Student portrait construction method and system based on large model agent

Abstract

The invention provides a student portrait construction method and system based on a large model agent. The method comprises the steps of collecting learning behavior data of students; generating an expert cognitive path map of each subject based on the large model, and constructing a standard path database containing typical tasks; performing semantic alignment on the student path and a standard path, and calculating a cognitive residual feature vector including step redundancy, a path deviation angle and concept confusion frequency; calculating an attention weight matrix by using a bidirectional Transform model, positioning a cognitive interruption point in combination with a residual feature, and generating a sticking point interpretation map containing knowledge fault positioning and repair suggestions; generating a dynamic cognitive map and a strategy evolution tree; and generating a student cognitive portrait containing a sticking point interpretation map, a cognitive map and a strategy evolution tree. According to the method, the student cognition portrait containing the sticking point interpretation map, the cognition map and the strategy evolution tree is generated, the student cognition features, learning obstacles and strategy preferences are comprehensively described, and a visual decision basis is provided for teaching intervention.

Description

Technical Field

[0001] The present invention relates to the technical field of student portrait construction, and particularly relates to a method and system for constructing a student portrait based on a large model agent. Background Art

[0002] In the field of education, the construction of student portraits is an important research direction. By analyzing the learning behavior data of students to construct student portraits, it is possible to facilitate the understanding of students' learning status and provide support for teaching.

[0003] Traditional methods for constructing student portraits often have the following deficiencies: mainly relying on structured data such as questionnaires and exam scores, lacking multi-dimensional analysis of students' learning behaviors, and being unable to comprehensively reflect students' learning status; due to the lack of in-depth understanding of students' cognitive status, being unable to provide personalized learning support for each student, resulting in poor teaching effects; the constructed student portraits are often static and unable to reflect changes in students' cognitive status, resulting in educators being unable to timely understand students' learning progress and adjust teaching plans; often ignoring changes in students' learning strategies during the learning process and being unable to provide educators with in-depth insights into students' learning strategies. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to propose a method and system for constructing a student portrait based on a large model agent to solve the above-mentioned problems.

[0005] According to a method for constructing a student portrait based on a large model agent proposed by the present invention, the method includes:

[0006] Collecting students' learning behavior data;

[0007] Generating an expert cognitive path map for each subject based on a large model, and extracting a standard problem-solving path from the large model parameters to construct a standard path database containing typical tasks;

[0008] Semantically aligning the student path with the standard path to obtain an aligned path pair, and calculating cognitive residual feature vectors, including a step redundancy vector, a path deviation angle matrix, and a concept confusion frequency vector;

[0009] The attention weight matrix is calculated using a bidirectional Transformer model, and combined with the residual features to locate cognitive breakpoints, generating a card point explanation map containing knowledge tomography location and repair suggestions;

[0010] [[ID=ID=33]] Generating a dynamic cognitive map and constructing a strategy evolution tree based on the residual features;

[0011] Generating a student cognitive portrait containing a card point explanation map, a cognitive map, and a strategy evolution tree.

[0012] Further, the steps of generating the expert cognitive path map for each subject based on the large model and extracting the standard problem-solving path from the large model parameters include:

[0013] Input the subject textbooks and classic question banks into the pre-trained large model to generate the cognitive paths for sub-knowledge points and sub-question types, and obtain the expert path map;

[0014] Verify and correct the path by domain experts to conform to the actual teaching situation, and obtain the verified path map;

[0015] Extract the problem-solving step sequence of typical tasks from the large model parameters, construct a standard path database, and each standard path includes a step ID, semantic label, difficulty coefficient, and associated knowledge points.

[0016] Further, the steps of calculating the cognitive residual feature vector include:

[0017] Calculate the step redundancy, generate a step redundancy vector, and the formula is: step redundancy = number of invalid steps / total number of steps, where the number of invalid steps is the number of student path steps with a semantic similarity lower than the similarity threshold to the standard path, and the total number of steps is the total number of all steps in the student path;

[0018] Calculate the path deviation angle, generate a path deviation angle matrix, and the formula is:

[0019]

[0020] where, v s is the student path vector, v s，i represents the component of v s in the i-th dimension, v t is the standard path vector, v t，i represents the component of v t in the i-th dimension;

[0021] Calculate the ratio of the number of times of misusing concepts to the total number of steps to obtain the concept confusion frequency, and generate a concept confusion frequency vector.

[0022] Further, the steps of generating the blockage point explanation map including knowledge fault location and repair suggestions include:

[0023] Input the step redundancy vector, path deviation angle matrix, concept confusion frequency vector and the aligned path pair into a bidirectional Transformer model to calculate the weight matrix;

[0024] Based on the weight matrix, screen the differential steps by presetting a weight threshold, and generate a key point explanation graph by combining the step redundancy vector and the concept confusion frequency vector. Mark the knowledge fault nodes in the key point explanation graph and associate micro-lesson videos and example questions recommendations.

[0025] Further, the steps for calculating the weight matrix include:

[0026] For each question, integrate the corresponding step redundancy vector, the step-by-step features of the path deviation angle matrix, the concept confusion frequency vector, and the embedded representation of the aligned path pair into an input sequence;

[0027] Input the input sequence into a bidirectional Transformer model, where the Transformer model includes multiple stacked encoder layers, and each layer consists of a multi-head self-attention mechanism and a feed-forward neural network;

[0028] For the vector representation of each step in the input sequence, generate a query vector, a key vector, and a value vector respectively through linear transformation;

[0029] For each query vector, calculate the dot product of it and all key vectors, and normalize it through a scaling and softmax function to obtain the attention weights of the query vector and each step in the input sequence;

[0030] Concatenate the weight matrices output by multiple attention heads, and aggregate them through linear transformation into a final weight matrix, which is used to reflect the semantic similarity and importance between the student path and the standard path at each step.

