Intelligent question setting and learning feedback method and system

Abstract

The invention provides an intelligent question setting and learning feedback method and system, and relates to the technical field of artificial intelligence. Constraints of question stem semantic consistency, content length and knowledge point coverage are embedded in a controlled language model, and controllable and verifiable AI question setting is achieved; mapping answer features of modes such as texts, voices, images and the like and corresponding standard answer features to a unified semantic space for scoring, so that cross-modal comparable scoring is realized; the knowledge mastery degree is dynamically updated based on the scoring result in combination with a time decay mechanism, and the retraining questions are recommended, so that a personalized recommendation closed loop is realized, and the question setting quality, the scoring consistency and the learning efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent question generation and learning feedback method and system. Background Technology

[0002] With the increasing application of artificial intelligence in education, automated question generation and intelligent assessment systems have become key capabilities for online education platforms. Common solutions in existing technologies include: Existing Solution A: A question generation system based on a manual question bank and batch processing. This solution typically involves administrators or teachers manually maintaining the question bank and generating test papers in batches according to a fixed template. While this type of system has a simple structure, it suffers from slow question updates, difficulty in matching questions to individual learners' levels, and an inability to adjust question difficulty based on real-time learning progress, resulting in low learning efficiency.

[0003] Existing Solution B: An automated question generation system based on a single-round API (Application Programming Interface) call. This system uses a pre-trained language model to directly generate questions via the API, but it lacks mechanisms for question quality control, knowledge point matching, and historical learning data-driven approaches. The generated content may deviate from the user's learning objectives, and the generated results are unstable, making it difficult to guarantee consistent question quality.

[0004] Existing Solution C: AI (Artificial Intelligence) assessment system based on single-modal scoring. Current AI assessments mostly process text input, such as answers to objective questions or short-answer questions. For oral, handwritten, or image-based responses, the lack of multimodal perception and a unified scoring mechanism leads to a single evaluation dimension, delayed feedback, and an inability to form an effective learning loop.

[0005] In summary, existing technologies cannot simultaneously meet the following requirements: 1. The intelligence and stability of the question generation (quality of generation and matching degree of knowledge points); 2. The multimodal uniformity and interpretability of the scoring; 3. Real-time and personalized adaptive learning feedback.

[0006] In other words, the core challenge that current technologies struggle to overcome is how to ensure the quality of generated questions while simultaneously achieving a closed loop of cross-modal scoring and individualized knowledge tracking. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent question generation and learning feedback method and system to improve question generation quality, scoring consistency and learning efficiency.

[0008] In a first aspect, the present invention provides an intelligent question generation and learning feedback method, comprising: The system obtains the question generation parameters from the question generation request. Based on the question generation parameters and the knowledge graph, it calls the controlled language model to generate structured questions under the constraints of semantic consistency of the question stem, content length, and knowledge point coverage. After quality testing, the generated questions are stored in the question bank database. The question generation parameters include the topic, knowledge point range, question type ratio, and difficulty level. Based on the scope of knowledge points, the proportion of question types, and the difficulty level, questions are extracted from the question bank database to construct the current test paper, which is then asynchronously sent to the user's terminal. The system retrieves multimodal responses to each question in the current test paper from the user's client, maps the response features of each modality in the response content to the corresponding standard answer features, and scores them in a unified semantic space to obtain the scoring results. The multimodal responses include one or more of text, speech, and image. Based on the scoring results of each question in the current test paper and the time decay mechanism, the system dynamically updates the user's mastery of the corresponding knowledge points and recommends retraining questions to the user based on the updated knowledge mastery data.

[0009] In an optional implementation, based on the question generation parameters and a knowledge graph, a controlled language model is invoked to generate structured questions under constraints of semantic consistency of the question stem, content length, and knowledge point coverage. The generated questions are then quality-checked and stored in a question bank database, including: Based on the scope of knowledge points, extract subgraphs of knowledge points related to the topic from the knowledge graph and generate task identifiers for question requests; Based on knowledge point subgraphs and question parameters, a controlled language model is invoked to generate structured questions by optimizing a preset objective function; The generated questions undergo quality checks, which include one or more of the following: answer uniqueness verification, option logical consistency verification, and semantic similarity clustering for deduplication. Questions that pass the quality check are stored in the question bank database, and question embedding vector indexes are generated for subsequent retrieval.

[0010] In an optional implementation, the objective function is: ; in, L ctrl Describe the objective function. L base This represents the basic loss function of the controlled language model. C semantic This represents a constraint used to control the semantic consistency between the question stem and the standard answer explanation. C length These represent constraints used to limit the length of the question stem and answer choices. C knowledgeThis refers to constraints used to ensure that the content of the questions covers the relevant knowledge points. c 1. c 2. c 3 represents the preset weighting coefficient.

[0011] In an optional implementation, the response features of each modality in the response content are mapped to a unified semantic space with the corresponding standard answer features for scoring, resulting in a scoring result, including: Modality recognition is performed on the answers, and the corresponding feature extraction function is called to perform feature mapping based on the recognized modality; The response features of each modality obtained by feature mapping and the corresponding standard answer features are input into the cross-modal alignment model to obtain the score of each modality; The scores of each modality are weighted, fused, and normalized to obtain a comprehensive score, and its confidence level is generated.

[0012] In an optional implementation, the weighted fusion is calculated using a weighted semantic similarity calculation formula, which is: ; in, Score Indicates the weighted fusion score. T , A , I They represent three modalities: text, speech, and image. w m Representing modes m The weight, sim Represents cosine similarity. E m ( user ) represents a mode m Response characteristics in a unified semantic space E m ( standard ) represents a mode m The standard answer features corresponding to the unified semantic space.

