A knowledge text-based cognitive diagnosis model and a cognitive diagnosis method thereof

Abstract

This invention relates to the field of artificial intelligence-based educational assessment, and discloses a cognitive diagnostic model and method based on knowledge text. The model includes a text feature extraction module for extracting and generating knowledge point association vectors representing the degree of association between questions and preset knowledge points; a knowledge proficiency embedding module for mapping knowledge proficiency vectors; a question attribute embedding module for mapping question knowledge difficulty vectors and question discrimination scalars; a knowledge association attention module for adjusting attention weights and generating weighted knowledge representations; and an adaptive nonlinear prediction module containing a multi-layer KANLinear structure composed of B-spline basis functions and outputting the prediction probability of correct student answers. This invention solves the problem of fragmented internal module functions in existing models and features high interpretability and independence from manual annotation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education assessment, and more specifically, to a cognitive diagnostic model and method based on knowledge text. Background Technology

[0002] Cognitive Diagnosis Models (CDMs) are a key type of model in the field of educational assessment, aiming to accurately infer students' potential mastery of various knowledge points by analyzing their response behavior. These models are the core technological foundation for personalized instruction, intelligent learning path recommendations, and adaptive assessment systems. However, existing cognitive diagnostic methods, whether traditional statistical models or modern deep learning models, generally suffer from the following deep-seated limitations:

[0003] First, existing models heavily rely on manual annotation, resulting in low automation and poor scalability. Most cognitive diagnostic models require a predefined "Q-matrix" (i.e., a question-knowledge point association matrix) as input. The construction of this matrix typically depends on manual annotation by educational experts. This process is not only labor-intensive and costly but also highly subjective, with annotation results potentially differing between experts. More seriously, when the question bank is updated or new questions are introduced, manual annotation must be performed again, making it difficult for the model to quickly adapt to dynamically changing teaching content and severely limiting its scalability.

[0004] Secondly, existing models suffer from fragmented internal mechanisms and lack the ability to collaboratively model cognitive processes. While some deep learning methods introduce attention mechanisms to capture connections between knowledge points or consider attributes such as question discrimination and difficulty, these factors are typically treated as independent, static modules that are simply combined. For example, the knowledge point connections uncovered by attention mechanisms are fixed and cannot be dynamically adjusted based on the characteristics of the questions themselves (such as high-discrimination questions versus low-discrimination questions). This "modular splicing" design philosophy ignores the complex and interdependent collaborative relationships among various elements in students' cognitive processes (knowledge mastery, question difficulty, and question discrimination ability), resulting in insufficient model fitting accuracy to real cognitive processes and limited generalization ability.

[0005] Finally, existing models generally suffer from the "black box" problem, offering superficial interpretability and failing to guide teaching practice. While traditional neural network models can improve prediction accuracy, their internal decision-making processes are opaque, making it impossible to explain why a particular prediction is made. Even when some models use attention weights to provide so-called "interpretability" (e.g., showing which knowledge points the model focuses on), this explanation remains superficial. It fails to reveal how students' knowledge proficiency interacts non-linearly with question attributes to ultimately influence their answers. This "knowing what, but not why" situation makes it difficult to translate model results into concrete and feasible teaching intervention strategies, significantly diminishing their practical value.

[0006] Therefore, there is an urgent need in this field for a novel cognitive diagnostic technology solution that can break free from the reliance on manual annotation and achieve end-to-end diagnosis starting from the original question text; furthermore, its internal mechanism should be able to simulate the synergy of the cognitive process, achieving deep coupling and dynamic adjustment of each module; finally, it should also have profound interpretability, making the decision-making process inside the model transparent, so that it becomes a truly "white box" diagnostic tool that can guide teaching practice. Summary of the Invention

[0007] To address the problem of fragmented functionality among internal modules in existing models, this invention provides a cognitive diagnostic model and method based on knowledge text, which features independence from manual annotation and strong interpretability.

[0008] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:

[0009] A cognitive diagnostic model based on knowledge text includes a text feature extraction module, a knowledge proficiency embedding module, a question attribute embedding module, a knowledge association attention module, and an adaptive nonlinear prediction module.

[0010] The text feature extraction module is used to receive student answer data and corresponding question data, and through a text processing network, extract and generate knowledge point association vectors that correlate students' knowledge status with knowledge points.

[0011] The knowledge proficiency embedding module is used to map student data into a knowledge proficiency vector;

[0012] The question attribute embedding module is used to map question data into a question knowledge difficulty vector and a question discrimination scalar.

