Personalized historical exercise recommendation and automatic correction system based on large model

Abstract

The invention relates to the technical field of intelligent education, and discloses a personalized historical exercise recommendation and automatic correction system based on a large model, and the system comprises a cross-modal data collection module which is used for collecting character input, time behavior and interaction track data of a student in an answering process, and outputting a collection result to a cognitive map construction module; and the cognitive map construction module receives the acquisition result output by the cross-modal data acquisition module, and is used for generating a cognitive map representing the mastering degree of the students on the historical knowledge points and the relationship thereof based on the acquisition result, and sending the cognitive map to a cognitive drive recommendation module and an automatic correction module. The cross-modal data acquisition module is introduced, so that dominant content answered by the students is acquired, diversified recessive behavior data such as answering duration, editing and modifying tracks, mouse hovering positions and answering pauses of the students are further acquired, and the system can surpass traditional simple answering right and wrong judgment due to the multi-dimensional data capture capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent education technology, specifically to a personalized history exercise recommendation and automatic correction system based on a large model. BACKGROUND

[0002] With the rapid development of information technology, the field of education is undergoing profound changes. Personalized learning, as a key path to improving teaching efficiency and student learning outcomes, is increasingly attracting widespread attention. Traditional education models typically use a "one-size-fits-all" teaching approach, where all students learn the same content and complete the same exercises, and the correction method is relatively simple. For subjective questions in subjects such as history, artificial correction is often relied upon, which is inefficient and difficult to ensure consistency in evaluation standards. This model fails to fully consider individual differences in students' knowledge background, learning habits, understanding ability, and cognitive weaknesses.

[0003] In existing technology, some computer-aided teaching systems have emerged to assist students in exercise practice and partially automated correction. For example, some systems can provide a question bank retrieval function, allowing students to select practice questions based on knowledge points; other systems attempt to automatically correct objective questions through keyword matching or preset rules. However, these existing solutions generally have significant technical flaws and limitations.

[0004] Specifically, existing technologies in exercise recommendation are often limited to simple extraction from a fixed question bank or coarse-grained recommendation based on students' historical scores. This approach fails to delve into the subtle flaws in students' knowledge structure, particularly the understanding of the relationship between knowledge points. Therefore, the recommended exercises may not accurately target students' individual weaknesses for effective intervention, leading to repeated practice of mastered knowledge points or failure to reach the true knowledge blind spot, thereby reducing learning efficiency and relevance.

[0005] In addition, existing systems generally lack a dynamic, closed-loop adaptive learning mechanism. After students complete exercises and receive correction results, this feedback information is often not effectively utilized to update students' individual cognitive models in real time, nor is it able to optimize the entire system's recommendation and correction strategies. This means that the system cannot adjust and evolve itself based on students' continuous learning and progress, making it difficult to form a truly personalized, continuously optimized intelligent learning environment. Students' learning activities in the system are often discrete and isolated, failing to form a virtuous cycle of knowledge from input to mastery to feedback reinforcement.

[0006] In summary, the current education technology field still has significant technical gaps and needs in personalized historical exercise recommendation and automatic correction, and there is an urgent need for a comprehensive solution that can overcome the above limitations, achieve deep cognitive modeling, dynamic content generation, and intelligent feedback. SUMMARY

[0007] To overcome the shortcomings of the prior art, the present application provides a large model-based personalized historical exercise recommendation and automatic correction system, which solves the problems of coarse granularity of student cognitive state modeling, non-personalization of exercise recommendation, lack of intelligence in subjective question correction, and lack of dynamic adaptive learning closed loop in existing computer-aided teaching systems.

[0008] To achieve the above purpose, the present application is implemented by the following technical solutions: a large model-based personalized historical exercise recommendation and automatic correction system, comprising: A cross-modal data acquisition module is used to acquire the student's text input, time behavior and interaction trajectory data during the answering process, and output the acquisition results to the cognitive graph construction module; A cognitive graph construction module receives the acquisition results output by the cross-modal data acquisition module, generates a cognitive graph representing the student's mastery of historical knowledge points and their relationships based on the acquisition results, and sends the cognitive graph to the cognitive-driven recommendation module and the automatic correction module; A cognitive-driven recommendation module receives the cognitive graph generated by the cognitive graph construction module, identifies knowledge mastery defects in the cognitive graph, and outputs a set of question recommendations to the control-generated question construction module; A control-generated question construction module receives the question recommendation set from the cognitive-driven recommendation module, constructs a prompt input based on the cognitive graph and the recommendation set, and calls a large language model to generate a target exercise, which is submitted to the student for answering via a system interaction interface; An automatic correction module receives the student's answer content, performs semantic scoring, causal consistency scoring and viewpoint rationality scoring on the student's answer based on the standard answer and the cognitive graph, and outputs the scoring results to the cognitive feedback control module; A cognitive feedback control module receives the scoring results, updates the cognitive graph, and outputs system optimization parameters to the cognitive graph construction module and the cognitive-driven recommendation module based on the changes in cognitive state and scoring fluctuations.

[0009] Preferably, the cross-modal data acquisition module comprises: A text acquisition unit is used to extract the student's text answer content; A behavior time acquisition unit is used to record the length of time and pause interval during the student's answering process; A behavior track collection unit is configured to collect input sequence, modification operation and interaction track information of a student on a question interface. A structure identification unit is configured to extract causality structure information between historical events from the text content.