[0013] In an optional implementation, the user's mastery of the corresponding knowledge points is dynamically updated based on the score of each question in the current test paper, combined with a time decay mechanism, including: Update the knowledge point mastery level according to the following formula: , ; or, , ; in, M t+1 ( k () indicates the current time after the initial update. t+1 knowledge point k The degree of mastery, M t ( k () indicates the last update time. t Knowledge points k The degree of mastery, α t The learning rate represents the confidence level of the rating result. Score k Representing knowledge points k The overall score in the scoring results for the corresponding question. Indicates the current moment after the final update. t +1 knowledge point k The degree of mastery, This represents the preset time decay factor. Indicates the current time t +1 and the last update time t The time interval between them.

[0014] In an optional implementation, retraining questions are recommended to the user based on the updated knowledge mastery data, including: From the updated knowledge mastery data, knowledge points with mastery levels below a preset threshold are identified and constituted as a set of weak knowledge points; The system retrieves candidate questions related to the set of weak knowledge points from the question bank database, and determines the recommended score for each candidate question based on the matching degree between the question difficulty and the user's mastery level, as well as the semantic similarity between the question and the knowledge point. Based on the calculated recommended scores, all candidate questions are sorted in descending order, and the first preset number of candidate questions are selected as retraining questions and recommended to the user.

[0015] In optional implementations, the intelligent question generation and learning feedback method also includes: Generate a visual learning report that includes knowledge mastery, learning curve, and weakness analysis, and call a large language model to generate structured explanations for error points, which are then pushed to the user to complete the learning loop.

[0016] Secondly, the present invention provides an intelligent question generation and learning feedback system, comprising: The question generation module is used to obtain the question generation parameters of the question generation request. Based on the question generation parameters and the knowledge graph, it calls the controlled language model to generate structured questions under the constraints of semantic consistency of the question stem, content length and knowledge point coverage. After quality checking, the generated questions are stored in the question bank database. The question generation parameters include topic, knowledge point range, question type ratio and difficulty level. The module is used to extract questions from the question bank database to construct the current test paper based on the scope of knowledge points, the proportion of question types, and the difficulty level, and then asynchronously send it to the user's terminal. The scoring module is used to obtain the multimodal answers returned by the user for each question in the current test paper, map the answer features of each modality in the answer content to the corresponding standard answer features to a unified semantic space for scoring, and obtain the scoring result; wherein, multimodality includes one or more of text, speech and image; The recommendation module is used to dynamically update the user's mastery of the corresponding knowledge points based on the score of each question in the current test paper and the time decay mechanism, and recommend retraining questions to the user based on the updated knowledge mastery data.

[0017] In an optional implementation, it further includes: The generation module is used to generate a visual learning report that includes knowledge mastery, learning curve, and weakness analysis. It also calls a large language model to generate structured explanations for error points and pushes them to the user to complete the learning loop.

[0018] The intelligent question generation and learning feedback method and system provided by this invention includes the following steps: obtaining question generation parameters from a question generation request; based on the question generation parameters and a knowledge graph, calling a controlled language model to generate structured questions under constraints of semantic consistency of the question stem, content length, and knowledge point coverage; and storing the generated questions in a question bank database after quality testing. The question generation parameters include the topic, knowledge point range, question type ratio, and difficulty level. Based on the knowledge point range, question type ratio, and difficulty level, questions are extracted from the question bank database to construct the current test paper, which is then asynchronously sent to the user. The user receives multimodal responses to each question in the current test paper; the response features of each modality in the response content are mapped to a unified semantic space with the corresponding standard answer features for scoring, resulting in a scoring result. The multimodal responses include one or more of text, speech, and image. Based on the scoring result of each question in the current test paper and a time decay mechanism, the user's mastery of the corresponding knowledge points is dynamically updated, and retraining questions are recommended to the user based on the updated knowledge mastery data. By embedding constraints on semantic consistency of question stems, content length, and knowledge point coverage into a controlled language model, controllable and verifiable AI question generation is achieved. Mapping the answer features of text, speech, and image modalities to the corresponding standard answer features into a unified semantic space for scoring enables cross-modal comparable scoring. Based on the scoring results and combined with a time decay mechanism, the knowledge mastery is dynamically updated, and retraining questions are recommended, realizing a personalized recommendation closed loop, thereby improving question generation quality, scoring consistency, and learning efficiency. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an intelligent question generation and learning feedback method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent question generation and learning feedback system provided in an embodiment of the present invention; Figure 3 This is an overall architecture diagram of an intelligent question generation and learning feedback system provided in an embodiment of the present invention; Figure 4 A flowchart of intelligent question generation and scoring provided in an embodiment of the present invention; Figure 5 This is a closed-loop flowchart for knowledge tracking and recommendation provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The embodiments of the present invention mainly address the following technical shortcomings of existing intelligent question generation and learning feedback systems: 1. Question generation lacks knowledge alignment and control. Existing question generation methods based on language models often result in semantic drift, mismatch between question stems and knowledge points, and a lack of constraint mechanisms tailored to different learner levels.

[0023] 2. The scoring process is unimodal and lacks consistency. Current AI scoring systems mainly target text input and cannot uniformly handle multimodal responses such as voice and images, resulting in inconsistent evaluation standards and low confidence levels.

[0024] 3. Learning feedback lacks adaptability. Existing systems cannot dynamically track learners' knowledge mastery, historical answer characteristics, and time decay patterns, lacking true knowledge tracking and personalized learning path recommendation capabilities.

[0025] 4. The system's task processing is highly coupled and difficult to expand. The question generation, scoring, and recommendation modules are usually executed serially, which can easily cause system latency and resource blockage when the algorithm is updated or when there is high concurrency of tasks.

[0026] Therefore, the embodiments of the present invention aim to provide an AI question generation method that enables precise control of knowledge points, a unified evaluation system that supports multimodal scoring, and a learning feedback closed-loop system that integrates knowledge tracking and personalized recommendation, so as to improve question generation quality, scoring consistency and learning efficiency.

[0027] This invention provides an intelligent question generation and learning feedback method and system based on a multimodal AI engine and a knowledge tracking closed loop. The intelligent question generation, answer scoring and learning feedback, which combines a multimodal artificial intelligence engine, knowledge tracking algorithm and personalized recommendation mechanism, can be applied to scenarios such as online examinations, corporate training and adaptive learning.