[0013] The knowledge association attention module takes the knowledge point association vector as its input and adjusts the attention weights to generate a weighted knowledge representation that highlights the dependencies of core knowledge points in the current question context.

[0014] The adaptive nonlinear prediction module contains a multi-layer KANLinear structure composed of B-spline basis functions. It receives the student's knowledge proficiency vector, the question's knowledge difficulty vector, the question's discrimination scalar, and the interactive features composed of the weighted knowledge representation. It adaptively adjusts the grid granularity of its internal B-spline basis functions according to the question's discrimination scalar, performs nonlinear mapping on the interactive features, and finally outputs the predicted probability of the student's correct answer.

[0015] Preferably, the text feature extraction module includes a pre-trained language model based on the Transformer architecture, which is used to deeply understand the semantic information of the question text and generate the knowledge point association vector.

[0016] Furthermore, the knowledge association attention module adjusts the attention weights by using a multi-head self-attention mechanism on the knowledge point association vector to model the potential dependencies between knowledge points.

[0017] Furthermore, the KANLinear structure consists of three layers. Its inputs are a student knowledge proficiency vector, a question knowledge difficulty vector, a question discrimination scalar, and the student-question interaction features composed of the weighted knowledge representation. The model is performed through KANLinear modules layer by layer, and the probability of answering correctly is finally output by the Sigmoid function.

[0018] Furthermore, the adaptive adjustment of the B-spline basis function grid granularity is achieved by mapping the value of the question discrimination scalar to the grid division density through a preset or learnable mapping function, thereby realizing nonlinear fitting for questions with high discrimination.

[0019] Furthermore, the discriminative scalar of the question is normalized by a sigmoid function and multiplied by a learnable scaling factor before being input into other parts of the model, in order to obtain a range of values ​​that are most suitable for model tuning.

[0020] Furthermore, the last layer of the adaptive nonlinear prediction module uses a Sigmoid activation function to constrain the output value to the interval [0,1], which serves as the final probability of a correct answer.

[0021] A cognitive diagnosis method based on a knowledge-text-based cognitive diagnosis model includes the following specific steps:

[0022] Input the data of students who answered the questions and the corresponding question data;

[0023] The student answer data and corresponding question data are input into the question text feature extraction module. Through the text processing network, the knowledge point association vectors that are related to the students' knowledge status and knowledge points are extracted and generated.

[0024] The knowledge proficiency embedding module maps student data into a knowledge proficiency vector; the question attribute embedding module maps question data into a question knowledge difficulty vector and a question discrimination scalar.

[0025] The knowledge point association vector is input into the knowledge association attention module to adjust the attention weight, thereby generating a weighted knowledge representation that can highlight the core knowledge point dependencies in the current question context;

[0026] The knowledge proficiency vector, the question knowledge difficulty vector, the question discrimination scalar, and the weighted knowledge representation are combined into interactive features and input into the adaptive nonlinear prediction module.

[0027] In the adaptive nonlinear prediction module, the grid granularity of its internal B-spline basis function is adaptively adjusted according to the question discrimination scalar, and the interaction features are nonlinearly mapped to finally output the predicted probability of the student answering correctly.

[0028] Preferably, it also includes an interpretable diagnostic step, which uses a visualization program to present the shape of the activated B-spline curve in the adaptive nonlinear prediction module, in order to analyze and reveal the nonlinear influence pattern between a particular student's knowledge mastery and the attributes of the question when answering a particular question.

[0029] Furthermore, before cognitive diagnosis, there is a model training step, which uses the negative log-likelihood loss function as the objective function and iteratively optimizes all trainable parameters in the model using the Adam optimizer.

[0030] The beneficial effects of this invention are as follows:

