Knowledge graph driving-based large-model personalized exercise recommendation method and related device

Abstract

The invention discloses a large-model personalized exercise recommendation method based on knowledge graph driving and a related device, and relates to the technical field of digital education, the method comprises the following steps: constructing an improved Bayesian knowledge tracking DP-BKT model fusing exercise difficulty and student types, and dynamically calculating mastery degree indexes of students on knowledge points; designing a student personal portrait knowledge graph based on the mastery degree index, and storing association relationships between students and knowledge points and exercises; a knowledge graph, an exercise database and a large language model LLM are integrated, and a personalized exercise recommendation sequence is generated through weak knowledge point determination, vector index and rule matching, large model reasoning and comprehensive multi-dimensional score reordering optimization. The problems that in a traditional recommendation system, information gets lost and a traditional Bayesian knowledge tracking model cannot be associated with exercise difficulty are solved, semantic level alignment and dynamic evolution recommendation of knowledge points and exercises are achieved, and the learning efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital education, in particular to a large model personalized exercise recommendation method based on a knowledge graph and related devices. BACKGROUND

[0002] Under the background of rapid development of digital education, the learning resources (such as textbook question banks, online courses, simulated test questions, etc.) available to students are growing explosively, but this has also led to the "information navigation" dilemma: students have difficulty locating content that matches their knowledge level among the vast number of exercises, and blindly brushing exercises results in wasted time and increased cognitive load. Moreover, the traditional Bayesian Knowledge Tracing (BKT) model has limitations: it cannot associate exercise difficulty, only divides knowledge states into "mastered" and "not mastered", and does not consider student forgetting characteristics, resulting in insufficient prediction accuracy of knowledge point mastery and difficulty supporting personalized exercise recommendation. Therefore, how to solve the problems existing in the prior art and achieve accurate and dynamic personalized exercise recommendation has become a difficult problem to be broken through in the field of education, and is crucial for the goal of efficiency and burden reduction in education reform. SUMMARY

[0003] The purpose of the present application is to provide a large model personalized exercise recommendation method based on a knowledge graph and related devices, which can solve the problems of "information navigation" and the inability of the traditional BKT model to associate exercise difficulty, and achieve accurate and dynamic personalized exercise recommendation.

[0004] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0005] In a first aspect, the present application provides a large model personalized exercise recommendation method based on a knowledge graph, comprising the following steps:

[0006] An improved Bayesian Knowledge Tracing model based on exercise difficulty and student type is constructed, the student's historical answer sequence is input, and the student's mastery index of knowledge points is iteratively calculated.

[0007] Based on the student's mastery index of knowledge points, a personal portrait knowledge graph containing the association relationship between students, knowledge points and exercises is constructed; the association relationship includes the student's mastery index of knowledge points and the student's completion degree of exercises.

[0008] Based on the personal portrait knowledge graph, the student's weak knowledge points are determined, and the weak knowledge points are filtered in the course exercise database according to the weak knowledge points. The weak knowledge point exercise list is determined by vector indexing and rule matching.

[0009] According to the weak knowledge point exercise list and the preset prompt word, the large language model is further screened in the course exercise database to obtain an implicit associated knowledge point exercise list.

[0010] For any exercise in the weak knowledge point exercise list and the implicit associated knowledge point exercise list, a comprehensive score of the exercise is calculated according to the historical exercise of the student, the difficulty of the exercise and the knowledge point involved in the exercise.

[0011] The exercises in the weak knowledge point exercise list and the implicit associated knowledge point exercise list are sorted in descending order according to the comprehensive score, and the first few exercises are taken as the final exercise recommendation list.

[0012] Optionally, in the improved Bayesian knowledge tracking model based on the exercise difficulty and the student type, the guessing probability of the student when answering the exercise t is calculated according to the following formula:

[0013]

[0014] Wherein, is the guessing probability of the student when answering the exercise t, P(G) k is the probability of the student guessing correctly on the knowledge point k to lead to a correct answer to the question, ω is a personal portrait coefficient representing the knowledge mastery level of the student, D t is the difficulty coefficient of the tth question.

[0015] The failure probability of the student when answering the exercise t is calculated according to the following formula:

[0016]

[0017] Wherein, is the failure probability of the student when answering the exercise t, P(S) k is the probability of the student failing on the knowledge point k to lead to an incorrect answer to the question.

[0018] Optionally, the course exercise database is a relational database, which stores the association relationship among courses, exercises and knowledge points, and the attributes of the exercises include exercise ID, stem, exercise type and exercise difficulty.

[0019] The weak knowledge points of the student are determined based on the personal portrait knowledge graph, and the weak knowledge points are screened in the course exercise database, and the weak knowledge point exercise list is determined through vector indexing and rule matching, which specifically includes the following steps:

[0020] The knowledge points with a mastery index lower than a threshold value are extracted from the personal portrait knowledge graph as weak knowledge points through a graph database query language.

[0021] The initial associated exercise list is obtained by matching exercise data associated with the weak knowledge points in the course exercise database through multi-table association of the relational database. The initial associated exercise list includes several exercises, exercise IDs, stems, types, and exercise difficulties.

[0022] The weak knowledge points and each exercise in the initial associated exercise list are converted into high-dimensional vectors using the Embedding technology, and the cosine similarity between the weak knowledge point vector and each exercise vector is calculated through the collaborative filtering algorithm.

[0023] According to the cosine similarity between the weak knowledge point vector and each exercise vector, the initial associated exercise list is double-screened according to the cosine similarity threshold and the preset rules, and the weak knowledge point exercise list is obtained.

