Cognitive diagnosis method based on automatic construction of q matrix, medium, equipment and product

Abstract

The application provides a cognitive diagnosis method, medium, equipment and product for automatic construction of a Q matrix, relates to the technical field of cognitive diagnosis, and the method comprises the following steps: calculating the correlation score of an exercise and a pre-defined knowledge point, and screening an initial knowledge point set of the exercise by using a large language model; generating a reasoning path for the exercise solution by using the large language model, screening necessary knowledge points on the reasoning path, generating an implicit knowledge point set of the exercise, and constructing a Q matrix by using the initial knowledge point set and the implicit knowledge point set of the exercise; and using the Q matrix for cognitive diagnosis by using a diagnosis model. The application realizes full-automatic construction of the Q matrix in a non-labeled scene, and reduces the dependence of a CDM on data in actual application.

Description

Technical Field

[0001] This invention relates to the field of cognitive diagnostics, and more particularly to cognitive diagnostic methods, media, devices, and products for the automated construction of Q matrices. Background Technology

[0002] With the deep integration of artificial intelligence technology and education, intelligent education has become a key driving force for promoting the modernization and personalization of education. Among numerous intelligent education applications, cognitive diagnosis aims to assess students' knowledge mastery and is the core and prerequisite for personalized learning path recommendations, adaptive practice delivery, and targeted teaching interventions. Accurate diagnostic results not only help students clearly recognize their knowledge weaknesses but also provide teachers and intelligent education systems with a scientific basis for decision-making, thereby significantly improving learning efficiency and teaching quality.

[0003] Cognitive Diagnosis Models (CDMs) are a key task in the field of smart education, aiming to assess students' mastery of knowledge points based on student response records and the association matrix between exercises and knowledge points. The performance of existing CDMs relies on accurately depicting the Q-matrix of the association between questions and knowledge points. By analyzing the Q-matrix, students' mastery of various knowledge concepts within a specific domain can be diagnosed. The construction of the Q-matrix mainly involves two methods: manual annotation and automated construction.

[0004] (1) Incomplete and biased manual annotation by experts: Traditionally, the construction of Q-matrix relies on domain experts to analyze and annotate each exercise. Compared with the simple collection of test-taking records, Q-matrix annotation has a high threshold, requiring experienced teachers to complete, making data acquisition extremely costly and time-consuming. In addition, the expert annotation process is highly dependent on the subjective judgment of experts. Different experts may have different understandings of the knowledge points tested in the same question, resulting in incomplete and subjective biases in the annotation results. This manual annotation method is difficult to obtain an effective Q-matrix when dealing with massive and dynamically updated question banks.

[0005] (2) Automated methods struggle to obtain deep associations: To address the challenges of manual annotation, researchers have begun exploring methods for automating the construction of Q-matrices. Text classification models, such as TextCNN, have been used to extract knowledge points from practice texts and construct their associations with those knowledge points. However, these methods have significant limitations. They primarily capture local patterns and keyword features of the text through convolutional kernels, enabling them to capture shallow semantic associations between exercises and knowledge points. However, they fall short in understanding the deeper semantic logic, reasoning relationships, and implicit, unstated knowledge concepts within the questions. For example, the key to solving a complex mathematical problem may not be a single word in the text, but rather the comprehensive application and abstract reasoning of multiple concepts. Models like TextCNN struggle with this deep semantic understanding task, resulting in Q-matrices that often contain omissions and mislabeling of knowledge points, failing to meet the demands of high-precision diagnosis.

[0006] Existing methods for constructing Q-matrices face a dual challenge: expert annotation is costly and subject to subjective bias, while automated annotation, limited by its representational nature, can only capture superficial semantic relationships. Therefore, how to break free from dependence on expensive and subjective expert resources and develop a new method capable of deeply understanding the meaning of exercises and constructing Q-matrices automatically and with high accuracy has become a critical scientific problem urgently needing to be solved in the field of cognitive diagnostics. Summary of the Invention

[0007] The purpose of this invention is to address the problems of high cost and subjective bias in manual annotation, and the fact that automated annotation can only capture shallow semantic relationships. This invention proposes a cognitive diagnostic method for automated Q-matrix construction, comprising the following steps:

[0008] S1. Scoring based on the relevance between calculation exercises and knowledge points;

[0009] S2. Based on the correlation between exercises and knowledge points, score the initial set of knowledge points for the exercises using a large language model;

[0010] S3. Use a large language model to generate reasoning paths for solving exercises, make necessity judgments on the knowledge points in the reasoning paths, select necessary knowledge points in the reasoning paths, and generate a set of implicit knowledge points for the exercises.