[0028] To facilitate understanding of this embodiment, a detailed description of the intelligent question generation and learning feedback method disclosed in this embodiment of the invention will be provided first.

[0029] This invention provides an intelligent question generation and learning feedback method, which can be executed by an electronic device with data processing capabilities. See also... Figure 1 The diagram shows a flowchart of an intelligent question generation and learning feedback method, which mainly includes the following steps S110 to S140: Step S110: Obtain the question generation parameters of the question generation request. Based on the question generation parameters and the knowledge graph, call the controlled language model to generate structured questions under the constraints of semantic consistency of the question stem, content length and knowledge point coverage. After quality detection of the generated questions, store them in the question bank database. The question generation parameters include topic, knowledge point range, question type ratio and difficulty level.

[0030] After receiving a question request that includes the topic, scope of knowledge points, proportion of question types, and difficulty level, the system first locates relevant knowledge points from the knowledge graph to clarify the scope of the examination. Then, it calls a controlled language model to generate structured questions under multiple constraints, ensuring that the question stem is semantically clear, the content is of appropriate length, and closely covers the target knowledge points. Finally, it performs quality checks such as logical verification and similarity deduplication on the generated questions, and stores them in the question bank after passing the checks.

[0031] In some possible embodiments, step S110 may include: extracting topic-related knowledge point subgraphs from the knowledge graph based on the knowledge point scope, and generating a task identifier for the question request; based on the knowledge point subgraphs and question parameters, calling a controlled language model to generate structured questions by optimizing a preset objective function; performing quality checks on the generated questions, including one or more of answer uniqueness checks, option logical consistency checks, and semantic similarity clustering deduplication; storing the questions that pass the quality checks into the question bank database, and generating a question embedding vector index for subsequent retrieval.

[0032] In one possible implementation, the above objective function can be: ; in, L ctrl Describe the objective function. L base This represents the basic loss function of the controlled language model. C semantic This represents a constraint used to control the semantic consistency between the question stem and the standard answer explanation. C length These represent constraints used to limit the length of the question stem and answer choices. C knowledge This refers to constraints used to ensure that the content of the questions covers the relevant knowledge points. c 1. c 2. c 3 is preset and respectively with C semantic , C length , C knowledge The corresponding weighting coefficients.

[0033] Step S120: Based on the scope of knowledge points, the proportion of question types, and the difficulty level, extract questions from the question bank database to construct the current test paper, and asynchronously send it to the user terminal.

[0034] Based on the user-defined knowledge point scope, question type ratio, and difficulty level, the system automatically selects suitable questions from a quality-tested question bank database. These questions are then combined into a comprehensive test paper with a reasonable question structure and appropriate difficulty distribution. The generated test paper is then delivered to the user's device via an asynchronous task mechanism, ensuring that the user can promptly access personalized assessment content. The user's device can be a web application or a mobile app.

[0035] Step S130: Obtain the multimodal answer content returned by the user terminal for each question in the current test paper, map the answer features of each modality in the answer content to the corresponding standard answer features to a unified semantic space for scoring, and obtain the scoring result; wherein, multimodality includes one or more of text, speech and image.

[0036] After receiving user responses submitted in various formats such as text, voice, or images, the system uses modality recognition and feature extraction technology to map the response information from different modalities (such as text semantics, speech-to-text and acoustic features, and image recognition results) together with the standard answer to the corresponding question into a unified semantic vector space. In this space, the feature similarity is calculated and weighted fusion is performed to generate the final comprehensive score, thereby achieving unified, interpretable, and automated scoring of multimodal responses.

[0037] In some possible embodiments, step S130 above may include: performing modality recognition on the answer content, and calling the corresponding feature extraction function to perform feature mapping based on the recognized modality; inputting the answer features of each modality obtained by feature mapping and the corresponding standard answer features into a cross-modality alignment model to obtain the score of each modality; performing weighted fusion and normalization on the scores of each modality to obtain a comprehensive score, and generating its confidence level.

[0038] In one possible implementation, weighted fusion is calculated using a weighted semantic similarity calculation formula, which can be: ; in, Score Indicates the weighted fusion score. T , A , I They represent three modalities: text, speech, and image. w m Representing modes m The weight, sim Represents cosine similarity. E m ( user ) represents a mode m Response characteristics in a unified semantic space E m ( standard ) represents a mode m The standard answer features corresponding to the unified semantic space.

[0039] Step S140: Based on the scoring results of each question in the current test paper and the time decay mechanism, dynamically update the user's mastery of the corresponding knowledge points, and recommend retraining questions to the user based on the updated knowledge mastery data.

[0040] Based on the scoring results of each question in the current test paper, and combined with the time decay factor reflecting the forgetting curve, the user's mastery of relevant knowledge points is dynamically updated. Subsequently, based on the updated knowledge mastery data, the user's weak knowledge points are automatically identified, and questions with matching difficulty and relevant content are accurately selected from the question bank to generate a personalized list of retraining questions to recommend to the user, thus completing the adaptive learning loop of assessment-update-recommendation.

[0041] In some possible implementations, the knowledge point mastery level can be updated first based on the scoring results, and then a time decay mechanism can be introduced to determine the final updated knowledge point mastery level; alternatively, the knowledge point mastery level can be updated first using a time decay mechanism, and then the final updated knowledge point mastery level can be determined based on the scoring results. Based on this, the knowledge point mastery level can be updated according to the following formula: , ; or, , ; in, M t+1 ( k () indicates the current time after the initial update. t +1 knowledge point k The degree of mastery, M t ( k () indicates the last update time. t Knowledge points k The degree of mastery, α t The learning rate represents the confidence level of the rating result. Score k Representing knowledge points k The overall score in the scoring results for the corresponding question. Indicates the current moment after the final update. t +1 knowledge point k The degree of mastery, This represents the preset time decay factor. Indicates the current time t +1 and the last update time t The time interval between them.