[0010] Preferably, in the cognitive map construction module, the step of generating a cognitive map representing the mastery degree of the student on the historical knowledge points and their relationships comprises: constructing a knowledge graph node set based on a preset historical knowledge point relationship network, wherein the nodes in the node set represent specific historical knowledge points, and the edges represent the relationships between the knowledge points; mapping the student's answering behavior data to the corresponding knowledge points, and constructing an original mastery degree vector by counting the accuracy, answering time and answering frequency of the student on each knowledge point ; normalizing the original mastery degree vector to obtain a standardized knowledge mastery degree vector , wherein each dimension in the standardized knowledge mastery degree vector represents the mastery degree of the student on a certain knowledge point, and the mathematical form of the normalization method is: ; wherein, is the normalized mastery degree of the student on the first knowledge point; is the original mastery degree of the student on the first knowledge point; is the minimum value of the mastery degree of the student on all knowledge points; is the maximum value of the mastery degree of the student on all knowledge points; based on the standardized knowledge mastery degree vector, the logical relationship, the causality relationship and the structural dependency relationship between the knowledge points, introducing a propagation mechanism to model the influence between the knowledge points, and using a relationship matrix to represent the connection strength between the knowledge points, the calculation formula of the knowledge mastery degree of the student after considering the relationship propagation is: ; wherein, is a weighting factor for balancing the original mastery degree of the student and the mastery degree after the relationship propagation between the knowledge points; is the mastery degree vector of the student after considering the relationship propagation between the knowledge points; is a relationship propagation matrix between the knowledge points; According to and knowledge point relationship network, build a cognitive map of each student, each node attached to the final mastery value of the student on the knowledge point, and combine the edge weight to represent the influence relationship between knowledge points, and generate a complete personalized historical knowledge cognitive map.

[0011] Preferably, the cognitive map construction module further comprises a map updating unit for updating the edge weight in the cognitive map according to the scoring results output by the automatic grading module.

[0012] Preferably, the cognitive-driven recommendation module comprises: a defect identification unit for identifying a set of edges in the cognitive map whose edge weights are below a preset threshold as a set of cognitive defect paths; a graph structure optimization unit for selecting a minimum set of questions covering the defect paths from a preset question bank as a set of recommended questions based on the set of cognitive defect paths.

[0013] Preferably, in the control-generated question construction module, the step of constructing a prompt input based on the cognitive map and the recommended set comprises: reading the student mastery vector obtained in the step of generating a cognitive map; sorting the values corresponding to each knowledge point in the student mastery vector and selecting a number of knowledge points with the lowest mastery level to form a recommended set; for each knowledge point in the recommended set, based on the adjacency nodes and edge weights of the knowledge point in the knowledge graph, extracting the associated concepts, key historical events and time information of the knowledge point to generate corresponding structured prompt content; organize the structured prompt information into a prompt input, which is used to provide personalized learning support in combination with the student's knowledge mastery state.

[0014] Preferably, the automatic grading module comprises: a semantic similarity scoring unit for calculating the semantic embedding similarity between the student's answer and the standard answer; a causal consistency scoring unit for identifying the time and causal logic structure between historical events in the student's answer and comparing it with the standard answer; a viewpoint rationality scoring unit for evaluating the completeness of the student's viewpoint and the rationality of the historical stance; a comprehensive scoring unit for calculating the final score according to the following formula: ; wherein, , and respectively are semantic similarity score weight coefficient, causal consistency score weight coefficient and viewpoint rationality score weight coefficient, and can be dynamically set according to the cognitive graph state; For students To the overall score of the test questions; , respectively are semantic similarity score, causal consistency score and viewpoint rationality score of the test questions. Preferably, the cognitive feedback regulation module is used for the step of updating the cognitive graph, comprising: Receiving the score results output by the automatic grading module;

[0015] Based on the score results, the following optimization objective function is used to adjust the system parameters: ; ; Wherein, is the cognitive graph before updating; is the cognitive graph after updating; is the student score fluctuation variance; is the balance parameter; is the system parameter set to be optimized; is minimized on the parameter set ; is the expected value of the student; is the difference between the original cognitive graph of the student and the reconstructed graph; On the basis of the above objective function, the parameter set is optimized through the graph neural network to generate the updated student cognitive graph, so as to more accurately represent the historical knowledge mastery state of the student. Preferably, the difference between the cognitive graph before and after updating is calculated by the following formula:

[0016] ; Wherein, is the value of the cognitive graph before updating at the index position ; is the value of the cognitive graph after updating at the index position ; By introducing this difference, the model distribution deviation can be constrained during the updating process, preventing overfitting and improving the stability and convergence effect of the model.

[0017] ​​​​​Preferably, the control generation type question construction module generates corresponding question content dynamically according to the question recommendation set, and the question type includes a selection question, a fill-in-the-blank question, a short answer question and an open generation question.

[0018] The application provides a personalized history exercise recommendation and automatic correction system based on a large model. 1、The application introduces a cross-modal data acquisition module, not only obtains the explicit content of the student's answer, but also deeply acquires the student's answer duration, editing modification track, mouse hovering position and answering pause and other diversified implicit behavior data, so that the system can surpass the traditional simple right or wrong judgment of the answer, more carefully and comprehensively understand the real cognitive load, thinking process and difficulties of the student in the answering process, and provide a more rich and higher granularity data basis for subsequent construction of a precise cognitive atlas.

[0019] 2、The application overcomes the isolated disadvantages of traditional knowledge point mastery evaluation by constructing a dynamic cognitive atlas and introducing a cognitive state propagation mechanism, in addition, the system can not only calculate the original mastery degree according to the direct answer performance of the student, but more importantly, it can infer the potential understanding level of the student to the knowledge points not directly examined based on the correlation such as causality and time sequence in the preset history knowledge network, and update the mastery degree of the associated knowledge points, so that the modeling of the student's knowledge state is more comprehensive, accurate and close to the real cognitive structure, thereby the weak links in the knowledge network can be more effectively identified.

[0020] 3、The application accurately identifies weak pathways in the student's cognitive atlas through the cognitive driving recommendation module, rather than only locating a single low-mastery knowledge point, and generates a minimum base question set for recommendation by using an optimization algorithm, so that the "diagnosis-prescription" type of accurate recommendation avoids the "sea of questions" or invalid repeated exercises that may occur in traditional systems, can focus the student's learning resources on the most needed knowledge structure defects, significantly improves the learning efficiency and the pertinence of personalized teaching, and enables the student to obtain the maximum cognitive improvement with the least time input.

[0021] 4、The application generates new and personalized exercises dynamically by using a large language model through the control generation type question construction module, and can accurately control the type, difficulty, target knowledge point and examination relationship of the generated question through structured prompts, so that the on-demand generation capability greatly enriches the diversity and adaptability of the exercises, and makes up for the deficiencies of the traditional static question bank, such as limited number, not timely updating and difficult to accurately match the personalized needs of the students.

