A cognitive analysis system that focuses on ability changes and item characteristics

Abstract

The application belongs to the technical field of education evaluation, and discloses a cognitive analysis system focusing on ability change and question characteristics, which comprises an interactive coding network module, an ability acquisition module, an ability prediction module, an ability rationality module, a question prediction module, a knowledge point prediction module and a multi-task fusion module. The ability prediction module, the question prediction module and the knowledge point prediction module are used to respectively predict the probability of answering the next exercise correctly. Finally, the multi-task fusion module balances the three types of prediction probabilities through a fusion training mode to obtain the final probability of answering the next exercise correctly, and the final probability is used for cognitive analysis of the subject. The application can track the ability change of the subject in real time, provide dynamic ability value evaluation, and better reflect the ability development in the learning process.

Description

Technical Field

[0001] This invention belongs to the field of educational assessment technology, and in particular relates to a cognitive analysis system that focuses on changes in ability and the characteristics of questions. Background Technology

[0002] While traditional Item Response Theory (IRT) models can provide accurate ability assessments, they are static models that cannot capture dynamic changes in ability during the learning process or reflect the real-time learning status of learners, thus lacking dynamism. Existing knowledge tracking methods (such as DKT and DKVMN) can dynamically track the learning process, but they lack constraints on the rationality of ability changes. They often exhibit the unreasonable phenomenon of ability scores increasing after incorrect answers, leading to unreliable learning analysis results and a significant lack of rationality.

[0003] Furthermore, existing methods primarily focus on predicting single responses, lacking the ability to summarize and analyze the overall learning performance of test takers. Additionally, recommendations based on unreasonable ability values ​​are often ineffective and fail to truly achieve personalized learning. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a cognitive analysis system that focuses on ability changes and question characteristics. It achieves real-time ability tracking by replacing static parameter estimation with dynamic sequence modeling; it constrains the direction of ability changes to conform to learning patterns by designing an ability change loss function; it performs knowledge tracking and ability assessment simultaneously through a multi-task learning framework; and it achieves accurate recommendations by matching question difficulty based on reasonable ability values.

[0005] This invention provides the following technical solution:

[0006] A cognitive analysis system that focuses on changes in abilities and characteristics of questions includes: an interactive coding network module, an ability acquisition module, an ability prediction module, an ability rationality module, a question prediction module, a knowledge point prediction module, and a multi-task fusion module.

[0007] The interactive coding network module is used to encode the question-answer sequence containing question-knowledge point-answer to obtain historical interactive feature coding, which includes ability-based historical interactive feature coding, question-based historical interactive feature coding and knowledge point-based historical interactive feature coding.

[0008] The ability acquisition module is used to obtain the ability value sequence of the subject after t practice sessions based on the historical interaction feature encoding of the subject's ability class during the t-th practice session. Among them, the ability value sequence This represents the probability of answering a question correctly in each practice session;

[0009] The ability prediction module is used to combine the ability value sequence after each historical practice session. And the knowledge points involved in the next question The probability of answering the next exercise correctly under the predictive ability dimension;

[0010] The ability rationality module is used to constrain the rationality of ability value changes by comparing the changes in ability values ​​between different practice sessions and whether the answers are correct.

[0011] The question prediction module is used to predict the probability of answering the question correctly in the next exercise based on the historical interaction feature encoding of the question type and the question-knowledge point embedding vector of the next exercise.

[0012] The knowledge point prediction module is used to predict the probability of answering the next exercise correctly under the knowledge point adaptation dimension based on the historical interaction feature encoding of the knowledge point class and the question-knowledge point embedding vector of the next exercise.

[0013] The multi-task fusion module is used to balance the prediction probabilities obtained by the ability prediction module, question prediction module, and knowledge point prediction module through fusion training to obtain the final probability of answering correctly in the next exercise. The final probability is used for the cognitive analysis of the subject. The total loss function of the multi-task fusion module in the fusion training stage consists of cross-entropy loss based on ability dimension prediction, cross-entropy loss based on question feature dimension prediction, cross-entropy loss based on knowledge point adaptation dimension prediction, and ability rationality loss. The weight parameters of each type of loss are constants between 0 and 1, and their sum is equal to 1.