[0042] Optionally, α t = α 0×(1+ d × confidence ), α 0 is the initial learning rate. d This is the preset confidence level adjustment coefficient. confidence Confidence level; λ= 0.02 / day, where day represents the time unit "day".

[0043] In some possible embodiments, the step of recommending retraining questions to users based on updated knowledge mastery data can be implemented as follows: identifying knowledge mastery levels below a preset threshold from the updated knowledge mastery data. i The knowledge points are used to form a set of weak knowledge points; candidate questions related to the set of weak knowledge points are retrieved from the question bank database, and the recommended score for each candidate question is determined based on the matching degree between the difficulty of the question and the user's mastery, as well as the semantic similarity between the question and the knowledge point; all candidate questions are sorted in descending order according to the calculated recommended scores, and the top preset number of candidate questions are selected as retraining questions and recommended to the user.

[0044] The above-mentioned preset threshold iBoth the preset quantity and the actual quantity can be set according to actual needs; this embodiment does not limit this. For example, i =0.6, preset quantity is 5.

[0045] Furthermore, the aforementioned intelligent question generation and learning feedback methods also include: generating a visual learning report that includes knowledge point mastery, learning curves, and weakness analysis, and calling a large language model to generate structured explanations for error points, which are then pushed to the user to complete the learning loop.

[0046] The system automatically generates a visual learning report based on the user's updated knowledge mastery data. The report can cover the mastery of each knowledge point, changes in historical learning curves, and a heatmap of the distribution of weaknesses. At the same time, it uses a large language model to intelligently analyze the error points in the answers and generate structured explanations that include the reasons for the errors, the correct approaches, and analogies. Finally, the report and explanations are pushed to the user's device, forming a complete learning loop of assessment, analysis, feedback, and consolidation.

[0047] In addition, the embodiments of the present invention adopt an asynchronous task orchestration architecture, in which the question generation process, the scoring process, and the recommendation process of retraining questions are all decoupled and executed, supporting high concurrency and scalable AI modular collaboration.

[0048] The intelligent question generation and learning feedback method provided in this invention embeds constraints on question stem semantic consistency, content length, and knowledge point coverage into a controlled language model, achieving controllable and verifiable AI question generation; it maps the answer features of text, voice, and image modalities to the corresponding standard answer features into a unified semantic space for scoring, achieving cross-modal comparable scoring; and it dynamically updates knowledge mastery based on the scoring results combined with a time decay mechanism, and recommends retraining questions, achieving a personalized recommendation closed loop, thereby improving question generation quality, scoring consistency, and learning efficiency.

[0049] Corresponding to the above-described intelligent question generation and learning feedback method, this embodiment of the invention also provides an intelligent question generation and learning feedback system. See also... Figure 2 The diagram shown illustrates the structure of an intelligent question generation and learning feedback system. The system includes: The question generation module 201 is used to obtain the question generation parameters of the question generation request. Based on the question generation parameters and the knowledge graph, it calls the controlled language model to generate structured questions under the constraints of semantic consistency of the question stem, content length and knowledge point coverage. After quality detection, the generated questions are stored in the question bank database. The question generation parameters include topic, knowledge point range, question type ratio and difficulty level. Module 202 is used to extract questions from the question bank database to construct the current test paper based on the scope of knowledge points, the proportion of question types, and the difficulty level, and then asynchronously send it to the user terminal. The scoring module 203 is used to obtain the multimodal answer content returned by the user for each question in the current test paper, map the answer features of each modality in the answer content to the corresponding standard answer features to a unified semantic space for scoring, and obtain the scoring result; wherein, multimodality includes one or more of text, speech and image; The recommendation module 204 is used to dynamically update the user's mastery of the corresponding knowledge points based on the score of each question in the current test paper and the time decay mechanism, and recommend retraining questions to the user based on the updated knowledge mastery data.

[0050] The intelligent question generation and learning feedback system provided in this invention embeds constraints on question stem semantic consistency, content length, and knowledge point coverage into a controlled language model, enabling controllable and verifiable AI question generation. It maps the answer features of text, voice, and image modalities to corresponding standard answer features in a unified semantic space for scoring, achieving cross-modal comparable scoring. Based on the scoring results and a time decay mechanism, it dynamically updates knowledge mastery and recommends retraining questions, realizing a personalized recommendation closed loop, thereby improving question generation quality, scoring consistency, and learning efficiency.

[0051] Furthermore, the question generation module 201 is specifically used for: extracting topic-related knowledge point subgraphs from the knowledge graph based on the knowledge point scope, and generating a task identifier for the question generation request; generating structured questions by calling a controlled language model to optimize a preset objective function based on the knowledge point subgraphs and question generation parameters; performing quality checks on the generated questions, including one or more of answer uniqueness checks, option logical consistency checks, and semantic similarity clustering deduplication; storing the questions that pass the quality checks into the question bank database, and generating a question embedding vector index for subsequent retrieval.

[0052] Furthermore, the objective function described above is: ; in, L ctrl Describe the objective function. L base This represents the basic loss function of the controlled language model. C semantic This represents a constraint used to control the semantic consistency between the question stem and the standard answer explanation. C length These represent constraints used to limit the length of the question stem and answer choices. C knowledge This refers to constraints used to ensure that the content of the questions covers the relevant knowledge points. c 1. c 2. c 3 represents the preset weighting coefficient.

[0053] Furthermore, the scoring module 203 is specifically used for: performing modality recognition on the answer content, and calling the corresponding feature extraction function to perform feature mapping based on the recognized modality; inputting the answer features of each modality obtained by feature mapping and the corresponding standard answer features into the cross-modality alignment model to obtain the score of each modality; performing weighted fusion and normalization on the scores of each modality to obtain a comprehensive score, and generating its confidence level.