[0031] This invention discloses a cognitive diagnostic model based on knowledge text. Through a built-in text feature extraction module, it can automatically learn the association between questions and knowledge points directly from question and student data, completely eliminating the reliance on manually labeled Q-matrices. This allows the model to seamlessly integrate with various dynamically updated question banks, greatly improving the system's automation level and scalability, and clearing obstacles for large-scale, low-cost practical applications. This invention uniquely uses "question discrimination" as the core control parameter throughout the model, enabling it to dynamically adjust the intensity of knowledge association mining by the upstream attention module and adaptively adjust the fineness of nonlinear fitting by the downstream KAN network. This deeply coupled "one parameter for two uses" design allows the model to flexibly change its internal working mode according to different question characteristics, achieving unprecedented fine-grained modeling of the student-question interaction process, thereby significantly improving prediction accuracy. This invention not only predicts whether students answer correctly but also presents the complex nonlinear decision-making process within the model in an intuitive B-spline function curve form through its core "interpretable diagnostic steps." This "white box" diagnostic capability can clearly reveal to teachers and students the specific strengths and weaknesses of their cognitive abilities (for example, at what critical point knowledge acquisition abruptly changes), providing unprecedented and reliable decision support for developing precise and efficient personalized learning plans, and has extremely high practical application value. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the module flow of the cognitive diagnostic model based on knowledge text of the present invention.

[0033] Figure 2 This is a schematic diagram of the computational process of the knowledge association attention module in the cognitive diagnostic model based on knowledge text of this invention.

[0034] Figure 3 This is a schematic diagram of the specific process of the cognitive diagnosis method of the present invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0036] Example 1

[0037] like Figure 1 As shown, a cognitive diagnostic model based on knowledge text includes a text feature extraction module, a knowledge proficiency embedding module, a question attribute embedding module, a knowledge association attention module, and an adaptive nonlinear prediction module.

[0038] The text feature extraction module is used to receive student answer data and corresponding question data, and through a text processing network, extract and generate knowledge point association vectors that correlate students' knowledge status with knowledge points.

[0039] In this embodiment, the text feature extraction module incorporates a pre-trained language model (such as BERT) based on the Transformer architecture and has been fine-tuned for educational text corpora. This module performs deep semantic analysis on the question text, automatically identifying the knowledge concepts contained within the text, and ultimately outputs a "knowledge point association vector" with a dimension equal to the preset total number of knowledge points. Each dimension of this vector represents the association strength between question B and the corresponding knowledge point. This step completely replaces the time-consuming and labor-intensive process of manually annotating the Q-matrix in traditional methods.

[0040] The knowledge proficiency embedding module is used to map student data into a knowledge proficiency vector;

[0041] In this embodiment, the knowledge proficiency embedding module models all students using an embedding matrix of size student_n, where the dimension of each embedding vector is the same as the total number of knowledge points (i.e., knowledge_dim). The embedding matrix maps indices to vectors through the nn.Embedding layer and uses the Sigmoid function to normalize the output to the [0,1] interval to represent the students' mastery of each knowledge point.

[0042] The question attribute embedding module is used to map question data into a question knowledge difficulty vector and a question discrimination scalar.

[0043] In this embodiment, the question attribute embedding module uses an embedding matrix of size exer_n to represent the knowledge difficulty information of each question. The dimension of this embedding vector is consistent with the number of knowledge points. After normalization by the Sigmoid function, a difficulty value falling within the [0,1] interval is obtained, which is used to characterize the relative difficulty of the question on each knowledge point. The question attribute embedding module also introduces a discrimination scalar of length 1 for each question, which is used to measure the ability of the question to distinguish between students with different levels of mastery. The larger the value, the more discriminative the question is. After this parameter is normalized to the [0,1] interval by the Sigmoid function, it is multiplied by a scaling factor (such as multiplying by 10) to obtain a suitable value range, so that its influence weight is automatically adjusted during model training to adapt to the characteristics of different questions.

[0044] In this embodiment, the discrimination is achieved by using a linear transformation layer of nn.Linear(1, 1) and combined with the Softmax function to generate weight coefficients. The discrimination of the questions is dynamically weighted in the student-question interactive calculation, thereby realizing personalized discrimination control for different questions.

[0045] The knowledge association attention module takes the knowledge point association vector as its input and adjusts the attention weights to generate a weighted knowledge representation that highlights the dependencies of core knowledge points in the current question context.

[0046] In this embodiment, as Figure 2 As shown, the associative attention module adopts the nn.MultiheadAttention structure. The diagram illustrates how the input "knowledge point association vector" is mapped to a query (Q), key (K), and value (V) matrix. Attention weights are calculated using scaling dot product operations and the Softmax function. These weights are then applied to the value (V) matrix to generate a weighted knowledge representation that incorporates the dependencies between knowledge points. The attention calculation process is dynamically adjusted by the acquired question discrimination scalar. Specifically, when calculating the attention score, this discrimination scalar acts as a moderating factor, multiplied into or integrated into the attention score calculation formula in a weighted fusion manner. For example, for questions with high discrimination, this moderating factor amplifies the weight of their core knowledge points in the attention calculation, enabling the model to generate a weighted knowledge representation that better highlights the dependencies between core knowledge points in the current question context.