[0014] ,

[0015] in, It is the total loss during the training process. , These are cross-entropy loss methods based on prediction probabilities, namely, ability dimension, question feature dimension, and knowledge point adaptation dimension. It is a loss of rationality. , , , They are , , , The corresponding weight parameters are all constants between 0 and 1, and satisfy the following conditions: ;

[0016] Based on the probabilities of answering correctly in the next practice session under the ability dimension, the probability of answering correctly in the next practice session under the question feature dimension, and the probability of answering correctly in the next practice session under the knowledge point adaptation dimension, the final probability of answering correctly in the next practice session is obtained by integrating the probabilities using the average method:

[0017] ,

[0018] in, It is the predicted final probability of answering correctly in the next practice session. These are the probabilities of answering correctly in the next practice session, estimated based on ability, the question, and the knowledge points.

[0019] Preferably, the method for obtaining historical interaction feature codes in the interaction coding network module includes:

[0020] Calculate the knowledge point vector for the t-th practice session;

[0021] Based on the knowledge point vector and the corresponding question vector at the t-th practice session, establish the question-knowledge point-answer interaction vector at the t-th practice session;

[0022] Using an LSTM network, the question-knowledge point-answer interaction vector is encoded to obtain the historical interaction feature encoding.

[0023] Preferred methods for calculating the knowledge point vector at the t-th practice session include:

[0024] The sum of all knowledge point embedding vectors corresponding to the question in the t-th practice session is obtained by dividing by the number of knowledge points; wherein, the number of knowledge points is obtained by statistical analysis of the knowledge point index contained in the question.

[0025] Preferably, the encoding form of the question-knowledge point-answer interaction vector is as follows:

[0026] When the answer is correct, the question vector, the knowledge point vector, and a 2d vector of all zeros are concatenated in sequence.

[0027] When an answer is incorrect, the vector consisting of all zeros, the question vector, and the knowledge point vector are concatenated in sequence.

[0028] Where d is the vector length of the question and the knowledge point.

[0029] Preferably, the interactive coding network module is equipped with three LSTM networks with identical structures but independent parameters. The input of the three LSTM networks is the question-knowledge point-answer interaction vector, and each network independently calculates feature encoding using its previous output feature value and gate mechanism parameters.

[0030] The first LSTM network outputs historical interaction feature encodings for ability categories, the second LSTM network outputs historical interaction feature encodings for question categories, and the third LSTM network outputs historical interaction feature encodings for knowledge point categories.

[0031] Preferably, the capability acquisition module includes several fully connected layers, which realize the capability value sequence through progressive processing. Calculation;

[0032] The ability acquisition module obtains the subject's ability value sequence after t training sessions. The methods include:

[0033] First, a linear rectified activation operation is performed on the output of the previous layer, then the learnable weight matrix is ​​transformed and a bias term is added, and then a second linear rectified activation operation is performed.

[0034] Then, the result after the second activation is multiplied element-wise with the feature transformation weight vector, and the result is summed and pooled.

[0035] Finally, the activation function outputs a normalized sequence of ability values ​​after t training sessions. .

[0036] Preferably, the method by which the ability prediction module predicts the probability of answering correctly in the next practice session includes:

[0037] First, based on the subject's ability value sequence after t training sessions. By performing dot product operations, the ability values ​​corresponding to the knowledge points involved in the next exercise are extracted, and a historical sequence of ability values ​​is constructed.

[0038] Based on the historical sequence of the ability values, an LSTM network is used, combined with a network gate mechanism to assist in the calculation, to obtain the historical ability feature vector corresponding to the t-th practice session.

[0039] Based on the historical ability feature vector, after passing through several fully connected layers, and finally through an activation function, the probability of answering correctly in the next exercise under the ability dimension is obtained.

[0040] Preferably, the method for judging changes in capability values ​​by the capability rationality module includes:

[0041] Based on the ability values ​​of two consecutive practice sessions, after eliminating the influence of numerical normalization through the inverse operation of the activation function, the difference between the next practice session and the previous practice session is calculated as the actual ability change value. An actual ability change value greater than 0 corresponds to an increase in ability, and a value less than 0 corresponds to a decrease in ability.

[0042] Preferably, in the ability rationality module, when the answer is correct, the ideal ability change value is the product of the one-hot vector of the knowledge point involved in the question and 0.5, superimposed with a vector of the same dimension and all elements of 0.5; when the answer is incorrect, the ideal ability change value is the product of the one-hot vector of the knowledge point involved in the question and -0.5, superimposed with a vector of the same dimension and all elements of -0.5.