[0054] Furthermore, the above weighted fusion is calculated using the weighted semantic similarity calculation formula, which is: ; in, Score Indicates the weighted fusion score. T , A , I They represent three modalities: text, speech, and image. w m Representing modes m The weight, sim Represents cosine similarity. E m ( user ) represents a mode m Response characteristics in a unified semantic space E m ( standard ) represents a mode m The standard answer features corresponding to the unified semantic space.

[0055] Furthermore, the aforementioned recommendation module 204 is specifically used to update the knowledge point mastery level according to the following formula: , ; or, , ; in, M t+1 ( k () indicates the current time after the initial update. t +1 knowledge point k The degree of mastery, M t ( k () indicates the last update time. t Knowledge points k The degree of mastery, α t The learning rate represents the confidence level of the rating result. Score k Representing knowledge points k The overall score in the scoring results for the corresponding question. Indicates the current moment after the final update. t +1 knowledge point k The degree of mastery, This represents the preset time decay factor. Indicates the current time t +1 and the last update time t The time interval between them.

[0056] Furthermore, the aforementioned recommendation module 204 is also used to: identify knowledge points with a mastery level below a preset threshold from the updated knowledge mastery data, forming a set of weak knowledge points; retrieve candidate questions related to the set of weak knowledge points from the question bank database, and determine the recommendation score for each candidate question based on the matching degree between the question difficulty and the user's mastery level, as well as the semantic similarity between the question and the knowledge point; sort all candidate questions in descending order according to the calculated recommendation scores, and select the top preset number of candidate questions as retraining questions to recommend to the user.

[0057] The question-generating module 201, the scoring module 203, and the recommendation module 204 can communicate with each other through an asynchronous message queue to support high concurrency and scalable AI modular collaboration.

[0058] Furthermore, the aforementioned intelligent question generation and learning feedback system also includes: The generation module is used to generate a visual learning report that includes knowledge mastery, learning curve, and weakness analysis. It also calls a large language model to generate structured explanations for error points and pushes them to the user to complete the learning loop.

[0059] The intelligent question generation and learning feedback system provided in this embodiment has the same implementation principle and technical effects as the aforementioned intelligent question generation and learning feedback method embodiment. For the sake of brevity, any parts not mentioned in the intelligent question generation and learning feedback system embodiment can be referred to the corresponding content in the aforementioned intelligent question generation and learning feedback method embodiment.

[0060] To facilitate understanding, the above-mentioned intelligent question generation and learning feedback methods and systems will be described in detail below.

[0061] The core concept of this invention lies in constructing an intelligent question generation and learning feedback system that supports asynchronous parallel task orchestration by introducing a multimodal AI engine, a knowledge graph-driven question generation mechanism, and a vectorized knowledge tracking algorithm. The system adopts a layered architecture, modularizing the "question generation, quality inspection, scoring, and recommendation" algorithms. It utilizes message queues to achieve asynchronous task delivery and combines vector databases and knowledge embedding technology to achieve the following key innovations: Controlled AI Question Generation Mechanism: Knowledge constraints and question type ratio constraints are embedded in the generative model to achieve controllable and verifiable AI question generation; Multimodal unified scoring system: Construct a unified feature embedding space for modalities such as text, speech, and image to achieve comparable scoring across modalities; Dynamic knowledge tracking algorithm: Based on the mastery vector + time decay model, dynamically update the learner's knowledge status to achieve a closed loop of personalized recommendation; Asynchronous task orchestration architecture: Decouples the tasks of question generation, scoring, and recommendation, supporting high concurrency and scalable modular AI collaboration.

[0062] Through the above combination scheme, the embodiments of the present invention realize an adaptive cycle of question generation-scoring-feedback-relearning, which solves the long-standing technical contradictions of "uncontrollable quality, untimely feedback, and rigid learning paths" in AI education systems.

[0063] This invention proposes an intelligent question generation and learning feedback method and system based on a multimodal AI engine and a knowledge tracking closed loop. The system uses a controlled language model (LLM) as its core, and constructs an intelligent learning closed loop of "question generation—scoring—profile creation—recommendation—relearning" through a knowledge graph-driven question generation mechanism, a multimodal scoring algorithm with a unified semantic space, a dynamic mastery update model, and a reinforcement learning recommendation strategy.

[0064] The intelligent question generation and learning feedback method of this invention may include the following steps: Step S1: Task initialization and knowledge analysis.

[0065] 1. The system receives user input parameters, including Topic, Knowledge Nodes, Type Ratio, and Difficulty Level.

[0066] 2. The system extracts topic-related knowledge nodes from the knowledge graph to form a knowledge structure subgraph (i.e., a knowledge point subgraph): , where V' is the set of selected knowledge points, and E' represents the prerequisite relationship and semantic similarity between knowledge points. The prerequisite relationship between knowledge points refers to other knowledge that needs to be mastered in advance before learning a certain knowledge.

[0067] 3. The system generates a unique task identifier (task_id) and delivers it to the message queue through the task orchestration module to achieve asynchronous task management.

[0068] Step S2: Controlled AI generates questions.

[0069] 1. The question generation module calls the controlled language model based on the knowledge node set, and optimizes the objective function using the following constraints: ; in, Csemantice Used to ensure consistency between the question stem and the parsing semantics. C length Used to limit the number of words in the question stem and options. C knowledge This is used to ensure that the content of the questions covers the knowledge nodes.

[0070] 2. The generated result is a structured JSON question, for example: { "question_id": "Q12345", "topic": "cybersecurity", "type": "Single Choice Question", "difficulty": 0.6, "knowledge_tag": "Password Management", "question_text": "...", "options": ["A...", "B...", "C...", "D..."], "answer": "B", "analysis": "..." } 3. The system performs quality checks on the generated questions: Answer uniqueness check; Option logical consistency check; Semantic similarity clustering for deduplication (e.g., based on Sentence-BERT, with a threshold > 0.9).

[0071] 4. Questions that pass the quality inspection are written into the question bank database, and question embedding vector indexes are generated for subsequent retrieval.

[0072] Step S3: Test paper construction and task release.