[0031] Further, the steps for generating a key point explanation graph by combining the step redundancy vector and the concept confusion frequency vector include:

[0032] Traverse the weight matrix, take the steps of the student path with weights lower than the weight threshold as differential steps, and record the indexes, the corresponding questions, and the corresponding weight values of the differential steps;

[0033] Combine the step redundancy vector and the concept confusion frequency vector to identify the frequently redundant differential steps as key points;

[0034] Perform semantic parsing on the key points, mark their associated knowledge points, and recommend micro-lesson videos and example questions;

[0035] Calculate the priority of the repair suggestions, and the formula is: repair priority = fault frequency × concept difficulty coefficient × step redundancy, where the fault frequency is the value of the corresponding knowledge point in the concept confusion frequency vector;

[0036] Taking the stuck point as a node, associating knowledge points, micro-lesson videos, and example recommendations as edges, construct a stuck point explanation graph. The attribute information of each node includes step redundancy, concept confusion frequency, weight value, and repair priority.

[0037] Furthermore, the steps of identifying the different steps with high-frequency redundancy as the stuck point by combining the step redundancy vector and the concept confusion frequency vector include:

[0038] Calculate the comprehensive score of each different step, and the formula is:

[0039] Comprehensive score = α · step redundancy + β · concept confusion frequency + γ · (1 - average weight), where α, β, and γ are weight coefficients;

[0040] Define the different steps with the comprehensive score exceeding the score threshold as the stuck point.

[0041] Furthermore, the steps of generating the dynamic cognitive map include:

[0042] Taking the subject knowledge points as nodes, calculate the mastery degree through the concept confusion frequency vector;

[0043] According to the step redundancy vector, the path deviation angle matrix, and the mastery degree, mark the cognitive fault area. The color depth of the mark = mastery degree × step redundancy × path deviation angle weight, and the path deviation angle weight is a coefficient dynamically adjusted according to the size of the path deviation angle;

[0044] Based on the repair suggestions of the stuck point explanation graph, generate a repair path from the fault node to the mastery node;

[0045] Dynamically update the cognitive map at preset time intervals to reflect the recent changes in the students' cognitive state.

[0046] Furthermore, the steps of constructing the strategy evolution tree include:

[0047] Based on the residual features and learning behavior data, discover the subtypes of learning strategies through K-Means clustering;

[0048] Taking time as the horizontal axis, construct a strategy evolution tree, with the nodes being the subtypes of strategies and the edges being the strategy switching events;

[0049] Mark the factors affecting strategy switching;

[0050] Calculate the strategy stability index, and the formula is:

[0051] Stability = strategy duration × stuck point repair success rate / total learning time.

[0052] The present invention also provides a student portrait construction system based on a large model agent, which is used to implement the above-mentioned student portrait construction method based on a large model agent. The system includes:

[0053] A data collection module: used to collect the learning behavior data of students;

[0054] A standard path module: used to generate an expert cognitive path map for each subject based on a large model, extract a standard problem-solving path from the large model parameters, and construct a standard path database containing typical tasks;

[0055] A residual feature module: used to semantically align the student path with the standard path to obtain an aligned path pair, and calculate cognitive residual feature vectors, including a step redundancy vector, a path deviation angle matrix, and a concept confusion frequency vector;

[0056] A bottleneck map module: used to calculate an attention weight matrix using a bidirectional Transformer model, and combine the residual features to locate cognitive breakpoints, and generate a bottleneck explanation map containing knowledge fault location and repair suggestions;

[0057] A cognitive map module: used to generate a dynamic cognitive map and construct a strategy evolution tree based on the residual features;

[0058] A cognitive portrait module: used to generate a student cognitive portrait containing a bottleneck explanation map, a cognitive map, and a strategy evolution tree.

[0059] In summary, the student portrait construction method based on a large model agent of the present invention collects the real learning behavior data of students; generates expert problem-solving paths for sub-knowledge points and sub-question types based on a large model, and constructs a standard path database containing typical tasks through verification and correction by domain experts to provide a benchmark for comparing student paths and ensure the objectivity of residual feature calculation and bottleneck location; aligns the student path with the standard path, and calculates multi-dimensional residual feature vectors, including a step redundancy vector, a path deviation angle matrix, and a concept confusion frequency vector, to quantify the student's cognitive deviation and provide multi-dimensional feature support for breakpoint location; uses a bidirectional Transformer model to calculate an attention weight matrix, and combines the residual features to locate cognitive breakpoints, and generates a bottleneck explanation map containing knowledge fault location and repair suggestions to realize problem-resource association and support precise intervention; reflects the change of the student's cognitive state in real time by generating a dynamic cognitive map, supports teachers to dynamically adjust teaching strategies, and reveals the evolution law of learning strategies by constructing a strategy evolution tree, which is convenient for optimizing the learning path recommendation; integrates the bottleneck explanation map, the cognitive map, and the strategy evolution tree into a student cognitive portrait, forming a closed loop from diagnosis to repair and then to optimization, comprehensively depicting the student's cognitive characteristics, learning obstacles, and strategy preferences, and providing a visual decision-making basis for teaching intervention.

[0060] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the embodiments of the present invention. Description of the Drawings

[0061] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the description of the embodiments in conjunction with the following drawings, where:

[0062] Figure 1 is a flowchart of a method for constructing a student portrait based on a large model agent according to Embodiment 1 of the present invention;

[0063] Figure 2 is a system block diagram of a system for constructing a student portrait based on a large model agent according to Embodiment 2 of the present invention. Detailed Embodiments

[0064] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0065] It should be noted that when an element is referred to as being "fixedly provided on" another element, it can be directly on the other element or there may also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0067] Embodiment 1

[0068] Please refer to Figure 1 , the present invention proposes a method for constructing a student portrait based on a large model agent, and the method includes steps S100 to S600:

[0069] S100, collect the learning behavior data of students.

[0070] Specifically, collecting students' learning behavior data can include data such as the digital homework process and online learning behavior records. Among them, collecting digital homework process data is to collect data on various types of homework, exercises, and tests completed by students on the digital homework platform. These data include, but are not limited to, answering time, answering order, answer correctness, problem-solving ideas (such as problem-solving steps, formulas or methods used, etc.), modification records, etc. Collecting online learning behavior data is to monitor and record students' browsing records, video viewing progress, online teaching interaction situations, resource download times, etc. on the online learning platform. These data reflect students' learning interests, learning habits, and learning progress.

