Question bank generation system and method based on large language model

Abstract

The invention discloses a question bank generation system and method based on a large language model, and the system comprises a student daily ability evaluation monitoring module which is used for collecting student learning data, decomposing question information, adding question difficulty labels, analyzing the mastering level of knowledge points, and forming original wrong question data; the question bank management and updating module is used for storing original wrong question data as a basic wrong question bank, storing a self-defined enhanced question bank and a wrong question variant question bank, and updating all question banks according to new questions; the question bank generation module is used for specifically generating new questions based on a large language model and a diffusion model to form a wrong question variant question bank; the question quality auditing and evaluating module is used for evaluating new questions and adding question difficulty labels to the questions in a dynamic difficulty adjusting mode; and the special practice tracking module is used for making a personalized question practice plan according to the learning condition of the student. Targeted question setting is carried out on each student, a high-quality question bank is generated, and the learning effect of the students is improved.

Description

Technical Field

[0001] The present invention relates to the fields of educational informatization and intelligent learning technology, particularly a question bank generation system that utilizes advanced technologies such as large language models and diffusion models to accurately assess student learning progress and provide personalized, intensive practice on incorrect questions. Specifically, it relates to a question bank generation system and method based on a large language model. Background Art

[0002] In traditional education models, the construction of student practice question banks relies primarily on teachers' experience and various teaching materials, making it difficult to meet students' personalized learning and specialized reinforcement needs. In particular, for certain rare and error-prone question types, students can only rely on repeated or highly similar questions for practice. While some intelligent question bank systems exist, they still need to improve the targetedness, diversity, and intelligence of question bank generation. Existing technologies are particularly deficient in generating high-quality, error-prone question variants that are tailored to students' actual knowledge levels.

[0003] Although some schools have already piloted intelligent systems such as learning analytics and electronic error collection systems, where teachers can collect students' answer sheets and enter them into the system, automatically recording students' errors and forming error collections, there is still insufficient information on how to utilize these errors. The main issues are as follows: (1) Existing technical means basically serve parents and teachers, providing them with information on students’ learning status at school, knowledge points learned, etc.

[0004] (2) The functions that serve students are basically focused on helping students organize wrong questions, matching the same knowledge points or similar wrong questions from the existing question bank, and providing students with learning guidance by combining some artificial intelligence analysis capabilities. However, for some rare and easy-to-make mistakes, when the existing question bank does not have enough questions, students can only practice classic questions repeatedly, making learning more inclined to mechanical memorization rather than truly mastering knowledge points and problem-solving skills.

[0005] (3) When assigning homework to students, teachers often draw from existing question banks. Especially during long holidays such as winter and summer vacations, they often use homework books compiled by major publishers to complete the homework assignments. However, such homework assignments pose the risk of being unable to control the quality of the assignments. Moreover, during holidays, students are most likely to forget knowledge points. Low-quality or scattered questions may have a counterproductive effect on students.

[0006] Therefore, the existing question bank generation system has problems such as being unable to generate personalized practice question banks, being unable to effectively track students' learning progress and locate weak knowledge points to provide targeted questions for each student. Summary of the Invention

[0007] The technical problem to be solved by the present invention is that existing question bank generation systems have problems such as the inability to generate personalized practice question banks, the inability to effectively track students' learning progress and identify weak points to provide targeted questions for each student, and other issues. The purpose of the present invention is to provide a question bank generation system and method based on a large language model. The present invention can generate personalized practice question banks, track students' learning progress and identify weak points to provide targeted questions for each student, generate high-quality question banks, and improve students' learning outcomes.

[0008] The present invention is achieved through the following technical solutions: In a first aspect, the present invention provides a question bank generation system based on a large language model, the system comprising: The student daily ability assessment monitoring module is used to collect student learning data, break down question information based on the student learning data, add question difficulty labels, and analyze the degree of mastery of knowledge points; based on the degree of mastery of knowledge points, calculate the knowledge point ability score to form the original wrong question data; The question bank management and update module is used to store the original wrong question data as the basic wrong question bank, store the customized enhanced question bank and wrong question variant question bank, and update all question banks according to new questions; The question bank generation module is used to generate new questions based on the basic wrong question bank, customized enhanced question bank and knowledge point distribution, using the large language model and diffusion model to form a wrong question variant question bank; The question quality review and evaluation module is used to evaluate the effectiveness and multimodal consistency of new questions, and to add question difficulty labels to questions using dynamic difficulty adjustment. New questions that pass the evaluation will be added to the wrong question variant question bank. The special practice tracking module is used to formulate personalized question practice plans based on students' learning status, and monitor students' question practice completion status and wrong question statistics; and feed back the question practice completion status and wrong question statistics to students' daily ability assessment monitoring module to form a closed-loop optimization.