[0022] 5、The application carries out multi-dimensional deep evaluation on the students' subjective question answers through the automatic correction module, including semantic similarity, causal consistency and viewpoint rationality, and adopts a weighted comprehensive scoring model to realize more objective and comprehensive evaluation on the students' understanding depth of historical knowledge, logical reasoning ability and correct historical viewpoint; meanwhile, the dimension weights can be dynamically adjusted, so that the correction can be more flexibly adapted to different question types and teaching targets to provide more guiding feedback for the students. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The system flowchart of the application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the application.

[0025] Please refer to the accompanying Figure 1 The embodiment of the application provides a personalized historical exercise recommendation and automatic correction system based on a large model, which comprises: A cross-modal data acquisition module is configured to acquire the text input, time behavior and interaction trajectory data of the student during the answering process and output the acquisition result to a cognitive graph construction module.

[0026] The cross-modal data acquisition module comprises: A text acquisition unit is configured to extract the text answering content input by the student. A behavior time acquisition unit is configured to record the time length and pause interval during the answering process of the student. A behavior trajectory acquisition unit is configured to acquire the input sequence, modification operation and interaction trajectory information of the student on the answering interface. A structure identification unit is configured to extract the causal structure information between historical events from the text content.

[0027] For the cross-modal data acquisition module, in the embodiment, the cross-modal data acquisition module serves as the perception layer of the system of the application, and its main purpose is to comprehensively and finely capture various behavior data of the student on the learning interaction interface to provide rich and multi-dimensional original input for the subsequent cognitive graph construction module. The module runs throughout the student answering behavior and aims to represent the invisible thinking process of the student through quantifiable data dimensions.

[0028] Specifically, the module realizes non-intrusive and accompanying collection of student learning behavior data through a front-end monitoring script or a deeply integrated application program interface (API) deployed on the user terminal. The collected data types are not limited to the final answer submitted by the student, but constitute a comprehensive data set including the behavior process and the final result.

[0029] The collection of cross-modal data specifically includes the following aspects: First, the module collects the explicit content data of the student's answer. This data is the most direct manifestation of the student's learning output. For example, for a multiple-choice question, the student's final selected option is recorded; for a fill-in-the-blank question, the student's input text content is recorded; for a short answer or essay question, the student's final submitted text passage is recorded in its entirety. This part of data constitutes the basis for semantic analysis by the subsequent automatic grading module.

[0030] Second, the module focuses on collecting interactive behavior process data that can reflect the cognitive load and thought trajectory of the student when answering the question. This part of data is the key support for the invention to realize deep modeling of the student's cognitive state, and the precision of the collection directly affects the accuracy of subsequent recommendation and grading.

[0031] Preferably, the interactive behavior process data includes time dimension data. The module not only records the total duration of the student's answer to a question, i.e. the complete time span from the completion of loading the question to the student's click on the submit button, but also records a series of key behavior nodes with high-precision timestamps, such as the time when the student starts reading the question stem, the time when the student first inputs characters in the answer box, and the time when the final submission is made. These segmented duration data, such as "first-response latency", can indirectly reflect the student's familiarity with the question and the fluency of knowledge extraction.

[0032] Preferably, the interactive behavior process data also includes the student's editing behavior data. During the student's process of answering subjective questions, the module will record and save all operation sequences in the text input box in real time, including but not limited to character addition, deletion, modification, and text cutting and pasting. Through sequential analysis of these editing behaviors, an "evolution path" of the student's answer from the first draft to the final draft can be constructed. This path reveals the student's thought revision process, for example, repeated modifications of a historical detail may indicate that the knowledge point is fuzzy in memory, while significant adjustments to the overall argument logic may reflect a deepening understanding of the causal relationship of the event.

[0033] Preferably, the interactive behavior process data further comprises interface interaction event data of the student. The module monitors the mouse or touch operation trajectory of the student on the answer interface. For example, the duration and frequency of the student's mouse cursor hovering over a specific keyword of the stem, a paragraph of historical materials or a selected item of a multiple-choice question are recorded. Long hovering or repeated clicking to review is usually a signal of the student's focused attention or understanding difficulties. These data provide direct evidence for the system to determine which specific knowledge elements the student has invested more cognitive resources or encountered difficulties.

[0034] In addition, the module also collects the student's answer pause data. By analyzing the character input time interval of the student during the text input process, the module can identify abnormal pauses. Unlike the natural pauses in the regular typing process, a long pause may correspond to the student's deep recall, reasoning or organization of complex knowledge structures. The quantitative analysis of these pause patterns provides a unique perspective for evaluating the degree of automation of students' knowledge application and the coherence of their thinking.

[0035] Finally, the cross-modal data collection module structures all the data collected above, including explicit answer content, time dimension data, editing behavior data, interface interaction event data and answer pause data. Each piece of data is marked with a unique student identifier (UserID), a unique question identifier (QuestionID) and an accurate timestamp, and integrated into a complete, multi-dimensional learning behavior record. This record is then securely transmitted to the cognitive graph construction module as the basis for dynamically updating the student's personal cognitive graph, ensuring the accuracy and effectiveness of all subsequent personalized services.

[0036] The cognitive graph construction module receives the collection results output by the cross-modal data collection module, and is configured to generate a cognitive graph representing the student's mastery of historical knowledge points and their relationships based on the collection results, and send the cognitive graph to the cognitive-driven recommendation module and the automatic grading module.

