Multi-agent based companion learning method and system

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and educational technology, and proposes a multi-agent-based companion learning method and system. The method includes: acquiring learning behavior data for learning situation analysis; calculating the probability of students mastering knowledge points based on item response theory; inferring students' learning emotional state through dialogue content in learning interactions; updating the memory module of the digital learning companion agent based on the learning situation analysis results, the probability of mastering knowledge points, and the learning emotional state, and then recommending matching practice questions; and generating personalized learning paths for students through a constructed digital tutor agent. This invention, based on the collaborative updating of the digital learning companion agent and the digital tutor agent, achieves dynamic generation of practice question recommendations and personalized learning paths, thereby improving the accuracy, continuity, and adaptability of learning support.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technologies of artificial intelligence and educational technology, specifically to a multi-agent-based accompanying learning method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of large language models and intelligent agent technology, educational intelligent agents are gradually being applied to educational scenarios such as teaching assistance, learning support, and process evaluation. Compared with traditional learning systems that mainly focus on curriculum content delivery, learning management, and outcome assessment, learners need continuous companionship, dynamic guidance, and precise feedback based on individual ability differences during the actual learning process. Especially in the entire process of pre-class preparation, in-class learning, and post-class consolidation, existing systems are unable to continuously adapt to students' cognitive states, learning progress, and emotional changes, and can no longer meet the development needs of personalized teaching and accompanying learning.

[0004] Existing educational intelligent agents or accompanying learning systems still have two problems in application: First, the mining and utilization of student learning process data is insufficient, and the identification of students' knowledge mastery, learning difficulties and status changes is relatively rough. It is difficult to accurately represent students' actual learning status in the accompanying learning process, resulting in insufficient targeting of learning support and teaching intervention. Second, existing learning path generation or recommendation methods mostly rely on preset rules, fixed processes or static relationships, lacking the ability to continuously respond to changes in students' learning status. It is difficult to dynamically adjust personalized learning paths according to changes in students' learning status, resulting in insufficient adaptability of learning content recommendation and learning pace arrangement. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a multi-agent-based companion learning method and system. Based on the collaborative updating of digital learning companion agents and digital tutor agents, it achieves dynamic generation of practice question recommendations and personalized learning paths, thereby improving the accuracy, continuity, and adaptability of learning support.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of this invention provides a multi-agent-based adjoint learning method, comprising the following steps:

[0008] Acquire learning behavior data, extract students' problem-solving reasoning chains for questions, compare them with standard knowledge reasoning paths, and obtain learning situation analysis results;

[0009] Based on the analysis results of the knowledge points corresponding to the target questions and the learning situation, and based on the item response theory, the probability of students mastering the knowledge points is calculated.

[0010] Based on the dialogue content of learning interactions in learning behavior data, infer students' learning emotional state;

[0011] Based on the analysis results of learning progress, the probability of mastering knowledge points, and the emotional state of learning, the memory module of the constructed digital learning companion intelligent agent is updated, and practice question recommendations that are adapted to the student's current learning status are generated based on the updated digital learning companion intelligent agent.

[0012] Based on the learning situation analysis results and the obtained learning task information, a personalized learning path for each student is generated through the constructed digital tutor intelligent agent.

[0013] Further technical solutions include building a digital learning companion intelligent agent, including student profiles and multi-layer memory modules;

[0014] Based on the results of learning situation analysis, the probability of mastering knowledge points, and data on learning emotional state, a multi-dimensional feature set is formed by comprehensively representing the current learning characteristics of students, which serves as a student profile.

[0015] The memory modules include factual memory modules, short-term memory modules, and long-term memory modules;

[0016] The fact memory module is used to store student learning records obtained from learning behavior data analysis in real time;

[0017] The short-term memory module is used to store the student's learning records obtained in the most recent short period of time;

[0018] The long-term memory module stores students' learning records that form long-term memories, simulating the stable ability structure and knowledge mastery patterns that students gradually develop during long-term learning.

[0019] A further technical solution, which involves acquiring students' learning behavior data, extracting students' problem-solving reasoning chains, comparing them with standard knowledge reasoning paths, and obtaining analysis results of students' learning progress, includes the following steps:

[0020] Extract students' answers to the target questions and the intermediate calculation steps from the learning behavior data;

[0021] The answer results and intermediate calculation steps are encoded to generate intermediate calculation step vectors. Through embedding operations and self-attention mechanisms, the dependencies and importance weights of each problem-solving step are obtained.

[0022] Based on the intermediate computation step vector, the dependencies between each problem-solving step, and the importance weight of each step, a problem-solving representation containing the step sequence relationship and key reasoning paths is constructed.

[0023] By using a subject-based educational language model to represent problem-solving, semantic analysis and causal inference are performed, mapping the student's problem-solving process into a structured problem-solving reasoning chain;

[0024] The rationality, consistency and completeness of each reasoning node in the structured problem-solving reasoning chain are verified and compared with the standard knowledge reasoning path to locate reasoning deviation nodes and determine the reasons for answering errors based on the reasoning deviation nodes.

[0025] A further technical solution, based on the analysis results of the knowledge points corresponding to the target question and the learning situation, and based on item response theory, calculates the probability of students mastering the knowledge points, including the following steps:

[0026] Based on learning behavior data, students' answers to target questions are extracted, and combined with the discrimination and difficulty parameters of the target questions, the parameters of students' answers are estimated based on the item response theory model to obtain student ability factors.

[0027] The questions are mapped to the corresponding knowledge points in the knowledge graph, and the probability of students mastering each knowledge point is calculated based on the student's ability factor.

[0028] Further technical solutions, including methods for calculating the probability of students mastering each knowledge point, include:

[0029] The intermediate calculation steps in student ability factors, questions, and learning behavior data are vectorized and concatenated to form a student behavior feature vector.

[0030] The student behavior feature vector is weighted and nonlinearly transformed to obtain the feature weight of each knowledge point. The feature weight is then mapped to the probability of the student mastering each knowledge point through a normalization function.

[0031] Further technical solutions involve updating the memory module of the constructed digital learning companion agent based on learning progress analysis results, the probability of mastering knowledge points, and the emotional state of learning. This includes the following steps:

[0032] Based on the learning situation analysis results, the mastery probability of knowledge points, and the learning emotional state data, the static profile features are updated. Using a large language model, the updated static profile features are mapped to the dynamic answer behavior results of students in the target question scenario, which serves as a new student learning record.

[0033] Based on the obtained student learning records, update the memory module of the digital learning companion agent:

[0034] Based on the obtained student learning records, update the fact memory module and reinforce the counting of similar historical records;

[0035] Based on the updated fact memory module, new student learning records are stored in the short-term memory module to preserve students' continuous learning behavior within a recently set time period.

[0036] Based on the reinforcement counting results and forgetting rules, the long-term memory module is updated to form a stable representation of students' knowledge mastery.

[0037] A further technical solution, based on the updated digital learning companion agent, generates practice question recommendations that are adapted to the student's current learning status, including the following steps:

[0038] Based on students’ knowledge point mastery probability, forgetting score and learning emotional state, identify knowledge points that need to be consolidated, improved or corrected, and use them as target knowledge points.

[0039] Based on the prompt word template, the target knowledge point, the probability of mastering the knowledge point, the time since the last practice of the knowledge point, and the forgetting score are input into the large language model to generate personalized practice question recommendations.

[0040] A further technical solution, based on the acquired learning task information and learning situation analysis results, generates a personalized learning path for students through the construction of a digital tutor intelligent agent, including the following steps:

[0041] The digital mentor intelligent agent constructs a problem graph based on a domain problem database;

[0042] The digital tutor intelligent agent constructs a knowledge graph based on domain knowledge;

[0043] Based on knowledge graphs, student learning analysis results, and current learning task information, candidate knowledge points are structurally enhanced and matched with learning states to determine the next learning objective and generate a personalized learning path that includes the target knowledge points and a set of questions extracted from the problem graph.

[0044] A further technical solution, based on knowledge graphs, student learning analysis results, and current learning task information, involves structurally enhanced representation and learning state matching of candidate knowledge points to determine the next learning objective and generate a personalized learning path containing the target knowledge points and a set of questions extracted from the problem graph. This method includes the following steps:

[0045] Based on the knowledge graph, the predecessor and successor knowledge points of the target knowledge point in the current learning task are obtained, and the target knowledge point is enhanced to obtain the enhanced text description sequence of the target knowledge point.