[0009] Furthermore, the student daily ability assessment and monitoring module includes: Data collection unit, used to collect student learning data, including student test scores, daily homework completion status and classroom exercise content; The mastery level analysis unit is used to decompose question information into three-dimensional vectors of knowledge point labels, cognitive levels, and error patterns based on student learning data. It also adds question difficulty labels to quantitatively analyze students' mastery of knowledge points. The ability score calculation unit is used to calculate the knowledge point ability score based on the temporal attention mechanism according to the degree of mastery of the knowledge points to form the original wrong question data.

[0010] Furthermore, the knowledge point ability score The expression is: ; in, is the total number of time points recorded; For time The attention weight of , which measures the impact of the learning situation at that time point on the ability score of the current knowledge point; For students In time Knowledge points the degree of mastery; is the forgetting rate parameter, which controls the forgetting speed; To simulate the forgetting curve, it means from time To current time The forgetting effect.

[0011] Furthermore, the student daily ability assessment and monitoring module is also used to generate statistical information for teachers and parents based on knowledge point ability scores, and to add weights according to the learning stage and the new and old knowledge to form a comprehensive evaluation report; it is also used to detect changes in grades and trigger early warnings for students whose grades are declining.

[0012] Furthermore, the customized enhanced question bank is a customized question bank that students, parents or teachers plan to enhance specifically based on the current learning stage or learning situation; the wrong question variant question bank is a wrong question variant question bank generated by the question bank generation module based on the large language model and diffusion model.

[0013] Furthermore, the question bank generation module includes: The question template setting unit is used to set preset examples for various question templates as question template prompts, guide the large language model to optimize prompts for common question types and question-answering methods, and control the large model to generate a special question bank; The large language model unit uses the large language model to identify original question examples based on the basic wrong question bank, customized enhanced question bank, and knowledge point distribution. It then generates a wrong question variant bank for students based on different knowledge point scopes and levels of difficulty. The text content of the wrong question variant bank includes the questions and reference answers. The diffusion model unit is used to generate a question image that meets the requirements of the question based on the diffusion model using a reverse process for the question containing an image.

[0014] Furthermore, the question quality review and assessment module includes: The validity assessment unit is used to verify whether the question has a solution based on the student role reasoning model. The analysis information is compared with the reference answer, difficulty label and knowledge point scope when the question was generated to conduct a comprehensive evaluation to confirm the validity of the question. The consistency evaluation unit is used to calculate the image-text similarity based on the CLIP model for questions containing images. Based on the image-text similarity, it determines the image and text comparison loss to align the latent feature space of the image and text. Based on multimodal reasoning and parsing, it compares the parsed information with the reference answer, difficulty label, and knowledge point scope when the question was generated to conduct a comprehensive evaluation to confirm the consistency of the image and text. The difficulty label adjustment unit is used to construct the difficulty matrix of the Bloom taxonomy using a dynamic difficulty adjustment method. It also uses item response theory combined with a large language model to simulate the answers of test-takers and real test-takers on specific questions, conducts a comprehensive dynamic assessment of the difficulty of the questions based on the students' abilities and the accuracy of the answers, and adds difficulty labels to the questions.

[0015] Furthermore, a personalized practice plan is developed based on the student's learning situation, including: Based on the students' learning status, teachers are provided with a function to generate special exercises for students. Specifically, when a teacher wants to conduct special exercises on a certain knowledge point for a student, the teacher will filter the knowledge point questions that need to be practiced from the system according to the classification tags, and set the wrong question ratio to form comprehensive special exercises. Based on students' learning status, we provide parents with a time series forecasting function. Specifically, we use the Prophet algorithm combined with holiday information and seasonal trends to provide parents with a predicted ability development curve for the next three months.

[0016] In a second aspect, the present invention further provides a method for generating a question bank based on a large language model. The method is based on the above-mentioned system for generating a question bank based on a large language model, and the method comprises: Collect student learning data, break down question information and add question difficulty labels based on the data, and analyze the degree of mastery of knowledge points; calculate the knowledge point ability score based on the degree of mastery of knowledge points to form the original wrong question data; Store the original wrong question data as the basic wrong question database, and store the customized enhanced question database and wrong question variant database; Based on the basic wrong question bank, customized enhanced question bank and knowledge point distribution, new questions are generated in a targeted manner based on the large language model and diffusion model to form a wrong question variant question bank; Conduct validity and multimodal consistency evaluations on new questions, and assign difficulty labels to them using dynamic difficulty adjustment. New questions that pass the evaluation are added to the wrong question variant database; and Develop personalized question practice plans based on students' learning status, monitor students' question practice completion status and wrong question statistics; and provide feedback on question practice completion status and wrong question statistics to form a closed-loop optimization.