[0071] S200. Generate an expert cognitive path map for each subject based on the large model, and extract the standard problem-solving path from the large model parameters to construct a standard path database containing typical tasks.

[0072] Generating an expert cognitive path map that reflects the subject knowledge structure through the large model, extracting the standard problem-solving path from it, and constructing a standard path database for comparing students' paths can provide a standardized benchmark for the comparison and difference analysis of students' paths, help identify deviations and deficiencies in students' paths, and thus support the scientific decision-making and implementation of personalized teaching. By constructing a standard path database, the cognitive patterns and learning obstacles of students can be more accurately characterized, providing targeted guidance and support for educators and learners.

[0073] Further optionally, the steps of generating an expert cognitive path map for each subject based on the large model and extracting the standard problem-solving path from the large model parameters include steps S210 to S230:

[0074] S210. Input subject textbooks and classic question banks into the pre-trained large model to generate cognitive paths for sub-knowledge points and sub-question types, and obtain an expert path map.

[0075] S220. Verify and correct the path by domain experts to conform to the actual teaching situation to obtain a verified path map. s

[0076] S230. Extract the problem-solving step sequence of typical tasks from the large model parameters to construct a standard path database. Each standard path includes a step ID, semantic label, difficulty coefficient, and associated knowledge points.

[0077] In specific implementation, input subject textbooks, classic question banks, etc. into the pre-trained large model, and utilize the semantic understanding and generation capabilities of the large model to generate cognitive paths for sub-knowledge points and sub-question types. These paths are presented in the form of "stage → sub-task → semantic label", which details the step-by-step learning process from basic knowledge to complex tasks.

[0078] Extract the sequence of problem-solving steps for typical tasks from the parameters of the large model. These step sequences constitute the core content of the standard path database. Each standard path includes step ID, semantic label, difficulty coefficient, associated knowledge points, etc., which can provide a standardized benchmark for the comparison of student paths.

[0079] Integrate the extracted standard problem-solving paths to build a standard path database containing typical tasks. This database not only includes the sequence of problem-solving steps for each subject, but also covers semantic labels, difficulty coefficients, and associated knowledge points related to the steps, providing rich data support for the construction of student portraits. With the update of subject knowledge and the improvement of teaching methods, it is necessary to regularly update and maintain the standard path database to ensure the accuracy and timeliness of its content.

[0080] S300, semantically align the student path with the standard path to obtain the aligned path pair, and calculate the cognitive residual feature vector, including the step redundancy vector, path deviation angle matrix, and concept confusion frequency vector.

[0081] Aligning the student path with the standard path ensures an accurate correspondence between the student path and the standard path, facilitating the calculation of cognitive residual features. Then, by calculating the cognitive residual feature vector, the differences between the student path and the standard path can be quantified, including redundant operations, path deviation, and concept confusion, etc., so as to accurately identify the student's learning state, as well as cognitive biases and obstacle points in the problem-solving process, providing targeted guidance and support for educators and learners, and improving teaching effectiveness and learning efficiency.

[0082] Among them, the step redundancy indicates the redundancy degree of each step in the student path, the path deviation angle indicates the deviation degree between the student path and the standard path at each step, and the concept confusion frequency indicates the concept confusion degree of each step in the student path.

[0083] Specifically, semantic alignment can be performed through the dynamic time warping (DTW) algorithm. The DTW algorithm allows for a certain degree of temporal flexibility in the path, and can more accurately reflect the correspondence between the student path and the standard path. Input the student path and the standard path into the DTW algorithm, and the algorithm finds the best alignment method between the two by calculating the minimum cumulative distance between the paths. The aligned path pair has the same number of steps, and each step corresponds to similar semantic content.

[0084] Further optionally, the steps of calculating the cognitive residual feature vector include steps SS310 to S340:

[0085] S310, align the student operation path with the standard path through the DTW algorithm to obtain the aligned path pair (student path', standard path');

[0086] S320. Calculate the redundancy of steps, generate a step redundancy vector. The formula is: step redundancy = number of invalid steps / total number of steps, where the number of invalid steps is the number of steps in the student path with a semantic similarity lower than the similarity threshold compared to the standard path, and the total number of steps is the total number of all steps in the student operation path.

[0087] S330. Calculate the path deviation angle, generate a path deviation angle matrix. The formula is:

[0088]

[0089] where, v s is the student path vector, v s，i represents the component of v s in the i-th dimension, v t is the standard path vector, v t，i represents the component of v t in the i-th dimension;

[0090] S340. Calculate the ratio of the number of times of misusing concepts to the total number of steps to obtain the concept confusion frequency, and generate a concept confusion frequency vector.

[0091] In specific implementation, calculate the redundancy degree of each step in the student path, that is, the ratio of the number of invalid steps to the total number of steps, where the invalid steps are those with a semantic similarity lower than the similarity threshold compared to the standard path. The step redundancy vector is a one-dimensional vector with the same length as the number of steps in the student path.

[0092] To calculate the deviation degree of the student path and the standard path at each step. By calculating the angle (deviation angle) between the two path vectors, a path deviation angle matrix is obtained. This is a two-dimensional matrix, where the rows and columns respectively correspond to the steps of the student path and the standard path. The larger the deviation angle, the higher the deviation degree of the two paths at this step.

[0093] Count the number of times of misusing concepts in the student path, and calculate the ratio of it to the total number of steps to obtain the concept confusion frequency. The concept confusion frequency vector is a one-dimensional vector, reflecting the confusion degree of the student in different concepts.

[0094] S400. Use a bidirectional Transformer model to calculate the attention weight matrix, and combine the residual features to locate the cognitive breakpoint, generating a checkpoint explanation graph containing knowledge fault location and repair suggestions.

[0095] Capture the semantic similarity and importance between the student path and the standard path at each step through the multi-head self-attention mechanism of the bidirectional Transformer model to obtain a weight matrix, providing a scientific basis for checkpoint identification. Combine the residual feature vector calculated in step S300 and the aligned path pair to generate a checkpoint explanation map containing knowledge fault location and repair suggestions. The generated checkpoint explanation map can accurately locate students' cognitive obstacles and provide repair suggestions.