[0011] S4. Merge the initial set of knowledge points and the set of implicit knowledge points of the exercises to obtain the final set of knowledge points, and use the final set of knowledge points to construct the Q matrix;

[0012] S5. Use the Q matrix in the diagnostic model for cognitive diagnosis.

[0013] Furthermore, the correlation score between the exercises and predefined knowledge points is calculated, specifically as follows:

[0014] The exercise text and predefined knowledge points are encoded into vectors. The vector similarity between the exercise and the knowledge points is calculated. Based on the vector similarity, N knowledge points are selected as a candidate knowledge point set.

[0015] Input the set of exercises and candidate knowledge points into the large language model to obtain the relevance score of each candidate knowledge point;

[0016] The relevance score and vector similarity are weighted and fused to calculate the relevance score of each knowledge point.

[0017] Furthermore, cosine similarity is used to calculate the vector similarity between exercises and knowledge points, using the following formula:

[0018]

[0019] in, express and Vector similarity, Indicates the first A set of exercise texts, Indicates the first One knowledge point, express The vector representation of , express The vector representation of , Represents the magnitude of a vector.

[0020] Furthermore, the initial set of knowledge points for the exercises is selected using a large language model, specifically as follows:

[0021] Select the K knowledge points with the highest relevance scores, use a large language model to evaluate the relevance scores s between the K knowledge points and the exercises, and retain the knowledge points with relevance scores s greater than a set threshold as the initial set of knowledge points.

[0022] Furthermore, the relevance score and vector similarity are weighted and fused to calculate the relevance score between the exercise and the candidate knowledge point, expressed by the formula:

[0023]

[0024] in, express and The correlation score Indicates the first A set of exercise texts, Indicates the first One knowledge point, To integrate the weighting coefficients, Indicates that LLM is obtained and Relevance score express and Vector similarity.

[0025] Furthermore, the method for selecting necessary knowledge points along the reasoning path is as follows: if knowledge points are removed... Under the counterfactual setting, if a certain reasoning path cannot be solved, then knowledge point k is a necessary knowledge point, where knowledge point k is a certain knowledge point on the reasoning path.

[0026] Furthermore, the Q-matrix is ​​constructed using the final set of knowledge points, and is represented as follows:

[0027]

[0028]

[0029] in, Represents the first in the Q matrix Line number Column elements, Indicates the first j One knowledge point, express The final collection of knowledge points, and Representing the exercises The initial set of knowledge points and the set of implicit knowledge points. Indicates the first Exercise text.

[0030] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described cognitive diagnostic method for automated Q-matrix construction.

[0031] The present invention also proposes an electronic device, including a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including computer-readable instructions, and the processor is configured to invoke the computer-readable instructions to execute the above-described cognitive diagnostic method for automated Q-matrix construction.

[0032] The present invention also proposes a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the cognitive diagnostic method for automated Q-matrix construction described above.

[0033] The beneficial effects of the technical solution provided by this invention are:

[0034] This invention utilizes the deep semantic understanding capabilities of Large Language Models (LLMs) to identify explicit knowledge concepts in exercises, and leverages the logical reasoning capabilities of LLMs to uncover implicit knowledge concepts. Compared to traditional text feature extraction methods, LLMs can parse the deep semantics of questions and generate knowledge associations that highly align with expert cognition and are superior. This invention achieves fully automated construction of the Q-matrix in unlabeled scenarios, significantly reducing the dependence of CDMs on data in practical applications. Attached Figure Description

[0035] Figure 1 This is a flowchart of a cognitive diagnostic method for automated Q-matrix construction according to an embodiment of the present invention;

[0036] Figure 2 This is a block diagram of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0038] Example 1:

[0039] To calculate the correlation score between exercises and knowledge points, and considering the high time complexity of using LLM when calculating the similarity between exercises and knowledge points, a vector recall mechanism is introduced to improve efficiency, selecting a set of candidate knowledge points for each exercise. For each exercise and all predefined knowledge points for each exercise, the embedding model BGE-m3 / text-embedding-ada-002 is used to transform the natural language text portion into a high-dimensional, dense vector representation. Specifically, the exercise text... Encoded as exercise vectors Each knowledge point Converted into knowledge point vectors These semantic vectors effectively capture the semantics of knowledge points and exercises, providing fundamental support for subsequent similarity calculation and matching tasks. Cosine similarity is used to calculate the vector similarity between exercises and knowledge points; the formula is:

[0040]

[0041] in, express and Vector similarity, Indicates the first A set of exercise texts, Indicates the first One knowledge point, express The vector representation of , express The vector representation of , Represents the magnitude of a vector.