[0046] Encode and aggregate multi-layer neighborhoods of the enhanced text description sequence of the target knowledge points to obtain a knowledge point representation with enhanced relationships.

[0047] Based on the analysis results of students' learning situation and the current knowledge points, the students' knowledge status is dynamically modeled to obtain the students' knowledge status representation at the time step. This representation is then concatenated with the relationship enhancement representation of the current knowledge points, the knowledge point embedding vector, and the answer result embedding vector to form a unified representation of learning behavior.

[0048] The sequence of learning behaviors, represented by the learning behaviors, is input into a sequence modeling network based on a self-attention mechanism to model historical learning behaviors and obtain the context-aware feature representation of the current moment.

[0049] Based on context-aware feature representation, candidate knowledge points are matched and calculated. The probability of each knowledge point being recommended as the next learning objective is calculated. The top-ranked knowledge points are selected as the recommendation results. The set of questions corresponding to the knowledge points is then selected in the hierarchical question graph. Personalized learning paths that match the student's current knowledge status and ability level are generated by combining learning resources.

[0050] A second aspect of the present invention provides a multi-agent-based accompanying learning system, comprising:

[0051] The learning situation analysis module is configured to acquire learning behavior data, extract students' problem-solving reasoning chains for questions, compare them with standard knowledge reasoning paths, and obtain learning situation analysis results.

[0052] The probability recognition module is configured to calculate the probability of a student's mastery of a knowledge point based on the knowledge points corresponding to the target question and the analysis results of the student's learning situation, using item response theory.

[0053] The emotion state inference module is configured to infer the student's learning emotion state based on the dialogue content of learning interactions in the learning behavior data;

[0054] The emotional state module is configured to update the memory module of the constructed digital learning companion agent based on the learning situation analysis results, the mastery probability of knowledge points, and the emotional state of learning. Based on the updated digital learning companion agent, it generates practice question recommendations that are adapted to the student's current learning state.

[0055] The personalized learning path generation module is configured to generate personalized learning paths for students based on the results of learning situation analysis and the obtained learning task information, through the constructed digital tutor intelligent agent.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] This invention acquires learning behavior data, extracts students' problem-solving reasoning chains for specific questions, and compares them with standard knowledge reasoning paths to identify students' knowledge gaps, error causes, and cognitive blockages, improving the accuracy of student learning analysis from the problem-solving process perspective. Furthermore, it combines the knowledge points corresponding to the target questions with item response theory to calculate the probability of students mastering those knowledge points, thereby quantifying and dynamically assessing students' knowledge acquisition. Simultaneously, it infers students' emotional state based on learning interaction dialogue content, incorporating emotional changes during the learning process into learning support decisions. Building on this, it updates the memory module of the digital learning companion agent to comprehensively reflect students' short-term learning status and long-term learning characteristics, generating practice question recommendations tailored to the student's current learning state. Finally, combined with learning task information, the digital tutor agent dynamically generates personalized learning paths, enabling continuous response to changes in students' learning status and adaptive adjustments to learning content, pace, and teaching intervention methods, thereby enhancing the personalization, continuity, and accuracy of the accompanying learning process.

[0058] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0059] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0060] Figure 1 This is a flowchart of the multi-agent-based adjoint learning method of Embodiment 1 of the present invention;

[0061] Figure 2 This is an interactive flowchart of the accompanying learning process in Embodiment 1 of the present invention;

[0062] Figure 3 This is a technical roadmap of the digital learning companion intelligent agent of Embodiment 1 of the present invention;

[0063] Figure 4 This is a technical roadmap of the digital tutor intelligent agent in Embodiment 1 of the present invention;

[0064] Figure 5 This is a teaching interaction framework diagram of teacher-student-machine-environment in Embodiment 1 of the present invention. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0067] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0068] Example 1

[0069] In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 5 As shown, a multi-agent-based adjoint learning method includes the following steps:

[0070] Step 1: Obtain learning behavior data, extract students' problem-solving reasoning chains for questions, compare them with standard knowledge reasoning paths, and obtain learning situation analysis results;

[0071] Step 2: Based on the analysis results of the knowledge points corresponding to the target questions and the learning situation, and based on the item response theory, calculate the probability of students mastering the knowledge points;

[0072] Step 3: Based on the dialogue content of learning interactions in the learning behavior data, infer the students' learning emotional state;

[0073] Step 4: Based on the learning situation analysis results, the probability of mastering knowledge points, and the emotional state of learning, update the memory module of the constructed digital learning companion intelligent agent, and generate practice question recommendations that are adapted to the student's current learning state based on the updated digital learning companion intelligent agent;

[0074] Step 5: Based on the learning situation analysis results and the obtained learning task information, generate personalized learning paths for students through the constructed digital tutor intelligent agent.

[0075] This embodiment acquires learning behavior data, extracts students' problem-solving reasoning chains for questions, and compares them with standard knowledge reasoning paths to identify students' knowledge gaps, error causes, and cognitive blockages, improving the accuracy of student learning analysis from the problem-solving process level. Furthermore, it combines the knowledge points corresponding to the target questions with item response theory to calculate the probability of students mastering the knowledge points, thereby quantifying and dynamically evaluating students' knowledge mastery. Simultaneously, it infers students' emotional state based on learning interaction dialogue content, incorporating students' emotional changes during the learning process into learning support decisions. Based on this, it updates the memory module of the digital learning companion agent to comprehensively reflect students' short-term learning status and long-term learning characteristics, generating practice question recommendations adapted to students' current learning status. Finally, combined with learning task information, the digital tutor agent dynamically generates personalized learning paths based on the comprehensive learning analysis results, thereby achieving continuous response to changes in students' learning status and adaptive adjustment of learning content, learning pace, and teaching intervention methods, ultimately improving the personalization, continuity, and accuracy of the accompanying learning process.

[0076] The accompanying learning method in this embodiment can be implemented in an accompanying learning system, which includes a digital learning companion agent and a digital tutor agent; it is feasible that steps 1 to 4 above can be implemented by constructing a digital learning companion agent.

[0077] Optionally, a digital learning companion intelligent agent can be constructed, including student profiles and multi-layer memory modules, to simulate students' learning abilities and cognitive processes, and to depict changes in students' abilities and knowledge mastery during the learning process;

[0078] In the digital learning companion intelligent agent, a multi-dimensional feature set is formed by comprehensively representing the student's current learning characteristics based on the learning situation analysis results, the mastery probability of knowledge points, and learning emotional state data. This student profile is used to comprehensively represent the student's current learning status.

[0079] The memory modules include factual memory, short-term memory, and long-term memory. Each type of memory is empty upon initialization.

[0080] The fact memory module is used to store student learning records obtained from learning behavior data analysis in real time;

[0081] The short-term memory module is used to store the student's learning records obtained in the most recent short period of time;

[0082] The long-term memory module is used to store students' learning records that form long-term memories, in order to simulate the stable ability structure and knowledge mastery pattern that students gradually form during the long-term learning process.

[0083] Among them, the updated and constructed memory module of the digital learning companion intelligent agent can dynamically synchronize the digital learning companion intelligent agent with the student's learning ability;

[0084] Step 1: Learning behavior data may include the answer results and intermediate problem-solving steps for the target question, as well as the dialogue content of the learning interaction; student learning situation analysis results may include learning effectiveness indicators, reasons for answering errors, answer accuracy rate, etc.

[0085] Accompanying learning systems include a set of students and the set of questions Each question The association includes textual information and knowledge point information, represented as follows: For any student Their learning behavior data is represented as a time-ordered sequence of tests. :

[0086] (1);

[0087] in, Students In the The title of the step This indicates the student's answer to the question. Indicate the student's answer steps; Representation and title Related knowledge point information;

[0088] Step 1 involves acquiring students' learning behavior data, extracting their problem-solving reasoning chains, and comparing them with standard knowledge reasoning paths to obtain the analysis results of students' learning progress. This method includes the following steps:

[0089] Step 11: Extract students' answers to the target questions and intermediate calculation steps from the learning behavior data;

[0090] Step 12: Encode the answer results and intermediate calculation steps to generate intermediate calculation step vectors, and process them through embedding operations and self-attention mechanisms to obtain the dependencies and importance weights of each solution step;

[0091] Using a subject-specific large language model, semantic parsing is performed on students' answers and intermediate calculation steps. Each intermediate calculation step is encoded into a corresponding intermediate calculation step vector. By generating intermediate calculation step vectors, the dependencies and importance weights between each problem-solving step are obtained through embedding operations and self-attention mechanisms to represent the correlation between problem-solving steps and to reinforce key steps and common error steps.