[0024] Optionally, a text-embedding-ada-002 text embedding model is used to map any length text to a 1536-dimensional dense vector space through a deep neural network architecture. The text-embedding-ada-002 text embedding model uses a multi-layer Transformer structure to dynamically calculate the correlation weight between text units through a self-attention mechanism. Each layer of the Transformer includes a multi-head attention and a feedforward network two sub-modules, which realize stable extraction of deep features through residual connection and layer normalization.

[0025] Optionally, according to the weak knowledge point exercise list and the preset prompt word, a large language model is used for further screening in the course exercise database to obtain an implicit associated knowledge point exercise list. Specifically, the weak knowledge point exercise list is input into the large language model to analyze the student's mistake mode, identify the potential knowledge gap logic chain, and further screen in the course exercise database based on the knowledge reasoning ability to supplement the implicit associated knowledge point exercises that are not captured by the vector index and rule matching method, and obtain the implicit associated knowledge point exercise list.

[0026] Optionally, the comprehensive score of the exercise is calculated according to the following formula:

[0027] S final = αS similarity + βS difficulty + γS LLM

[0028] Wherein, S final is the comprehensive score, α is the weight of the vector similarity score, S similarity is the vector similarity score, β is the weight of the difficulty matching score, S difficulty is the difficulty matching score, γ is the weight of the large language model score, S LLM is the large language model score.

[0029] The vector similarity score is calculated according to the following formula:

[0030] S similarity = 1-S

[0031] Wherein, S is the cosine similarity between the exercise vector and the weak knowledge point vector.

[0032] The difficulty matching score is calculated according to the following formula:

[0033] S difficulty = 1-|d target -d exercise |

[0034] Wherein, d tar get is the current best difficulty of the student, d exercise is the difficulty value of the exercise.

[0035] In a second aspect, the application provides a large model personalized exercise recommendation system driven by a knowledge graph, comprising:

[0036] The mastery index calculation module is used to construct an improved Bayesian knowledge tracking model based on exercise difficulty and student type, input the student's historical answering sequence, and iteratively calculate the student's mastery index of the knowledge point.

[0037] The personal portrait knowledge graph construction module is used to construct a personal portrait knowledge graph containing the association relationship between the student, the knowledge point and the exercise based on the student's mastery index of the knowledge point; the association relationship includes the student's mastery index of the knowledge point and the student's completion degree of the exercise.

[0038] The weak knowledge point exercise determination module is used to determine the student's weak knowledge point based on the personal portrait knowledge graph, and determine the weak knowledge point exercise list according to the weak knowledge point in the course exercise database.

[0039] The implicit association knowledge point exercise determination module is used to further filter the implicit association knowledge point exercise list in the course exercise database according to the weak knowledge point exercise list and the preset prompt word by using a large language model.

[0040] The exercise comprehensive score calculation module is used to calculate the comprehensive score of any exercise in the weak knowledge point exercise list and the implicit association knowledge point exercise list according to the student's historical exercise situation, the difficulty of the exercise and the knowledge point involved in the exercise.

[0041] The final exercise list recommendation module is used to sort the exercises in the weak knowledge point exercise list and the implicit association knowledge point exercise list in descending order according to the comprehensive score, and take the first few exercises as the final exercise recommendation list. The final exercise list recommendation module is used to sort the exercises in the weak knowledge point exercise list and the implicit association knowledge point exercise list in descending order according to the comprehensive score, and take the first few exercises as the final exercise recommendation list.

[0042] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the knowledge graph driven large model personalized exercise recommendation method described above.

[0043] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of the knowledge graph driven large model personalized exercise recommendation method described above.

[0044] In a fifth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the steps of the knowledge graph driven large model personalized exercise recommendation method described above.

[0045] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0046] The present application provides a knowledge graph driven large model personalized exercise recommendation method and related devices, which introduces an exercise difficulty coefficient and a student type coefficient based on a Bayesian knowledge tracking model, solves the problem that the traditional model cannot associate the exercise difficulty, and the dynamic iterative calculation mechanism can reflect the change of the student's knowledge mastery in real time, improves the accuracy of the mastery index, and lays a reliable foundation for subsequent recommendation. And through the structured storage of the knowledge graph, the multi-dimensional association of students, knowledge points and exercises is intuitively presented, not only can the student's knowledge short board be clearly positioned, but also the logical dependency relationship between knowledge points can be revealed, providing systematic data support for personalized recommendation. Then, combined with the vector index technology and the pre-set rule matching, the irrelevant exercises are effectively filtered, avoiding the students from falling into the "information maze", improving the pertinence of exercise screening. On this basis, the large language model is used to mine and capture the implicit association between knowledge points that traditional methods have not found, supplement the extension exercises of weak knowledge points, and perfect the comprehensive training of the student's knowledge system. Finally, the scoring mechanism and the sorting dynamic generation of the final recommendation sequence are generated by comprehensively considering the multi-dimensional factors, balancing the adaptability of the exercises and the student's level, the comprehensiveness of the knowledge point coverage, and ensuring that the optimal exercises are presented first, focusing on weak points for intensive training, and also considering the knowledge correlation for expansion, effectively solving the problem of blind exercise, and helping students learn efficiently. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0048] Figure 1 A flow chart of a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0049] Figure 2 A topology structure diagram of a standard Bayesian knowledge tracking model in a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0050] Figure 3 A topology structure diagram of an improved Bayesian knowledge tracking model in a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0051] Figure 4 A schematic diagram of a personal portrait knowledge graph in a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0052] Figure 5 A schematic diagram of a few-sample prompt word template used in a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0053] Figure 6 A schematic diagram of a thinking chain prompt word template used in a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0054] Figure 7 An effect diagram of an exercise recommendation function in a large model personalized exercise recommendation method based on knowledge graph driving provided by an embodiment of the present application.