[0096] Specifically, assume that the operation path of a student when solving a trigonometric function problem (student path) is collected and aligned with the standard problem-solving path of the same problem (the expert cognitive path generated by the large model). In step S300, calculate the step redundancy vector of the student path (such as the proportion of redundant operations such as repeatedly viewing formulas and trying wrong solutions multiple times), the path deviation angle matrix (the deviation degree between the student path and the standard path at each step), and the concept confusion frequency vector (such as confusing the concepts of sine and cosine functions).

[0097] Input these input data (aligned path pair, step redundancy vector, path deviation angle matrix, concept confusion frequency vector) into the bidirectional Transformer model. Through the multi-head self-attention mechanism of the bidirectional Transformer model, calculate the semantic similarity and importance between the student path and the standard path at each step to obtain a weight matrix. The weight matrix can be a two-dimensional matrix, where the rows represent the steps of the student path, the columns represent the steps of the standard path, and each element in the matrix represents the semantic similarity and importance between the corresponding steps.

[0098] In the weight matrix, if the weights of steps such as "understand the properties of the sine function", "understand the properties of the sine function", and "apply trigonometric function formulas to solve problems" are all lower than the weight threshold (such as 0.3), then mark them as different steps. Combine the step redundancy vector and the concept confusion frequency vector to calculate the comprehensive score of each different step. If the comprehensive score exceeds the score threshold (such as 0.7), then determine the corresponding different step as a checkpoint.

[0099] Perform semantic analysis on each checkpoint, label the associated knowledge points (such as the properties of the sine function, the properties of the cosine function, the application of trigonometric function formulas, etc.), and associate repair suggestions such as micro-lesson videos and example recommendations. For example, for the checkpoint of understanding the properties of sine and cosine functions, a micro-lesson video that explains the properties of sine and cosine functions in detail and several related examples for students to practice can be associated.

[0100] Integrate the identified bottlenecks, associated knowledge points, and repair suggestions into a bottleneck explanation graph. The bottleneck explanation graph can be presented in the form of a tree diagram or a flowchart. The root node is "Trigonometric function problem-solving obstacles", and the sub-nodes are bottlenecks such as "Insufficient understanding of the properties of the sine function" and "Insufficient understanding of the properties of the cosine function". Specific knowledge points and repair suggestions are associated with each sub-node. Educators can provide targeted guidance and support to the student based on the bottleneck explanation graph to help them overcome cognitive obstacles and improve learning outcomes.

[0101] Further optionally, the steps of generating a bottleneck explanation graph including knowledge gap positioning and repair suggestions include steps S410 to S420:

[0102] S410, Input the step redundancy vector, path deviation angle matrix, concept confusion frequency vector, and aligned path pairs into a bidirectional Transformer model to calculate a weight matrix, which is used to reflect the semantic similarity and importance between the student path and the standard path at each step;

[0103] S420, Based on the weight matrix, screen out the different steps through a preset weight threshold, and generate a bottleneck explanation graph in combination with the step redundancy vector and the concept confusion frequency vector. Mark the knowledge gap nodes in the bottleneck explanation graph and associate micro-lesson videos and example recommendations.

[0104] Specifically, input the step redundancy vector, path deviation angle matrix, concept confusion frequency vector, and the aligned path pairs (student path' and standard path') calculated in step S300 as input data into a bidirectional Transformer model. The bidirectional Transformer model calculates a weight matrix through the multi-head self-attention mechanism, which is used to reflect the semantic similarity and importance between the student path and the standard path at each step.

[0105] Based on the weight matrix, screen out the different steps through a preset weight threshold to identify the key obstacle points, i.e., bottlenecks, in the student path. Perform semantic analysis on each bottleneck, mark the associated knowledge points, and associate repair suggestions such as micro-lesson videos and example recommendations. These repair suggestions can help students overcome cognitive obstacles and improve learning outcomes.

[0106] Integrate the identified bottlenecks, associated knowledge points, and repair suggestions into a bottleneck explanation graph. This graph graphically shows the students' cognitive obstacles and repair paths, facilitating intuitive understanding and application by educators and learners.

[0107] Through key point recognition and annotation, this embodiment can accurately identify the key obstacle points in the student's path and provide targeted repair suggestions, which helps students overcome cognitive obstacles. The generated key point explanation graph can visually display the students' cognitive obstacles and repair paths in a graphical manner, facilitating understanding and application by educators and learners, thereby improving the pertinence and effectiveness of personalized teaching.

[0108] Further optionally, the step of calculating the weight matrix includes steps S411 to S415:

[0109] S411, for each question, integrate the corresponding step redundancy vector, the step-by-step features of the path deviation angle matrix, the concept confusion frequency vector, and the embedded representation of the aligned path pair into an input sequence;

[0110] S412, input the input sequence into a bidirectional Transformer model, where the Transformer model includes multiple stacked encoder layers, and each layer consists of a multi-head self-attention mechanism and a feed-forward neural network;

[0111] S413, for the vector representation of each step in the input sequence, generate query vectors, key vectors, and value vectors respectively through linear transformation;

[0112] S414, for each query vector, calculate the dot product of it and all key vectors, and normalize it through a scaling and softmax function to obtain the attention weights of the query vector and each step in the input sequence;

[0113] S415, splice the weight matrices output by multiple attention heads, and aggregate them into a final weight matrix through linear transformation, which is used to reflect the semantic similarity and importance between the student path and the standard path at each step.

[0114] Specifically, when integrating the input sequence, the step redundancy vector, the step-by-step features of the path deviation angle matrix, and the concept confusion frequency vector are mapped into high-dimensional vectors through linear transformation or an embedding layer; the embedded representations of the aligned path pairs (student path', standard path') are generated through the embedding layer of the path steps, and positional encoding is added to retain the sequence order information; splicing is performed to obtain the final input sequence, which is used to reflect the path differences and semantic information at the question level.