[0037] In the cognitive graph construction module, the step of generating a cognitive graph representing the student's mastery of historical knowledge points and their relationships includes: Building a knowledge graph node set based on a pre-set historical knowledge point relationship network, the nodes in the node set representing specific historical knowledge points, and the edges representing the relationships between the knowledge points; Mapping the student's answer behavior data to the corresponding knowledge points, and constructing an original mastery vector by counting the student's accuracy, answer time and answer frequency at each knowledge point ; Normalizing the original mastery vector to obtain a standardized knowledge mastery vector , each dimension in the standardized knowledge mastery vector represents the student's mastery of a certain knowledge point. In addition, the mathematical form of the normalization method is: ; in, For students For the first Normalized mastery of each knowledge point; For students For the first The original mastery of each knowledge point; For students The minimum mastery level among all knowledge points; For students The maximum mastery level among all knowledge points; Based on the standardized knowledge mastery vector, the logical relationship between knowledge points, the causal relationship, and the structural dependency relationship, a propagation mechanism is introduced to model the influence between knowledge points, and a relationship matrix is ​​used to represent the connection strength between knowledge points. The formula for calculating the student's knowledge mastery after considering relationship propagation is: ; in, is a weighting factor used to balance the students’ original mastery level with the mastery level after the relationship between knowledge points is propagated; For students The mastery vector after considering the spread of knowledge point relationships; is the relationship propagation matrix between knowledge points; according to And the knowledge point relationship network is used to construct the student's individual cognitive map, in which each node is accompanied by the student's final mastery value of the knowledge point, and the edge weight is combined to represent the influence relationship between knowledge points, thereby generating a complete personalized historical knowledge cognitive map.

[0038] The cognitive graph construction module further includes a graph updating unit, which is used to update the edge weights in the cognitive graph according to the scoring results output by the automatic correction module.

[0039] In this embodiment, the cognitive map construction module is the core processing hub of the system described in this invention. Its primary purpose is to construct a refined and dynamic personal cognitive map for each individual student user. This map, in the form of a networked graph structure, intuitively and quantitatively depicts the student's mastery of various knowledge points in the history discipline and their understanding of the internal connections between these knowledge points.

[0040] Specifically, the operation of the module is based on a pre-constructed, authoritative basic historical knowledge network. The basic network is calibrated by experts in the field of history education, or generated by knowledge extraction from a large number of history teaching materials, academic literature and authoritative historical materials through natural language processing technology. In the network, each "node" represents an atomized historical knowledge point, such as "the establishment of the imperial examination system", "the Sarajevo incident" or "the New Deal of Roosevelt"; the "edge" between the nodes defines the objectively existing and teaching meaningful association between the knowledge points.

[0041] Preferably, the relationship type at least includes: causal relationship, used to represent the cause and effect between historical events, such as "the second industrial revolution" pointing to "the transition of major capitalist countries to imperialism"; time sequence relationship, used to represent the order of occurrence of events; and concept subordination relationship, used to represent the hierarchical attribution of knowledge points, such as "the reign of Zhen Guo" and "the prosperity of Kai Yuan Dynasty" belong to "the prosperity of Tang Dynasty" which is a higher level concept. The basic knowledge network provides a unified and stable initial framework for the individualized cognitive modeling of all students.

[0042] When the module receives the student learning behavior record transmitted by the aforementioned cross-modal data collection module, it starts the initialization or dynamic updating process of the student's personal cognitive graph. For students who use the system for the first time, the system will generate a copy of the basic historical knowledge network as the initial state of their personal cognitive graph. Subsequently, the module adjusts the attributes of the nodes and edges in the graph according to the continuously incoming behavior data.

[0043] In the updating process, the module first calculates the original mastery degree of the specific knowledge point involved in the student's answer behavior. This value is a preliminary quantitative evaluation based on the student's direct behavior performance. Its calculation can be a comprehensive function that integrates multiple features extracted from cross-modal data, such as the student's historical answer accuracy, the total number of times the student answers questions related to the knowledge point, the average response time, and the number of editing and modifying times during the answering process. If a knowledge point is answered multiple times, quickly and accurately by the student, the corresponding original mastery degree value will be relatively high.

[0044] However, evaluating the mastery degree of each knowledge point in isolation cannot truly reflect the student's structured knowledge system. Therefore, a key technical feature of the invention is to introduce a cognitive state propagation mechanism. This mechanism simulates the characteristic of knowledge transfer in the human learning process, that is, the understanding of a knowledge point will affect other knowledge points associated with it. The module realizes this propagation process through the following mathematical formula: ; In this formula: Representative knowledge points The final mastery degree of a knowledge point in a student's cognitive map after the cognitive propagation update. This value reflects the student's mastery of the knowledge point in the overall knowledge network.

[0045] The original mastery degree calculated based on direct behavior.

[0046] is a relationship matrix, and its definition comes from the basic historical knowledge network. This matrix quantifies the strength and type of association between knowledge points. For example, if knowledge point has a direct and close causal relationship with , the value of the corresponding element in the matrix will be higher.

[0047] The product operation is essentially using the relationship matrix to weight and sum the mastery degrees of all other knowledge points related to knowledge point , thereby calculating the "influence degree" of the surrounding knowledge network on knowledge point .

[0048] is an adjustable balance hyperparameter, with a value range of 0 to 1. This parameter is used to balance the importance of students' direct performance and knowledge association reasoning in the model. When is higher, the system tends to accept the original mastery degree reflected by the students' answer behavior; conversely, when is lower, the system focuses more on the association propagation effect within the knowledge network. This parameter can be dynamically adjusted according to the teaching objectives or the learning stage of the students.

[0049] Through the execution of the above cognitive propagation algorithm, the updating process of the student's cognitive map goes beyond simple single-point information accumulation. For example, even if a student has never directly practiced a question related to "The Treaty of Versailles", if they show a high mastery degree for knowledge points with strong associations such as "the end of World War I" and "the establishment of the League of Nations", this module can infer that the student has a certain potential understanding of "The Treaty of Versailles" and accordingly increase their final mastery degree .

[0050] Finally, the student's personal cognitive map generated after processing by this module has a quantitative final mastery degree The attribute accurately reflects the knowledge state of the student at the current time point. The dynamically evolving cognitive map is then output to a subsequent cognitive-driven recommendation module as the fundamental basis for diagnosing weak knowledge links and making accurate exercise recommendations.

[0051] The cognitive-driven recommendation module receives the cognitive map generated by the cognitive map construction module, identifies knowledge mastery defects in the cognitive map, and outputs a set of exercise recommendations to the control-generated exercise construction module.

