and the corresponding feedback knowledge point name. Preprocess the answer record data to obtain the user's answering correct rate, answer error type code, and quantified learning progress. The answer error type code is sorted from difficult to easy and from large to small according to the correction difficulty of the wrong answer judged by experts. The quantified learning progress is the percentage of the current learning progress mark in the total learning task.

[0045] The reinforcement learning unit in the optimization and continuous learning module has a state space of user status vectors, including answering correct rate, answer error type code, answering duration, and quantified learning progress; its action space includes adjusting the question difficulty weight and adjusting the knowledge point coverage. Continuously adjust the question generation strategy based on the user's answer record data, and use the student's performance feedback to optimize the difficulty and type of questions.

[0046] Set a reward function in the reinforcement learning unit according to the user status vector to drive the adjustment of the question generation strategy. The reward function has the following expression:

[0047] ,

[0048] where , , and are reward weights; is the answering correct rate; is the error type code; is the total number of error types; is the answering duration; is the quantified learning progress; is the label of the current user status vector. The reward function is calculated based on data from multiple dimensions shown by the user during the answering process, used to measure the user's performance on specific questions, and provide a basis for subsequent personalized question generation strategies. By comprehensively analyzing the user's performance, dynamically adjust the difficulty and type of questions according to the learning status to improve the effect of personalized teaching.

[0049] The adjustment of the question generation strategy is carried out through the value function, and the expression is:

[0050] ,

[0051] Among them, is the current strategy; is the strategy parameter at the next moment; is the step amplitude parameter used to control the adjustment of the test question generation strategy; is the reward function; is the user state vector value function.

[0052] The transfer learning in S2 uses a long short-term memory network to predict the current knowledge weak points of the user based on the user's historical answering record data, and feeds them back to the reinforcement learning unit based on the knowledge weak points. Clustering is performed according to the subject selected by the user. When a user with insufficient historical answering record data enters, the cosine similarity of the user state vector between the current user and each cluster is calculated based on the existing answering record data, and the most similar cluster is selected for transfer; the judgment basis for insufficient historical answering record data is that the number of operations of this user is lower than the first threshold. When the number of times of using transfer learning data accumulates to the first threshold, the transfer learning weight is reduced according to the preset step size, and it transitions to pure reinforcement learning. Transfer learning can transfer the learning experience of other users to new users, helping the system to adapt to the learning progress and weak links of new users faster. In this process, the incremental update mechanism plays a key role. It converts the user's answering record data into the basis for system update, and adjusts the personalized learning path of the student through the updated data to ensure continuous optimization and adaptation to the changing needs of users during long-term operation, improving the user experience.

[0053] In S3, the nodes in the knowledge graph include the knowledge point name and the difficulty of the test questions, and the edge attributes include the relationship type and the similarity score between nodes. The relationship type in the knowledge graph is established by comparing the similarity score between nodes with the second threshold. The relationship types include belonging to knowledge points, associated wrong questions, easily confused knowledge points, prerequisite knowledge points, and extended knowledge points. The construction of the knowledge graph is based on textbooks, question banks, and reference materials in the knowledge base and is obtained through natural language processing and relationship extraction.

[0054] Calculate the user similarity between the user state vector and the nodes, and expand the similarity of the nodes corresponding to the feedback labels by a preset ratio. Sort according to the obtained user similarity, and select a preset number of nodes with the largest similarity to form a retrieval set as the retrieval result. Calculate the user similarity The expression of

[0055] ,

[0056] Among them, is the current user state vector, is the embedded representation of the nodes in the knowledge graph; is the node set The elements in it. The higher the similarity, the more relevant the node is to the user's current learning state, and questions related to these nodes are recommended first.

[0057] The test questions generated in S4 include the question stem, answers, difficulty levels, and corresponding knowledge points. The generated test questions also need to go through a difficulty verification, and the verified difficulty level number needs to be lower than the preset multiple multiplied by the answering correct rate in the user state vector. The generated test questions will meet the personalized needs of users, ensure that each user receives appropriate challenges in the knowledge points they have mastered, and promote their improvement in weak links. Through this test question generation method based on user data, feedback, and knowledge graphs, this embodiment can provide accurate and personalized educational services, improve the learning effect of users, while reducing the burden of teachers in creating questions and promoting the implementation of personalized teaching.

[0058] Embodiment 2

[0059] This embodiment provides a system for adaptive question generation based on a retrieval enhancement algorithm. The system includes a data collection module, an optimization and continuous learning module, a retrieval enhancement module, and a test question generation module.

[0060] The data collection module collects the user's answering record data, filters out outliers, and performs data preprocessing to obtain the user state vector and feedback labels and inputs them into the optimization and continuous learning module. These data are used to evaluate the student's mastery of questions, common error types, and problem-solving speed, etc., so as to provide a basis for personalized question generation. The collected data will be cleaned and preprocessed to ensure its consistency and accuracy, so as to provide high-quality data support for the subsequent test question generation module. All the collected data are uniformly stored in the learning database, supporting efficient retrieval and management, providing data support for the system, and ensuring that the subsequent modules can generate personalized test questions through accurate data analysis.