[0042] By sorting vector similarities from largest to smallest, and selecting the predefined knowledge points corresponding to the top N vector similarities as the candidate knowledge point set for each exercise, the processing scope of LLM is effectively narrowed.

[0043] Vector similarity can only capture the surface-level semantic relationships between exercises and knowledge points, and cannot perform complex reasoning. Therefore, a large-scale language model (LLM) is introduced for deep semantic analysis. A structured prompt is constructed, with the following input structure:

[0044] (1) Exercise content: The complete text including the question stem and options;

[0045] (2) Candidate knowledge point list: obtained from the initial recall stage, each knowledge point is accompanied by a number and name.

[0046] User prompts in this embodiment of the invention can be:

[0047] Please analyze the relevance of the following exercises to the candidate knowledge points:

[0048] Exercise content: {Exercise text};

[0049] Candidate knowledge point list: {List of knowledge point numbers and names}.

[0050] Meanwhile, the system prompts the model role to be set as "education expert in a specific field," limiting the judgment scope to the candidate set, thereby improving professionalism and focus.

[0051] The model output adopts the following structured format:

[0052] Knowledge Point <Knowledge Point ID> (<Knowledge Point Name>)

[0053] Relevance:

[0054] Evidence: <Key text extracted from exercises>

[0055] Explanation: <Briefly explain the connection between this knowledge point and the exercises>

[0056] To enhance model consistency and stability, the Prompt provides Few Shot examples to help the model understand the target format. The Prompt also includes clear behavioral guidelines, requiring objective scoring, traceable evidence, concise explanations, and prohibiting the output of irrelevant content, significantly reducing the risk of hallucinations. In summary, this module integrates role setting, task constraints, structural reduction, example guidance, and behavioral constraints, fully leveraging the semantic understanding capabilities of LLM to effectively generate high-quality relevance scores. This provides strong support for subsequent knowledge point prediction.

[0057] The exercises and candidate knowledge points are input into a large language model to obtain a relevance score for each candidate knowledge point. To comprehensively and robustly evaluate the correlation between exercises and knowledge points, the relevance scores obtained from the LLM are... Similarity to vectors Weighted fusion is performed to calculate the correlation score between the exercises and the candidate knowledge points. The formula is:

[0058]

[0059] in, For fusion weighting coefficients.

[0060] By setting a relevance scoring threshold or selecting the K knowledge points with the highest relevance scores, we can obtain the set of knowledge points most relevant to the exercises. .

[0061] Based on the set of knowledge points most relevant to the exercises Then, the initial set of knowledge points for the exercises is selected again using an inference-based large language model. The inference-based LLM applies this to the set of exercises and knowledge points. The correlation between each knowledge point is quantitatively evaluated to obtain an evaluation score, and only the set is retained. The initial set of knowledge points is obtained by identifying knowledge points with a score of 3 or higher in the assessment. .

[0062] Using the initial set of knowledge points Construct the Q matrix, represented as:

[0063]

[0064] in, Represents the first in the Q matrix Line number Column elements, Indicates the first j One knowledge point, Representing exercises The final collection of knowledge points, Representing exercises The initial set of knowledge points, Indicates the first Exercise text.

[0065] The above method combines vector similarity and LLM semantic relevance for dual screening, guiding LLM to perform in-depth analysis of practice texts, predict knowledge concepts associated with the texts, and automatically generate a Q-matrix without annotation. The Q-matrix automatically generated by the above method is then used in a diagnostic model for cognitive diagnosis.