[0052] The cognitive-driven recommendation module includes: The defect identification unit identifies a set of edges in the cognitive map whose weights are below a preset threshold as a set of cognitive defect paths. The graph structure optimization unit selects a minimum set of exercises that cover the defect paths from a preset exercise library based on the set of cognitive defect paths as the set of recommended exercises.

[0053] For the cognitive-driven recommendation module, in this embodiment, the cognitive-driven recommendation module plays the role of "intelligent diagnosis and prescription" in the entire system process. Its core task is to receive and deeply analyze the student's personal cognitive map generated by the aforementioned cognitive map construction module, accurately locate the knowledge structural defects reflected therein, and generate an efficient and clearly directed exercise recommendation scheme based on the diagnosis results.

[0054] Specifically, the function implementation of this module mainly includes two closely connected steps: accurate identification of cognitive defects and optimized generation of the set of recommended exercises.

[0055] First, the module performs a cognitive defect positioning operation. The input of this operation is the personal cognitive map that carries the current complete knowledge state of the student. One of the significant features of this invention is that the defect positioning process is not limited to simply retrieving isolated knowledge nodes with low mastery values. More deeply, the module aims to excavate the "weak paths" in the student's knowledge system, i.e., the "edges" with low connection strength.

[0056] For example, a student's cognitive map may show that he or she has a high mastery of two independent knowledge nodes (Node) "Westernization Movement" and "Sino-Japanese War of 1894-1895", but the weight of the "edge" connecting the two is very low. This reveals an implicit cognitive defect: the student understands the historical facts but fails to establish a deep internal logical connection between the two. To identify such structural problems, the module can use related algorithms in graph theory.

[0057] Preferably, the module can automatically and systematically identify all similar weak connections by running algorithms such as finding "minimum cut" or identifying "weakest path" on the student's personal cognitive map. These identified low-weight nodes and low-weight edges together form a structured "cognitive deficiency list", which is the direct basis for subsequent targeted intervention.

[0058] After completing the accurate identification of cognitive deficiencies, the module then performs the operation of generating a recommended question set. The goal of the module is not to simply list all deficiency-related questions, but to pursue "precision" and "efficiency" in recommendations.

[0059] Based on the cognitive deficiency list generated in the previous step, the module will search in the system's question bank resources or request from the subsequent control-generated question construction module to match exercises that can effectively cover the identified deficiencies.

[0060] In this process, the core optimization goal of the module is to find a minimum cardinality question set that can achieve maximum coverage of all current cognitive deficiencies with the least amount of practice. This design aims to reduce the learning burden of students and avoid inefficient and repetitive training. In algorithm implementation, this problem can be modeled as a classic "set cover problem", where each cognitive deficiency is an element that needs to be covered, and each question is a set that can cover one or more elements. The goal is to select the fewest sets to cover all elements. The module can use known approximation algorithms to solve this problem, thereby obtaining a question recommendation list close to the optimal solution.

[0061] Finally, the output of the module is not a simple question list, but a complete and information-rich recommendation plan. In addition to one or more questions selected through optimization, the plan also includes a "recommendation reason" for each recommended question. The reason explicitly tells the student what the main learning goal of the question is. For example, the system may prompt: "This question is recommended to help you understand the ideological inheritance and evolution relationship between 'New Culture Movement' and 'May 4th Movement'." This makes the student's practice more purposeful and reflective.

[0062] The control-generated question construction module receives the question recommendation set from the cognitive-driven recommendation module, constructs a prompt input based on the cognitive map and the recommendation set, and calls a large language model to generate a target exercise. The target exercise is submitted to the student for answering through the system interaction interface.

[0063] In the control-generated question construction module, the step of constructing a prompt input based on the cognitive map and the recommendation set includes: reading the student mastery degree vector obtained in the generating cognitive map step; sorting the values corresponding to each knowledge point in the student mastery degree vector, and selecting a number of knowledge points with the lowest mastery degree to form a recommendation set; For each knowledge point in the recommendation set, based on the adjacent nodes and edge weight relationship of the knowledge point in the knowledge graph, the associated concepts, key historical events and time information of the knowledge point are extracted, and the corresponding structured prompt content is generated; Organize the structured prompt information into a prompt input, which is used to provide personalized learning support in combination with the student's knowledge mastery state.

[0064] The control generation question construction module dynamically generates the corresponding question content according to the question recommendation set, and the question types include multiple choice questions, fill-in-the-blank questions, short answer questions and open-ended generation questions.

[0065] For the control generation question construction module, the control generation question construction module is the key execution unit to realize the true personalization and adaptability of the system described in the present application. Its main purpose is to break through the inherent limitations of traditional static question banks in quantity, dimension and pertinence, and dynamically construct new and fully matched exercises for students according to their current cognitive state by using the generation ability of large language models (LLM).

[0066] Specifically, this module plays a transitional role in the system workflow. It receives the recommendation scheme containing the student's specific cognitive defects output by the aforementioned cognitive-driven recommendation module, and converts it into an instruction that has precise constraints and guidance for the large language model. This process ensures that the final generated question is neither random nor generalized, but a "surgical knife" type of precise intervention with a clear teaching goal.

[0067] The core technology of this module lies in its structured prompt (Prompt) construction mechanism. This mechanism does not simply pass the name of the cognitive defect directly to the large language model, but deeply processes and organizes it into a multi-dimensional, information-rich structured instruction. This instruction provides a clear framework, detailed constraints and specific goals for the generation process of the large language model.

[0068] Preferably, the construction process of the structured prompt includes the following steps: This module first analyzes the received cognitive deficit list and takes the knowledge points or knowledge point associations to be strengthened as the core target knowledge points for generating questions. For example, if the recommendation module identifies that the student lacks understanding of the comparison between "Meiji Restoration" and "Guguzi Reform", then these two knowledge points are set as the core target for this generation task.

[0069] Subsequently, the module defines the specific relationship type that needs to be examined according to the nature of the cognitive deficiency. If the deficiency lies in the logical connection between knowledge points, the prompt will explicitly require the examination of "cause and effect relationship", "contrast relationship" or "impact relationship" and the like. This is directly derived from the analysis of the weak "edges" in the student's cognitive map, aiming to guide students to build a more solid knowledge network structure.