[0055] Figure 8 A functional module diagram of a large model personalized exercise recommendation system based on knowledge graph driving provided by an embodiment of the present application.

[0056] Figure 9 A structural schematic diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0058] In order to make the above objectives, characteristics and advantages of the present application more apparent, further and specific embodiments will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] The method for personalized exercise recommendation of a large model based on a knowledge graph provided by the embodiments of the present application comprises the following steps in an exemplary embodiment, as shown in Figure 1

[0060] A1, an improved Bayesian knowledge tracking model based on exercise difficulty and student type is constructed, and a student historical answer sequence is input to iteratively calculate a mastery index of a student on a knowledge point.

[0061] The standard BKT model models the learning process of a student as a Markov process, and the basic idea is that the learning of a student is not fixed, but dynamically changes as the practice proceeds, and the topological structure is as shown in Figure 2 The K node is a hidden node, representing the knowledge state of a student, such as K0 representing the mastery state P(L0) of a student on a knowledge point at the initial time. The O node is a performance node, representing the answer behavior of a student, such as O0 representing the correct or incorrect situation of a student at the first question, i.e. the probability P(G)0 of a student guessing correctly at the initial time and the probability P(S)0 of a student making a mistake, and the edges between the nodes represent the conditional dependence relationship. The parameters in the model and their detailed meanings are shown in Table 1. It can be seen from Figure 2 that the knowledge state of a student at the current time only depends on the knowledge state at the previous time, and the answer performance at the current time only depends on the knowledge state at the current time. For example, the mastery state of a student on a knowledge point when answering the second question is only determined by the mastery state when answering the first question, and is irrelevant to other factors; the answer performance on the second question is only related to the current knowledge state, and is irrelevant to other factors.

[0062] Table 1: Parameters and meanings of the standard BKT model

[0063]

[0064] The standard BKT model has five basic assumptions:

[0065] 1) The knowledge state of a student is a binary variable, i.e. a student has only two states of “mastered” and “not mastered” for each knowledge point. Through learning, a student can be transferred from the “not mastered” state to the “mastered” state.

[0066] 2) A student can forget, i.e. a student can be transferred from the “mastered” state to another state for a knowledge point.

[0067] ​3) The performance of the student answering the question depends on the current knowledge state, but there is still a possibility of error and guessing. Even if the student's knowledge state is "mastered", there may be errors due to mistakes; even if the student's knowledge state is "not mastered", there may be correct answers due to guessing.

[0068] 4) The knowledge state of the student is constantly updated during the learning process, i.e. each time the knowledge point is applied to solve a new problem, the knowledge point mastery state of the student changes.

[0069] 5) The probability of a student transitioning from a "not mastered" state to a "mastered" state is fixed and independent of whether the student answers the question correctly.

[0070] The BKT model is a special first-order Markov chain that also satisfies the Markov property assumption and the observation independence assumption. Combining the parameter definitions of the model, assuming t is the number of exercises, the state space can be set as {K t ,0<t<T} and the observation space as {O t ,0<t<T}, i.e. as shown in equation (1.1):

[0071]

[0072] And the following three probability distribution tables can be obtained.

[0073] (1) The initial state probability table lists the probabilities of the student's knowledge states being "mastered" and "not mastered" before the initial learning, a total of two cases, as shown in Table 2.

[0074] Table 2 Initial state probability table

[0075] Knowledge State Probability Mastered [P(L0)] Not Mastered 1 - P(L0)

[0076] (2) The state transition probability table, corresponding to the Markov property assumption, lists the probabilities of the student's mutual conversion between the "mastered" and "not mastered" knowledge states under the observation state O, a total of four cases, as shown in Table 3.

[0077] Table 3 State transition probability table

[0078] Knowledge State Mastered Not Mastered Mastered 1 - P(F) P(F) Not Mastered P(T) 1 - P(T)

[0079] (3) The emission probability table, corresponding to the observation independence assumption, lists the probabilities of the student's different answer conditions under the two different knowledge states, a total of four cases, as shown in Table 4.

[0080] Table 4 Emission probability table

[0081]

[0082]

[0083] Bayesian knowledge tracking includes two stages: learning stage and prediction stage. The learning stage includes learning all parameters of the model, generally using EM algorithm or gradient descent algorithm. The prediction stage refers to predicting the probability that the student has mastered knowledge point k and the probability of answering the next question correctly according to the trained model and the real answering situation, and the specific process is as follows:

[0084] First, according to the probability distribution table, the probability that the student u correctly answers the tth question involving knowledge point k is defined as: the probability that the student does not make a mistake under the condition of having mastered knowledge point k, and the probability of guessing correctly under the condition of not having mastered the knowledge point, as shown in equation (1.2):

[0085]

[0086] Where k represents the knowledge point, P (k | u, t) represents the probability that the student u masters the knowledge point k when answering the tth question, P (error | k, u, t) represents the error probability of the student on knowledge point k, P (guess | k, u, t) represents the guessing probability of the student on knowledge point k.