[0047] The adaptive nonlinear prediction module contains a multi-layer KANLinear structure composed of B-spline basis functions. It receives the student's knowledge proficiency vector, the question's knowledge difficulty vector, the question's discrimination scalar, and the interactive features composed of the weighted knowledge representation. It adaptively adjusts the grid granularity of its internal B-spline basis functions according to the question's discrimination scalar, performs nonlinear mapping on the interactive features, and finally outputs the predicted probability of the student's correct answer.

[0048] In this embodiment, the nonlinear fitting capability of the adaptive nonlinear prediction module is adaptive. The grid partitioning of the B-spline basis functions within the module is not static, but rather correlated with the "question discrimination scalar" through a preset mapping function. When processing questions with high discrimination, the module automatically uses a finer, denser grid to fit the interaction features to capture more subtle differences in students' abilities; conversely, when processing questions with low discrimination, a coarser grid is used.

[0049] After being processed layer by layer by the KANLinear module with adaptive fitting capabilities, the final output is passed through a Sigmoid activation function to obtain a probability value between [0,1], which is the probability that student A answers question B correctly as predicted by the model.

[0050] Example 2

[0051] In this embodiment, student answer log data is obtained, including student ID, question ID, and original text data;

[0052] This embodiment uses two datasets to train and validate the knowledge-text-based cognitive diagnostic model: a self-built real-world dataset AIFC and a publicly available educational dataset ASSIST. The AIFC data comes from student answer records collected during actual teaching in the "Fundamentals of Artificial Intelligence" course, containing logs of 86 students' answers to 150 multiple-choice questions covering 16 core knowledge points, totaling 40,800 samples. The ASSIST data is publicly available student answer sequence data, widely used in cognitive modeling tasks. Before training, all data were divided into training, validation, and test sets according to student dimensions, with proportions of 70%, 10%, and 20%, respectively, to ensure the independence of sample distribution and the objectivity of evaluation results.

[0053] In the feature construction phase, the model performs embedding modeling for students, questions, and knowledge points respectively. The student embedding represents their proficiency vector for each knowledge point, and the question embedding includes the difficulty and discrimination of the knowledge point. Among them, the question discrimination is linearly mapped and then input into the Softmax function to generate attention weights, which are used to dynamically adjust the contribution of each question in the student-question interaction, thereby achieving a more granular prediction mechanism.

[0054] The knowledge-related multi-head attention module uses a multi-head self-attention mechanism on the vectors of knowledge points associated with the question to model the potential dependencies between knowledge points, thereby enhancing the model's ability to express knowledge structures.

[0055] The final interaction characteristics between the student and the question are constituted by the following expression:

[0056] input_x = e_discrimination × tanh(stu_emb − k_difficulty) ×attended_knowledge

[0057] Wherein, stu_emb represents the student's knowledge proficiency vector, k_difficulty represents the difficulty vector of the knowledge points covered by the questions, e_discrimination is the discrimination weight, and attended_knowledge is the knowledge relevance after attention mechanism processing. This expression comprehensively considers the student's mastery, question difficulty, discrimination ability, and knowledge structure information, constituting the core interactive feature representation of the model of this invention.

[0058] The KANLinear multi-layer nonlinear modeling structure inputs the aforementioned interactive features, `input_x`, into multiple KANLinear layers, extracting higher-order nonlinear features layer by layer. Each KANLinear layer is constructed based on the B-spline interpolation function to achieve a smooth mapping from the input space to the output space, exhibiting stronger nonlinear modeling capabilities and interpretability compared to traditional fully connected layers. While maintaining numerical stability and a controllable range of expression, KANLinear can effectively fit the complex cognitive relationship between students and knowledge points, and progressively extracts more discriminative feature representations within the network layers. Finally, it outputs the probability of a student answering a question correctly through the Sigmoid activation function.

[0059] During training, the model uses a negative log-likelihood loss function as the training objective function to minimize the difference between the predicted probability distribution and the actual labels. In each training round, the model calculates the loss based on the current batch of samples, backpropagates and updates the parameters, uses the Adam optimizer for parameter learning, and implements an Early Stopping mechanism on the validation set to prevent overfitting. Training automatically stops when the model's performance on the validation set no longer improves, and its generalization ability is evaluated on the test set.