[0043] Preferably, the method for constraining changes in capability values ​​by the capability rationality module includes:

[0044] Based on the actual ability change sequence from the second practice session to the tth practice session, and the corresponding ideal ability change sequence, a mean squared error loss function is used. By minimizing the calculation result of this loss function, the actual ability change value is forced to approach the ideal ability change value, thus ultimately constraining the ability change value.

[0045] Preferably, the method by which the question prediction module predicts the probability of answering the next question correctly includes:

[0046] The question-knowledge point embedding vector of the next practice session is concatenated with the historical interaction feature encoding of the question class to form the question class prediction feature vector.

[0047] The predicted feature vectors of the question type are input into several fully connected layers, and the probability of answering correctly in the next practice session is output through the activation function.

[0048] Preferably, the method by which the knowledge point prediction module predicts the probability of answering the next question correctly includes:

[0049] The knowledge point embedding vector of the next practice question is concatenated with the historical interaction feature encoding of the knowledge point class to form the knowledge point class prediction feature vector.

[0050] The prediction feature vectors of knowledge points are input into several fully connected layers, and the prediction accuracy of the subjects for each knowledge point is output through the activation function.

[0051] Calculate the average of all prediction accuracy rates to obtain the probability of answering the next exercise correctly under the knowledge point adaptation dimension.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention provides a cognitive analysis system that focuses on changes in ability and the characteristics of questions, which has the following advantages compared to traditional solutions:

[0054] 1. Achieving dynamic ability assessment: Compared with the static IRT model, this invention can track changes in the subject's ability in real time, provide dynamic ability value assessment, and better reflect the ability development during the learning process.

[0055] 2. Ensuring the rationality of ability values: Compared with traditional knowledge tracking models, this invention ensures that the predicted ability values ​​conform to the learning law through the ability change loss function. When answering questions correctly, the ability increases, and when answering questions incorrectly, the ability decreases, providing reliable learning analysis results.

[0056] 3. Provide a comprehensive summary of learning progress: Through a multi-task learning mechanism, it not only predicts the results of answering questions, but also analyzes the mastery of questions and the degree of understanding of concepts.

[0057] 4. Optimize question recommendation: Based on reasonable ability value assessment, achieve more accurate personalized question recommendations and improve learning effectiveness. Attached Figure Description

[0058] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the overall structure of the cognitive analysis system according to Embodiment 1 of the present invention;

[0060] Figure 2 This is a schematic diagram of the overall process of the cognitive analysis system according to Embodiment 1 of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Example 1

[0063] This invention provides a cognitive analysis system that focuses on changes in ability and the characteristics of questions, such as... Figure 1 As shown, it mainly includes an interactive coding network module, a capability acquisition module, a capability prediction module, a capability rationality module, a question prediction module, a knowledge point prediction module, and a multi-task fusion module. Specifically:

[0064] The interactive coding network module is used to encode the question-knowledge point-answer sequence to obtain historical interactive feature codes, which include ability-based historical interactive feature codes, question-based historical interactive feature codes, and knowledge point-based historical interactive feature codes.

[0065] The ability acquisition module is used to obtain the ability value sequence of the subject after t practice sessions based on the historical interaction feature encoding of the subject's ability class during the t-th practice session. .

[0066] The ability prediction module is used to combine the ability value sequence after each historical practice session. And the knowledge points involved in the next question The probability of answering the next exercise correctly under the predictive ability dimension.

[0067] The ability rationality module is used to constrain the rationality of ability value changes by comparing the changes in ability values ​​between different practice sessions and whether the answers are correct.

[0068] The question prediction module is used to predict the probability of answering the question correctly in the next exercise based on the historical interaction feature encoding of the question class and the embedding vector of the next exercise.

[0069] The knowledge point prediction module is used to predict the probability of answering the next exercise correctly under the knowledge point adaptation dimension, based on the historical interaction feature encoding of the knowledge point class and the embedding vector of the next question.

[0070] The multi-task fusion module is used to balance the predicted probabilities obtained from the ability prediction module, question prediction module, and knowledge point prediction module through fusion training, and obtain the final probability of answering correctly in the next exercise. The final probability is used for cognitive analysis of the subjects.