[0092] Optionally, the answer result can be used as a supervision signal input to the large language model to guide the large language model in determining the correctness and error type of the answer.

[0093] For example, in solving a quadratic equation At that time, the student's step sequence can be represented as The results obtained in step 12 are shown in Table 1.

[0094] Table 1 shows the dependencies and importance weights among the problem-solving steps;

[0095]

[0096] Step 13: Based on the intermediate calculation step vector, the dependencies between each solution step, and the importance weight of each step, construct a solution representation that includes the step sequence relationship and key reasoning path, including the following steps;

[0097] Specifically, intermediate computational step vectors are constructed as nodes, directed connections are built based on the dependencies between each problem-solving step, and temporal connections are supplemented according to the order of the steps, forming a graph structure representation of the student's problem-solving process. Path searching is performed starting from the initial conditions and ending at the final result to generate a problem-solving representation, i.e., a standardized chain sequence of problem-solving steps.

[0098] Based on the results in Table 1, the graph structure representation of the student's problem-solving process can be represented as: V1→V2→V3; the final generated solution representation is: containing step nodes V1, V2, and V3, with node connection relationships of V1→V2 and V2→V3, and the weights of each node are V1:0.7, V2:0.9, and V3:0.6, respectively. The key reasoning path is V1→V2→V3.

[0099] Step 13 represents and models the student's problem-solving process at the structural level, forming a problem-solving representation that includes step nodes, dependencies, and critical paths, but without semantic information;

[0100] Step 14: Utilize the subject-specific educational language model to perform semantic analysis and causal inference on the problem-solving representation, mapping the student's problem-solving process into a structured problem-solving reasoning chain. The process includes the following:

[0101] Step 141: Using the solution representation generated in Step 13 as input, use the subject-specific large language model to perform semantic parsing and information extraction on each solution step to obtain the semantic representation of each step. The semantic representation can identify the mathematical concepts, solution basis, conditions used, operations and conclusion types corresponding to each step.

[0102] Step 142: Map each problem-solving step to a reasoning node in a structured reasoning chain, and add a corresponding semantic representation to each reasoning node;

[0103] Step 143: Construct a directed graph with the sequential or logical dependencies between adjacent steps as edges and the reasoning nodes as nodes. Extract information about each node in the problem-solving process and the relationships between nodes based on the directed graph to generate a structured problem-solving reasoning chain.

[0104] Specifically, a large language model specifically designed for subject education can be used to perform reasoning analysis, semantic parsing, and causal inference on students' problem-solving process, mapping the problem-solving process into a structured reasoning chain. It is feasible to fine-tune the pre-trained large language model specifically designed for subject education by taking standardized chain-like problem-solving steps as input and structured problem-solving reasoning chains as output.

[0105] The subject-specific large language model is a large language model that is pre-trained or adapted based on specific subject-specific educational knowledge, teaching resources, and student learning behavior data. The large language model can be a GPT series model (such as GPT-3, GPT-4 or their subsequent versions), a GLM series model (such as ChatGLM, GLM-4), etc. The subject-specific large language model is used to perform reasoning analysis on students' answer results and intermediate calculation steps. Through semantic parsing and causal inference reasoning paths, the student's problem-solving process is mapped into a structured reasoning chain.

[0106] Step 14, based on the problem-solving representation in Step 13, further assigns semantic meaning to each step and constructs causal relationships through semantic parsing and causal inference, thereby obtaining a structured problem-solving reasoning chain with clear inference semantics. The example structured problem-solving reasoning chain obtained in Table 1 is shown below:

[0107] Node V1: Equation x²-5x+6=0, Type: Problem Modeling;

[0108] Node V2: Factorize (x-2)(x-3)=0; Type: Formula Transformation; Basis: Factorization Method;

[0109] Node V3: x=2; Type: Solution result; Problem: Incomplete solution set;

[0110] Step 15: Verify the rationality, consistency and completeness of each reasoning node in the structured problem-solving reasoning chain, compare it with the standard knowledge reasoning path, locate the reasoning deviation node, and determine the reason for the answer error based on the reasoning deviation node;

[0111] Regarding the specific example above, in solving a quadratic equation... At that time, the student's step sequence can be represented as A standard inference chain should contain two solutions. If the model detects a missing step or an incomplete conclusion, it determines that the node has a reasoning bias of "incomplete solution set", thereby locating the logical derivation error.

[0112] Optionally, in step 15, a large language model can be used to progressively verify the rationality, consistency, and completeness of each reasoning node in the reasoning chain, identify the deviation from the standard knowledge reasoning path, and locate the reasoning deviations that students may produce in terms of concept understanding, condition usage, operation rules, or logical deduction. The answer results can be used as a supervision signal to be input into the large language model multiple times, and the model can be iteratively guided to identify and determine the correctness and error type of the answer.

[0113] Furthermore, it also includes step 16: calculating learning effectiveness indicators based on learning behavior data, and using the learning effectiveness indicators and the reasons for students' answer errors as the results of student learning situation analysis.

[0114] In step 16, learning effectiveness indicators are calculated based on learning behavior data. These indicators include learning input indicators, learning habit indicators, and answer performance indicators.

[0115] In some embodiments, learning performance indicators characterize students' overall learning characteristics, which are typically stable and derived from long-term learning experience. After obtaining students' historical learning behavior data, learning performance indicators can also be used to analyze students' learning progress. Construct corresponding digital learning companion intelligent agents The profile is used to describe the student's multi-dimensional characteristics, such as ability level, mastery of knowledge points, and emotional state during the learning process.

[0116] The profile serves as the core state representation of the digital learning companion agent, driving the agent to simulate, predict, and respond to student behavior. This ensures that the digital learning companion agent matches or slightly surpasses the student's current ability level in terms of problem-solving skills, error patterns, and learning performance, thereby achieving targeted guidance and feedback.

[0117] In some embodiments, student learning engagement metrics are used to characterize their investment of time, energy, and cognitive resources during the learning process, including: learning intensity metrics, learning time and learning labor intensity metrics.

[0118] Learning intensity index Specifically, it is the ratio of the number of effective learning tasks completed within a learning statistical period to the total duration of the corresponding learning statistical period. It is used to characterize students' enthusiasm, initiative, effectiveness, and persistence in participating in learning activities. The calculation formula is as follows:

[0119] (2);

[0120] in, This indicates the number of effective learning tasks completed by students within the learning statistics period. This indicates the total duration of the corresponding statistical period.

[0121] Learning time refers to the effective time students spend on learning tasks both inside and outside the classroom, measured in minutes. It is collected and defined through learning and practice-related system platforms. for:

[0122] (3);

[0123] in, Indicates that the student is in Effective learning time in each learning interaction;

[0124] The learning workload intensity index is specifically defined as the sum of standardized input intensity across all learning activities. It measures the cognitive and behavioral load of students in learning activities such as previewing, listening, recording, reviewing, practice, and creation. Its value ranges from [0,1]. Workload intensity is defined as... for:

[0125] (4);

[0126] in, Indicates that the student is in Standardized input intensity in similar learning activities This represents the total number of learning activity types.

[0127] In some embodiments, student learning habit indicators are used to characterize students' self-regulation and collaboration abilities during the learning process, including self-monitoring indicators, learning design indicators, self-reflection indicators, and collaborative learning indicators.

[0128] Self-monitoring indicators are used to characterize students' ability to self-manage and monitor their learning progress, as well as their continuous improvement behavior based on self-monitoring.

[0129] Learning design metrics are used to describe students' planning abilities in areas such as setting academic goals, developing learning plans, time management, and improving learning efficiency.

[0130] Self-reflection indicators are used to reflect students' ability to summarize, evaluate, and continuously improve their own learning process and outcomes.