[0066] Example 2:

[0067] Example 2, based on Example 1, optimizes the Q-matrix generated in Example 1, uncovers and supplements any missing implicit knowledge points, and enhances its intelligent capabilities to further improve diagnostic accuracy. Details are as follows:

[0068] The flowchart of the cognitive diagnosis method for automated Q-matrix construction according to an embodiment of the present invention is as follows: Figure 1 Specifically, it includes the following steps:

[0069] S1. Calculate the correlation score between exercises and knowledge points. Considering the high time complexity of using LLM when calculating the similarity between exercises and knowledge points, a vector recall mechanism is introduced to improve efficiency, filtering out the candidate knowledge point set for each exercise. For each exercise and all predefined knowledge points for each exercise, the embedding model BGE-m3 / text-embedding-ada-002 is used to transform its natural language text portion into a high-dimensional, dense vector representation. Specifically, the exercise text... Encoded as exercise vectors Each knowledge point Converted into knowledge point vectors These semantic vectors effectively capture the semantics of knowledge points and exercises, providing fundamental support for subsequent similarity calculation and matching tasks. Cosine similarity is used to calculate the vector similarity between exercises and knowledge points; the formula is:

[0070]

[0071] in, express and Vector similarity, Indicates the first A set of exercise texts, Indicates the first One knowledge point, express The vector representation of , express The vector representation of , Represents the magnitude of a vector.

[0072] By sorting vector similarities from largest to smallest, and selecting the predefined knowledge points corresponding to the top N vector similarities as the candidate knowledge point set for each exercise, the processing scope of LLM is effectively narrowed.

[0073] Vector similarity can only capture the surface-level semantic relationships between exercises and knowledge points, and cannot perform complex reasoning. Therefore, a large-scale language model (LLM) is introduced for deep semantic analysis. A structured prompt is constructed, with the following input structure:

[0074] (1) Exercise content: The complete text including the question stem and options;

[0075] (2) Candidate knowledge point list: obtained from the initial recall stage, each knowledge point is accompanied by a number and name.

[0076] User prompts in this embodiment of the invention can be:

[0077] Please analyze the relevance of the following exercises to the candidate knowledge points:

[0078] Exercise content: {Exercise text};

[0079] Candidate knowledge point list: {List of knowledge point numbers and names}.

[0080] Meanwhile, the system prompts the model role to be set as "education expert in a specific field," limiting the judgment scope to the candidate set, thereby improving professionalism and focus.

[0081] The model output adopts the following structured format:

[0082] Knowledge Point <Knowledge Point ID> (<Knowledge Point Name>)

[0083] Relevance:

[0084] Evidence: <Key text extracted from exercises>

[0085] Explanation: <Briefly explain the connection between this knowledge point and the exercises>

[0086] To enhance model consistency and stability, the Prompt provides Few Shot examples to help the model understand the target format. The Prompt also includes clear behavioral guidelines, requiring objective scoring, traceable evidence, concise explanations, and prohibiting the output of irrelevant content, significantly reducing the risk of hallucinations. In summary, this module integrates role setting, task constraints, structural reduction, example guidance, and behavioral constraints, fully leveraging the semantic understanding capabilities of LLM to effectively generate high-quality relevance scores. This provides strong support for subsequent knowledge point prediction.

[0087] The exercises and candidate knowledge points are input into a large language model to obtain a relevance score for each candidate knowledge point. To comprehensively and robustly evaluate the correlation between exercises and knowledge points, the LLM relevance scores are... Similarity to vectors Weighted fusion is performed to calculate the correlation score between the exercises and the candidate knowledge points. The formula is:

[0088]

[0089] in, To integrate the weighting coefficients, calculations are performed on 1917 knowledge points in the embodiments of the invention. α The value is 0.4, representing the similarity between a specific exercise and a specific knowledge point obtained through LLM. Vector similarity A correlation score was obtained. Different fusion weight coefficients result in different output Q matrices, leading to varying performance of the diagnostic model. Therefore, selecting the appropriate coefficient based on the model's performance is crucial. value.

[0090] S2. Based on the correlation score between exercises and knowledge points, an initial set of knowledge points for the exercises is selected using a large language model. Specifically, by setting a correlation score threshold or selecting the K knowledge points with the highest correlation scores, the set of knowledge points most relevant to the exercises is obtained. .

[0091] Based on the set of knowledge points most relevant to the exercises Then, the initial set of knowledge points for the exercises is selected again using an inference-based large language model. The inference-based LLM applies this to the set of exercises and knowledge points. The correlation between each knowledge point is quantitatively evaluated to obtain an evaluation score, and only the set is retained. The initial set of knowledge points is obtained by identifying knowledge points with a score of 3 or higher in the assessment. .

[0092] S3. Input the exercises into the large language model, and use the large language model to generate reasoning paths for the solutions. Employ structured chain-of-thought (CoT) hints, requiring the model to "reason step-by-step and demonstrate the complete process," resulting in step-by-step reasoning text. Perform semantic parsing and step segmentation on the step-by-step reasoning text, representing each step as a triple:

[0093]

[0094] in, This indicates the j-th step. This represents the solution method for the j-th step. This represents the knowledge point reference relationship in the j-th step. Let represent the axiom of the j-th step.