[0071] Below, in conjunction with Figure 1 The overall structural diagram shown and Figure 2 The overall flowchart shown illustrates the functional implementation and operational logic of each module:

[0072] 1) Interactive coding network module:

[0073] In the scenario of predicting test-takers' answers, the core features of test-takers' abilities, question characteristics, and knowledge point suitability are all deeply related to the historical test-taking interaction process. Relying solely on a single embedded vector of a question or knowledge point (such as using only fixed labels to identify questions and knowledge points) cannot reflect key information such as "the impact of test-takers' long-term answering habits on their abilities," "the reference value of historical accuracy rates for similar questions for current questions," and "changes in mastery of knowledge points after repeated testing." Existing encoding methods do not integrate this historical interaction data, resulting in a lack of comprehensive feature support for predictions in the dimensions of ability, question characteristics, and knowledge point suitability, thus limiting prediction accuracy. Therefore, it is necessary to integrate historical interaction information into feature encoding to provide more targeted and high-quality input for each prediction module.

[0074] In view of this, in this embodiment, the historical interaction feature codes generated by the interaction coding network module include three categories: ability-based historical interaction feature codes, question-based historical interaction feature codes, and knowledge-point-based historical interaction feature codes. The specific method is as follows:

[0075] S1.1. Calculate the knowledge point vector for the t-th practice session. Specifically:

[0076] The sum of the embedding vectors of all knowledge points corresponding to the question at the t-th practice session is obtained by dividing by the number of knowledge points. The number of knowledge points is obtained by counting the knowledge point indices contained in the question.

[0077] ;

[0078] in, It is the problem vector The average of the knowledge points involved This represents the number of knowledge points covered in the t-th exercise. This represents the index of the knowledge points included in the t-th exercise. The first Embedding vectors of knowledge points.

[0079] S1.2. Based on the knowledge point vector and corresponding question vector at the t-th practice session, establish the question-knowledge point-answer interaction vector for the t-th practice session. Specifically:

[0080] ;

[0081] in, This is the question-knowledge point-answer interaction vector for the t-th practice session, with a length of 4d, where d is the question vector. The mean of the vector of related knowledge points The length of in this formula It is a vector of length 2d with all elements equal to 0. This represents vector concatenation. Indicates whether the answer is correct; if correct... =1, error occurs =0.

[0082] S1.3. Using an LSTM network, the question-knowledge point-answer interaction vector is encoded to obtain the historical interaction feature encoding.

[0083] The encoding format for the question-knowledge point-answer interaction vector is as follows:

[0084] When the answer is correct, the vector is formed by concatenating the question vector, the knowledge point vector, and a 2d vector of all zeros. When the answer is incorrect, the vector is formed by concatenating the 2d vector of all zeros, the question vector, and the knowledge point vector. Here, d represents the length of the question and knowledge point vectors.

[0085] ;

[0086] in, These are the interaction feature values ​​from the previous t training sessions. yes Network gate mechanisms are used to assist in computation. It does not participate in subsequent calculations.

[0087] The interactive coding network module generates historical interactive feature codes for ability, question, and knowledge points, corresponding to the ability dimension, question feature dimension, and knowledge point adaptation dimension, respectively. Their core objectives differ: ability dimension prediction focuses on the "trajectory of subject ability changes," question feature dimension prediction focuses on the "history of question feature adaptation," and knowledge point adaptation dimension prediction focuses on the "evolution of knowledge point mastery." If the three types of predictions share the same historical interactive features, the feature requirements of different dimensions will interfere with each other during training (e.g., ability-related features are diluted by question features), leading to insufficient feature specificity and affecting the prediction accuracy of subsequent ability prediction, question prediction, and knowledge point prediction modules. Therefore, it is necessary to generate dedicated historical interactive features separately for each dimension of prediction to ensure accurate matching between features and prediction targets.

[0088] Therefore, in this embodiment, the interactive coding network module includes three LSTM networks with identical structures but independent parameters. The input to each LSTM network is a question-knowledge point-answer interaction vector, and each network independently calculates its feature encoding using its previous output feature value and gate mechanism parameters. Specifically, the first LSTM network focuses on extracting historical interaction feature encodings for ability categories, outputting historical interaction feature encodings for ability categories, which are adapted for ability dimension prediction. The second LSTM network focuses on extracting historical interaction feature encodings for question categories, outputting historical interaction feature encodings for question categories, which are adapted for question feature dimension prediction. The third LSTM network focuses on extracting historical interaction feature encodings for knowledge point categories, outputting historical interaction feature encodings for knowledge point categories, which are adapted for knowledge point adaptation dimension prediction. That is:

[0089] ,

[0090] ,

[0091] .