[0017] Furthermore, we collect student learning data, break down question information based on the data, add question difficulty labels, and analyze the degree of knowledge point mastery. Based on the degree of knowledge point mastery, we calculate the knowledge point ability score to form the original wrong question data, including: Collect student learning data, including test scores, daily homework completion, and classroom exercise content; Based on student learning data, the question information is decomposed into three-dimensional vectors of knowledge point labels, cognitive levels, and error patterns. After adding the question difficulty label, the degree of students' mastery of the knowledge points is quantitatively analyzed; According to the mastery of knowledge points, the knowledge point ability score is calculated based on the temporal attention mechanism to form the original wrong question data; Among them, the knowledge point ability score The expression is: ; in, is the total number of time points recorded; For time The attention weight of , which measures the impact of the learning situation at that time point on the ability score of the current knowledge point; For students In time Knowledge points the degree of mastery; is the forgetting rate parameter, which controls the forgetting speed; To simulate the forgetting curve, it means from time To current time The forgetting effect.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: The present invention provides a system and method for generating a question bank based on a large language model. Based on the existing technology, the present invention expands the pain points of teachers setting questions and students practicing questions through the fusion of multimodal features and the capabilities of the large language model, and determines a set of variant generation processes based on a wrong question bank. The present invention can generate personalized practice question banks in a targeted manner, and can also track students' learning status and locate weak knowledge points to set targeted questions for each student, generate high-quality question banks, and improve students' learning effects.

[0019] (1) Personalized question bank generation: Monitor students' daily ability assessments and intelligently generate personalized practice question banks based on their learning situations and weak knowledge points to improve learning outcomes. (2) High-quality question generation: Utilizing advanced technologies such as large language models and multimodal reasoning models, we generate high-quality and diverse questions and enrich the question bank resources; (3) Intelligent question bank management: Provides automated question bank management and quality control mechanisms to ensure the accuracy and effectiveness of the question bank. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings: Figure 1 This is a structural diagram of a question bank generation system based on a large language model according to the present invention; Figure 2 This is a flow chart of the student daily ability assessment monitoring module of the present invention; Figure 3 This is a flowchart of the question bank management and update module of the present invention; Figure 4 A flowchart for generating text-based questions in the question bank generation module of the present invention (humanities example); Figure 5 Generate a flow chart for the multimodal question type of the present invention (science example); Figure 6 This is a flowchart of the quality review and assessment module of the present invention; Figure 7 This is a flow chart of a method for generating a question bank based on a large language model according to the present invention. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0022] This invention aims to provide a system that leverages the power of large language models to intelligently generate personalized practice question banks. Specifically, it aims to provide a system and method for generating question banks based on large language models. This system also tracks students' learning progress and identifies weak points, thereby providing targeted questions for each student. The main objectives are as follows: 1. Daily assessment and ability monitoring of students: Based on students' learning data, accurately analyze their knowledge weaknesses and provide an accurate basis for question bank generation. Data such as test scores, daily homework, and exercises are used to create a heat map tracking system based on the correctness of each test question. The system also intelligently and dynamically assigns weights to each knowledge point based on the learning stage.

[0023] 2. Create and maintain a student's personal question bank. Each student's personal question bank consists of a collection of original wrong questions automatically recorded by the system and a custom question bank. The custom question bank allows students, parents, or teachers to add questions they want to enhance their practice. Once a personal question bank is created, it is categorized and stored according to the knowledge points covered and tagged by difficulty level. When a student's error rate for a certain knowledge point or difficulty level drops to a threshold, the importance of that question type is reduced, and vice versa.

[0024] 3. Leveraging advanced large language models and diffusion models, we generate high-quality, diverse variants of incorrect questions, enriching the question bank. We use large language models to identify original question examples and generate high-quality variant text content for students, including questions and reference answers, tailored to the scope of knowledge points and difficulty levels. For questions with image examples, we use a diffusion model to generate images that meet the requirements through a reverse process.

[0025] 4. Provide intelligent question bank management and quality control mechanisms to ensure the accuracy and effectiveness of the question bank. This involves introducing a multimodal large language model to identify generated questions, and using prompt word engineering to provide information about the answer, question type, and difficulty level. A comprehensive comparison using a large inference model confirms that the questions meet the requirements for non-toxicity, eliminating any that do not meet these requirements. Risky questions are clearly marked when provided to teachers, parents, or students.

[0026] The core innovation of this invention lies in the use of advanced technologies such as large language models and multimodal fusion models of images and text to achieve intelligent generation and management of personalized exercise question banks. The main technical points are as follows: Combining students' daily learning data and wrong question records, we can accurately locate knowledge weaknesses and provide a basis for question bank generation.

[0027] Utilize large language models and image-text multimodal fusion models to generate high-quality and diverse wrong question variants to meet personalized learning needs.

[0028] Develop intelligent question bank management and quality control mechanisms to ensure the accuracy and effectiveness of the question bank.