[0087] Similarly, the probability that the student u incorrectly answers the tth question involving knowledge point k is defined as: the probability that the student makes a mistake under the condition of having mastered knowledge point k, and the probability of guessing incorrectly under the condition of not having mastered the knowledge point, as shown in equation (1.3):

[0088]

[0089] According to Bayes' theorem, the probability that the student u masters the knowledge point k under the condition of correctly answering the tth question can be calculated as shown in equation (1.4):

[0090]

[0091] Similarly, the probability that the student u masters the knowledge point k under the condition of incorrectly answering the tth question can be calculated as shown in equation (1.5):

[0092]

[0093] After obtaining the transition probability of the tth question, the probability that the student masters the knowledge point when answering the t+1th question can be calculated as: the probability that the knowledge point is not forgotten when answering the tth question and the probability that it is learned under the condition of not having mastered the knowledge point, as shown in equation (1.6):

[0094]

[0095] The probability of correctly answering the t+1 question is the sum of the probability of not missing the point under the condition of mastering the knowledge point and the probability of guessing correctly under the condition of not mastering the knowledge point, as shown in equation (1.7):

[0096]

[0097] The improved DP-BKT model in this embodiment introduces a new problem node D to the standard BKT model to introduce the difficulty coefficient of the problem into the knowledge tracking model, and the topological structure is as shown in Figure 3 At this time, the value of the performance node O is no longer determined by the knowledge state K at the current time, but also related to the problem node D at the current time. The difficulty of the problem D directly affects the values of the guessing probability P(G) and the missing probability P(S), and further indirectly affects the value of the performance node O. Taking the second question as an example, the answer performance O1 is determined by the problem difficulty D1 and the knowledge state K1. The problem difficulty D1 affects the values of the guessing probability P(G)1 and the missing probability P(S)1. The larger d1 is, the larger P(S)1 is, and the smaller the probability of correct answer is; the smaller D1 is, the larger P(G)1 is, and the larger the probability of correct answer is.

[0098] In the improved Bayesian knowledge tracking model based on problem difficulty and student type in this embodiment, the guessing probability of the student when answering the problem t is calculated according to equation (1.8) as follows:

[0099]

[0100] wherein, is the guessing probability of the student when answering the problem t, P(G) k is the probability of the student guessing correctly on the knowledge point k to lead to the correct answer of the question, ω is the personal portrait coefficient, representing the knowledge mastery level of the student, D t is the difficulty coefficient of the t question.

[0101] The missing probability of the student when answering the problem t is calculated according to equation (1.9) as follows:

[0102]

[0103] wherein, is the missing probability of the student when answering the problem t, P(S) k is the probability of the student missing on the knowledge point k to lead to the incorrect answer of the question. For students of different types, the difficulty coefficient is adjusted by introducing the personal portrait coefficient ω to be closer to the real level.

[0104] For knowledge point k, corresponding to the DP-BKT model, when the answer is correct, two cases are considered: the first is that the student has mastered the knowledge point and made no mistakes; the second is that the student has not yet mastered the knowledge point but is able to correctly answer the exercise by guessing. Similarly, when the answer is incorrect, there are also two cases: the first is that the student has mastered the knowledge point but made mistakes when answering; the second is that the student has not yet mastered the knowledge point and did not guess the answer correctly. The specific formulas are shown in (1.10) and (1.11):

[0105]

[0106] in, represents the probability that student u answers the tth question under knowledge point k correctly; represents the probability that student u gives an incorrect answer to question t under knowledge point k; represents the probability that a student makes a mistake in answering the tth question on knowledge k; It represents the probability that a student guesses correctly on question t at knowledge point k and gives the correct answer.

[0107] According to the student's answer sequence, the learning prior probability is updated, which can be expressed by formula (1.12):

[0108]

[0109] in, represents the probability that student u has mastered knowledge point k based on answering the tth question under knowledge point k correctly; It represents the probability that student u will master knowledge point k based on the incorrect answer to question t under knowledge point k.

[0110] The formula for updating the student's knowledge status for the next question is shown in formula (1.13):

[0111]

[0112] Among them, P(T) k It represents the probability that a student changes from a state of not mastering knowledge point k to a state of mastering it through learning.

[0113] The predicted probability formula of the student answering the next question correctly is shown in formula (14):

[0114]

[0115] In summary, the probability of guessing and the probability of error for each exercise in the DP-BKT model are jointly determined by the knowledge points and the exercises, while the difficulty coefficient of each exercise is fixed, and its difficulty coefficient is dynamically adjusted according to different types of students to better fit their actual level.

[0116] A2, based on the index of the students' mastery of knowledge points, a personal portrait knowledge graph containing the correlation between students, knowledge points and exercises is constructed; the correlation includes the index of the students' mastery of knowledge points and the completion degree of the students to exercises. The knowledge graph aims to systematically store the students' mastery of each knowledge point in the learning process, thereby providing data support and intelligent support for subsequent exercise recommendation. The node and relationship design of the knowledge graph is shown in Table 5 and Table 6, covering core entities such as students, knowledge points, exercises, and their correlation, such as the mastery of students and knowledge points, the correlation between knowledge points and exercises, etc. Through this structured design, the knowledge graph can fully reflect the learning state of students and the construction of knowledge system.

[0117] Table 5 Knowledge graph node description

[0118]

[0119] Table 6 Personal portrait graph relationship description

[0120] Relationship Type Property Name Description Relate Relate (Course) -[:Relate]-> (Knowledge Point) Contain Contain (Exercise) -[:Contain]-> (Knowledge Point) Mastery Level Mastery Level (Student) -[:Mastery_Level]-> (Knowledge Point) Finish Level Finish Level (Student) -[:Finish_Level]-> (Exercise) Belong to Belong to (Exercise) -[:Belong_to]-> (Exercise Type)

[0121] According to the set knowledge graph node relationship, the personal portrait knowledge graph as shown in Figure 4 is obtained, wherein each sub-node covers the knowledge point mastery index of the student obtained by step A1, and the color of the node is set from light to deep according to the index from high to low, so that the node with deeper color can better reflect the weak knowledge points mastered by the student.