[0131] Collaborative learning indicators are used to characterize students' ability to conduct collaborative learning and practice with fellow dormitory members, club members, project team members, and teachers in a group learning environment. They primarily reflect students' participation and collaboration levels in mutual assistance, exchange and discussion, and collective practical activities. These indicators are obtained through comprehensive analysis of data such as students' behavior records on the learning platform, task completion status, and periodic summary and reflection reports.

[0132] The aforementioned learning habit indicators are obtained through comprehensive analysis of data such as students' behavior records on the learning platform, task completion status, and periodic summary and reflection reports. Since these indicators are not numerically quantified, they can be constructed as prompt words and used as input for a large language model.

[0133] In some embodiments, the student's performance indicators are derived from the student's performance. Available practice records The statistical characteristics obtained from the statistics are shown in the figure, including: learning activity, success rate of answering questions, knowledge coverage, answering stability, answering efficiency, and mastery of knowledge points.

[0134] Learning activity level is the ratio between the number of practice questions completed by students and the total number of practice questions in the question bank.

[0135] Learning activity level is used to represent the degree of student engagement in learning, reflecting the level of enthusiasm students have for participating in learning activities and providing clues for simulating student practice behavior. For example, students with higher learning activity levels generally perform better. Student activity level is defined as:

[0136] (5);

[0137] in, This indicates the number of practice problems the student has completed. This indicates the total number of practice questions in the question bank. (Degree)

[0138] The success rate is the ratio of the number of correctly answered questions in a student's practice question record to the total number of practice questions completed by the student. It is used to characterize the student's level of accuracy in answering questions and is defined as the success rate. for:

[0139] (6);

[0140] in, Indicates the student's opinion on the first The success rate of answering the practice questions. It is used to characterize the student's overall answer level at the current stage.

[0141] Knowledge coverage is the ratio between the set of knowledge points that students have practiced and the pre-set set of all knowledge points. It is used to characterize the breadth of students' mastery of knowledge points and is defined as knowledge coverage. for:

[0142] (7);

[0143] in, This is a collection of knowledge points that students have already practiced. This represents the complete set of knowledge points on the platform. The diversity of exercises reflects the breadth of knowledge the student has covered within the knowledge system.

[0144] Answer stability is calculated based on the degree of change in students' success rate across different time periods or questions of varying difficulty. It reflects the fluctuation in students' accuracy during continuous answering. The core idea is: the smaller the fluctuation in accuracy, the higher the stability. The student's answer sequence is divided by time. Calculate the success rate within each of the consecutive time periods. :

[0145] (8);

[0146] in, Indicates the first The set of answers within a time period. Indicates the first Is the answer correct?

[0147] Furthermore, the variance of students' success rate in answering questions over different time periods was calculated:

[0148] (9);

[0149] in, This represents the average accuracy rate of responses.

[0150] Based on the above variance, the stability of students' responses is defined as follows:

[0151] (10);

[0152] in, The larger the value, the smaller the fluctuation in the accuracy of the student's answers and the more stable the knowledge mastery.

[0153] Answering efficiency is defined as the number of effective practice questions a student completes per unit of time. It is used to measure a student's time performance in completing practice tasks. Answering efficiency is defined as:

[0154] (11);

[0155] in, This represents the number of valid practice questions completed by students within a preset statistical time window. The total time spent by students completing all the above exercises within the statistical time window;

[0156] Knowledge point mastery is a numerical value that comprehensively assesses students' long-term retention of learned knowledge points by analyzing their correct answers to the same knowledge point multiple times, the results of repeated answers at intervals, and changes in the forgetting rate. It is used to characterize the long-term retention effect of students on any given knowledge point. The mastery of knowledge points is defined as:

[0157] (12);

[0158] in, Indicates the first This focuses on knowledge points. The answer result, This indicates the time interval between the current response and the previous response. The forgetting factor is used to characterize the decay effect of time interval on the degree of knowledge retention. Indicates the knowledge point The cumulative number of times a response was made;

[0159] Furthermore, a feature mapping model is constructed to map numerical values ​​such as learning activity, answer success rate, and knowledge coverage to different levels;

[0160] Optionally, the feature mapping model can use an interval determination algorithm to map the above learning performance indicators into three levels: high, medium, or low by comparing the relative size and degree of difference between the input value and a preset threshold; each level is assigned a different value.

[0161] In this embodiment, learning effectiveness indicators and error reasons are used together to characterize students' ability level as the result of learning situation analysis; learning effectiveness indicators are used to reflect the statistical characteristics of students in the process of answering questions, and error reasons are used to characterize the cognitive biases of students in the process of answering questions.

[0162] Step 2, based on the analysis results of the knowledge points corresponding to the target question and the student's learning situation, uses the item response theory to calculate the probability of the student's mastery of the knowledge points, including the following steps:

[0163] Step 21: Extract students' answers to the target questions based on learning behavior data, and combine the discrimination and difficulty parameters of the target questions to estimate the parameters of the students' answers based on the item response theory model to obtain the students' ability factors.

[0164] Specifically, the project response theory model is used to estimate the parameters of students' responses and infer student ability factors. The inference formula is as follows:

[0165] (13);

[0166] in, This indicates the discrimination and difficulty parameters of the questions. For the Sigmoid function; The ability factor represents the learner's potential ability level. Indicate learner In the title The answer results on Indicate learner In the title The probability of answering correctly.

[0167] Step 22: Map the questions to the corresponding knowledge points in the knowledge graph, and calculate the probability of students mastering each knowledge point based on the student's ability factor.

[0168] Step 22, the method for calculating the probability of students mastering each knowledge point, includes:

[0169] Step 221: Vectorize the intermediate calculation steps in the student ability factors, questions, and learning behavior data, and concatenate them to form a student behavior feature vector.

[0170] Among them, the target question is encoded to obtain the question encoding vector;

[0171] Step 222: Perform weighted calculation and nonlinear transformation on the student behavior feature vector to obtain the feature weights of each knowledge point, and map the feature weights to the probability of students mastering each knowledge point through a normalization function. Based on this, a knowledge point mastery matrix for students is constructed. :

[0172] (14);

[0173] in, Each row represents the student's level of mastery of each knowledge point.

[0174] Optionally, the student behavior feature vector is input into the feature fusion module, and the student behavior feature vector is weighted and nonlinearly transformed. The feature fusion module can use an attention mechanism or a multi-layer neural network structure to model the nonlinear relationship between student ability and knowledge points.

[0175] In step 3, based on the dialogue content of learning interactions in the learning behavior data, the students' learning emotional state is inferred. The large language model specifically for subject education can be used to infer and analyze the students' learning emotional state, which includes emotional factors and emotional state.

[0176] In this embodiment, by comparing and analyzing the students' actual answers with the model's inference results, the emotional factors of the students are characterized, including anxiety, overconfidence, or negative emotions. At the same time, by combining the dialogue process between the students and the large language model, the students' learning engagement, learning motivation, and attitude towards the learning content are inferred and analyzed.

[0177] In step 4, based on the learning progress analysis results, the probability of mastering knowledge points, and the emotional state of learning, the memory module of the constructed digital learning companion agent is updated, including the following steps:

[0178] Step 401: Based on the learning situation analysis results, the mastery probability of knowledge points, and the learning emotional state data, update the static profile features. Using a large language model, map the updated static profile features to the dynamic answer behavior results of students in the target question scenario, as a new student learning record.

[0179] Specifically, based on the multi-dimensional feature parameters in the student profile, a prompt word template is preset and input into a subject-specific large language model in combination with the current question content. This simulates the student's cognitive processing and answering process in the current learning state, generating knowledge point identification results, problem-solving paths, final answer results, and accuracy prediction results that match the student's current ability level, knowledge mastery, and emotional state. In this way, the static student profile is transformed into dynamic answering behavior prediction results that can be used for subsequent learning analysis and interactive decision-making.

[0180] One specific implementation involves pre-setting a prompt word template based on multi-dimensional feature parameters in the student profile, which may include defining labels for student ability level, knowledge point mastery, learning emotional state, and questions.

[0181] Student ability level labels can be determined based on the specific calculated values ​​of students' answer performance indicators in the analysis results of students' learning situation. These can include learning activity labels, answer success rate labels, knowledge coverage labels, and error reason labels.