[0029] Example 1 like Figure 1 As shown, the present invention is a question bank generation system based on a large language model. The main goal of the system is to analyze students' learning data and wrong question records, and pass the analysis results to the question bank generation module for targeted question generation. The system includes: The student daily ability assessment monitoring module is used to collect student learning data, break down question information based on the student learning data, add question difficulty labels, and analyze the degree of mastery of knowledge points; based on the degree of mastery of knowledge points, calculate the knowledge point ability score to form the original wrong question data; The question bank management and update module is used to store the original wrong question data as the basic wrong question bank, store the customized enhanced question bank and wrong question variant question bank, and update all question banks according to new questions; The question bank generation module is used to generate new questions based on the basic wrong question bank, customized enhanced question bank and knowledge point distribution, using the large language model and diffusion model to form a wrong question variant question bank; The question quality review and evaluation module is used to evaluate the effectiveness and multimodal consistency of new questions, and to add question difficulty labels to questions using dynamic difficulty adjustment. New questions that pass the evaluation will be added to the wrong question variant question bank. The special practice tracking module is used to formulate personalized question practice plans based on students' learning status, and monitor students' question practice completion status and wrong question statistics; and feed back the question practice completion status and wrong question statistics to students' daily ability assessment monitoring module to form a closed-loop optimization.

[0030] In this embodiment, the student daily ability assessment monitoring module is responsible for collecting all learning data of students throughout the entire learning period. It then intelligently analyzes and determines daily ability based on data such as test scores, daily homework completion, and classroom exercise content submitted by students. The knowledge points involved in the students' incorrect answers and their answers are then archived to form a raw record of incorrect answer data.

[0031] Specifically, the student daily ability assessment and monitoring module includes: The data collection unit is used to collect student learning data, which includes students' test scores, daily homework completion status and classroom exercise content; for example, a summary record of students' test scores, daily homework completion status and classroom exercise content for each semester, and also records each student's wrong questions.

[0032] The mastery level analysis unit is used to decompose the question information into three-dimensional vectors of knowledge point label (K), cognitive level (L), and error pattern (M) based on the student learning data, and then quantitatively analyze the student's mastery of the knowledge point by adding the question difficulty label; For each student S, the question information can be represented as a three-dimensional vector V=(K,L,M), where K is the knowledge point label, K={k1,k2,...,k n}, where k i Indicates specific knowledge points; L is the cognitive level, L={l1,l2,...,l m}, where lj represents the cognitive level (such as memory, understanding, application, etc.); M is the error pattern, M={m1,m2,...,m p}, where m k Indicates the type of error (such as calculation error, conceptual error, carelessness, etc.).

[0033] student For knowledge point k i Degree of mastery It can be quantified as: ; in, For students Knowledge points involved The total number of topics, For the The accuracy of the questions (correct is 1, wrong is 0), For the The weight of the questions (based on factors such as question difficulty and cognitive level).

[0034] The ability score calculation unit is used to calculate the knowledge point ability score based on the degree of knowledge mastery using the temporal attention mechanism to generate raw wrong question data. This invention introduces the temporal attention mechanism to capture the characteristics of the knowledge forgetting curve, allowing students' knowledge point ability scores to fluctuate appropriately based on the forgetting curve.

[0035] Introducing the temporal attention mechanism, students S In time t Knowledge points k i Knowledge point ability score The expression is: ; in, is the total number of time points recorded; For time The attention weight of , which measures the impact of the learning situation at that time point on the ability score of the current knowledge point; For students In time Knowledge points the degree of mastery; is the forgetting rate parameter, which controls the forgetting speed; To simulate the forgetting curve, it means from time To current time The forgetting effect.

[0036] As a further implementation, the student daily ability assessment monitoring module is also used to collect students' various data and wrong questions, and then generate statistical information for teachers and parents based on knowledge point ability scores. It also adds weights according to the learning stage and the new and old knowledge. Teachers can also mark important knowledge points and change the weight level according to teaching needs to form a comprehensive evaluation report. As a further implementation, the student daily ability assessment monitoring module is also used to analyze each student's mastery of knowledge points since the beginning of the school year through the grade change detection unit after each grade and homework is entered. The fluctuations in grades and accuracy will finally form a learning outcome analysis report for teachers and parents. For students whose grades have continuously declined or there is an obvious decline in grades, early warnings will be triggered. For example, teachers can be warned and the students' current learning situation and knowledge weaknesses can be analyzed to provide a basis for teachers to improve teaching priorities and parents to provide tutoring.

[0037] In this embodiment, the question bank management and update module is responsible for processing and archiving the question bank and generating a variant question bank using the capabilities of the large language model. The purpose is to enrich the breadth and depth of students' practice and to obtain sufficiently novel variant wrong question sets of appropriate difficulty for improvement.