[0092] in, Encoding historical interaction features for capability categories, Encoding the historical interaction features of the question type, Encode the historical interaction features of knowledge point categories.

[0093] The three LSTM networks independently calculate feature codes using their previous output feature values ​​and gate mechanism parameters, without interfering with each other. This ensures that each type of feature code can accurately support the prediction logic of the corresponding module, thereby improving the overall prediction performance.

[0094] (ii) Capability Acquisition Module:

[0095] The core objective of existing cognitive analysis systems is mostly focused on "predicting whether a subject can answer a specific question correctly." Ability values ​​are only used as intermediate calculation variables to achieve this prediction and are not regarded as key results that need to be accurately output. This leads to a simplification of the calculation logic of ability values ​​(such as processing them through simple weighting or a single-layer network). This results in the inability to accurately reflect the subject's true mastery of knowledge points and to support the practical application needs such as subsequent personalized ability assessment and learning trajectory tracking.

[0096] Therefore, in this embodiment, the capability acquisition module is equipped with several fully connected layers, and the capability value sequence is realized through progressive processing. The calculation involves obtaining the subject's ability value sequence after t training sessions. The methods include:

[0097] First, a linear rectified activation operation is performed on the output of the previous layer, then the learnable weight matrix is ​​transformed and a bias term is superimposed, and then a second linear rectified activation operation is performed.

[0098] Then, the result after the second activation is multiplied element-wise with the feature transformation weight vector, and the result is summed and pooled.

[0099] Finally, the activation function outputs a normalized sequence of ability values ​​after t training sessions. This ensures that the capability values ​​are within a uniform range (0-1) and have horizontal comparability.

[0100] ;

[0101] in, It is an activation function; , This is a learnable bias term for feature transformation; , The learnable weight matrix for feature transformation; It is the feature transformation weight vector after summation and pooling; It is a linear rectification activation function; This is element-wise multiplication (Hadamard product). It is a summation pooling operation.

[0102] (iii) Capacity Prediction Module:

[0103] The accurate ability values ​​generated by the ability acquisition module (which have been transformed from "intermediate state" to "reusable assessment results") are the core and effective data for assisting in predicting test takers' answers. However, these ability values ​​are subject to update lags—the actual improvement of a test taker's ability is a gradual process (e.g., when answering questions correctly consecutively, the ability value will gradually increase from 0.1 to 0.4), rather than an instantaneous leap. Existing prediction methods that rely solely on the current single ability value cannot reflect this dynamic growth trajectory, leading to prediction bias (e.g., the test taker's actual ability has improved through practice with multiple questions, but the current ability value has not yet fully reflected this; using only this value for prediction will underestimate the probability of answering questions). Therefore, it is necessary to combine the characteristics of historical ability value sequences and fully explore the trend of ability changes to maximize the predictive value of ability values ​​and achieve accurate prediction of answer probabilities.

[0104] Therefore, in this embodiment, the ability prediction module first bases its prediction on the subject's ability value sequence after t training sessions. By performing dot product operations, the ability values ​​corresponding to the knowledge points involved in the next exercise are extracted, and a historical sequence of ability values ​​is constructed. (The next practice topic will be identified by the one-hot vector):

[0105] ,

[0106] in, It is a one-hot vector representing the knowledge points involved in the (t+1)th exercise. It is the ability value of all knowledge points obtained after completing t exercises, as obtained from the ability acquisition module. yes and The Hadamard product represents the ability value of the knowledge points involved in the (t+1)th practice question.

[0107] Then, based on the historical sequence of ability values, an LSTM network is used, combined with a network gating mechanism to assist in the calculation, to obtain the historical ability feature vector corresponding to the t-th practice session:

[0108] ,

[0109] in, It is the ability feature vector from the previous t practice sessions. yes Network gate mechanisms are used to assist in computation. .

[0110] Finally, based on the historical ability feature vector, through several fully connected layers, and finally through an activation function, the probability of answering correctly in the next exercise under the ability dimension is obtained, which is used to assist in the final prediction:

[0111] ;

[0112] in, This represents the probability of correctly answering the (t+1)th practice question, predicted based on the ability value.