[0073] 1. The system generates test papers based on the proportion of question types and the coverage of knowledge: ,in T The total number of questions. r i This represents the percentage of question types.

[0074] 2. The system guarantees a test paper coverage rate of ≥90% and generates a unique exam_id (test paper identifier); 3. The task scheduling module pushes the test paper to the user's terminal (Web or App) to achieve asynchronous distribution.

[0075] Step S4: Multimodal response and feature extraction.

[0076] 1. Learners respond via user interface using text, voice, and images; 2. The system automatically identifies the answer modality and uses corresponding feature extraction methods to extract features from the answer content of different modalities. For example, text is parsed using NLP (Natural Language Processing); speech is recognized using ASR (Automatic Speech Recognition) + acoustic feature extraction; and images are recognized using OCR (Optical Character Recognition) + CNN (Convolutional Neural Network) visual encoding.

[0077] 3. Perform feature mapping on different modal inputs: ,in, f m For modality m The feature extraction function.

[0078] Step S5: Unify multimodal scoring.

[0079] 1. The system inputs the user's answer and the standard answer into a cross-modal alignment model (such as CLIP (Contrastive Language–Image Pretraining) or Contrastive Learning Transformer) and maps them to a unified semantic space; 2. Calculate weighted semantic similarity: ,in, w m Modal weights (e.g., 0.5 for text, 0.3 for speech, and 0.2 for image). sim Cosine similarity; 3. Normalize the scoring results and generate a confidence index for use by the knowledge tracking module; 4. Output the results, for example: { "question_id": "Q12345", "score": 0.86, "confidence": 0.92, "explanation": "The semantic expression is correct, but the logic is slightly flawed." } Step S6: Update the knowledge profile.

[0080] 1. The system updates the user's knowledge mastery vector based on the score results: M t+1 ( k )= M t ( k )+ α t ( Score k - M t ( k )),in: α t = α 0×(1+ d × confidence ), α 0 is the initial learning rate. d This is the preset confidence level adjustment coefficient. confidence Confidence level; Score k Score for the current question; 2. The system incorporates a time decay mechanism to reflect the forgetting effect: ,in l =0.02 / day; 3. The results are written to the knowledge profile database in real time for personalized recommendations.

[0081] Step S7: Personalized recommendation and reinforcement learning optimization.

[0082] 1. The system recognizes that the level of knowledge mastery is below the threshold. i Key points: ,in, W This represents the set of weak knowledge points. M ( k This indicates the user's understanding of the knowledge points. k The degree of mastery; 2. Perform similar question retrieval and difficulty difference ranking for weak knowledge points: ,in, Score r Indicates the recommended score. Diff ( q () indicates the topic q Difficulty level, Sim ( q , k () indicates the topic q With knowledge points k semantic similarity; 3. The system updates its recommendation strategy through reinforcement learning: ,in, a For recommended actions, R tRewards for learning and improvement; or Step size factor; π t ( a | s ) for time step t The strategy, representing the state given s At that time, the agent selects an action. a The probability distribution can be understood as t The old strategy of timing; π t+1 ( a | s ) for time step t The +1 strategy is a new strategy obtained after updating based on the reward feedback at the current moment; The benchmark return (or average return) is used to measure... R t Good or bad, log It is the natural logarithm.

[0083] 4. Finally, output the Top-N recommended questions and personalized learning paths.

[0084] Step S8: Learning feedback and visualization closed loop.

[0085] 1. The system generates a visual learning report, including: mastery level of each knowledge point; changes in the learning curve; and a heatmap of weaknesses. 2. The AI ​​explanation module calls LLM to generate structured feedback, for example: "1. Error: The core concept of 'multimodal embedding' was not mentioned;" 2. Correct approach: The alignment relationship between different modalities should be specified in the description; 3. Analogy Question: Please explain the application of "semantic alignment" in image recognition. 3. Learning reports are pushed to the user's end in real time via WebSocket, realizing a closed learning loop.

[0086] The system structure is described below. The aforementioned intelligent question generation and learning feedback system may include the following modules: 1. Task orchestration module: Receives requests, generates task_ids, submits them to the message queue, and monitors their status; 2. AI Question Generation Module: Performs controlled generation and constraint optimization; 3. Quality Inspection Module: Performs logic verification, similarity clustering, and question deduplication; 4. Scoring Module: Identifies modalities and performs unified semantic embedding scoring; 5. Knowledge Tracking Module: Dynamically updates the mastery vector; 6. Recommendation Module: Performs reinforcement learning recommendation and Top-N retrieval; 7. Explanation Module: Generates explanations and review materials for incorrect answers; 8. Data storage module: Stores the question bank, answer records, and knowledge profiles; 9. Notification module: Real-time push notifications of learning results and reports.

[0087] The modules interact with each other through a message bus and database to achieve asynchronous decoupling and parallel computing.

[0088] like Figure 3 As shown, the intelligent question generation and learning feedback system provided in this embodiment of the invention involves a user terminal, a task orchestration module, an AI engine layer, a notification service, and object storage. The AI ​​engine layer is the core of the system, responsible for question generation, scoring, and recommendation. The task orchestration module is responsible for scheduling and status tracking, and all data is persisted in the database and object storage layers. Specifically, the user terminal sends an API request to the task orchestration module; the task orchestration module generates a task and sends an asynchronous message of the task to the AI ​​engine layer; the AI ​​engine layer stores the asynchronous message in the object storage. The AI ​​engine layer includes a question generation engine, a scoring engine, and a recommendation engine. The questions generated by the question generation engine are stored in the question bank database after quality inspection, the scoring results generated by the scoring engine are stored in the answer record database, and the knowledge mastery data generated by the recommendation engine are stored in the knowledge profile database. The task orchestration module can also detect when a task is completed and push the result to the user terminal through the notification service.

[0089] like Figure 4 As shown, the intelligent question generation and scoring process includes: knowledge extraction and topic analysis → AI-generated questions → quality control and deduplication → test paper release → user answers → multimodal scoring → knowledge profile update → personalized recommendation. This process demonstrates a complete closed loop from topic analysis to personalized recommendation, ensuring dynamic adjustment of the learning path.