[0095] A reasoning path Defined as an ordered sequence of steps:

[0096]

[0097] This invention uses an example to illustrate the exercises. One line of reasoning:

[0098] exercise :prove ;

[0099] Reasoning path: → → → → .

[0100] The triplet elements of the steps in this reasoning path are represented using Table 1:

[0101] Table 1

[0102]

[0103] exercise The set consists of all possible reasoning paths:

[0104]

[0105] calculate The similarity between each step and predefined knowledge points ,in, express and Vector similarity function, For text embedding functions, Indicating the reasoning path The j-th step in This is a predefined knowledge point. The exercises are selected from the predefined knowledge points whose similarity scores exceed a set similarity threshold. Reasoning path The collection of knowledge points that are called or applied in the middle .

[0106] For exercises Reasoning path The collection of knowledge points in If the set is removed Under the counterfactual setting of a knowledge point k, the reasoning path If the solution cannot be completed or requires a fundamental rewrite, then it is considered that knowledge point k is relevant to the exercise. It is necessary; knowledge point k is a necessary knowledge point. Here, based on the large language model, the necessity of knowledge points in the reasoning path of the exercises is judged to obtain the exercises. The set of implicit knowledge points .

[0107] S4. Merge the initial set of knowledge points and the implicit set of knowledge points for the exercises to obtain the final set of knowledge points. Construct the Q matrix using the final set of knowledge points, represented as:

[0108]

[0109]

[0110] in, express Q The first in the matrix Line number Column elements, Indicates the first j One knowledge point, Representing exercises The final collection of knowledge points, and Representing the exercises The initial set of knowledge points and the set of implicit knowledge points. Indicates the first Exercise text.

[0111] Ultimately, the zero-sample generation function of the Q matrix was achieved, that is, the automatic generation of a high-quality Q matrix without any manual annotation. Representing exercises Related knowledge points This method eliminates the reliance on subjective expert resources, greatly improving the efficiency and accuracy of Q-matrix construction and providing a solid foundation for subsequent cognitive diagnostic tasks.

[0112] S5. Use the Q matrix in the diagnostic model for cognitive diagnosis.

[0113] The generated Q-matrix is ​​combined with existing cognitive diagnostic models, using the NCD model as an example. The formula is as follows:

[0114]

[0115] in, It is an element-wise product. This indicates a three-layer fully connected layer. This is the association vector of exercise knowledge points extracted from the Q matrix. For student ability vectors, Let the difficulty vector of the exercises be denoted as . This is the discrimination scalar for the exercises.

[0116] in, , and In line with the calculation method of traditional cognitive diagnostic models, this method applies to the knowledge point association vectors in exercises. This invention no longer relies on a pre-defined Q-matrix based on human experience, but instead directly utilizes an automatically constructed Q-matrix. (To distinguish the Q-matrix generated by this invention from the Q-matrix preset based on human experience, the Q-matrix generated by this invention is represented as follows) )therefore, It can be represented as:

[0117]

[0118] in , It's a set of exercises. vector, It refers to the number of knowledge points. It refers to the number of exercises.

[0119] The cognitive diagnosis method for automatically constructing the Q-matrix in this invention changes the inherent paradigm of cognitive diagnosis. The Q-matrix is ​​no longer an artificial construct that needs to be provided in advance, but rather an intermediate component that can be automatically generated on demand. This invention designs an innovative prompt engineering method and combines vector similarity and LLM semantic relevance for dual screening, guiding LLM to perform deep analysis of practice texts and accurately predict their associated knowledge concepts. Compared with shallow models such as TextCNN, LLM can better understand the context of the questions, distinguish easily confused concepts, and identify implicit knowledge points that require comprehensive application to solve the problems, thereby producing Q-matrixes of far superior quality to traditional automated methods. In addition, the method of this invention has high flexibility. It can generate a completely new Q-matrix from scratch for datasets that have no Q-matrix at all, or it can intelligently enhance existing expert-annotated Q-matrixes by mining and supplementing them with missing implicit knowledge points, further improving diagnostic accuracy.

[0120] To verify the effectiveness of the method of the present invention, experiments were conducted on two cognitive diagnostic datasets, MOOPer and MOOCRadar. The experiment was set up to use 80% of the samples in the datasets as training data and 20% as the test set for evaluation.