[0113] (iv) Capability Rationality Module:

[0114] The ability values ​​generated by the ability acquisition module have been upgraded from "intermediate calculation variables" to "precise and reusable core assessment results." The rationality of their changes directly determines the reliability of subsequent learning analysis and ability tracking. However, existing systems, because they do not use ability values ​​as the final output target and lack standardized requirements for their change logic, often exhibit abnormal fluctuations in ability values ​​that contradict learning patterns (e.g., ability increases after answering questions incorrectly and decreases after answering questions correctly). Without clearly defining and constraining reasonable standards for ability changes, the ability values ​​generated by the ability acquisition module will lose their practical application value and cannot support personalized learning guidance. Therefore, it is necessary to establish clear ideal requirements for ability changes to provide a basis for judging the rationality of the ability assessment module.

[0115] Therefore, in this embodiment, based on the ability values ​​generated by the ability acquisition module for two consecutive practice sessions, after eliminating the influence of numerical normalization through the inverse operation of the activation function, the difference between the subsequent practice session and the previous practice session is calculated as the actual ability change value. An actual ability change value greater than 0 corresponds to an increase in ability, and a value less than 0 corresponds to a decrease in ability.

[0116] ;

[0117] in, It is the value of the rationality of the t-th exercise. It is the inverse operation of the activation function. This is the value for the reasonableness of the second practice session.

[0118] The rules for setting the ideal ability change value are as follows: When the answer is correct, the ideal ability change value is the product of the one-hot vector of the knowledge point involved in the question and 0.5, superimposed with a vector of the same dimension and all elements of 0.5 (ensuring a slight increase in ability). When the answer is incorrect, the ideal ability change value is the product of the one-hot vector of the knowledge point involved in the question and -0.5, superimposed with a vector of the same dimension and all elements of -0.5 (ensuring a slight decrease in ability).

[0119] ,

[0120] in, It is a one-hot vector representing the knowledge points involved in the t-th exercise. Indicates and Vectors with the same dimension and all elements being 1; This represents the change in ability at the ideal t-th practice session. Indicates whether the answer is correct; if correct... =1, error occurs =0.

[0121] Based on the aforementioned clearly defined ideal ability change requirements, the ability rationality module already has a judgment benchmark, but it lacks a quantitative constraint mechanism. Qualitative comparison alone cannot accurately correct the deviation between actual ability changes and ideal requirements, leading to potential "small abnormal fluctuations" in ability values ​​(such as an increase in ability far exceeding the ideal value after answering questions correctly, or a decrease in ability insufficient after answering questions incorrectly). Without establishing quantitative constraint logic, the aforementioned ideal standards cannot be effectively implemented, and the rationality of the ability values ​​generated by the ability acquisition module remains difficult to guarantee. Therefore, a specialized loss function is needed to transform the ideal ability change requirements into calculable constraint rules, enabling precise correction of actual ability changes.

[0122] In view of this, this embodiment further clarifies the quantitative evaluation method of the capability rationality module, adopting the mean squared error loss function as the core constraint tool. This function focuses on the deviation between the actual capability change value and the ideal capability change value. The calculation object of this function is the sequence of actual capability change values ​​from the second practice session to the t-th practice session, and the corresponding sequence of ideal capability change values. During model training, by minimizing the calculation result of this loss function, the actual capability change value is forced to converge towards the ideal capability change value, ultimately achieving precise constraints on capability value changes. This ensures that the capability values ​​generated by the capability acquisition module not only conform to qualitative laws but also meet quantitative rationality requirements, providing highly reliable data support for subsequent applications.

[0123] ,

[0124] Wherein, MSE is the mean squared error loss function, used to measure the deviation between the actual change in capability and the ideal change in capability.

[0125] (v) Question Prediction Module:

[0126] The ability scores generated by the ability acquisition module primarily reflect the subject's mastery of knowledge points and are an important basis for predicting answer results. However, the accuracy of answering questions is not solely determined by the subject's ability—the characteristics of the questions themselves (such as difficulty level, complexity of knowledge point connections, and examination format) are equally crucial. Current prediction methods that rely solely on the subject's ability scores from the ability acquisition module ignore the influencing factors at the question level (e.g., a subject with strong abilities may still answer questions beyond the syllabus incorrectly, while a subject with weak abilities may answer basic questions correctly), leading to significant discrepancies between predicted results and actual outcomes. Therefore, it is necessary to supplement this prediction with an independent prediction dimension based on question characteristics to improve prediction accuracy.

[0127] In view of this, in this embodiment, the question prediction module makes predictions based on the historical interaction feature encoding of question types and combined with the embedded features of the next question (including information such as question difficulty and knowledge point relevance), including:

[0128] S5.1. Concatenate the question-knowledge point embedding vector of the next practice session with the historical interaction feature encoding of the question class to form the question class prediction feature vector.