[0182] Learning emotional state tags can include emotional factor tags and attitude tendency tags;

[0183] The correspondence between tags and various indicators can be summarized as follows: Learning activity tag, denoted as: [Learning Activity]; Answer success rate tag, denoted as: [Answer Success Rate]; Knowledge coverage tag, denoted as: [Knowledge Coverage]; Error reason tag, denoted as: [Error Reason]; Knowledge point mastery tag, denoted as: [Knowledge Point Mastery]; Emotional factor tag, denoted as: [Emotional Factor]; Attitude bias tag, denoted as: [Attitude Bias]; Question tag, denoted as: [Question];

[0184] A specific example of the prompt template content is: "You are a student with [learning activity level], an overall answer success rate of [answer success rate], a knowledge point coverage rate of [knowledge coverage rate], and an error reason of [error reason]; you have a [knowledge mastery] level of mastery of the relevant knowledge points; at the same time, your current learning emotional state is characterized by [emotional factors], and your learning attitude tendency is [attitude tendency]. Given that the currently recommended practice question is [question], please simulate the student's answering behavior based on the above profile, and perform the following tasks in sequence: 1) Identify the knowledge points tested in the current question from the knowledge graph; 2) Generate a problem-solving approach and provide the final answer; 3) Predict whether the student can answer correctly."

[0185] Based on the above template, the corresponding actual indicator content is filled into the corresponding label positions in the prompt word template to complete the prompt word instantiation and obtain the complete prompt word. The complete prompt word is then input into a subject-specific large language model to guide the model to generate problem-solving strategies and answers that match the student's current ability level, and to comprehensively evaluate the student's ability level, knowledge mastery, and learning emotional state. The question text, corresponding knowledge points, and answer results are extracted from the content generated by the large language model as the student's dynamic answer behavior results in the target question scenario, serving as a new student learning record.

[0186] To enable the digital learning companion to continuously simulate the student's learning process, a memory module is constructed for it in this embodiment. The memory module allows the agent, based on a large language model, to gradually observe and summarize the historical practice experience of its corresponding student, providing important clues for simulating responses to unseen practice questions. Based on human learning mechanisms, this embodiment designs three types of memory for the digital learning companion agent: factual memory, short-term memory, and long-term memory, each type of memory being empty upon initialization.

[0187] Step 402: Based on the obtained student learning records, update the memory module of the digital learning companion agent;

[0188] Step 4021: Based on the obtained student learning records, update the fact memory module and reinforce the counting of similar historical records;

[0189] When the agent is in the th Step to gain students When a new student response record is entered into a student's learning record, the new response record is transmitted to factual memory for processing. It can be represented as:

[0190] (15);

[0191] When students repeatedly practice similar problems or knowledge points, their memory is strengthened. This example illustrates each record in factual memory. Introduce an additional counter (Initially set to 1), used to track the number of times the memory is reinforced. Formally, for the agent... Assuming it has been observed This factual memory is represented as:

[0192] (16);

[0193] Based on the similarity between new responses and historical responses, a reinforcement count is performed on historical records in factual memory. Specifically: when a new record is received... First, compare the record with each record in the current factual memory. The similarity between two records. If two records involve similar knowledge concepts, then the two records are considered similar, and in this case, [the following will be considered]. Increment the counter by one, and... Incorporate factual memories; otherwise, directly... The record remains unchanged, and no reinforcement is applied. After the new record is processed and stored, the factual memory triggers the updating of short-term and long-term memory.

[0194] Step 4022: Based on the updated fact memory module, store the new student learning records in the short-term memory module to save the student's continuous learning behavior within a recently set time period;

[0195] Human short-term memory refers to recent information that can be retained and recalled within a relatively short period of time. The short-term memory of the digital learning companion consists of several short-term learning state units, each defined as follows:

[0196] (17);

[0197] in, This indicates the number of times the student has completed the current task. Practice questions; This represents the set of knowledge points involved in the exercise. Indicates the student's opinion on the first The results of answering the practice questions; This indicates the time it took the student to complete the exercise. This indicates that the digital learning companion agent models the first [aspect / aspect] based on its capabilities. Confidence estimation of the answers to practice questions;

[0198] Optionally, the digital learning companion's short-term memory employs a fixed-capacity sliding window mechanism, with its capacity set as follows: ; at time step Short-term memory is represented as:

[0199] (18);

[0200] When a student completes a new learning task, the digital learning companion writes the corresponding short-term learning status unit into the short-term memory module; if the short-term memory has reached its capacity limit, the oldest record is automatically removed, thus ensuring that the short-term memory always reflects the student's most recent continuous learning behavior.

[0201] Step 4023: Based on the reinforcement counting results and forgetting rules, update the long-term memory module to form a stable knowledge mastery representation for students;

[0202] The long-term memory module of the digital learning companion simulates the stable ability structure and knowledge mastery pattern that students gradually develop during long-term learning. Compared to short-term memory, the long-term memory module is formed through repeated practice and self-reflection, simulating the human long-term memory mechanism. The updating process of the long-term memory module includes the following steps:

[0203] Step 40231: During each long-term memory update, the agent traverses the current factual memory. Determine the number of times each learning record is reinforced;

[0204] Learning Records Count Exceeding the preset enhancement threshold This indicates that the memory has been reinforced. Next, it is transferred to the long-term memory module.

[0205] Step 40232: Utilize the large language model within the intelligent agent to summarize short-term and long-term memories through memory reflection, generate new summary information, and replace the old summary information with the new summary information generated in each step;

[0206] It is possible to summarize the practice process and new insights gained by the agent in linguistic form, ignoring specific practice details to filter out noise, irrelevant content, or potentially misleading information.

[0207] Optionally, long-term memory modules can be stored using knowledge points as the basic unit, with each long-term memory entry represented as follows:

[0208] (19);

[0209] in, This represents the k-th knowledge point; This indicates the probability of a student's long-term mastery of this knowledge point; This indicates the time when the knowledge point was last reinforced or updated.

[0210] Step 40233: Calculate the forgetting score of each learning record. When the forgetting score exceeds the forgetting threshold, set the number of reinforcements to a set value, store the learning records whose number of reinforcements exceeds the set reinforcement threshold and the new summary information in the long-term memory module, and update the long-term memory module.

[0211] Long-term memory stores repeatedly reinforced learning information and high-level summaries, including behavioral data that has been repeatedly accessed or reinforced over multiple learning cycles, representing a student's mastery of knowledge points. Each learning record in long-term memory may be forgotten according to the human forgetting curve theory, meaning that memory decays rapidly in the initial stages and then gradually slows down. The forgetting function, defined over time and observation steps, calculates the forgetting score as follows:

[0212] (20);

[0213] Among them, when When the record is forgotten, the number of times it is reinforced in factual memory... Reset to 1, The threshold for forgetting is represented by i, which represents the number of learning interactions at the current time point; and n represents the set critical time point, which is the critical time point from when the memory has never been forgotten to when it is forgotten.

[0214] Step 4 involves generating practice question recommendations tailored to the student's current learning status based on the updated digital learning companion agent, including the following steps:

[0215] Step 421: Based on students' knowledge mastery probability, forgetting score, and learning emotional state, identify the knowledge points to be consolidated, improved, or corrected, and use them as target knowledge points.

[0216] Step 422: Based on the prompt word template, input the target knowledge point, the probability of mastering the knowledge point, the time since the last practice of the target knowledge point, and the forgetting score into the large language model to generate personalized practice question recommendation results.

[0217] This embodiment uses a large language model to select candidate questions from a pre-set question bank that correspond to the target knowledge points and match the student's current ability level;

[0218] Furthermore, it also includes a challenge-based interaction mechanism with the digital learning companion agent, comprising the following steps:

[0219] Step 431: Receive challenge requests from students to the digital learning companion agent;

[0220] Step 432: Based on the recommended practice questions, the digital learning companion agent outputs practice questions that are suitable for the student's current learning status.

[0221] Step 433: Obtain students' answers to the practice questions and generate learning behavior data, including records of incorrect answers, based on the answers to proceed to the next round of memory update and practice question recommendation process for the digital learning companion agent.