[0090] like Figure 5 As shown, the closed-loop process of knowledge tracking and recommendation includes: answer data → scoring module → knowledge profile update → mastery prediction model → weakness detection → recommendation engine → personalized questions / explanation content → user practice again → answer data. This forms a cyclical iterative path of "answering questions—scoring—updating—recommending—answering questions again," achieving adaptive learning.

[0091] The application process of the present invention will be described below through two specific embodiments.

[0092] Example 1: Adaptive question generation and feedback in an enterprise intelligent training system.

[0093] (1) Application scenarios.

[0094] The internal security training and examination system of a large enterprise adopts the intelligent question generation and learning feedback system of this invention to realize personalized knowledge assessment and retraining recommendations for employees.

[0095] (2) Implementation process.

[0096] Step 1: Topic and Rule Setting. The administrator enters the topic "Cybersecurity Awareness" in the system front end, sets the question type ratio (40% multiple choice, 30% true / false, 30% short answer), difficulty distribution (easy:medium:difficult = 3:5:2), and selects the "Support Voice Questions" option. The system generates tasks and calls the task orchestration module to produce a unique task_id, which is then written to the task table.

[0097] Step 2: Knowledge Extraction and AI Question Generation. The task orchestration module calls the AI ​​question generation engine. The AI ​​engine retrieves knowledge nodes such as "information leakage," "password security," and "phishing emails" from the enterprise's internal knowledge base and constructs a Prompt template: "Based on the knowledge point of [cryptographic security], generate a medium-difficulty multiple-choice question with 4 options, a question stem of no more than 50 words, and a standard answer and explanation." The generated results are then processed by the quality control module: Determine the uniqueness of the options; Check the logical consistency of the answer; The question similarity matrix is ​​calculated by embedding the model, and duplicate items are removed if the cosine similarity is greater than 0.9.

[0098] Questions that pass the quality inspection are written into the question bank database and marked as approved.

[0099] Step 3: Exam Paper Generation and Distribution. The system selects questions from the approved question bank according to a set ratio, ensuring that the knowledge point coverage rate is ≥90%; it automatically generates exam paper records (exam_id) and distributes the exam link to employees through the IM (Instant Messaging) interface.

[0100] Step 4: Multimodal Questioning and Scoring. Employees answer questions via webpage or mobile device: Text-based question: Enter text; Voice recording task: Record your narration; Image-based question: Upload a screenshot (e.g., a screenshot showing the configuration).

[0101] The scoring module executes according to modal branches: Text → LLM score; Speech → ASR + articulation features; Image → OCR+CNN matching.

[0102] The scores of each modality are weighted and fused to calculate the total score: ,in, w T =0.5, w A =0.3, w I =0.2.

[0103] Step 5: Knowledge Profile Update and Feedback Generation. The system updates the knowledge point mastery level for each question using the following formula: ,in, α Dynamically adjusted based on confidence level: α =0.05+0.15× confidence ; The system simultaneously calculates the time decay function. To model the forgetting effect.

[0104] The AI ​​explanation module uses LLM to generate explanation templates based on incorrect answers, for example: "1. Error point: {misconception};" 2. Correct approach: {rationale}; 3. Hint: {hint}; 4. Analogy question: {generated_question}. The learning report includes: mastery rate of each knowledge point; radar chart showing the "mastery gap"; and 5 recommended review questions for weak areas.

[0105] Step 6: System Closed Loop and Iterative Learning. After employees complete the refresher training, the system scores them again and calculates the improvement Δ=M. new -M old If Δ > 0.15, the system determines that the learning target has been met and records the learning event, which is then included in the long-term performance analysis.

[0106] Example 2: Student adaptive practice on an online learning platform.

[0107] (1) Scene description.

[0108] An online education platform deployed this system for high school mathematics courses. After each student answers a question, the system automatically adjusts the difficulty and recommends new questions.

[0109] (2) Technical process.

[0110] 1. Upon the student's initial response, the system estimates the knowledge mastery vector based on 20 introductory questions. M 0).

[0111] 2. The AI-generated question module dynamically generates the next set of questions based on the mastery matrix: like M (k If the value is less than 0.5, generate basic questions. If 0.5≤ M ( k If the value is less than 0.8, a comprehensive problem will be generated. like M ( k ≥0.8 → Generate challenge questions.

[0112] 3. The knowledge tracking module is updated after answering questions. M ( k And predict the probability of mastering the material in the next exam.

[0113] 4. The system plots a learning curve based on the changing trend and generates reinforcement learning suggestions.

[0114] This embodiment demonstrates the adaptability and intelligent adjustment capabilities of the present invention.

[0115] In summary, examples of relevant parameters and algorithms of the modules in the system of this invention are shown in Table 1 below.

[0116] Table 1

[0117] The core functionalities of this invention are shown in Table 2 below.

[0118] Table 2

[0119] The embodiments of the present invention have the following advantages: Question generation process: A controlled generation algorithm is introduced to ensure knowledge alignment; Scoring process: A multimodal unified embedding model is used to achieve semantic consistency; Feedback process: An adaptive closed loop is formed through dynamic knowledge tracking and reinforcement learning recommendation; System architecture: Based on asynchronous task orchestration and message queues, it achieves high concurrency and low coupling.

[0120] This invention also provides a computer-readable storage medium storing a computer program. When a processor runs the computer program, it executes the intelligent question generation and learning feedback method described in the preceding method embodiments. The computer-readable storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk.