[0038] The question banks of this invention are mainly divided into three categories. The first category is the basic wrong question bank based on the results of students' daily practice. The question bank contains the students' original answers and attributes such as three-dimensional vectors analyzed by the student's daily ability assessment monitoring module. The second category is the customized enhanced question bank that students, parents, or teachers want to enhance specifically for their current learning stage or learning situation. The third category is a high-quality wrong question variant question bank generated using a large language model and diffusion model. The search method for all question banks is based on type and three-dimensional vector label.

[0039] The dynamic update mechanism of the question bank of the present invention aims to provide students with a variety of wrong questions with different difficulty levels and styles for the same knowledge points. When new wrong questions or custom questions are entered, the question bank will expand and generate questions for the fewer tagged questions, ensuring that the question bank remains up to date while giving priority to generating fewer questions, thereby increasing the richness of the entire question bank and maintaining updates.

[0040] In this embodiment, the question bank generation module is responsible for the generation and maintenance of the wrong question variant question bank. The main purpose is to use advanced large language models and diffusion models to generate high-quality and diverse wrong question variants and enrich the question bank resources.

[0041] Specifically, the question bank generation module includes: The question template setting unit is used to set preset examples for various question templates as question template prompts before generating the question bank, guiding the large language model to optimize prompts for common question types and question-answering methods, and controlling the large model to achieve results specifically for question bank generation; The large language model unit uses the large language model to identify original question examples based on the basic wrong question bank, customized enhanced question bank, and knowledge point distribution, and generates a wrong question variant question bank for students according to different knowledge point scopes and different levels of difficulty. The text content of the wrong question variant question bank includes questions and reference answers. Specifically, by utilizing the reasoning self-correction capability of the inference large model, the large language model can automatically filter out most incorrect generation methods when generating the question bank. The correct questions, answers, and incorrect questions and answers are used as fine-tuning datasets to fine-tune the large language model that needs to be used. While strengthening the large language model's ability to generate questions, it also reduces the illusion of generated questions and improves the robustness of the generated question bank. The dataset covers questions from different subjects and different test points to ensure that the fine-tuned model can meet the teaching requirements of multiple scenarios and disciplines.

[0042] like Figure 4 As shown, for questions in liberal arts subjects like history, Chinese, and politics, we use the RAG project to build a subject knowledge base. This data is categorized by age, supplementary book type, and other factors, storing reference book data. When the Big Language Model generates an extended question bank for these incorrect questions, it first verifies the validity of each premise from the RAG knowledge base before incorporating it into the Big Language Model's prompts. This ensures that false positives are minimized and the accuracy of the questions is improved.

[0043] The diffusion model unit is used to generate a question image that meets the requirements of the question based on the diffusion model using a reverse process for the question containing an image.

[0044] like Figure 5 As shown, for questions with images, a large language model is first used to parse the semantics of the original image. Prompt engineering is then used to extract core knowledge points, graphic elements, and difficulty parameters. LangChain is then used to build a processing pipeline. A structured data generation layer is generated for different question types. For example, for geometry questions, SVG vector parameters (coordinate points, angles, and line segment relationships) are generated; for chemistry questions, SMILES expressions are converted; and for geography questions, DEM parameters (elevation point matrices) are generated. Multimodal processing is then used to process this structured data and generate question images. For example, GeoGebra and Matlab are used for geometry; RDKit and StableDiffusion and Pymatgen are used for chemistry; QGIS and Platec are used for geography; and SchemDraw and Simupy are used for physics. These methods complete the initial generation of multimodal data for various subjects.

[0045] Among them, LangChain: is used to build workflows for large language model applications, supporting modular integration of external tools and data sources; SMILES expression: a chemical symbol system that describes molecular structures using ASCII strings; DEM digital elevation model parameters: a spatial distribution model representing ground elevation using a digital matrix; GeoGebra: dynamic mathematical software that supports geometric drawing, algebraic calculations, and coordinate system generation; Matlab: numerical calculation, plotting functions and data, verifying algorithms, etc.; RDKit: A cheminformatics toolkit for molecular structure processing and descriptor calculation.

[0046] Stable Diffusion: Diffusion-driven image generation technology that improves efficiency through latent space compression.

[0047] Pymatgen: Open source materials genomics library for processing crystal structures, chemical group analysis, etc.

[0048] DEM: Digital Elevation Model, which realizes digital simulation of ground terrain through limited terrain elevation data.

[0049] QGIS: An open source geographic information system for processing DEM data, generating topographic maps, and performing spatial analysis.

[0050] Platec: A 2D terrain generator based on plate tectonics.

[0051] SchemDraw: A library for drawing high-quality circuit schematics.

[0052] Simupy: A library for modeling and simulation of multibody system dynamics, used to model underlying physics.

[0053] In this embodiment, the question quality review and evaluation module is to introduce an intelligent question bank management and quality control mechanism, with the purpose of confirming the validity and accuracy of the generated questions and helping students to conduct systematic practice under the correct questions.