[0222] Furthermore, a digital tutor intelligent agent is constructed, including a problem graph construction module, a knowledge graph construction module, and a personalized learning path generation module, to implement the steps in step 5;

[0223] Step 5: The personalized learning path includes the target knowledge points to be learned and the corresponding set of questions for those knowledge points. Based on the acquired learning task information and the analysis results of the learning situation, the method for generating a personalized learning path for the student through the constructed digital tutor intelligent agent includes the following steps:

[0224] Step 51: The digital tutor agent constructs a problem graph based on the domain problem library, which is configured to be implemented in the problem graph construction module.

[0225] Step 511: Construct a question set based on a pre-defined domain knowledge question base, and encode the question text based on a large language model;

[0226] The digital tutor agent constructs a set of questions based on a pre-defined domain knowledge question base, denoted as:

[0227] (twenty one);

[0228] Each of the questions Each question includes the question text and its corresponding knowledge point annotations, represented as follows:

[0229] (twenty two);

[0230] in, Indicate the problem The title text content, This represents the set of knowledge points associated with the question.

[0231] To obtain the semantic representation of the question, a pre-trained large language model is used to analyze the question text content. Encoding is performed, and the formula is expressed as:

[0232] (twenty three);

[0233] in, This represents a vector encoded from the text content of the question. Indicates a prompt word template;

[0234] Step 512: Obtain historical data of students' answers to the questions, calculate the accuracy rate to determine the level of difficulty and divide the questions into levels, and obtain a hierarchical question map;

[0235] To achieve tiered question difficulty, historical student response data is introduced. Let the student set be... ,student The answer sequence is The students' answers were For each question Statistical average accuracy :

[0236] (twenty four);

[0237] Based on the accuracy rate of questions as an indicator of difficulty, the problem graph can be divided into three layers:

[0238] (25);

[0239] in, and For a preset threshold, and ; , and These represent low-order problems, intermediate-order problems, and high-order problems, respectively.

[0240] The final hierarchical problem map is as follows: .

[0241] Based on the obtained problem graph, when generating subsequent learning paths, the digital tutor agent can select problems from different difficulty levels that match the student's current ability.

[0242] Step 52: The digital tutor intelligent agent constructs a knowledge graph based on domain knowledge. This process is configured and implemented in the knowledge graph construction module.

[0243] Based on a domain knowledge base, let the set of knowledge point nodes be . The structural dependencies between knowledge points are represented as Constructing a knowledge graph:

[0244] (26);

[0245] By constructing knowledge graphs Ensure the collection of knowledge points A comprehensive coverage of domain knowledge points provides structural constraints for subsequent learning path planning.

[0246] Step 53: Based on the knowledge graph, student learning status analysis results, and current learning task information, perform structural enhancement representation and learning state matching on candidate knowledge points to determine the next learning objective and generate a personalized learning path containing the target knowledge points and a set of questions extracted from the question graph. This is configured to be implemented in the personalized learning path generation module and includes the following steps:

[0247] Step 531: Based on the knowledge graph, obtain the predecessor and successor knowledge points of the target knowledge point in the current learning task, enhance the target knowledge point, and obtain the enhanced text description sequence of the target knowledge point;

[0248] In a knowledge graph, based on the target knowledge points in the current learning task Extract its predecessor node set and the set of successor nodes :

[0249] (27);

[0250] (28);

[0251] in, This indicates the preceding knowledge points; This refers to rear-wheel drive knowledge points. It represents the set of directed edges between knowledge point nodes in a knowledge graph, used to represent the dependencies, prerequisites, or content connections between knowledge points.

[0252] Design a prompt word template and define the current knowledge point label, the predecessor node list label, and the successor node list label, which are respectively denoted as: [knowledge point], [predecessor node list], and [successor node list].

[0253] An example prompt template is as follows: Please explain the current knowledge point [knowledge point]. The predecessor knowledge points are known to be [predecessor node list], and the successor knowledge points are known to be [successor node list]. Based on the above context, please generate an explanatory text that includes its definition and logical relationship with the preceding and following knowledge points. ;

[0254] The generated explanatory text is concatenated with the target knowledge point name, predecessor node name, and successor node name to form an enhanced text description sequence for the target knowledge point. , denoted as:

[0255] (29);

[0256] Step 532: Encode and perform multi-layer neighborhood aggregation on the enhanced text description sequence of the target knowledge points to obtain the knowledge point representation with enhanced relationships;

[0257] Specifically, a special marker is added to the beginning of the sequence, which is then input into a pre-trained encoder for encoding. The formula is as follows:

[0258] (30);

[0259] The initial semantic vector As the initial features of nodes in a knowledge graph, the input graph neural network performs structure-aware representation learning, and obtains relation-enhanced knowledge point representations through multi-layer neighborhood aggregation. :

[0260] (31);

[0261] in, Representing knowledge points The set of neighboring nodes is used to aggregate the knowledge point representation by injecting the prerequisite and dependency relationships between knowledge points into the knowledge point representation, thereby obtaining a relation-enhanced knowledge point representation.

[0262] It is represented as a nonlinear transformation function (usually implemented by an MLP); This is represented as an aggregation function for neighboring nodes; This represents the fusion operation between the current representation and the aggregation results of its neighbors;

[0263] Step 533: Based on the analysis results of students' learning progress and current knowledge points, dynamically model the students' knowledge status to obtain the students' knowledge status at each time step. Knowledge state representation, and knowledge point representation that enhances the relationship with the current knowledge point. The knowledge point embedding vector and the answer result embedding vector are concatenated to form a unified representation of learning behavior;

[0264] Specifically, for students' historical learning sequence, a time series consisting of knowledge points and answers is constructed, and a sequence model is used to dynamically model students' knowledge status, obtaining the students' progress at each time step. Knowledge state representation:

[0265] (32);

[0266] Knowledge point embedding vectors are obtained through mapping via an embedding layer. Specifically, a knowledge point embedding matrix is ​​constructed. ,in, This represents the total number of knowledge points. This represents the embedding dimension for the current knowledge point. The embedding vector is obtained by looking up a table.

[0267] The answer embedding vector is obtained from the answer embedding matrix. Construct the answer embedding matrix. This is used to indicate whether the answer is correct or incorrect.

[0268] At time step Enhance the representation of relationships between current knowledge points. Student knowledge status representation Knowledge point embedding vector and the embedding vector of the answer results The data is then concatenated to form a unified representation of the learning behavior, using the following formula:

[0269] (33);

[0270] Step 534: Input the sequence of learning behaviors, composed of learning behavior representations, into a sequence modeling network based on a self-attention mechanism to model historical learning behaviors and obtain the context-aware feature representation of the current moment, expressed by the formula:

[0271] (34);

[0272] in, ( ) indicates feedforward neural network processing; This indicates the operation of the self-attention mechanism;

[0273] Step 535: Based on context-aware feature representation, perform matching calculations on candidate knowledge points, calculate the recommendation probability of each knowledge point as the next learning target, select the top-ranked knowledge points as the recommendation results, and filter the set of questions corresponding to the knowledge points in the hierarchical question graph. Combine learning resources to generate a personalized learning path that matches the student's current knowledge status and ability level.

[0274] Specifically, based on the recommendation probability of the next learning objective, the knowledge points are ranked, and the top-ranked knowledge points are selected as the recommendation results.

[0275] Furthermore, this embodiment also includes a teaching interaction framework for teachers, students, machines, and environments, including student-student interaction mode, teacher-student interaction mode, student-machine interaction mode, and machine-environment interaction mode;

[0276] The peer-to-peer interaction model is used to enable interaction among students, including self-directed learning, collaborative practice, and communication. Through peer interaction, students share knowledge and transfer experience in the process of communicating and collaborating with their peers, thereby promoting the understanding and internalization of learning content.

[0277] In this implementation, a student collaboration participation index can be constructed based on the collaborative behavior data generated by students during peer-to-peer interactions. It can be represented as:

[0278] (35);

[0279] in, Students In time period The number of times of active participation within the organization. Indicates the number of times they engage in collaborative communication. These are the weighting coefficients, This is for the statistical time window length.

[0280] The teacher-student interaction mode is used to realize the interaction between teachers and students, including: learning assistance, guidance, value shaping and emotional support. Teachers guide and intervene in the learning process based on the student learning data generated in the learning environment.