[0121] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0122] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent question generation and learning feedback, characterized in that, include: The question generation parameters of the question generation request are obtained. Based on the question generation parameters and the knowledge graph, a controlled language model is invoked to generate structured questions under the constraints of semantic consistency of the question stem, content length, and knowledge point coverage. The generated questions are then stored in the question bank database after quality inspection. The question generation parameters include the topic, knowledge point range, question type ratio, and difficulty level. Based on the scope of knowledge points, the proportion of question types, and the difficulty level, questions are extracted from the question bank database to construct the current test paper, which is then asynchronously sent to the user's terminal. The system obtains the multimodal response content returned by the user terminal for each question in the current test paper, maps the response features of each modality in the response content to the corresponding standard answer features to a unified semantic space for scoring, and obtains the scoring result; wherein, the multimodality includes one or more of text, speech and image; Based on the scoring results of each question in the current test paper and the time decay mechanism, the user's mastery of the corresponding knowledge points is dynamically updated, and retraining questions are recommended to the user based on the updated knowledge mastery data.

2. The intelligent question generation and learning feedback method according to claim 1, characterized in that, The process involves using the question generation parameters and a knowledge graph to call a controlled language model to generate structured questions under constraints of semantic consistency, content length, and knowledge point coverage. The generated questions are then quality-checked and stored in a question bank database. This includes: Based on the scope of knowledge points, extract subgraphs of knowledge points related to the topic from the knowledge graph and generate a task identifier for the question request; Based on the knowledge point subgraph and the question generation parameters, the controlled language model is invoked to generate structured questions by optimizing a preset objective function; The generated questions are subjected to quality checks, which include one or more of the following: answer uniqueness check, option logical consistency check, and semantic similarity clustering deduplication. Questions that pass the quality check are stored in the question bank database, and question embedding vector indexes are generated for subsequent retrieval.

3. The intelligent question generation and learning feedback method according to claim 2, characterized in that, The objective function is: ； in, L ctrl Describe the objective function. L base This represents the basic loss function of the controlled language model. C semantic This represents a constraint used to control the semantic consistency between the question stem and the standard answer explanation. C length These represent constraints used to limit the length of the question stem and answer choices. C knowledge This represents a constraint used to ensure that the content of the question covers the scope of the stated knowledge points. γ 1. γ 2. γ 3 represents the preset weighting coefficient.

4. The intelligent question generation and learning feedback method according to claim 1, characterized in that, The process of mapping the response features of each modality in the response content to the corresponding standard answer features and then scoring them in a unified semantic space to obtain the scoring results includes: Modality recognition is performed on the response content, and feature mapping is performed by calling the corresponding feature extraction function based on the recognized modality; The response features of each modality obtained by feature mapping and the corresponding standard answer features are input into the cross-modal alignment model to obtain the score of each modality; The scores of each modality are weighted, fused, and normalized to obtain a comprehensive score, and its confidence level is generated.

5. The intelligent question generation and learning feedback method according to claim 4, characterized in that, The weighted fusion is calculated using a weighted semantic similarity calculation formula, which is as follows: ； in, Score Indicates the weighted fusion score. T , A , I They represent three modalities: text, speech, and image. w m Representing modes m The weight, sim Represents cosine similarity. E m ( user ) represents a mode m Response characteristics in a unified semantic space E m ( std ) represents a mode m The standard answer features corresponding to the unified semantic space.

6. The intelligent question generation and learning feedback method according to claim 1, characterized in that, The method of dynamically updating the user's mastery of the corresponding knowledge points based on the scoring results of each question in the current test paper and a time decay mechanism includes: Update the knowledge point mastery level according to the following formula: ， ； or, , ; in, M t+1 ( k () indicates the current time after the initial update. t +1 knowledge point k The degree of mastery, M t ( k () indicates the last update time t Knowledge points k The degree of mastery, α t The learning rate represents the confidence level of the rating result. Score k Representing knowledge points k The overall score in the scoring results for the corresponding question. Indicates the current moment after the final update. t +1 knowledge point k The degree of mastery, This represents the preset time decay factor. Indicates the current time t +1 and the last update time t The time interval between them.

7. The intelligent question generation and learning feedback method according to claim 1, characterized in that, The method of recommending retraining questions to users based on updated knowledge mastery data includes: From the updated knowledge mastery data, knowledge points with mastery levels below a preset threshold are identified, forming a set of weak knowledge points; Candidate questions related to the set of weak knowledge points are retrieved from the question bank database, and a recommended score for each candidate question is determined based on the matching degree between the difficulty of the question and the user's mastery, as well as the semantic similarity between the question and the knowledge point. Based on the calculated recommendation scores, all candidate questions are sorted in descending order, and the top preset number of candidate questions are selected as retraining questions and recommended to the user.

8. The intelligent question generation and learning feedback method according to claim 1, characterized in that, The intelligent question generation and learning feedback method also includes: Generate a visual learning report that includes knowledge mastery, learning curve, and weakness analysis, and call a large language model to generate structured explanations for error points, which are then pushed to the user to complete the learning loop.

9. An intelligent question generation and learning feedback system, characterized in that, include: The question generation module is used to obtain the question generation parameters of the question generation request, and based on the question generation parameters and the knowledge graph, call the controlled language model to generate structured questions under the constraints of semantic consistency of the question stem, content length and knowledge point coverage, and store the generated questions in the question bank database after quality detection; wherein, the question generation parameters include topic, knowledge point range, question type ratio and difficulty level; The construction module is used to extract questions from the question bank database to construct the current test paper based on the knowledge point range, the question type ratio and the difficulty level, and asynchronously send it to the user terminal; The scoring module is used to obtain the multimodal answer content returned by the user terminal for each question in the current test paper, map the answer features of each modality in the answer content to the corresponding standard answer features to a unified semantic space for scoring, and obtain the scoring result; wherein, the multimodality includes one or more of text, speech and image; The recommendation module is used to dynamically update the user's mastery of the corresponding knowledge points based on the score of each question in the current test paper and the time decay mechanism, and recommend retraining questions to the user based on the updated knowledge mastery data.

10. The intelligent question generation and learning feedback system according to claim 9, characterized in that, Also includes: The generation module is used to generate a visual learning report that includes knowledge mastery, learning curve, and weakness analysis. It also calls a large language model to generate structured explanations for error points and pushes them to the user to complete the learning loop.

Images