[0070] Next, the module determines the type of question to be generated according to the teaching strategy or the type of deficiency. For example, for memory deficiency of basic historical facts, "multiple choice questions" or "fill-in-the-blank questions" can be generated; for logical deficiency that requires in-depth reasoning, "material analysis questions", "short answer questions" or "open discussion questions" can be generated to match different examination dimensions.

[0071] At the same time, the module will dynamically set a reasonable difficulty level according to the mastery degree of related nodes in the student's cognitive map and the overall historical performance of the student. This ensures that the generated questions can effectively challenge the students without discouraging their learning enthusiasm due to excessive difficulty, achieving adaptive adjustment of difficulty.

[0072] A key feature of the present application is that the structured prompt also contains a description of the student's current cognitive level. The module will extract the student's existing understanding of the target knowledge point from the cognitive map (for example, "has basic knowledge of the Meiji Restoration, but does not clearly understand the specific economic measures") and write it into the prompt as context information. This information can guide the large language model to better grasp the starting point and focus when generating questions, making the questioning method more targeted.

[0073] In addition, for question types that require situational learning (such as material analysis questions), the module can also be configured to dynamically select or generate context information from an external knowledge base. That is, the system can automatically retrieve a historical material related to the target knowledge point that the student has not previously encountered, or instruct the large language model to generate a short background material, and integrate it into the prompt as the stem of the generated question, thereby ensuring the novelty and contextuality of the question.

[0074] After completing the construction of the above structured prompt, the module sends it to the preset large language model service interface. Preferably, the large language model can be a general-purpose model pre-trained on a large amount of general text, or a domain-specific model that has been fine-tuned on a corpus of historical education materials, papers, historical materials and the like. The latter has a deeper understanding of the terminology, logic and value judgment standards of the history discipline, and can generate more professional, accurate and teaching-standard-compliant content.

[0075] Finally, the module receives the returned result of the large language model. The result is a complete, directly presentable to students, exercise package, which not only contains carefully designed stems and questions, but also carries a detailed standard answer reference or multi-dimensional scoring points. The new question generated by this module is transmitted to the user interface for the students to answer, thus completing the complete closed loop from cognitive diagnosis to personalized content generation, providing core content support for the realization of the self-adaptive learning path of thousands of people.

[0076] The automatic correction module receives the student's answer content, and is used for semantic scoring, causal consistency scoring and viewpoint rationality scoring of the student's answer based on the standard answer and the cognitive map, and outputs the scoring result to the cognitive feedback regulation module.

[0077] The automatic correction module comprises: a semantic similarity scoring unit for calculating the semantic embedding similarity between the student's answer and the standard answer; a causal consistency scoring unit for identifying the time and causal logic structure between historical events in the student's answer and comparing with the standard answer; a viewpoint rationality scoring unit for evaluating the integrity of the student's viewpoint and the rationality of the historical position; a comprehensive scoring unit for calculating the final score according to the following formula: wherein, , and are the semantic similarity scoring weight coefficient, the causal consistency scoring weight coefficient and the viewpoint rationality scoring weight coefficient respectively, and can be dynamically set according to the cognitive map state; is the overall score of the student to the test question ; , and are the semantic similarity score, the causal consistency score and the viewpoint rationality score of the student to the test question .

[0078] For the automatic correction module, in this embodiment, the automatic correction module is the key feedback link for the system of the application to realize the teaching closed loop, and its main purpose is to automatically, multi-dimensionally and deeply intelligently evaluate the student's submitted answer, especially the subjective questions such as short answer questions and discussion questions which contain complex semantics and logic. The design concept of this module is to go beyond the traditional correction method based on keyword matching or shallow semantic comparison, so as to provide more accurate, objective and diagnostic scoring results.

[0079] ​Specifically, the module receives the student's answer content delivered by the user interface, and calls the built-in multi-dimensional comprehensive scoring model to quantitatively evaluate the answer. One key technical feature of the invention is that the final score is not the product of a single dimension, but is composed of multiple orthogonal evaluation dimensions weighted, and the core scoring model is as follows: ; In this scoring model: represent the student the final comprehensive score of the question.

[0080] represent the semantic similarity score, which aims to evaluate the degree of fit between the student's answer and the key points of the standard answer in content.

[0081] represent the causal consistency score, which aims to evaluate whether the logical relationship between the historical facts stated in the student's answer is accurate.

[0082] represent the view rationality score, which aims to evaluate the quality of argumentation and historical perspective exhibited by the student in open-ended questions.

[0083] is the dynamic weight coefficient of each score, and the sum is 1. These weights can be adaptively adjusted according to the examination objectives of the question and the cognitive state of the student, so as to achieve flexible and focused correction.

[0084] The specific implementation of each evaluation dimension is described in detail as follows: Regarding the implementation of semantic similarity scoring: To accurately grasp the deep meaning of the answer content, the module uses text representation technology based on deep learning. First, the system will pre-construct a set of core, non-divisible scoring points for the standard answer. When correcting, the module uses a pre-trained sentence embedding model (preferably, a model such as Sentence-BERT that has advantages in representing sentence-level semantics) to map the student's answer text and each scoring point of the standard answer into a high-dimensional semantic vector space. By calculating the cosine similarity or other distance measures between the student's answer vector and each scoring point vector, the module can quantify which points the student's answer covers and the degree of coverage. The final score is the result of comprehensive calculation (e.g., taking the maximum value or weighted average) based on these similarities, which accurately reflects the student's mastery of the scoring points.

[0085] Regarding the implementation of causal consistency scoring: ​The core of the history discipline lies in understanding the logical connections between events. To this end, the module is specially designed to evaluate the consistency of causality. The implementation process is as follows: first, through the named entity recognition (NER) and relation extraction (RelationExtraction) techniques in natural language processing, the student's answer text is parsed, and the historical entities mentioned in the text (such as events, characters, and time) and the relationships between them described by the student (such as "A leads to B" and "C is the background of D") are automatically identified. Subsequently, the module compares and verifies these extracted "fact-relation" triples with the authoritative relationships stored in the aforementioned basic history knowledge network of the invention. If the student's statement of causality or chronology is consistent with the definition in the knowledge network, the score will be higher; if there are logical fallacies such as causality inversion or time disorder, the score will be reduced accordingly. This mechanism enables the grading to go deep into the structural level of knowledge.