[0281] The interactive mode is used to enable interaction between students and digital learning companion agents and digital tutor agents in the learning environment;

[0282] In the interaction between students and the digital learning companion intelligent agent, students complete questions, and process data of students' answering behavior is collected to comprehensively evaluate students' ability level, mastery of knowledge points and learning emotional state, build student learning profiles, and update the intelligent agent's memory module to simulate students' ability level; students compete in problem-solving ability by challenging the digital learning companion intelligent agent, and generate wrong questions based on the answer results, and recommend related questions to students based on this.

[0283] In the interaction between students and the digital tutor agent, when students encounter difficulties in completing the current learning task, they seek guidance from the digital tutor agent. The digital tutor agent extracts relevant knowledge points and questions based on the student's current learning task and recommends personalized question sets, knowledge points and learning resources to the student.

[0284] The machine-environment interaction mode is used to realize the interaction process between the intelligent agent and the learning environment. By deploying sensing devices in the learning environment, facial expression features and head posture information of students during the learning process are collected. The student's head-raising behavior, gaze direction and facial expression changes are identified in real time to represent the student's current learning focus and level of engagement.

[0285] The facial expression features include at least states such as focused attention, confusion, distraction, and fatigue; the head posture features include at least behavioral features such as head tilt angle, duration of head tilt, and frequency of head tilt.

[0286] By jointly modeling the above features, a student learning focus index is generated. , can be represented as:

[0287] ;

[0288] in, This represents the attention sub-index calculated based on facial expression features. This represents the attention sub-index calculated based on head posture features. These are the weighting coefficients.

[0289] This embodiment utilizes the teacher-student-machine-environment interaction framework to enhance students' self-directed learning and collaborative practice abilities through peer-to-peer interaction; through continuous challenges and feedback with the intelligent agent in the student-machine interaction, students continuously improve their abilities, stimulate their interest in proactive exploratory learning, and enhance the immersion and interactivity of the learning process; teachers provide learning guidance and emotional support to students in the teacher-student interaction; the intelligent agent simulates students' learning abilities and generates personalized learning paths in the student-machine interaction; and the intelligent agent identifies and assesses students' learning focus during the machine-environment interaction, providing corresponding reminders or intervention prompts when a decline in focus is detected to guide them back to a learning state. This constructs a new four-element collaborative model of teacher-student-machine-environment, with students as the main body, teachers as guides, and intelligent agents as technical support, enabling an adaptive, personalized learning process of learning, thinking, and practicing.

[0290] Furthermore, in order to provide data to digital learning companion agents and digital tutor agents to support teaching Q&A, learning assessment, practice question recommendation, personalized learning path generation, and learning process intervention, it also includes the process of building teacher corpora and student corpora.

[0291] The teacher corpus provides the intelligent agent with course knowledge content, teaching resources, and logical relationships supporting the teaching process. It is created by collecting, cleaning, annotating, encoding, and structurally storing multi-source teaching corpora, including textbook documents, courseware files, teaching videos, project materials, and teaching platform data, forming a knowledge resource base for course instruction.

[0292] The teacher corpus serves two purposes: firstly, it supports the intelligent agent in content generation and reasoning in scenarios such as teaching question and answer, teaching resource recommendation, learning assessment, and project research support; secondly, by connecting with the knowledge graph constructed by the digital tutor intelligent agent, it provides a data foundation for extracting the predecessor, successor, dependency, and hierarchical relationships between knowledge points, thereby providing teaching content and structural logic basis for subsequent exercise recommendation, knowledge point recommendation, and personalized learning path generation.

[0293] The student corpus is used to record and represent students' learning behaviors, responses, interactions, and reflections during the learning and research process, thereby forming a continuous characterization of the students' learning process.

[0294] The student corpus collects, classifies, annotates, and structures learning process corpus, research practice corpus, reflection and achievement corpus, and interactive behavior corpus to establish the relationship between learning content, learning behavior, and knowledge points, forming a learning data foundation for modeling students' learning status.

[0295] Furthermore, the method for constructing a teacher corpus includes the following steps:

[0296] Step 61: Determine the construction goals and application scenarios: Based on the course teaching objectives and teaching organization forms, clarify the construction goals of the teacher corpus and determine the application scenarios of the intelligent agent, including but not limited to intelligent Q&A, teaching resource recommendation, learning process evaluation and project research support;

[0297] Step 62, Corpus Classification and Multi-Source Data Extraction: According to the preset corpus types, the teaching-related data are classified and collected. The corpus types include at least professional knowledge corpus, practical teaching corpus, teaching management corpus, and research project corpus. The sources of the corpus include textbook documents, courseware files, teaching videos, project materials, and teaching platform data.

[0298] Step 63, Corpus Cleaning and Knowledge Labeling: The collected corpus is processed for noise reduction, deduplication, and format standardization. Knowledge points, skill elements, and teaching objectives in the corpus are labeled to form structured or semi-structured corpus data.

[0299] Step 64, Semantic encoding and vectorization of the corpus: Semantic encoding is performed on the cleaned and labeled corpus to convert text, image or video data into vector representations to support subsequent semantic retrieval and reasoning calculations;

[0300] Step 65, Structured Storage and Index Construction: Store the vectorized corpus in a vector database or graph database, and construct the corresponding metadata index to achieve efficient retrieval and access of the corpus;

[0301] Step 66: Integration with Knowledge Graph and Logic Extraction: Integrate the teacher corpus with the knowledge graph to extract the core logical structure and knowledge relationships of the course, which will be used to support the agent's reasoning and teaching decisions.

[0302] Step 67, Dynamic Updates and Continuous Optimization: Based on teaching feedback, course content updates, and the operational performance of the digital learning companion and digital tutor intelligent agents, the teacher corpus is continuously supplemented, corrected, and optimized;

[0303] Furthermore, the method for constructing the student corpus includes the following steps:

[0304] Step 71: Modeling Objectives: Based on students' learning needs, determine the application objectives of the student corpus, including learning process recording, research practice archiving, and learner feature modeling.

[0305] Step 72: Learning data collection and classification: Collect and organize the learning behavior data generated by students during the learning and research process. The data should include at least the learning process data, research practice data, reflection and results data, and interaction behavior data, and be classified and stored according to the theme or time dimension.

[0306] Step 73, Preliminary annotation and tag generation: Add tags to the student corpus that reflect the learning status and learning content, including but not limited to tags for difficulties, key concepts, and unresolved problems.

[0307] Step 74: Simple structuring and association modeling: Use tables or knowledge graph tools to structure the student corpus and establish the association between learning content, learning behavior and knowledge points.

[0308] Step 75: Integrate the corpus into the intelligent agent: Integrate the student corpus into the intelligent agent, and verify the effectiveness and support capabilities of the corpus through question-and-answer interaction, learning recommendation, and feedback analysis.

[0309] Step 76, Feedback-driven iterative optimization: Based on the agent analysis results and teacher evaluations, continuously supplement and adjust the student corpus to improve the accuracy of the student corpus in depicting learning behavior.

[0310] In the above implementation, the teacher corpus and the student corpus form a closed loop of knowledge supply and behavioral feedback, jointly supporting the realization of the multi-agent system: the teacher corpus provides a standardized knowledge base, while the student corpus provides personalized behavioral data. Combined with item response theory and affective reasoning, this enables dynamic adaptation of practice question recommendations and learning paths. Through the combination of structured corpora and knowledge graphs, the digital tutor agent can generate an integrated learning path of knowledge points, questions, and resources based on the student's real-time status, enhancing the continuity and adaptability of the accompanying learning process.

[0311] Example 2

[0312] Based on Embodiment 1, this embodiment provides a multi-agent-based accompanying learning system, including:

[0313] The learning situation analysis module is configured to acquire learning behavior data, extract students' problem-solving reasoning chains for questions, compare them with standard knowledge reasoning paths, and obtain learning situation analysis results.

[0314] The probability recognition module is configured to calculate the probability of a student's mastery of a knowledge point based on the knowledge points corresponding to the target question and the analysis results of the student's learning situation, using item response theory.

[0315] The emotion state inference module is configured to infer the student's learning emotion state based on the dialogue content of learning interactions in the learning behavior data;

[0316] The emotional state module is configured to update the memory module of the constructed digital learning companion agent based on the learning situation analysis results, the mastery probability of knowledge points, and the emotional state of learning. Based on the updated digital learning companion agent, it generates practice question recommendations that are adapted to the student's current learning state.