[0060] To comprehensively evaluate the performance of the model of this invention, this study uses the following three evaluation metrics:

[0061] Accuracy%: Represents the proportion of correctly predicted samples out of the total number of samples, and measures the overall classification accuracy of the model.

[0062] Root Mean Square Error (RMSE): Reflects the degree of deviation between the model's predicted value and the true label. The smaller the value, the closer the model's prediction is to the true value.

[0063] Area Under Curve (AUC%): Measures the model’s ability to distinguish between positive examples (correct answers) and negative examples (incorrect answers). A higher AUC value indicates that the model has a good discriminative ability under different probability thresholds.

[0064] Experimental results show that the cognitive diagnostic model proposed in this invention achieves excellent performance on both the AIFC and ASSIST datasets, with high prediction accuracy, small error, and good interpretability and practical application value.

[0065] Example 3

[0066] like Figure 3 As shown, a cognitive diagnosis method based on a knowledge-text-based cognitive diagnosis model includes the following specific steps:

[0067] Input the data of students who answered the questions and the corresponding question data;

[0068] The student answer data and corresponding question data are input into the question text feature extraction module. Through the text processing network, the knowledge point association vectors that are related to the students' knowledge status and knowledge points are extracted and generated.

[0069] The knowledge proficiency embedding module maps student data into a knowledge proficiency vector; the question attribute embedding module maps question data into a question knowledge difficulty vector and a question discrimination scalar.

[0070] The knowledge point association vector is input into the knowledge association attention module to adjust the attention weight, thereby generating a weighted knowledge representation that can highlight the core knowledge point dependencies in the current question context;

[0071] The knowledge proficiency vector, the question knowledge difficulty vector, the question discrimination scalar, and the weighted knowledge representation are combined into interactive features and input into the adaptive nonlinear prediction module.

[0072] In the adaptive nonlinear prediction module, the grid granularity of its internal B-spline basis function is adaptively adjusted according to the question discrimination scalar, and the interaction features are nonlinearly mapped to finally output the predicted probability of the student answering correctly.

[0073] In one specific embodiment, an interpretable diagnostic step is also included, which uses a visualization program to present the shape of the activated B-spline curve in the adaptive nonlinear prediction module, in order to analyze and reveal the nonlinear influence pattern between a particular student's knowledge mastery and the attributes of the question when answering a particular question.

[0074] In one specific embodiment, prior to cognitive diagnosis, a model training step is included, which uses the negative log-likelihood loss function as the objective function and iteratively optimizes all trainable parameters in the model using the Adam optimizer.

[0075] In this embodiment, the present invention automatically learns the interaction patterns between knowledge points through a multi-head self-attention mechanism, effectively modeling the hierarchical structure of knowledge and contextual relationships, thus improving the model's accuracy in recognizing students' knowledge mastery status. The present invention designs a dynamic weighting module to achieve personalized question modeling. Traditional models often ignore the differences in the ability of questions to differentiate between students, resulting in insufficient generalization ability of prediction results. The present invention uses a weighting mechanism after Softmax normalization to dynamically adjust the distinguishing influence of questions on different students, making the model more flexible and accurate in scenarios with large variations in question difficulty and strong heterogeneity in the question bank. The present invention adopts a multi-layer KANLinear nonlinear modeling structure to improve the model's expressive power and interpretability. Although traditional neural networks have strong expressive power, it is difficult to explain their internal mechanisms. The present invention introduces a KAN module based on B-spline interpolation to achieve smooth nonlinear transformation of input features, which can better fit the complex cognitive relationship between students and knowledge points, and visualize the feature mapping process, enhancing the model's transparency and educational applicability.

[0076] In summary, the advantages of this invention are: strong adaptability: through structured attention and dynamic weighting mechanisms, it can adapt to different question types, knowledge structures, and student differences; strong expressive ability: the KAN module has high-order nonlinear fitting ability, which can comprehensively model students' answering behavior; good interpretability: each module corresponds to a clear educational meaning, which is helpful for result analysis and teaching feedback.

[0077] This invention can be widely applied to intelligent teaching systems such as online education platforms, adaptive assessment systems, learning path recommendations, and intelligent test paper generation, providing a reliable model foundation and decision support for personalized learning and ability assessment.