[0086] Implementation of the scoring of the rationality of the opinion: For open and critical questions, the module introduces a large language model enhanced by specific domain knowledge for evaluation. Preferably, the model is fine-tuned using a large amount of professional historical academic papers, authoritative journal reviews, and high-scoring papers reviewed by experienced teachers, etc. After this process, the model learns the expert knowledge of judging historical argumentation questions and can evaluate the student's opinion from multiple aspects, including: the closeness of the combination of history and argumentation, whether the argument is supported by sufficient historical facts; the rigor and completeness of the argumentation process; whether the opinion is objective and dialectical, and whether it conforms to the correct historical values, avoiding one-sided or extreme conclusions. The model outputs a quantitative opinion rationality score based on the comprehensive judgment Finally, the module weights and sums the scores of the above three dimensions according to the preset or dynamically adjusted weight coefficients 、 and of the current question to obtain the final comprehensive score.

[0087] The module outputs not only a single total score, but also a structured evaluation report containing the scores of each item. This detailed report will be transmitted to the cognitive feedback regulation module, serving as the most direct and accurate basis for updating the student's personal cognitive map, thereby effectively driving the operation and iterative optimization of the entire adaptive learning system.

[0088] The cognitive feedback regulation module receives the scoring results, updates the cognitive map, and outputs system optimization parameters to the cognitive map construction module and the cognitive-driven recommendation module according to the changes in cognitive state and scoring fluctuations.

[0089] The steps for the cognitive feedback control module to update the cognitive map include: Receive the scoring results output by the automatic grading module; Based on the scoring results, the following optimization objective function is used to adjust the system parameters: ; in, To update the cognitive map before; This is the updated cognitive map; The volatility variance of student ratings; is the balance parameter; is the set of system parameters to be optimized; For the parameter set Minimize on; For students expected value; For students The difference between the original cognitive map and the reconstructed map; Based on the above objective function, the parameter set is optimized through the graph neural network to generate an updated student cognitive map to more accurately represent the students' mastery of historical knowledge.

[0090] The difference between the graph before and after the update is calculated using the following formula: ; in, The cognitive map before updating is at the index position The value of The updated cognitive map is at the index position The value at .

[0091] By introducing this difference, the model distribution deviation can be constrained during the update process, preventing overfitting and improving the stability and convergence of the model.

[0092] The cognitive feedback control module, in this embodiment, is the key to the system's "closed-loop adaptive learning" capabilities. Its core function is to receive and analyze the student evaluation results output by the automatic grading module, dynamically updating the student's personal cognitive map based on these results. Furthermore, by optimizing the system's internal parameters, it ensures that the recommended and grading strategies consistently match the student's actual learning status and outcomes, enabling the system's continuous evolution and self-improvement.

[0093] Specifically, the running mechanism of the module is based on a well-designed optimization objective function. This function not only focuses on the accurate modeling of students' cognitive states, but also aims to improve the stability and effectiveness of the learning process. The module minimizes this objective function by continuously adjusting the system parameters, thereby driving the optimization process of the entire system.

[0094] The optimization objective function is as follows: ; In this objective function: It means that the goal of the module is to minimize the value of the entire function by adjusting the adjustable parameter set within the system . The parameter set widely covers the configurable items in various modules of the system, such as the parameters in the cognitive map construction module, the weight coefficients , , , in the automatic grading module, and the parameters used in the set cover algorithm in the cognitive-driven recommendation module, etc. It means that the expectation value is taken for the learning process of all users (students), which means that the optimization of the system is for the general improvement of all students, rather than for a specific student; It represents the cognitive map that the student has before this learning activity; It represents the updated cognitive map inferred from the latest performance of the student after completing this answer and being evaluated by the automatic grading module. Based on the grading results, the module will make fine adjustments to the mastery of knowledge points and the strength of knowledge point connections involved in , thereby obtaining ; It represents the KL divergence (Kullback-Leibler Divergence), which is used to measure the difference between the updated cognitive map and the updated cognitive map. KL divergence is a measure of the difference between two probability distributions in information theory. In this invention, it is used to evaluate the "magnitude" or "stability" of the cognitive map update. The purpose of minimizing this item is to ensure that the update of the cognitive map is smooth and gradual, avoiding the dramatic changes in the modeling of students' cognitive states due to the fluctuations of single answer results. This helps to maintain the coherence and stability of the evaluation of students' cognitive states, preventing "overfitting" of single performance; It represents the cognitive map that the student The variance of scores in a series of recent learning activities. One of the important goals of this module is to minimize this. A lower score variance means that the student's performance is more stable over time and across different types of questions, avoiding large fluctuations in scores. This directly reflects the "steady improvement" and "firm grasp of knowledge" of learning effectiveness, rather than accidental correctness or error. By optimizing this, the system encourages students to form a lasting and balanced mastery of knowledge; is an adjustable balance hyperparameter that adjusts the relative importance between the cognitive graph stability goal and the student score stability goal. When is high, the system tends to optimize the stability of the student's learning effectiveness; when it is low, it focuses more on maintaining the smoothness of the cognitive graph update. The setting of this parameter can be configured according to the overall teaching goal or the learning characteristics of different education stages.

[0095] This module continuously and iteratively applies optimization algorithms (preferably, gradient descent or its variants) to calculate the gradient of the objective function according to the real-time grading results, and fine-tunes the system parameters along the negative gradient direction. Each adjustment of the parameters will affect the strategy of subsequent question recommendation, the focus of question generation, and the weight distribution of automatic grading, thus forming a continuous improvement feedback loop.