[0122] In order to efficiently store and manage the data in the knowledge graph, Neo4j graph database is used as the storage engine in this embodiment. As a high-performance graph database, Neo4j can quickly process complex graph structure data, support efficient query and analysis operations, and is particularly suitable for storing and managing multi-dimensional relationship data in the personal portrait knowledge graph of students. Through the graph query language (Cypher) of Neo4j, the query, update and graph structure-based analysis of the students' knowledge point mastery can be easily realized.

[0123] A3, based on the personal portrait knowledge graph, the weak knowledge points of the student are determined, and the weak knowledge points are filtered in the course exercise database to determine the weak knowledge point exercise list through vector index and rule matching. The course exercise database is a relational database, which stores the correlation of courses, exercises and knowledge points, and the attributes of exercises include exercise ID, stem, exercise type and exercise difficulty. Step A3 specifically includes the following steps:

[0124] A31, extract knowledge points with mastery index below a threshold value from the personal portrait knowledge graph through a graph database query language as weak knowledge points.

[0125] A32, match exercise data associated with the weak knowledge points in the course exercise database through multi-table association of a relational database to obtain an initial associated exercise list; the initial associated exercise list includes several exercises, exercise IDs, stems, types, and exercise difficulties.

[0126] A33, use Embedding technology to convert the weak knowledge points and each exercise in the initial associated exercise list into high-dimensional vectors, and calculate the cosine similarity of the weak knowledge point vector and each exercise vector through a collaborative filtering algorithm.

[0127] A34, according to the cosine similarity of the weak knowledge point vector and each exercise vector, double-screen the initial associated exercise list according to a cosine similarity threshold and a preset rule to obtain a weak knowledge point exercise list.

[0128] Step A3 can be divided into two processes of weak knowledge point exercise screening and vector indexing and rule matching. The process of weak knowledge point exercise screening realizes the accurate positioning of students' weak knowledge points and corresponding exercises by integrating the advantages of graph databases and relational databases in two stages. In the first stage, based on the association query characteristics of the Neo4j graph database, the relationship between students and knowledge points is taken as the breakthrough point, the Cypher query language is used to match the mastery level relationship type between the student node (Student) and the knowledge point node (Knowledge Point), the mastery index (MasteryIndex) is set to be less than the threshold value to extract weak knowledge points, the structured output is realized, and an ordered list containing the knowledge point name and the mastery index is returned to establish a data basis for subsequent analysis. Through dynamic adjustment of the threshold value parameter, different teaching stages can be flexibly adapted. In the second stage, relying on the relational data storage advantage of the SQL database, exercise resources are matched through multi-table association. According to the knowledge point list obtained in the previous stage, in the association query of the exercise table (questions) and the knowledge point label table (question_knowledge_point), the exercise data bound to the target knowledge point is dynamically screened. The query result set is structured to form a standardized data format containing exercise ID, stem content, type classification, and difficulty classification, which provides associated exercises for the learning path recommendation system.

[0129] The two-stage processing forms a closed-loop data flow: the graph database focuses on revealing the topological relationship between knowledge points and the cognitive state of students, and the relational database efficiently manages structured exercise resources. The final output structured data set not only presents the quantitative characteristics of weak knowledge points, but also provides targeted reinforcement resources, establishing a complete data support chain for personalized learning.

[0130] In the vector index and rule matching process, when text vectorization is performed, the text-embedding-ada-002 text embedding model is used in the embodiment. Through a deep neural network architecture, any length of text is mapped to a 1536-dimensional dense vector space. The model performs excellently in both semantic preservation and computational efficiency. The model uses a multi-layer Transformer structure and dynamically calculates the correlation weight between text units through a self-attention mechanism (Self-Attention). This design breaks through the sequential processing limitation of traditional RNN and supports parallel capture of long-distance dependencies. In particular, each layer of the Transformer contains two sub-modules: multi-head attention (Multi-head Attention) and feedforward network (FFN). Through residual connection and layer normalization, stable extraction of deep features is achieved.

[0131] Unlike static word vector methods such as Word2Vec, the model is based on a pre-trained language model (Pre-trained Language Model) architecture and uses training strategies such as Masked Language Modeling (MLM) and Next Sentence Prediction (NSP). This pre-training mechanism enables the model to simultaneously integrate left and right context information for words. When dealing with polysemous words, it can dynamically generate differentiated vector representations based on context, significantly improving semantic disambiguation capabilities.

[0132] The 1536-dimensional vector output by the model is mapped to a special space through a contrastive learning (Contrastive Learning) optimization of the vector space distribution. In specific implementation, a cosine similarity loss function is used to ensure that the angle between similar text vectors tends to 0°, and the angle between unrelated text vectors tends to 90°.

[0133] After obtaining the knowledge point vector representation, the cosine similarity is used to measure the similarity between the exercise vector and the knowledge point vector, and then the vector index search is used to find the most similar exercise vector to the vector, thereby finding the exercise related to the knowledge point. The cosine similarity algorithm is used to quantify the similarity between problems, as shown in equation (3.1):

[0134]

[0135] where A·B is the dot product of vectors A and B, ||A|| and ||B|| are the Euclidean norms of vectors A and B, respectively, A i and B i are the values of the respective indices of the exercise vectors and the knowledge point vectors.