[0078] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A knowledge text-based cognitive diagnosis model, characterized in that: It includes a text feature extraction module, a knowledge proficiency embedding module, a question attribute embedding module, a knowledge association attention module, and an adaptive nonlinear prediction module; The text feature extraction module is used to receive student answer data and corresponding question data, and through a text processing network, extract and generate knowledge point association vectors that correlate students' knowledge status with knowledge points. The knowledge proficiency embedding module is used to map student data into a knowledge proficiency vector; The question attribute embedding module is used to map question data into a question knowledge difficulty vector and a question discrimination scalar. The knowledge association attention module takes the knowledge point association vector and the question discrimination scalar as inputs and is used to adjust the attention weights. The question discrimination scalar is used as an adjustment factor in the calculation of the attention score, thereby generating a weighted knowledge representation that can highlight the core knowledge point dependencies in the current question context. The adaptive nonlinear prediction module contains a multi-layer KANLinear structure composed of B-spline basis functions. It receives the student's knowledge proficiency vector, the question's knowledge difficulty vector, the question's discrimination scalar, and the interactive features composed of the weighted knowledge representation. Based on the question's discrimination scalar, it adaptively adjusts the grid granularity of its internal B-spline basis functions and performs a nonlinear mapping on the interactive features, ultimately outputting the predicted probability of the student's correct answer. The adaptive adjustment of the B-spline basis function grid granularity is achieved by mapping the value of the question's discrimination scalar to the grid's partition density using a preset or learnable mapping function, thus realizing nonlinear fitting for high-discrimination questions.

2. The cognitive diagnostic model based on knowledge text according to claim 1, characterized in that: The text feature extraction module includes a pre-trained language model based on the Transformer architecture, which is used to deeply understand the semantic information of the question text and generate the knowledge point association vector.

3. The cognitive diagnostic model based on knowledge text according to claim 1, characterized in that: The knowledge association attention module adjusts the attention weights by using a multi-head self-attention mechanism on the knowledge point association vector to model the potential dependencies between knowledge points.

4. The cognitive diagnostic model based on knowledge text according to claim 1, characterized in that: The KANLinear structure consists of three layers. Its inputs are a student knowledge proficiency vector, a question knowledge difficulty vector, a question discrimination scalar, and the student-question interaction features composed of the weighted knowledge representation. The model is performed through KANLinear modules layer by layer, and the probability of answering correctly is finally output by the Sigmoid function.

5. The cognitive diagnostic model based on knowledge text according to claim 1, characterized in that: The discriminant scalar value is normalized by a sigmoid function and multiplied by a learnable scaling factor before being input into other parts of the model, in order to obtain a range of values ​​that are most suitable for model tuning.

6. The cognitive diagnostic model based on knowledge text according to claim 1, characterized in that: The last layer of the adaptive nonlinear prediction module uses a Sigmoid activation function to constrain the output value to the interval [0,1], which serves as the final probability of a correct answer.

7. A cognitive diagnosis method based on a knowledge-text-based cognitive diagnosis model, characterized in that: The method based on any one of claims 1 to 6 includes the following specific steps: Input the data of students who answered the questions and the corresponding question data; The student answer data and corresponding question data are input into the question text feature extraction module. Through the text processing network, the knowledge point association vectors that are related to the students' knowledge status and knowledge points are extracted and generated. The knowledge proficiency embedding module maps student data into a knowledge proficiency vector; the question attribute embedding module maps question data into a question knowledge difficulty vector and a question discrimination scalar. The knowledge point association vector is input into the knowledge association attention module to adjust the attention weight, thereby generating a weighted knowledge representation that can highlight the core knowledge point dependencies in the current question context; The knowledge proficiency vector, the question knowledge difficulty vector, the question discrimination scalar, and the weighted knowledge representation are combined into interactive features and input into the adaptive nonlinear prediction module. In the adaptive nonlinear prediction module, the grid granularity of its internal B-spline basis function is adaptively adjusted according to the question discrimination scalar, and the interaction features are nonlinearly mapped to finally output the predicted probability of the student answering correctly.

8. The cognitive diagnostic method according to claim 7, characterized in that: It also includes an interpretable diagnostic step that uses a visualization program to present the shape of the activated B-spline curves in the adaptive nonlinear prediction module, revealing the nonlinear influence pattern between a particular student's knowledge mastery and the attributes of a particular question when answering a particular question.

9. The cognitive diagnostic method according to claim 7, characterized in that: Before cognitive diagnosis, there is also a model training step, which uses the negative log-likelihood loss function as the objective function and uses the Adam optimizer to iteratively optimize all trainable parameters in the model.

Images