[0115] The Transformer model of this embodiment can dynamically capture the fine-grained differences between the student path and the standard path at the semantic level for each question, and accurately locate the cognitive breakpoint by combining multi-dimensional residual features (redundancy, deviation angle, confusion frequency), providing an accurate quantitative basis for generating the key point explanation graph and the dynamic cognitive map.

[0116] Further optionally, the step of generating the bottleneck explanation graph by combining the step redundancy vector and the concept confusion frequency vector includes steps S421 to S425:

[0117] S421, traverse the weight matrix, take the student path steps with weights lower than the weight threshold as the differential steps, and record the indexes, the corresponding questions and the corresponding weight values of the differential steps;

[0118] S422, combine the step redundancy vector and the concept confusion frequency vector, and identify the frequently redundant differential steps as bottlenecks;

[0119] S423, perform semantic parsing on the bottlenecks, label their associated knowledge points, and recommend micro-lesson videos and example questions;

[0120] S424, calculate the priority of the repair suggestions, and the formula is: repair priority = fault frequency × concept difficulty coefficient × step redundancy, where the fault frequency is the value of the corresponding knowledge point in the concept confusion frequency vector;

[0121] S425, use the bottlenecks as nodes, and the associated knowledge points, micro-lesson video recommendations and example question recommendations as edges to construct a bottleneck explanation graph. The attribute information of each node includes step redundancy, concept confusion frequency, weight value and repair priority.

[0122] Specifically, traverse the weight matrix, and filter out the student path steps with weights lower than the weight threshold as the differential steps to identify the steps where the students have significant differences from the standard path during the problem-solving process, that is, the steps that may have cognitive obstacles. Combine the step redundancy vector and the concept confusion frequency vector, and identify the frequently redundant differential steps as bottlenecks, that is, further analyze the differential steps to find out the key steps that not only have large differences from the standard path, but also have redundant student operations and frequent concept confusion, that is, the real cognitive obstacle points, so as to more accurately locate the students' cognitive obstacles. Perform semantic parsing on the bottlenecks, label their associated knowledge points, and recommend micro-lesson videos and example questions, which can help students understand the knowledge points behind the bottlenecks and provide specific learning resources to help them overcome cognitive obstacles. By calculating the priority of the repair suggestions, the formula is: repair priority = fault frequency × concept difficulty coefficient × step redundancy. It is possible to judge the severity of the bottlenecks, so as to recommend the priority order of the repair suggestions for the students. The students can give priority to dealing with the bottlenecks that have the greatest impact on the learning effect, overcome the cognitive obstacles more efficiently, and improve the learning efficiency.

[0123] Further optionally, the step of combining the step redundancy vector and the concept confusion frequency vector and identifying the frequently redundant differential steps as bottlenecks includes steps S422.1 to S422.2:

[0124] S422.1, calculate the comprehensive score of each differential step, and the formula is:

[0125] Comprehensive score = α·Step redundancy + β·Concept confusion frequency + γ·(1 - Average weight), where α, β, and γ are weight coefficients;

[0126] S422.2, Define the differential steps with comprehensive scores exceeding the score threshold as bottleneck points.

[0127] Specifically, by comprehensively considering three dimensions: step redundancy, concept confusion frequency, and average weight, calculating the comprehensive score of each differential step can more comprehensively evaluate the severity of each differential step. Among them, step redundancy reflects the frequency of ineffective operations of students during the problem-solving process. The higher the redundancy, the less efficient the operations of students at this step; concept confusion frequency reflects the degree of confusion of students about knowledge points during the problem-solving process. The higher the confusion frequency, the more ambiguous the understanding of students on this knowledge point; average weight reflects the semantic similarity and importance between the student's path and the standard path at each step. The lower the average weight, the greater the deviation of the student from the standard path at this step.

[0128] Set a score threshold according to actual needs, define the differential steps with comprehensive scores exceeding the score threshold as bottleneck points, and record them. Through bottleneck point identification, the cognitive obstacles of students can be more accurately located, providing more targeted teaching support for the follow-up. At the same time, the identification of bottleneck points also helps educators more comprehensively understand the learning status and needs of students, so as to formulate more reasonable teaching plans.

[0129] S500, Generate a dynamic cognitive map and construct a strategy evolution tree based on residual features.

[0130] By integrating and analyzing the residual features of students during the learning process (such as step redundancy, path deviation angle, concept confusion frequency, etc.), generate a dynamic cognitive map and construct a strategy evolution tree to comprehensively reflect the cognitive state and learning strategy changes of students, providing scientific decision-making support and personalized teaching guidance for educators.

[0131] Further optionally, the steps of generating the dynamic cognitive map include steps S510 to S540:

[0132] S510, Take subject knowledge points as nodes and calculate the mastery degree through the concept confusion frequency vector;

[0133] S520, According to the step redundancy vector, path deviation angle matrix, and mastery degree, mark the cognitive fault area, and its color depth = mastery degree × step redundancy × path deviation angle weight, where the path deviation angle weight is a coefficient dynamically adjusted according to the size of the path deviation angle;

[0134] S530, Generate a repair path from the fault node to the mastery node based on the repair suggestions of the bottleneck point interpretation map;

[0135] The S540 updates the map dynamically at preset time intervals to reflect the recent changes in the cognitive state of students.

[0136] Specifically, taking subject knowledge points as nodes, the mastery degree of each knowledge point is calculated using the concept confusion frequency vector to quantitatively evaluate the understanding degree of students on each knowledge point. The value range of the mastery degree is from 0 to 1, where 1 represents complete mastery. Based on the step redundancy vector, path deviation angle matrix, and mastery degree, the cognitive fault regions existing in the process of students solving problems are marked. The depth of the region color can comprehensively reflect the cognitive fault situation of students in the process of solving problems, enabling educators to more intuitively identify which knowledge points students have greater cognitive obstacles in. Based on the repair suggestions of the stuck point explanation atlas, a repair path from the fault node to the mastery node is generated. The generation of the repair path helps students plan the path from the current learning difficulties to the complete mastery of knowledge points, which not only provides students with a clear learning direction and goal but also enhances their confidence in overcoming learning obstacles. The cognitive map is dynamically updated at preset time intervals (such as 7 days) to timely reflect and understand the learning progress and cognitive state changes of students, facilitating the adjustment of teaching plans according to the learning situation of students to ensure the pertinence and effectiveness of the teaching process.