[0121] The experimental results are shown in Table 2. Experimental validation was conducted based on two mainstream cognitive diagnostic methods: NCDM (Interpretable Cognitive Diagnosis with Neural Network) and RCD (Relation Map Driven Cognitive Diagnosis for Intelligent Education Systems). NCDM-F represents the diagnostic result using the NCDM basic diagnostic model, based on the Q-matrix constructed from the initial knowledge point set. NCDM-Q represents the diagnostic result using the NCDM basic diagnostic model, based on the final Q-matrix constructed by fusing the initial and implicit knowledge point sets. Evaluation metrics were: AUC (probability that the predicted value of a positive example is greater than the predicted value of a negative example), ACC (accuracy), and RMSE (root mean square error). Higher ACC and AUC values ​​were better, while lower RMSE values ​​were better.

[0122] Table 2

[0123]

[0124] In one exemplary embodiment, a computer-readable storage medium is included, which stores a computer program that, when executed by a processor, implements the cognitive diagnostic method for automated Q-matrix construction described above.

[0125] Please see Figure 2 In one exemplary embodiment, the device further includes an electronic device including at least one processor, at least one memory, and at least one communication bus.

[0126] The memory stores a computer program, which includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the aforementioned cognitive diagnostic method for automated Q-matrix construction.

[0127] In one exemplary embodiment, a computer program product is proposed, including a computer program / instructions that, when executed by a processor, implement the steps of the cognitive diagnostic method for automated Q-matrix construction described above.

[0128] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cognitive diagnosis method of Q matrix automatic construction, characterized in that, The method comprises the following steps: S1, calculating the relevance score of the exercise and the knowledge point; specifically: The exercise text and the pre-defined knowledge points are coded into vectors, the vector similarity of the exercise and the pre-defined knowledge points is calculated, and N knowledge points are selected from the pre-defined knowledge points as a candidate knowledge point set based on the vector similarity; The exercise and the candidate knowledge point set are input into a large language model to obtain a relevance score of the exercise and the candidate knowledge points; The relevance score and the vector similarity are weighted and fused to calculate the relevance score of the exercise and the candidate knowledge points; S2, based on the relevance score of the exercise and the knowledge point, the initial knowledge point set of the exercise is screened by using a large language model; Specifically: Select the top K knowledge points with the highest relevance score, evaluate the relevance score s of the K knowledge points and the exercise by using a large language model, and retain the knowledge points with a relevance score s greater than a set threshold as the initial knowledge point set; S3, using a large language model to generate a reasoning path for the exercise answer, judging the necessity of the knowledge points in the exercise reasoning path, screening the necessary knowledge points on the reasoning path, and generating a set of implicit knowledge points of the exercise; the method for screening the necessary knowledge points on the reasoning path is: if the solution of a certain reasoning path cannot be completed under the counterfactual setting of removing knowledge point k , then the knowledge point k is a necessary knowledge point, wherein the knowledge point k is a certain knowledge point on the reasoning path; S4, the initial knowledge point set of the exercise and the implicit knowledge point set are combined to obtain a final knowledge point set, and the final knowledge point set is used to construct a Q matrix; S5, the Q matrix is used for cognitive diagnosis in a diagnosis model.

2. The cognitive diagnosis method of claim 1, wherein, The vector similarity of the exercise and the knowledge point is calculated by using cosine similarity, and the formula is: wherein, denotes and a vector similarity, denotes the th exercise text, denotes the th knowledge point, denotes a vector representation of , denotes a vector representation of , denotes a vector norm.

3. The cognitive diagnosis method of claim 1, wherein, The relevance score and the vector similarity are weighted and fused to calculate the relevance score of the exercise and the candidate knowledge points, which is represented by the formula: wherein, represents and the relevance score of the association, represents the th exercise text, represents the th knowledge point, is a fusion weight coefficient, represents the relevance score of and obtained by the LLM, represents and the vector similarity.

4. The cognitive diagnosis method of claim 1, wherein, The final knowledge point set is used to construct a Q matrix, which is represented by: in, Represents the first in the Q matrix Line number Column elements, Indicates the first j One knowledge point, express The final collection of knowledge points, and Representing the exercises The initial set of knowledge points and the set of implicit knowledge points. Indicates the first Exercise text.

5. A computer readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to realize the method of any one of claims 1-4.

6. An electronic device, comprising: The computer program comprises computer readable instructions, and the processor is configured to call the computer readable instructions to execute the method of any one of claims 1-4.

7. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of any one of claims 1-4.

Images