[0136] By calculating the similarity between the student's knowledge points and exercises, each knowledge point can be matched with the exercises that have a higher similarity. Then, the candidate recommended exercise set can be obtained by adopting the corresponding recommendation rules, as shown in Table 7.

[0137] Table 7. Example of recommendation rules

[0138]

[0139] A4. According to the weak knowledge point exercise list and the preset prompt word, a large language model is used to further screen in the course exercise database to obtain an implicit associated knowledge point exercise list. Specifically, the weak knowledge point exercise list is input into the large language model, the student's mistake mode is analyzed, the potential knowledge gap logic chain is identified, and based on the knowledge reasoning ability, the course exercise database is further screened to supplement the implicit associated knowledge point exercises that are not captured by the vector index and rule matching method, and an implicit associated knowledge point exercise list is obtained.

[0140] The present application utilizes the strong reasoning ability of the large model to design prompt words from three levels of Few-Shot Prompting, Chain-of-Thought (CoT), and Role-Playing. These methods can help the large model more accurately match the student's weak knowledge points and suitable exercises, thereby improving the learning effect.

[0141] The core idea of Few-Shot Prompting is to provide a few high-quality examples to let the large model imitate the recommendation logic in the examples to generate exercise recommendations that meet the characteristics of the circuit course. The knowledge points of the circuit course usually have a hierarchical structure, and students may have different degrees of mastery at different levels of knowledge points. Therefore, by providing examples, the large model learns how to recommend reasonable exercises according to the student's mastery. First, a few examples of knowledge points-student mastery-exercise recommendations are given, and the large model is required to follow the same format to generate new recommendation results. A Few-Shot Prompting word template used in this embodiment is shown as Figure 5 .

[0142] The core idea of the Chain-of-Thought prompt is to let the large model recommend exercises in a step-by-step reasoning manner, rather than directly generating an exercise list. This way, the large model can first analyze the core concepts of the knowledge points, then think about how to design reasonable exercises, and finally generate the final recommendations. A Chain-of-Thought prompt word template used in this embodiment is shown asFigure 6 As shown. Thinking chain prompts are particularly suitable for more difficult circuit knowledge points, allowing large models to think in the following steps:

[0143] Analyze the core concepts of knowledge points: What are the basic principles of knowledge points? What are the application scenarios?

[0144] Identify students' potential learning difficulties: Where do students tend to make mistakes? What misconceptions are most common?

[0145] Determine the goal of the exercise design: How should the exercise help students understand the knowledge points? Do they need to be combined with real-world engineering applications?

[0146] Generate exercises that meet your goals: clarify the content, question types, and difficulty levels of the questions.

[0147] The core idea behind role-playing prompts is to allow the big model to simulate different roles to adapt to different teaching scenarios. In personalized exercise recommendations, the big model can play different roles and implement different recommendation strategies. For example, playing the role of a teacher analyzes students' knowledge from a teaching perspective and recommends the most appropriate exercises; playing the role of a student simulates the student's perspective and allows the model to actively select appropriate exercises based on their learning needs.

[0148] A5. For any exercise in the list of weak knowledge points and the list of implicitly associated knowledge points, calculate the comprehensive score of the exercise based on the student's past practice, the difficulty of the exercise, and the knowledge points involved. In this embodiment, the exercise score is calculated by combining multiple factors, and a comprehensive score is defined for each candidate exercise in the total recommendation list. The comprehensive score of the exercise is calculated according to formula (5.1):

[0149] S final =αS similarity +βS difficulty +γS LLM (5.1)

[0150] Among them, S final is the comprehensive score, α is the weight of the vector similarity score, S similarity is the vector similarity score, β is the weight of the difficulty matching score, S difficulty is the difficulty matching score, which is introduced to avoid recommending too difficult or too easy questions. γ is the weight of the large language model score. S LLM Scoring the large language model. Taking into account multiple scoring factors, the weight parameters for each score in formula (5.1) are set to α = 0.6, β = 0.3, and γ = 0.1. The weight values ​​can be adjusted based on teaching feedback to optimize the final recommendation effect.

[0151] Calculate the vector similarity score according to formula (5.2):

[0152] S similarity = 1 - S (5.2)

[0153] wherein S is the cosine similarity between the exercise vector and the weak knowledge point vector.

[0154] Calculate the difficulty matching score according to formula (5.3):

[0155] S difficulty = 1 - |d target -d exercise | (5.3)

[0156] wherein d target is the current best difficulty of the student, which is evaluated according to historical answer data, and the evaluation standard is the highest difficulty of the exercises that the student currently practices correctly, d exercise is the difficulty value of the exercise.

[0157] The large language model score S LLM then evaluates whether the exercise is in line with the learning progress through the large language model, and the specific evaluation process is that the large language model comprehensively considers the historical exercise situation of the student and the knowledge points, types, and difficulty of the candidate exercise to perform reasoning and calculation to obtain a recommended value of 0 to 1.

[0158] A6, sort the exercises in the weak knowledge point exercise list and the implicitly related knowledge point exercise list in descending order according to the comprehensive score, and take the first few exercises as the final exercise recommendation list. After calculating the comprehensive score of all exercises, sort the exercises in descending order according to the comprehensive score to ensure that the optimal exercises are ranked first, and return the first few exercises with the highest score as the final recommendation result.