[0317] The personalized learning path generation module is configured to generate personalized learning paths for students based on the results of learning situation analysis and the obtained learning task information, through the constructed digital tutor intelligent agent.

[0318] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0319] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0320] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-agent based companion learning method, characterized by, Includes the following steps: Acquire learning behavior data, extract students' problem-solving reasoning chains for questions, compare them with standard knowledge reasoning paths, and obtain learning situation analysis results; Based on the analysis results of the knowledge points corresponding to the target questions and the learning situation, and based on the item response theory, the probability of students mastering the knowledge points is calculated. Based on the dialogue content of learning interactions in learning behavior data, infer students' learning emotional state; Based on the analysis results of learning progress, the probability of mastering knowledge points, and the emotional state of learning, the memory module of the constructed digital learning companion intelligent agent is updated, and practice question recommendations that are adapted to the student's current learning status are generated based on the updated digital learning companion intelligent agent. Construct a digital learning companion intelligent agent, including student profiles and multi-layer memory modules; Based on the results of learning situation analysis, the probability of mastering knowledge points, and data on learning emotional state, a multi-dimensional feature set is formed by comprehensively representing the current learning characteristics of students, which serves as a student profile. The memory modules include factual memory modules, short-term memory modules, and long-term memory modules; The fact memory module is used to store student learning records obtained from learning behavior data analysis in real time; The short-term memory module is used to store the student's learning records obtained in the most recent short period of time; The long-term memory module is used to store students' learning records that form long-term memories, in order to simulate the stable ability structure and knowledge mastery pattern that students gradually form during the long-term learning process. Based on the learning progress analysis results, the probability of mastering knowledge points, and the emotional state of learning, the memory module of the constructed digital learning companion agent is updated, including the following steps: Based on the learning situation analysis results, the mastery probability of knowledge points, and the learning emotional state data, the static profile features are updated. A large language model is used to map the updated static profile features to the dynamic answer behavior results of students in the target question scenario, which serves as a new student learning record. Based on the obtained student learning records, update the memory module of the digital learning companion agent: Based on the obtained student learning records, update the fact memory module and reinforce the counting of similar historical records; Based on the updated fact memory module, new student learning records are stored in the short-term memory module to preserve students' continuous learning behavior within a recently set time period. Based on the reinforcement counting results and forgetting rules, the long-term memory module is updated to form a stable representation of students' knowledge mastery. Based on the learning situation analysis results and the obtained learning task information, a personalized learning path for each student is generated through the constructed digital tutor intelligent agent.

2. The multi-agent-based accompanying learning method as described in claim 1, characterized in that, The method for obtaining student learning behavior data, extracting students' problem-solving reasoning chains for questions, comparing them with standard knowledge reasoning paths, and obtaining the results of student learning analysis includes the following steps: Extract students' answers to the target questions and the intermediate calculation steps from the learning behavior data; The answer results and intermediate calculation steps are encoded to generate intermediate calculation step vectors. Through embedding operations and self-attention mechanisms, the dependencies and importance weights of each problem-solving step are obtained. Based on the intermediate computation step vector, the dependencies between each problem-solving step, and the importance weight of each step, a problem-solving representation containing the step sequence relationship and key reasoning paths is constructed. By using a subject-based educational language model to represent problem-solving, semantic analysis and causal inference are performed, mapping the student's problem-solving process into a structured problem-solving reasoning chain; The rationality, consistency and completeness of each reasoning node in the structured problem-solving reasoning chain are verified and compared with the standard knowledge reasoning path to locate reasoning deviation nodes and determine the reasons for answering errors based on the reasoning deviation nodes.

3. The multi-agent-based companion learning method as described in claim 1, characterized in that, Based on the analysis of the knowledge points corresponding to the target questions and the students' learning progress, and based on item response theory, the method for calculating the probability of students mastering the knowledge points includes the following steps: Based on learning behavior data, students' answers to target questions are extracted, and combined with the discrimination and difficulty parameters of the target questions, the parameters of students' answers are estimated based on the item response theory model to obtain student ability factors. The questions are mapped to the corresponding knowledge points in the knowledge graph, and the probability of students mastering each knowledge point is calculated based on the student's ability factor.

4. The multi-agent-based accompanying learning method as described in claim 3, characterized in that, Methods for calculating the probability of students mastering each knowledge point include: The intermediate calculation steps in student ability factors, questions, and learning behavior data are vectorized and concatenated to form a student behavior feature vector. The student behavior feature vector is weighted and nonlinearly transformed to obtain the feature weight of each knowledge point. The feature weight is then mapped to the probability of the student mastering each knowledge point through a normalization function.

5. The multi-agent-based companion learning method as described in claim 1, characterized in that, Based on the updated digital learning companion agent, practice question recommendations are generated that are adapted to the student's current learning status, including the following steps: Based on students’ knowledge point mastery probability, forgetting score and learning emotional state, identify knowledge points that need to be consolidated, improved or corrected, and use them as target knowledge points. Based on the prompt word template, the target knowledge point, the probability of mastering the knowledge point, the time since the last practice of the knowledge point, and the forgetting score are input into the large language model to generate personalized practice question recommendations.

6. The multi-agent-based accompanying learning method as described in claim 1, characterized in that, Based on the acquired learning task information and learning situation analysis results, a method for generating personalized learning paths for students through a constructed digital tutor intelligent agent includes the following steps: The digital mentor intelligent agent constructs a problem graph based on a domain problem database; The digital tutor intelligent agent constructs a knowledge graph based on domain knowledge; Based on knowledge graphs, student learning analysis results, and current learning task information, candidate knowledge points are structurally enhanced and matched with learning states to determine the next learning objective and generate a personalized learning path that includes the target knowledge points and a set of questions extracted from the problem graph.

7. The multi-agent-based accompanying learning method as described in claim 6, characterized in that, Based on knowledge graphs, student learning analysis results, and current learning task information, this method performs structural enhancement representation and learning state matching on candidate knowledge points to determine the next learning objective and generate a personalized learning path containing the target knowledge points and a set of questions extracted from the problem graph. The method includes the following steps: Based on the knowledge graph, the predecessor and successor knowledge points of the target knowledge point in the current learning task are obtained, and the target knowledge point is enhanced to obtain the enhanced text description sequence of the target knowledge point. Encode and aggregate multi-layer neighborhoods of the enhanced text description sequence of the target knowledge points to obtain a knowledge point representation with enhanced relationships. Based on the analysis of students' learning progress and their current knowledge points, a dynamic model is constructed to determine students' knowledge status at each time step. The knowledge state representation, and the enhanced representation of its relationship with the current knowledge point. The knowledge point embedding vector and the answer result embedding vector are concatenated to form a unified representation of learning behavior; The sequence of learning behaviors, represented by the learning behaviors, is input into a sequence modeling network based on a self-attention mechanism to model historical learning behaviors and obtain the context-aware feature representation of the current moment. Based on context-aware feature representation, candidate knowledge points are matched and calculated. The probability of each knowledge point being recommended as the next learning objective is calculated. The top-ranked knowledge points are selected as the recommendation results. The set of questions corresponding to the knowledge points is then selected in the hierarchical question graph. Personalized learning paths that match the student's current knowledge status and ability level are generated by combining learning resources.

8. A multi-agent-based accompanying learning system, employing the multi-agent-based accompanying learning method as described in any one of claims 1-7, characterized in that, include: The learning situation analysis module is configured to acquire learning behavior data, extract students' problem-solving reasoning chains for questions, compare them with standard knowledge reasoning paths, and obtain learning situation analysis results. The probability recognition module is configured to calculate the probability of a student's mastery of a knowledge point based on the knowledge points corresponding to the target question and the analysis results of the student's learning situation, using item response theory. The emotion state inference module is configured to infer the student's learning emotion state based on the dialogue content of learning interactions in the learning behavior data; The emotional state module is configured to update the memory module of the constructed digital learning companion agent based on the learning situation analysis results, the mastery probability of knowledge points, and the emotional state of learning. Based on the updated digital learning companion agent, it generates practice question recommendations that are adapted to the student's current learning state. The personalized learning path generation module is configured to generate personalized learning paths for students based on the results of learning situation analysis and the obtained learning task information, through the constructed digital tutor intelligent agent.

Images