[0061] The optimization and continuous learning module includes a reinforcement learning unit and a transfer learning unit, which are responsible for continuously optimizing the question generation strategy by analyzing the user's answer record data, and improving the accuracy and adaptability of personalized question generation. Based on machine learning algorithms, this module can continuously learn during the system operation, gradually adapt to the personalized needs of different students, ensure the continuous optimization of the question generation process, and thus improve the educational effect. Through the reinforcement learning unit, the question generation strategy is adjusted according to the user state vector, and through the transfer learning unit, the learning experience of other users is transferred to the adjustment of the question generation strategy for new users, realizing the dynamic adjustment of the user's personalized learning state, thereby driving the optimization of subsequent question generation and ensuring the accuracy and adaptability of question generation. Reinforcement learning technology enables the system to automatically adjust the question difficulty and knowledge point coverage according to the user's performance, while transfer learning enables the system to quickly adapt to the learning progress and weak links of new users, reducing the dependence on a large amount of data. Through periodic training and evaluation, the system continuously optimizes the question generation strategy, avoiding the limitations of traditional static question generation systems and being able to always adapt to the personalized needs of each student. The incremental update mechanism provides the system with the ability of continuous learning based on real-time feedback, making the question generation process always in an optimized state, thereby improving the accuracy of personalized teaching and the educational effect. Through the efficient retrieval of the retrieval enhancement algorithm module and the dynamic combination with the optimization and continuous learning module, it provides the question content support that fits the user's personalized learning needs.

[0062] The retrieval enhancement algorithm module uses the retrieval enhancement algorithm to retrieve through the knowledge graph in the knowledge base according to the user state vector and feedback label, calculates the similarity between the user state and the nodes in the knowledge graph, and selects the node with the highest similarity as the retrieval result. The retrieval enhancement algorithm module deeply associates relevant knowledge points through the relationships of the knowledge graph, thus improving the diversity and accuracy of the retrieval results. According to the user's learning situation and practice needs, the retrieval enhancement module quickly retrieves the most relevant questions or learning resources from the knowledge base. These retrieval results are not just simple questions or resources, but graph-enhanced relevant text fragments, questions, knowledge point descriptions and other information, which can help the system accurately understand the user's learning progress and weak links. The retrieval process of the retrieval enhancement algorithm module is dynamic, and the retrieval strategy is adjusted in real time according to the user's learning behavior and feedback. For example, the system will timely identify the user's weak links in specific fields or knowledge points according to the user's answer correct rate and error types, and select relevant content from the knowledge base for feedback. By combining the user's performance feedback, the system continuously adjusts and optimizes the retrieval strategy to ensure that the content retrieved each time can help the user progress to the greatest extent.

[0063] The question generation module generates candidate questions using a graph-based retrieval-augmented generation algorithm and a large language model based on the user status vector, question generation strategy, and retrieval results, and pushes them to the user and stores them in the knowledge base.

[0064] The system performance of this embodiment was comprehensively evaluated through the feedback and testing of actual users. The experimental results are shown in Table 1. According to the subjective evaluations of power industry experts and learners, the system in this embodiment performs excellently in several key aspects. First, regarding the question quality, the average score of experts is 4.7, the average score of learners is 4.5, and the overall average score is 4.6, indicating that the questions generated by the system are widely recognized in terms of quality and can effectively meet the needs of different users. Second, in terms of difficulty alignment, the average score of experts is 4.6, that of learners is 4.3, and the overall score is 4.45, showing that the difficulty of the questions generated by the system is highly aligned with the actual level of users, avoiding overly simple or complex questions, thus enhancing the learning experience of users. In addition, the acceptance of the system has been highly recognized by experts and learners. The average score of experts is 4.8, that of learners is 4.6, and the overall score is 4.7, reflecting the wide recognition and high user satisfaction of the system among the user group.

[0065] Table 1 Subjective evaluations of users of the adaptive question generation system

[0066]

[0067] The differences in the scores of experts and learners in each item are not significant, indicating that the system in this embodiment can better meet the needs of different user groups in terms of question quality, difficulty adjustment, and acceptance, and has received unanimous praise from users. The system in this embodiment can generate high-quality questions automatically and is precise and effective in question design, capable of providing high-quality content that meets user expectations. In addition, the system in this embodiment performs excellently in intelligent difficulty adjustment, capable of automatically adjusting the difficulty of questions according to the learning progress of different users to promote more effective learning, making the system not only have wide applicability but also be highly accepted among different groups, enhancing the pertinence and efficiency of learning. At the same time, the automatic generation of questions reduces the cost of manual question making and improves the efficiency of training and learning.