[0159] In actual application, the exercise recommendation function effect diagram is as shown in Figure 7 After the student clicks the AI exercise button in the applet, the server runs the algorithm in the background to generate an exercise recommendation sequence, and then recommends the corresponding exercises on the front end.

[0160] The application scheme innovatively adopts an improved Bayesian knowledge tracking model (DP-BKT) algorithm, fuses the exercise difficulty comprehensive prediction of the student's mastery index of knowledge points, successfully solves the key problem that the traditional Bayesian knowledge tracking model (BKT) cannot associate the exercise difficulty. On this basis, the problem of accurate acquisition of the student's knowledge point mastery index is further solved, and the knowledge graph and vector retrieval enhanced RAG technology are further fused to realize the semantic level accurate alignment of weak knowledge points and exercise knowledge points. The method combines the reordering algorithm, comprehensively considers the progressive difficulty of the question, the problem solving mode preference and the student type characteristics, and dynamically generates a continuously evolving recommendation sequence. This breakthrough scheme completely gets rid of the static limitations of the traditional recommendation system, enables the exercise resources to be intelligently recommended on demand, effectively solves the information navigation dilemma, and helps students to achieve efficient learning.

[0161] Based on the same inventive concept, the embodiments of the present application also provide a system for implementing the knowledge graph driven large model personalized exercise recommendation method described above. The implementation scheme of the system for solving the problem is similar to the implementation scheme described in the above method. In an exemplary embodiment, as shown in Figure 8 , a knowledge graph driven large model personalized exercise recommendation system is provided, including the following functional modules:

[0162] The mastery index calculation module is used to construct an improved Bayesian knowledge tracking model based on exercise difficulty and student type, input the student's historical answer sequence, and iteratively calculate the student's mastery index of knowledge points.

[0163] The personal portrait knowledge graph construction module is used to construct a personal portrait knowledge graph containing the association relationship among students, knowledge points and exercises based on the student's mastery index of knowledge points; the association relationship includes the student's mastery index of knowledge points and the student's completion degree of exercises.

[0164] The weak knowledge point exercise determination module is used to determine the student's weak knowledge points based on the personal portrait knowledge graph, and determine the weak knowledge point exercise list according to the weak knowledge points in the course exercise database.

[0165] The implicit association knowledge point exercise determination module is used to further screen the course exercise database by using a large language model according to the weak knowledge point exercise list and the preset prompt word, to obtain an implicit association knowledge point exercise list.

[0166] The exercise comprehensive score calculation module is used to calculate the comprehensive score of any exercise in the weak knowledge point exercise list and the implicit association knowledge point exercise list according to the student's historical exercise situation, the difficulty of the exercise and the knowledge points involved in the exercise.

[0167] The final exercise list recommendation module is configured to sort exercises in the weak knowledge point exercise list and the implicit associated knowledge point exercise list in descending order according to comprehensive scores, and take the first several exercises as a final exercise recommendation list.

[0168] Of course, Figure 8 The illustrated architecture is only exemplary, and in the implementation of different functions, according to the actual needs, one or at least two components in the system can be omitted Figure 8 The illustrated architecture is only exemplary, and in the implementation of different functions, according to the actual needs, one or at least two components in the system can be omitted

[0169] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and its internal structure diagram can be as shown. Figure 9 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through the system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control ability. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with the terminal outside through the network connection. The computer program is executed by the processor, which can realize the knowledge graph driven large model personalized exercise recommendation method provided in the above embodiment.

[0170] Those skilled in the art can understand, Figure 9 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0171] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and its internal structure diagram can be as shown.

[0172] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and its internal structure diagram can be as shown.

[0173] In an example embodiment, a computer program product is provided, including a computer program which, when executed by a processor, implements the steps of any of the above method embodiments.

[0174] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0175] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0176] The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0177] Any technical features in the above embodiments can be combined, and for the sake of brevity, not all possible combinations are described above, however, it should be understood that the application encompasses all possible combinations of the technical features unless such a combination is not technically possible.

[0178] The principles and implementation manners of the present application are described herein by using specific examples, and the above embodiments are only used to help understand the method of the present application and its core idea; meanwhile, according to the idea of the present application, the specific implementation manners and application scopes will be changed by those skilled in the art. In conclusion, the content of the present specification should not be understood as a limitation of the present application.

Claims

1. A large-model personalized exercise recommendation method based on knowledge graph, characterized by: include: An improved Bayesian knowledge tracking model based on question difficulty and student type is constructed. The student's historical answer sequence is input and the student's mastery index of the knowledge points is iteratively calculated. Based on the student's mastery index of knowledge points, a personal portrait knowledge graph is constructed containing the relationship between students, knowledge points, and exercises; the relationship includes the student's mastery index of knowledge points and the student's completion degree of exercises; Determine the student's weak knowledge points based on the personal portrait knowledge graph, screen the course exercise database based on the weak knowledge points, and determine a list of weak knowledge point exercises through vector index and rule matching; According to the list of exercises for weak knowledge points and the preset prompt words, a large language model is used to further filter in the course exercise database to obtain the list of exercises for implicitly associated knowledge points; For any exercise in the list of weak knowledge point exercises and the list of implicitly associated knowledge point exercises, a comprehensive score of the exercise is calculated based on the student's past practice, the difficulty of the exercise, and the knowledge points involved in the exercise; The exercises in the weak knowledge point exercise list and the implicitly associated knowledge point exercise list are sorted in descending order according to the comprehensive scores, and the first several exercises are used as the final exercise recommendation list.