[0096] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A personalized history exercise recommendation and automatic correction system based on a large model, characterized by: include: The cross-modal data collection module is used to collect students' text input, time behavior and interaction trajectory data during the test-taking process, and output the collection results to the cognitive map construction module; a cognitive map construction module, which receives the collection results output by the cross-modal data collection module, generates a cognitive map representing the student's mastery of historical knowledge points and their relationships based on the collection results, and sends the cognitive map to the cognitive-driven recommendation module and the automatic correction module; a cognitive-driven recommendation module, which receives the cognitive map generated by the cognitive map construction module, identifies knowledge gaps in the cognitive map, and outputs a set of recommended questions to the control-generation question construction module; Controlling the generative question construction module to receive the question recommendation set from the cognitive-driven recommendation module, construct prompt input based on the cognitive map and the recommendation set, and call the large language model to generate target exercises, which are submitted to the student for answer via the system interactive interface; An automatic grading module receives student responses and is used to perform semantic scoring, causal consistency scoring, and viewpoint rationality scoring on the student responses based on the standard answers and the cognitive map, and outputs the scoring results to the cognitive feedback control module; The cognitive feedback control module receives the scoring results, is used to update the cognitive map, and outputs system optimization parameters to the cognitive map construction module and the cognitive drive recommendation module according to the changes in cognitive state and the fluctuations in the scores.

2. The large-model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: The cross-modal data acquisition module includes: The text collection unit is used to extract the text answer content entered by students; Behavior time collection unit, used to record the length of time and pause intervals during students' answering process; Behavior track collection unit, used to collect students' input sequence, modification operations and interaction track information on the answering interface; The structure recognition unit is used to extract causal structure information between historical events from the text content.

3. The large-model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: In the cognitive map construction module, the step of generating a cognitive map representing the student's mastery of historical knowledge points and their relationships includes: Construct a knowledge graph node set based on a preset historical knowledge point relationship network, where the nodes in the node set represent specific historical knowledge points and the edges represent the relationships between the knowledge points; Map the students' answer behavior data to the corresponding knowledge points, and construct the original mastery vector by counting the students' accuracy, answer time and number of answers on each knowledge point. ; Normalize the original mastery vector to obtain the standardized knowledge mastery vector , each dimension in the standardized knowledge mastery vector represents the student's mastery of a certain knowledge point. In addition, the mathematical form of the normalization method is: ； in, For students For the first Normalized mastery of each knowledge point; For students For the first The original mastery of each knowledge point; For students The minimum mastery level among all knowledge points; For students The maximum mastery level among all knowledge points; Based on the standardized knowledge mastery vector, the logical relationship between knowledge points, the causal relationship, and the structural dependency relationship, a propagation mechanism is introduced to model the influence between knowledge points, and a relationship matrix is ​​used to represent the connection strength between knowledge points. The formula for calculating the student's knowledge mastery after considering relationship propagation is: ； in, is a weighting factor used to balance the students’ original mastery level with the mastery level after the relationship between knowledge points is propagated; For students The mastery vector after considering the spread of knowledge point relationships; is the relationship propagation matrix between knowledge points; according to And the knowledge point relationship network is used to construct the student's individual cognitive map, in which each node is accompanied by the student's final mastery value of the knowledge point, and the edge weight is combined to represent the influence relationship between knowledge points, thereby generating a complete personalized historical knowledge cognitive map.

4. The large-scale model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: The cognitive graph construction module further includes a graph updating unit, which is used to update the edge weights in the cognitive graph according to the scoring results output by the automatic correction module.

5. The large-model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: The cognitive driven recommendation module includes: a defect identification unit, configured to identify a set of edges in the cognitive graph whose edge weights are lower than a preset threshold as a set of cognitive defect paths; The graph structure optimization unit is used to select a minimum set of questions covering the defect paths from a preset question bank based on the cognitive defect path set as a recommended question set.

6. The large-model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: In the control generation question construction module, the step of constructing prompt input based on the cognitive map and the recommendation set includes: Reading the student mastery vector obtained in the step of generating the cognitive map; Sort the values ​​corresponding to each knowledge point in the student's mastery vector, and select several knowledge points with the lowest mastery level to form a recommendation set; For each knowledge point in the recommendation set, based on the adjacent nodes and edge weight relationships of the knowledge point in the knowledge graph, the associated concepts, key historical events and time information of the knowledge point are extracted to generate corresponding structured prompt content; The structured prompt information is organized into prompt input, and the prompt input is used to provide personalized learning support in combination with the student's knowledge mastery status.

7. The large-scale model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: The automatic correction module includes: Semantic similarity scoring unit, used to calculate the semantic embedding similarity between the student's answer and the standard answer; The causal consistency scoring unit is used to identify the temporal and causal logical structure between historical events in students' answers and compare them with the standard answers; The unit of scoring on the rationality of viewpoints is used to assess the completeness of students’ viewpoints and the rationality of their historical positions; The comprehensive scoring unit is used to calculate the final score according to the following formula: ； in, 、 and They are the semantic similarity score weight coefficient, the causal consistency score weight coefficient and the opinion rationality score weight coefficient, and can be dynamically set according to the cognitive map state; For students Test questions Overall rating of 、 and For students Test questions semantic similarity score, causal consistency score and opinion rationality score.

8. The large-scale model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: The step of updating the cognitive map by the cognitive feedback control module includes: Receiving the scoring result output by the automatic correction module; Based on the scoring results, the following optimization objective function is used to adjust the system parameters: ； in, To update the cognitive map before; This is the updated cognitive map; The volatility variance of student ratings; is the balance parameter; is the set of system parameters to be optimized; For the parameter set Minimize on; For students expected value; For students The difference between the original cognitive map and the reconstructed map; Based on the above objective function, the parameter set is optimized through the graph neural network to generate an updated student cognitive map to more accurately represent the students' mastery of historical knowledge.

9. The large-scale model-based personalized history exercise recommendation and automatic correction system according to claim 8 is characterized in that: The difference between the graph before and after updating is calculated using the following formula: ； in, The cognitive map before updating is at the index position The value of The updated cognitive map is at the index position The value of By introducing this difference, the model distribution deviation can be constrained during the update process, preventing overfitting and improving the stability and convergence of the model.

10. The large-scale model-based personalized history exercise recommendation and automatic correction system according to claim 1 is characterized in that: The control generation question construction module dynamically generates corresponding question content according to the question recommendation set. Question types include multiple-choice questions, fill-in-the-blank questions, short-answer questions and open-ended generation questions.

Images