[0054] Specifically, the topic quality review and assessment module includes: The validity assessment unit is used to verify whether a question has a solution based on the student role reasoning model. It also compares the analysis information with the reference answer, difficulty level, and knowledge point range when the question was generated to conduct a comprehensive evaluation to confirm the validity of the question. The specific implementation of this validity assessment unit is as follows: for the generated questions, the student role reasoning model, which is guided by a prompt project dedicated to common question types, will "play" the question and verify the question. The student role reasoning model will first verify whether the question has a solution, then analyze and answer it. The information obtained from its own analysis will be compared with the reference answer, difficulty level, and knowledge point range data when the question was generated to conduct a comprehensive evaluation to confirm the validity of the question. For questions that cannot be judged, a mark requiring manual intervention will be added and they will not be used before review.

[0055] The consistency assessment unit calculates image-text similarity for questions containing images using the CLIP model. Based on this image-text similarity, it uses the image-text contrast loss to align the latent feature spaces of the image and text. Furthermore, based on multimodal reasoning and parsing, it compares the parsed information with the reference answers, difficulty level labels, and knowledge point scope used during question generation for comprehensive evaluation to confirm image-text consistency. Specifically, this consistency assessment unit performs the following: For questions containing images, the CLIP model is first used to calculate image-text similarity, confirming that the similarity meets the specified value. The BLIP-2 model is then used to unify understanding and generation, and the image-text contrast loss is used to align the latent feature spaces of the image and text. After multimodal alignment, reasoning and parsing are performed using powerful multimodal reasoning and parsing models such as Skywork-R1V and YieldFCP, combined with subject-specific tools such as SymPy and RDKit. Similarly, a comprehensive evaluation is performed by comparing the reference answers, difficulty level, and knowledge point scope data used during question generation to confirm image-text consistency. Questions that cannot be determined are uniformly marked as requiring manual intervention and are not used before review.

[0056] The difficulty label adjustment unit is used to construct a difficulty matrix based on the Bloom taxonomy using a dynamic difficulty adjustment method. It also uses Item Response Theory (IRT) combined with a large language model to simulate the responses of test-takers and real test-takers on specific questions, and dynamically assesses the difficulty of questions based on student ability and answer accuracy, and adds difficulty labels to the questions. Dynamic difficulty assessment based on Item Response Theory (IRT): ; Constructing Bloom Taxonomy Difficulty Matrix : ; Among them, D=1.702 is a constant parameter, For students ability, For the title Difficulty. Indicates the Difficulty level The number of questions for each knowledge point.

[0057] In this embodiment, the purpose of the special practice tracking module is to provide teachers and parents with a special practice tracking function based on the monitoring situation and the wrong question database. Specifically, it includes: (1) Based on the students’ learning status, teachers are provided with a function to generate special exercises for students. Specifically, when a teacher wants to conduct special exercises on a knowledge point for a student, the teacher selects the knowledge point questions that need to be practiced from the system according to the classification labels, and sets the wrong question ratio to form comprehensive special exercises. The answer status can also be entered into the student’s daily ability assessment monitoring module.

[0058] (2) Provide parents with a time series prediction function based on the students' learning status. Specifically, the Prophet algorithm is used in combination with holiday information and seasonal trends to provide parents with a predicted ability development curve for the next three months. When students are in winter and summer vacations, without school supervision, if the gap between the student's curve and the predicted winter and summer vacation curve is too large, parents can find and correct it in time at home, and at this time, they can generate protective special connections from the question bank with one click to consolidate the students' knowledge points. Prevent the emergence of back-to-school syndrome and reduce the burden of re-strengthening learning quality management for teachers after school starts. Prediction value It can be expressed as: ; Among them, the Prophet algorithm open-sourced by Facebook is a decomposition method for time series forecasting. By decomposing time series into components such as trend, seasonality, and holiday effects, it can effectively process data with seasonal and nonlinear trends. It is particularly suitable for forecasting tasks involving special events such as holidays.

[0059] The gap between students' actual ability and predicted value It can be expressed as: ; in, Indicates that students are in time When students are in winter or summer vacation, if the gap Exceeding the set threshold triggers the generation of special exercises.

[0060] In the specific implementation, taking students' mathematics learning as an example, the implementation process of the system of the present invention is demonstrated: 1. Collection of students' daily learning data: Collect students' math test scores, daily homework completion status, and classroom exercise content, and record each wrong question; 2. Knowledge point mastery assessment: The system scores each mathematics knowledge point based on the student's answers, and determines the student's mastery level of different mathematics knowledge points: knowledge point mastery degree and knowledge point ability score; 3. Question template design: Teachers or question bank administrators design a series of math question templates based on the syllabus and exam requirements, covering different knowledge points and difficulty levels; 4. Personalized exercise generation: The system intelligently selects suitable question templates from the question bank based on students' ability scores and wrong question records. It also uses a large language model and a multimodal fusion model of images and texts, combined with students' knowledge weaknesses and learning progress, to generate personalized math exercises.