[0137] Further optionally, the steps of constructing the strategy evolution tree include steps S501 to S504:

[0138] S501, based on the residual features and learning behavior data, discover learning strategy subtypes through K-Means clustering;

[0139] S502, with time as the horizontal axis, construct a strategy evolution tree, where the nodes are strategy subtypes and the edges are strategy switching events;

[0140] S503, mark the factors affecting strategy switching;

[0141] S504, calculate the strategy stability index, and the formula is:

[0142] Stability = Strategy duration × Success rate of stuck point repair / Total learning time.

[0143] Specifically, based on residual features (such as step redundancy, concept confusion frequency, etc.) and learning behavior data (such as trial-and-error frequency, tool usage preference, etc.), different subtypes of learning strategies that students may adopt during the learning process (such as rapid trial-and-error type, in-depth derivation type, etc.) are discovered through methods such as K-Means clustering, so as to understand the learning styles and behavior patterns of students more deeply and provide a basis for formulating personalized teaching plans. Taking time as the horizontal axis, a strategy evolution tree is constructed, with nodes being strategy subtypes and edges being strategy switching events. The constructed strategy evolution tree can intuitively observe the dynamic changes of students' learning strategies during the learning process. Factors affecting strategy switching, such as the occurrence of bottlenecks and micro-lesson interventions, are marked on the strategy evolution tree. These factors may trigger students to adjust their learning strategies to meet the current learning needs. Marking the factors affecting strategy switching can insight into the motivation of students to switch learning strategies during the learning process, facilitating the formulation of targeted teaching intervention measures. By calculating the strategy stability index, the stability of students' learning strategies is quantitatively evaluated, which can well identify those students with unstable or ineffective learning strategies and provide them with additional support and guidance. Stability = strategy duration × bottleneck repair success rate / total learning time.

[0144] S600, generate a student cognitive portrait that includes a bottleneck explanation map, a cognitive map, and a strategy evolution tree.

[0145] Integrate the analysis results of the bottleneck explanation map, the cognitive map, and the strategy evolution tree. Among them, the bottleneck explanation map shows the cognitive obstacles and repair suggestions of students during the problem-solving process; the cognitive map visually shows the mastery degree and cognitive fault areas of students on subject knowledge points through visualization; the strategy evolution tree reveals the change trajectory and stability of students' learning strategies.

[0146] Present the integrated analysis results in a visual way to form a student cognitive portrait. This portrait may be presented in forms such as charts, graphs, and interactive interfaces, enabling educators and students to intuitively and deeply understand the learning status of students. Educators can quickly identify students' learning problems by observing the portrait and formulate targeted teaching plans; students can also understand their own learning status by viewing the portrait and plan their own learning plans.

[0147] In summary, the method for constructing a student portrait based on a large model agent of the present invention collects real learning behavior data of students; generates expert problem-solving paths for sub-knowledge points and sub-question types based on the large model, and constructs a standard path database containing typical tasks through verification and correction by domain experts to provide a benchmark for comparing student paths and ensure the objectivity of residual feature calculation and bottleneck point positioning; aligns the student path with the standard path, and calculates multi-dimensional residual feature vectors, including step redundancy vectors, path deviation angle matrices, and concept confusion frequency vectors, to quantify students' cognitive biases and provide multi-dimensional feature support for breakpoint positioning; uses a bidirectional Transformer model to calculate the attention weight matrix, combines the residual features to locate cognitive breakpoints, generates a bottleneck point explanation map containing knowledge fault location and repair suggestions, realizes problem-resource association, supports precise intervention; reflects the changes in students' cognitive states in real time by generating a dynamic cognitive map, supports teachers to dynamically adjust teaching strategies, and reveals the evolution law of learning strategies by constructing a strategy evolution tree, facilitating the optimization of learning path recommendations; integrates the bottleneck point explanation map, cognitive map, and strategy evolution tree into a student cognitive portrait, forming a closed loop from diagnosis to repair to optimization, comprehensively depicting students' cognitive characteristics, learning obstacles, and strategy preferences, and providing a visual decision-making basis for teaching intervention.

[0148] Embodiment 2

[0149] Please refer to Figure 2 , a student portrait construction system based on a large model agent proposed by the present invention, the system includes:

[0150] Data acquisition module: used to collect students' learning behavior data;

[0151] Standard path module: used to generate an expert cognitive path map for each subject based on the large model, extract the standard problem-solving path from the large model parameters, and construct a standard path database containing typical tasks;

[0152] Residual feature module: used to semantically align the student path with the standard path to obtain an aligned path pair, and calculate cognitive residual feature vectors, including step redundancy vectors, path deviation angle matrices, and concept confusion frequency vectors;

[0153] Bottleneck point map module: used to calculate the attention weight matrix using a bidirectional Transformer model, and combine the residual features to locate cognitive breakpoints, and generate a bottleneck point explanation map containing knowledge fault location and repair suggestions;

[0154] Cognitive map module: used to generate a dynamic cognitive map and construct a strategy evolution tree based on the residual features;

[0155] Cognitive portrait module: used to generate a student cognitive portrait containing a bottleneck point explanation map, a cognitive map, and a strategy evolution tree.

[0156] Further optionally, the standard path module is further configured to:

[0157] Input subject textbooks and classic question banks into a pre-trained large model to generate cognitive paths for sub-knowledge points and sub-question types, and obtain an expert path map;

[0158] Verify and correct the path by domain experts to conform to the actual teaching situation, and obtain a verified path map;

[0159] Extract the problem-solving step sequence of typical tasks from the large model parameters, construct a standard path database, and each standard path includes a step ID, a semantic label, a difficulty coefficient, and associated knowledge points.