2. The large-scale personalized exercise recommendation method based on knowledge graph drive according to claim 1 is characterized in that: In the improved Bayesian knowledge tracing model based on exercise difficulty and student type, the probability of a student guessing when answering exercise t is calculated according to the following formula: in, is the probability of a student guessing when answering exercise t, P(G) k is the probability that a student guesses correctly on knowledge point k and answers the question correctly, ω is the personal profile coefficient, which represents the student’s knowledge mastery level, and D t is the difficulty coefficient of question t; The probability of a student making a mistake when answering exercise t is calculated according to the following formula: in, is the probability of a student making a mistake when answering exercise t, P(S) k is the probability that a student makes a mistake on knowledge point k and answers the question incorrectly.

3. The large-scale personalized exercise recommendation method based on knowledge graph drive according to claim 1 is characterized in that: The course exercise database is a relational database that stores the relationship between courses, exercises, and knowledge points. The attributes of the exercises include exercise ID, question stem, exercise type, and exercise difficulty. Based on the personal portrait knowledge graph, the student's weak knowledge points are determined, and the weak knowledge points are screened in the course exercise database. The weak knowledge point exercise list is determined by vector index and rule matching, which specifically includes: Extracting knowledge points with a mastery index lower than a threshold from the personal portrait knowledge graph using a graph database query language as weak knowledge points; By performing a multi-table joint query in a relational database, the exercise data associated with the weak knowledge point is matched in the course exercise database to obtain an initial associated exercise list; the initial associated exercise list includes a number of exercises, exercise IDs, question stems, question types, and exercise difficulty; Using embedding technology to convert weak knowledge points and each exercise in the initial associated exercise list into a high-dimensional vector, and using a collaborative filtering algorithm to calculate the cosine similarity between the weak knowledge point vector and each exercise vector; According to the cosine similarity between the weak knowledge point vector and each exercise vector, the initial associated exercise list is double-screened according to the cosine similarity threshold and preset rules to obtain a weak knowledge point exercise list.

4. The large-scale personalized exercise recommendation method based on knowledge graph drive according to claim 3 is characterized in that: The text-embedding-ada-002 text embedding model is used to map text of arbitrary length into a 1536-dimensional dense vector space through a deep neural network architecture. The text-embedding-ada-002 text embedding model adopts a multi-layer Transformer structure and dynamically calculates the association weights between text units through a self-attention mechanism. Each Transformer layer contains two sub-modules: multi-head attention and feedforward network, and stable extraction of deep features is achieved through residual connections and layer normalization.

5. The large-scale personalized exercise recommendation method based on knowledge graph drive according to claim 1 is characterized in that: According to the list of exercises on weak knowledge points and preset prompt words, a large language model is used to further screen the course exercise database to obtain the list of exercises on implicitly associated knowledge points. Specifically, the list of exercises on weak knowledge points is input into the large language model, the students' wrong question patterns are analyzed, potential logical chains of knowledge gaps are identified, and based on knowledge reasoning ability, further screening is carried out in the course exercise database to supplement the implicitly associated knowledge point exercises that are not captured by the vector index and rule matching methods to obtain the list of exercises on implicitly associated knowledge points.

6. The large-scale personalized exercise recommendation method based on knowledge graph drive according to claim 1 is characterized in that: The overall score for the exercise is calculated according to the following formula: S final =αS similarity +βS difficulty +γS LLM Among them, S final is the comprehensive score, α is the weight of the vector similarity score, S similarity is the vector similarity score, β is the weight of the difficulty matching score, S difficulty is the difficulty matching score, γ is the weight of the large language model score, S LLM Scoring large language models; The vector similarity score is calculated according to the following formula: S similarity =1-S Where S is the cosine similarity between the exercise vector and the weak knowledge point vector; The difficulty matching score is calculated according to the following formula: S difficulty =1-|d target -d exercise | Among them, d target is the student’s current optimal difficulty, d exercise is the difficulty value of the exercise.

7. A large-scale personalized exercise recommendation system based on knowledge graph, characterized by: include: The mastery index calculation module is used to build an improved Bayesian knowledge tracking model based on the difficulty of the exercises and the type of students. It inputs the student's historical answer sequence and iteratively calculates the student's mastery index of the knowledge points. A personal portrait knowledge graph construction module is used to construct a personal portrait knowledge graph containing the associations between students, knowledge points, and exercises based on the students' mastery index of knowledge points; the associations include the students' mastery index of knowledge points and the students' completion degree of exercises; A weak knowledge point exercise determination module is used to determine the student's weak knowledge points based on the personal portrait knowledge graph, and determine a weak knowledge point exercise list in the course exercise database based on the weak knowledge points; An implicitly associated knowledge point exercise determination module is configured to further filter the list of exercises for the weak knowledge points and preset prompt words using a large language model in the course exercise database to obtain the list of exercises for the implicitly associated knowledge points; an exercise comprehensive score calculation module, for calculating a comprehensive score for any exercise in the weak knowledge point exercise list and the implicitly associated knowledge point exercise list based on the student's past practice, the difficulty of the exercise, and the knowledge points involved in the exercise; The final exercise list recommendation module is used to sort the exercises in the weak knowledge point exercise list and the implicitly associated knowledge point exercise list in descending order according to the comprehensive scores, and use the first several exercises as the final exercise recommendation list.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the large-model personalized exercise recommendation method based on knowledge graph drive as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the knowledge graph-driven large-model personalized exercise recommendation method described in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the knowledge graph-driven large-model personalized exercise recommendation method described in any one of claims 1 to 6.

Images