[0061] 5. Special exercises and ability assessment: Special exercises are generated based on students' weak mathematical knowledge points. After students complete the exercises, the system reassesses their abilities, analyzes their progress, provides a comprehensive evaluation, and generates a learning report for reference by teachers and parents.

[0062] Example 2 like Figure 7 As shown, the difference between this embodiment and embodiment 1 is that this embodiment provides a method for generating a question bank based on a large language model. This method is based on a system for generating a question bank based on a large language model in embodiment 1. The method includes: Collect student learning data, break down question information and add question difficulty labels based on the data, and analyze the degree of mastery of knowledge points; calculate the knowledge point ability score based on the degree of mastery of knowledge points to form the original wrong question data; Store the original wrong question data as the basic wrong question database, and store the customized enhanced question database and wrong question variant database; Based on the basic wrong question bank, customized enhanced question bank and knowledge point distribution, new questions are generated in a targeted manner based on the large language model and diffusion model to form a wrong question variant question bank; Conduct validity and multimodal consistency evaluations on new questions, and assign difficulty labels to them using dynamic difficulty adjustment. New questions that pass the evaluation are added to the wrong question variant database; and Develop personalized question practice plans based on students' learning status, monitor students' question practice completion status and wrong question statistics; and provide feedback on question practice completion status and wrong question statistics to form a closed-loop optimization.

[0063] As a further implementation, we collected student learning data, decomposed the question information based on the student learning data, added question difficulty labels, and analyzed the degree of knowledge point mastery. Based on the degree of knowledge point mastery, we calculated the knowledge point ability scores to form the original wrong question data, including: Collect student learning data, including test scores, daily homework completion, and classroom exercise content; Based on student learning data, the question information is decomposed into three-dimensional vectors of knowledge point labels, cognitive levels, and error patterns. After adding the question difficulty label, the degree of students' mastery of the knowledge points is quantitatively analyzed; According to the mastery of knowledge points, the knowledge point ability score is calculated based on the temporal attention mechanism to form the original wrong question data; Among them, the knowledge point ability score The expression is: ; in, is the total number of time points recorded; For time The attention weight of , which measures the impact of the learning situation at that time point on the ability score of the current knowledge point; For students In time Knowledge points the degree of mastery; is the forgetting rate parameter, which controls the forgetting speed; To simulate the forgetting curve, it means from time To current time The forgetting effect.

[0064] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0066] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0067] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0068] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A question bank generation system based on a large language model, characterized in that: The system includes: The student daily ability assessment monitoring module is used to collect student learning data. Based on the student learning data, the module decomposes the question information and adds question difficulty labels to analyze the degree of mastery of knowledge points. Based on the degree of mastery of knowledge points, the module calculates the knowledge point ability scores to form the original wrong question data. The question bank management and update module is used to store the original wrong question data as the basic wrong question bank, store the customized enhanced question bank and wrong question variant question bank, and update all question banks according to new questions; The question bank generation module is used to generate new questions based on the basic wrong question bank, customized enhanced question bank and knowledge point distribution, using the large language model and diffusion model to form a wrong question variant question bank; The question quality review and evaluation module is used to evaluate the effectiveness and multimodal consistency of new questions, and to add question difficulty labels to questions using dynamic difficulty adjustment. New questions that pass the evaluation will be added to the wrong question variant question bank. The special practice tracking module is used to formulate personalized question practice plans based on students' learning status, and monitor students' question practice completion status and wrong question statistics; and feed back the question practice completion status and wrong question statistics to students' daily ability assessment monitoring module to form a closed-loop optimization.

2. A question bank generation system based on a large language model according to claim 1, characterized in that: The student daily ability assessment and monitoring module includes: A data collection unit is used to collect student learning data, wherein the student learning data includes the student's test scores, daily homework completion status and classroom exercise content; The mastery level analysis unit is used to decompose question information into three-dimensional vectors of knowledge point labels, cognitive levels, and error patterns based on student learning data. It also adds question difficulty labels to quantitatively analyze students' mastery of knowledge points. The ability score calculation unit is used to calculate the knowledge point ability score based on the temporal attention mechanism according to the degree of mastery of the knowledge points to form the original wrong question data.

3. A question bank generation system based on a large language model according to claim 2, characterized in that: The knowledge point ability score The expression is: ； in, is the total number of time points recorded; For time The attention weight of , which measures the impact of the learning situation at that time point on the ability score of the current knowledge point; For students In time Knowledge points the degree of mastery; is the forgetting rate parameter, which controls the forgetting speed; To simulate the forgetting curve, it means from time To current time The forgetting effect.