[0160] Further optionally, the residual feature module is further configured to:

[0161] Calculate the step redundancy, generate a step redundancy vector, and the formula is: step redundancy = number of invalid steps / total number of steps, where the number of invalid steps is the number of student path steps with a semantic similarity lower than the similarity threshold to the standard path, and the total number of steps is the total number of all steps in the student path;

[0162] Calculate the path deviation angle, generate a path deviation angle matrix, and the formula is:

[0163]

[0164] where, v s is the student path vector, v s，i represents the component of v s in the i-th dimension, v t is the standard path vector, v t，i represents the component of v t in the i-th dimension;

[0165] Calculate the ratio of the number of times of misusing concepts to the total number of steps to obtain the concept confusion frequency, and generate a concept confusion frequency vector.

[0166] Further optionally, the stuck point map module is further configured to:

[0167] Input the step redundancy vector, the path deviation angle matrix, the concept confusion frequency vector, and the aligned path pair into a bidirectional Transformer model to calculate a weight matrix;

[0168] Based on the weight matrix, screen out different steps through a preset weight threshold, and combine the step redundancy vector and the concept confusion frequency vector to generate a stuck point explanation map, and mark knowledge fault nodes in the stuck point explanation map and associate micro-lesson videos and example recommendations.

[0169] Further optionally, the key point graph module is further configured to:

[0170] For each question, integrate the corresponding step redundancy vector, the step-by-step features of the path deviation angle matrix, the concept confusion frequency vector, and the embedding representation of the aligned path pair into an input sequence;

[0171] Input the input sequence into a bidirectional Transformer model, where the Transformer model includes multiple stacked encoder layers, and each layer is composed of a multi-head self-attention mechanism and a feed-forward neural network;

[0172] For the vector representation of each step in the input sequence, generate a query vector, a key vector, and a value vector respectively through linear transformation;

[0173] For each query vector, calculate the dot product of it and all key vectors, and normalize it through a scaling and softmax function to obtain the attention weights of the query vector and each step in the input sequence;

[0174] Concatenate the weight matrices output by multiple attention heads, and aggregate them into a final weight matrix through linear transformation, which is used to reflect the semantic similarity and importance between the student path and the standard path at each step.

[0175] Further optionally, the key point graph module is further configured to:

[0176] Traverse the weight matrix, regard the student path steps with weights lower than the weight threshold as different steps, and record the indexes, the corresponding questions, and the corresponding weight values of the different steps;

[0177] Combine the step redundancy vector and the concept confusion frequency vector to identify the frequently redundant different steps as key points;

[0178] Perform semantic parsing on the key points, label their associated knowledge points, and recommend micro-lesson videos and example questions;

[0179] Calculate the priority of the repair suggestion, and the formula is: repair priority = fault frequency × concept difficulty coefficient × step redundancy, where the fault frequency is the value of the corresponding knowledge point in the concept confusion frequency vector;

[0180] Taking the key points as nodes, and the associated knowledge points, micro-lesson video recommendations, and example question recommendations as edges, construct a key point explanation graph, and the attribute information of each node includes step redundancy, concept confusion frequency, weight value, and repair priority.

[0181] Further optionally, the key point graph module is further configured to:

[0182] Calculate the comprehensive score of each different step, and the formula is:

[0183] Comprehensive score = α · step redundancy + β · concept confusion frequency + γ · (1 - average weight), where α, β, and γ are weight coefficients;

[0184] Define the differential steps with a comprehensive score exceeding the score threshold as the bottleneck points.

[0185] Further optionally, the cognitive map module is further configured to:

[0186] Use subject knowledge points as nodes and calculate the mastery degree through the concept confusion frequency vector;

[0187] Mark the cognitive fault region according to the step redundancy vector, path deviation angle matrix, and mastery degree, where the marked color depth = mastery degree × step redundancy × path deviation angle weight, and the path deviation angle weight is a coefficient dynamically adjusted according to the size of the path deviation angle;

[0188] Generate a repair path from the fault node to the mastery node based on the repair suggestions of the bottleneck explanation map;

[0189] Dynamically update the cognitive map at preset time intervals to reflect the recent changes in the student's cognitive state.

[0190] Further optionally, the cognitive map module is further configured to:

[0191] Discover learning strategy subtypes through K-Means clustering based on residual features and learning behavior data;

[0192] Construct a strategy evolution tree with time as the horizontal axis, where the nodes are strategy subtypes and the edges are strategy switching events;

[0193] Mark the factors affecting strategy switching;

[0194] Calculate the strategy stability index, and the formula is:

[0195] Stability = strategy duration × bottleneck repair success rate / total learning time.

[0196] The above embodiments only represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

Claims

1. A method for constructing a student portrait based on large model agents, characterized in that, The method includes: Collecting the learning behavior data of students; Generating an expert cognitive path map for each subject based on a large model, extracting standard problem-solving paths from the large model parameters, and constructing a standard path database containing typical tasks; Semantically aligning the student paths with the standard paths to obtain aligned path pairs, and calculating cognitive residual feature vectors, including step redundancy vectors, path deviation angle matrices, and concept confusion frequency vectors; Calculating an attention weight matrix using a bidirectional Transformer model, and combining the residual features to locate cognitive breakpoints, generating a bottleneck explanation map containing knowledge fault location and repair suggestions; Generating a dynamic cognitive map and constructing a strategy evolution tree based on the residual features; Generating a student cognitive portrait containing the bottleneck explanation map, cognitive map, and strategy evolution tree.

2. The method for constructing a student portrait based on a large model agent according to claim 1, wherein The steps of generating an expert cognitive path map for each subject based on a large model and extracting standard problem-solving paths from the large model parameters include: Inputting subject textbooks and classic question banks into a pre-trained large model to generate cognitive paths for sub-knowledge points and sub-question types, obtaining an expert path map; Verifying and correcting the paths through domain experts to conform to teaching practice, obtaining a verified path map; Extracting the problem-solving step sequences of typical tasks from the large model parameters, constructing a standard path database, and each standard path includes a step ID, semantic label, difficulty coefficient, and associated knowledge points.