[0068] If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this 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. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

Claims

1. An adaptive question generation method based on a retrieval enhancement algorithm, characterized in that, The specific steps are as follows: S1. The data collection module collects the user's answer record data, filters out outliers and preprocesses the data to obtain the user status vector and feedback tags, and inputs them into the optimization and continuous learning module; the answer record data includes user ID, answers, answer correctness, answering duration, current learning progress marker, and subjective feedback text; feedback keywords are extracted from the subjective feedback text and converted into structured feedback tags, and the feedback tags include the feedback question difficulty and the corresponding feedback knowledge point name; the answer record data is preprocessed to obtain the user's answer correct rate, answer error type code, and quantified learning progress; S2. The optimization and continuous learning module adjusts the question generation strategy according to the user status vector through the reinforcement learning unit, and transfers the learning experience of other users to the adjustment of the question generation strategy of new users through the transfer learning unit; the state space of the reinforcement learning unit is the user status vector, including the answer correct rate, answer error type code, answering duration, and quantified learning progress; the action space of the reinforcement learning unit includes adjusting the question difficulty weight and adjusting the knowledge point coverage; In the reinforcement learning unit, a reward function is set according to the user state vector to drive the adjustment of the test question generation strategy. The reward function has the following expression: ， Among them, , , and are reward weights; is the answering accuracy rate; is the error type code; is the total number of error types; is the answering duration; is the quantified learning progress; is the label of the current user status vector; The adjustment of the question generation strategy is carried out through the value function, and the expression is: ， Among them, is the current strategy; is the strategy parameter at the next moment; is the step amplitude parameter used to control the adjustment of the test question generation strategy; is the reward function; is the user state vector is the value function; The transfer learning is carried out through a long short-term memory network. Based on the user's historical answer record data, the current knowledge weak points of the user are predicted and fed back to the reinforcement learning unit based on the knowledge weak points; clustering is carried out according to the subject selected by the user. When a user with insufficient historical answer record data enters, according to the existing answer record data, the cosine similarity of the user status vector between the current user and each cluster is calculated, and the most similar cluster is selected for transfer; the judgment basis for insufficient historical answer record data is that the number of operations of this user is lower than the first threshold. After the number of times of using transfer learning data accumulates to the first threshold, the transfer learning weight is reduced according to the preset step length, and it transitions to pure reinforcement learning; S3. According to the user status vector and feedback tags, the retrieval enhancement algorithm is used to retrieve through the knowledge graph in the knowledge base, calculate the similarity between the user status and the nodes in the knowledge graph, and select the node with the highest similarity as the retrieval result; S4. The question generation module generates questions according to the user status vector, question generation strategy, and retrieval result, uses the graph-based retrieval enhancement generation algorithm and large language model to generate questions, pushes them to the user, and stores them in the knowledge base.

2. The adaptive question generation method based on a retrieval enhancement algorithm according to claim 1, wherein In S3, the nodes in the knowledge graph include knowledge point names and question difficulties, the edge attributes include relationship types and node similarity scores, and relationship types are established in the knowledge graph by comparing the node similarity scores with the second threshold. The relationship types include belonging to knowledge points, associated wrong questions, easily confused knowledge points, prerequisite knowledge points, and extended knowledge points.

3. The adaptive question generation method based on a retrieval enhancement algorithm according to claim 2, wherein Calculate the user similarity between the user status vector and the node, expand the similarity of the node corresponding to the feedback label by a preset ratio, sort according to the obtained user similarity, select the nodes with the preset number of the largest similarities to form a retrieval set as the retrieval result, and calculate the user similarity The expression of which is: ， Among them, is the current user state vector, is the embedded representation of nodes in the knowledge graph; is the node set is an element in it.

4. An adaptive question generation method based on a retrieval enhancement algorithm according to claim 1, characterized in that The questions generated in S4 include question stems, answers, difficulty levels, and corresponding knowledge points. The generated questions also need to undergo difficulty verification, and the verification is that the level of the difficulty level needs to be lower than the answer correct rate in the user status vector multiplied by a preset multiple.

5. A system for the adaptive question generation method based on the retrieval enhancement algorithm according to any one of claims 1 to 4, characterized in that, The system includes a data acquisition module, an optimization and continuous learning module, a retrieval enhancement module, and a test question generation module; The data acquisition module collects the user's answer record data, filters outliers and preprocesses the data to obtain the user state vector and feedback label, and inputs them into the optimization and continuous learning module; The optimization and continuous learning module adjusts the test question generation strategy according to the user state vector through the reinforcement learning unit, and transfers the learning experience of other users to the adjustment of the test question generation strategy of new users through the transfer learning unit; The retrieval enhancement algorithm module retrieves through the knowledge graph in the knowledge base according to the user state vector and feedback label, uses the retrieval enhancement algorithm to calculate the similarity between the user state and the nodes in the knowledge graph, and selects the node with the highest similarity as the retrieval result; The test question generation module generates candidate test questions using the graph-based retrieval enhancement generation algorithm and large language model according to the user state vector, test question generation strategy, and retrieval result, and pushes them to the user and stores them in the knowledge base.

Images