4. The question bank generation system based on a large language model according to claim 1, characterized in that: The student daily ability assessment and monitoring module is also used to generate statistical information for teachers and parents based on knowledge point ability scores, and to add weights according to the learning stage and the new and old knowledge to form a comprehensive evaluation report; it is also used to detect changes in grades and trigger early warnings for students whose grades have declined.

5. The question bank generation system based on a large language model according to claim 1, characterized in that: The customized enhanced question bank is a customized question bank that students, parents or teachers plan to enhance specifically based on the current learning stage or learning situation; the wrong question variant question bank is a wrong question variant question bank generated by the question bank generation module based on the large language model and the diffusion model.

6. The question bank generation system based on a large language model according to claim 1, characterized in that: The question bank generation module includes: The question template setting unit is used to set preset examples for multiple question templates as question template prompts, guide the large language model to optimize prompts based on question types and question-answering methods, and control the large model to generate a question bank; The large language model unit uses the large language model to identify original question examples based on the basic wrong question bank, the customized enhanced question bank, and the knowledge point distribution, and generates a wrong question variant bank for students based on different knowledge point ranges and different levels of difficulty. The text content of the wrong question variant bank includes questions and reference answers; The diffusion model unit is used to generate a question image that meets the requirements of the question based on the diffusion model using a reverse process for the question containing an image.

7. The question bank generation system based on a large language model according to claim 1, characterized in that: The topic quality review and assessment module includes: The validity assessment unit is used to verify whether the question has a solution based on the student role reasoning model. The analysis information is compared with the reference answer, difficulty label and knowledge point scope when the question was generated to conduct a comprehensive evaluation to confirm the validity of the question. The consistency evaluation unit is used to calculate the image-text similarity based on the CLIP model for questions containing images. Based on the image-text similarity, it determines the image and text comparison loss to align the latent feature space of the image and text. Based on multimodal reasoning and parsing, it compares the parsed information with the reference answer, difficulty label, and knowledge point scope when the question was generated to conduct a comprehensive evaluation to confirm the consistency of the image and text. The difficulty label adjustment unit is used to construct the difficulty matrix of the Bloom taxonomy using a dynamic difficulty adjustment method. It also uses item response theory combined with a large language model to simulate the answers of test-takers and real test-takers on specific questions, conducts a comprehensive dynamic assessment of the difficulty of the questions based on the students' abilities and the accuracy of the answers, and adds difficulty labels to the questions.

8. The question bank generation system based on a large language model according to claim 1, characterized in that: Develop a personalized practice plan based on the student's learning situation, including: Based on the students' learning status, teachers are provided with a function to generate special exercises for students. Specifically, when a teacher wants to conduct special exercises on a certain knowledge point for a student, the teacher will filter the knowledge point questions that need to be practiced from the system according to the classification tags, and set the wrong question ratio to form comprehensive special exercises. Based on students' learning status, we provide parents with a time series forecasting function. Specifically, we use the Prophet algorithm combined with holiday information and seasonal trends to provide parents with a predicted ability development curve for the next three months.

9. A method for generating a question bank based on a large language model, characterized in that: The method includes: Collect student learning data, break down question information and add question difficulty labels based on the data, and analyze the degree of mastery of knowledge points; calculate the knowledge point ability score based on the degree of mastery of knowledge points to form the original wrong question data; Store the original wrong question data as the basic wrong question database, and store the customized enhanced question database and wrong question variant database; Based on the basic wrong question bank, customized enhanced question bank and knowledge point distribution, new questions are generated in a targeted manner based on the large language model and diffusion model to form a wrong question variant question bank; Conduct validity and multimodal consistency evaluations on new questions, and assign difficulty labels to them using dynamic difficulty adjustment. New questions that pass the evaluation are added to the wrong question variant database; and Develop personalized question practice plans based on students' learning status, monitor students' question practice completion status and wrong question statistics; and provide feedback on question practice completion status and wrong question statistics to form a closed-loop optimization.

10. The method for generating a question bank based on a large language model according to claim 9, characterized in that: Collect student learning data, break down question information and add question difficulty labels based on the data, and analyze the degree of mastery of knowledge points; Based on the degree of knowledge mastery, the knowledge point ability score is calculated to form the original wrong question data, including: Collecting student learning data, including student test scores, daily homework completion status, and classroom exercise content; Based on student learning data, the question information is decomposed into three-dimensional vectors: knowledge point label, cognitive level, and error pattern. After adding the question difficulty label, the student's mastery of the knowledge point is quantitatively analyzed; According to the mastery of knowledge points, the knowledge point ability score is calculated based on the temporal attention mechanism to form the original wrong question data; Among them, the knowledge point ability score The expression is: ； in, is the total number of time points recorded; For time The attention weight of , which measures the impact of the learning situation at that time point on the ability score of the current knowledge point; For students In time Knowledge points the degree of mastery; is the forgetting rate parameter, which controls the forgetting speed; To simulate the forgetting curve, it means from time To current time The forgetting effect.

Images