3. The method for constructing a student portrait based on a large model agent according to claim 1, wherein The steps of calculating the cognitive residual feature vectors include: Calculating the step redundancy, generating a step redundancy vector, and the formula is: step redundancy = number of invalid steps / total number of steps, where the number of invalid steps is the number of student path steps with a semantic similarity lower than the similarity threshold to the standard path, and the total number of steps is the total number of all steps in the student path; Calculating the path deviation angle, generating a path deviation angle matrix, and the formula is: Among them, v s is the student path vector, and v s，i represents the component of v s in the i-th dimension. v t is the standard path vector, and v t，i represents the component of v t in the i-th dimension; Calculating the ratio of the number of times of misusing concepts to the total number of steps to obtain the concept confusion frequency, and generating a concept confusion frequency vector.

4. The method for constructing a student portrait based on a large model agent according to claim 1, characterized in that The steps of generating a bottleneck explanation map containing knowledge fault location and repair suggestions include: Inputting the step redundancy vector, path deviation angle matrix, concept confusion frequency vector, and the aligned path pairs into a bidirectional Transformer model to calculate the weight matrix; Based on the weight matrix, screening out different steps through a preset weight threshold, and combining the step redundancy vector and concept confusion frequency vector to generate a bottleneck explanation map, marking knowledge fault nodes in the bottleneck explanation map and associating micro-lesson videos and example recommendations.

5. The method for constructing a student portrait based on a large model agent according to claim 4, wherein The steps of calculating the weight matrix include: For each question, integrating the corresponding step redundancy vector, step-by-step features of the path deviation angle matrix, concept confusion frequency vector, and the embedded representation of the aligned path pairs into an input sequence; Inputting the input sequence into a bidirectional Transformer model, where the Transformer model includes multiple stacked encoder layers, and each layer consists of a multi-head self-attention mechanism and a feed-forward neural network; For the vector representation of each step in the input sequence, generating query vectors, key vectors, and value vectors respectively through linear transformation; For each query vector, calculate its dot product with all key vectors, and normalize it through a scaling and softmax function to obtain the attention weights of the query vector at each step in the input sequence; Concatenate the weight matrices output by multiple attention heads and aggregate them through a linear transformation into a final weight matrix, which is used to reflect the semantic similarity and importance between the student path and the standard path at each step.

6. The method for constructing a student portrait based on a large model agent according to claim 4, wherein The step of generating the bottleneck explanation graph by combining the step redundancy vector and the concept confusion frequency vector includes: Traverse the weight matrix, take the steps of the student path with weights lower than the weight threshold as the differential steps, and record the indexes, the corresponding questions, and the corresponding weight values of the differential steps; Combine the step redundancy vector and the concept confusion frequency vector to identify the frequently redundant differential steps as bottlenecks; Perform semantic parsing on the bottlenecks, label their associated knowledge points, and recommend micro-lesson videos and example questions; Calculate the priority of the repair suggestions. The formula is: repair priority = fault frequency × concept difficulty coefficient × step redundancy, where the fault frequency is the value of the corresponding knowledge point in the concept confusion frequency vector; Construct a bottleneck explanation graph with the bottlenecks as nodes, and the associated knowledge points, micro-lesson video recommendations, and example question recommendations as edges. The attribute information of each node includes step redundancy, concept confusion frequency, weight value, and repair priority.

7. The method for constructing a student portrait based on a large model agent according to claim 6, wherein The step of combining the step redundancy vector and the concept confusion frequency vector to identify the frequently redundant differential steps as bottlenecks includes: Calculate the comprehensive score of each differential step. The formula is: Comprehensive score = α · step redundancy + β · concept confusion frequency + γ · (1 - average weight), where α, β, and γ are weight coefficients; Set the differential steps with comprehensive scores exceeding the score threshold as bottlenecks.

8. The method for constructing a student portrait based on a large model agent according to claim 1, wherein The step of generating the dynamic cognitive map includes: Take the subject knowledge points as nodes and calculate the mastery degree through the concept confusion frequency vector; According to the step redundancy vector, the path deviation angle matrix, and the mastery degree, mark the cognitive fault regions. The color depth of the marking = mastery degree × step redundancy × path deviation angle weight, and the path deviation angle weight is a coefficient dynamically adjusted according to the size of the path deviation angle; Generate a repair path from the fault nodes to the mastery nodes based on the repair suggestions of the bottleneck explanation graph; Dynamically update the cognitive map at preset time intervals to reflect the recent changes in the student's cognitive state.

9. The method for constructing a student portrait based on a large model agent according to claim 1, wherein, The step of constructing the strategy evolution tree includes: Based on the residual features and learning behavior data, discover the subtypes of learning strategies through K-Means clustering; Construct a strategy evolution tree with time as the horizontal axis, where the nodes are the subtypes of strategies and the edges are the strategy switching events; Mark the factors affecting the strategy switching; Calculate the strategy stability index. The formula is: Stability = strategy duration × bottleneck repair success rate / total learning time.

10. A student portrait construction system based on a large model agent, which is used to implement the method for constructing a student portrait based on a large model agent described in any one of claims 1 to 9, and is characterized in that, The system includes: Data acquisition module: used to acquire the learning behavior data of students; Standard path module: used to generate the expert cognitive path graphs of each subject based on the large model, extract the standard problem-solving paths from the large model parameters, and construct a standard path database containing typical tasks; Residual Feature Module: It is used to semantically align the student path with the standard path to obtain an aligned path pair and calculate the cognitive residual feature vector, including the step redundancy vector, the path deviation angle matrix, and the concept confusion frequency vector; Stuck Point Atlas Module: It is used to calculate the attention weight matrix using the bidirectional Transformer model, and combine the residual features to locate the cognitive breakpoints, generating a stuck point explanation atlas containing knowledge fault location and repair suggestions; Cognitive Map Module: It is used to generate a dynamic cognitive map and construct a strategy evolution tree based on the residual features; Cognitive Portrait Module: It is used to generate a student cognitive portrait containing the stuck point explanation atlas, the cognitive map, and the strategy evolution tree.

Images