[0129] S5.2. Input the predicted feature vector of the question class into several fully connected layers, and output the probability of answering correctly in the next practice under the question feature dimension through the activation function to assist the final prediction:

[0130] ;

[0131] in, It is the probability of correctly answering the question in the (t+1)th practice session, based on historical answer records, the next practice session, and knowledge point information. This represents the interaction vector of practice questions and knowledge points in the (t+1)th exercise.

[0132] (vi) Knowledge Point Prediction Module:

[0133] Besides the subject's ability and question characteristics, the core factors influencing the accuracy of answers are also directly related to the matching degree of the knowledge points involved in the questions. The same subject can have significantly different accuracy rates when faced with questions on different knowledge points (e.g., a subject who excels at addition but is weak in subtraction will have a high accuracy rate on addition questions but a low accuracy rate on subtraction questions). Existing prediction methods, which only consider the subject's ability and question characteristics without focusing on the suitability at the knowledge point level, will still exhibit prediction bias (e.g., the question difficulty is moderate and the subject's ability is sufficient, but the subject answers incorrectly due to weak knowledge points). Therefore, it is necessary to supplement this with an independent prediction dimension based on knowledge point mastery, forming a three-dimensional prediction logic of "subject ability + question characteristics + knowledge point suitability" to further improve prediction accuracy.

[0134] In view of this, in this embodiment, the knowledge point prediction module makes predictions based on the historical interaction feature encoding of knowledge point classes and in combination with the embedding vector of the next practice question, including:

[0135] S6.1. Concatenate the embedding vector of the next practice question with the historical interactive feature encoding of the knowledge point class to form the predicted feature vector of the knowledge point class.

[0136] S6.2. Input the knowledge point class prediction feature vector into several fully connected layers, and output the subject's prediction accuracy for each knowledge point through an activation function:

[0137] ;

[0138] in, It is an intermediate value calculated based on the historical answer record, the next practice and knowledge point information, that is, the probability that the subject answers each knowledge point correctly in the t+1th practice session;

[0139] S6.3. Calculate the average of all prediction accuracy rates to obtain the probability of answering correctly in the next exercise under the knowledge point adaptation dimension:

[0140] ;

[0141] in, It is the mean probability of answering the knowledge points covered in the (t+1)th practice session correctly, where This represents the index of the knowledge points included in the (t+1)th exercise. Indicates that the answer is correct. The probability of each knowledge point. This represents the number of knowledge points involved in the (t+1)th exercise.

[0142] VII) Multi-task fusion module:

[0143] Based on the ability prediction module, question prediction module, and knowledge point prediction module, a three-dimensional prediction logic of "subject ability + question features + knowledge point adaptation" has been formed. The ability rationality module ensures the rationality of ability values ​​through quantitative constraints. However, without a unified training balance mechanism, model training may become biased towards a single task (e.g., overemphasizing prediction accuracy while neglecting ability value rationality, or vice versa). An imbalance in the loss weights of different tasks will prevent the prediction accuracy and ability rationality of the ability prediction, question prediction, and knowledge point prediction modules from being balanced, ultimately affecting the overall system performance. Therefore, it is necessary to establish a multi-task fusion loss balance rule to achieve synergistic optimization of each objective.

[0144] In view of this, the training mechanism of the multi-task fusion module is clearly defined in this embodiment:

[0145] 1. The total loss function integrates four core loss types: cross-entropy loss based on ability dimension prediction, cross-entropy loss based on question feature dimension prediction, cross-entropy loss based on knowledge point adaptation dimension prediction, and ability rationality loss. Independent weight parameters (α, β, γ, δ) are assigned to each type of loss. All weight parameters are constants between 0 and 1, and their sum equals 1.

[0146] ,

[0147] in, It is the total loss during the training process. , These are cross-entropy loss methods based on prediction probabilities, namely, ability dimension, question feature dimension, and knowledge point adaptation dimension. It is a loss of rationality. , , , They are , , , The corresponding weight parameters, .

[0148] 2. The training direction can be flexibly balanced by adjusting the weight parameters: if the credibility of the ability assessment needs to be enhanced, the weight of the ability rationality loss (δ) can be increased; if the prediction accuracy needs to be improved, the weights of the three types of prediction loss (α, β, γ) can be appropriately increased.

[0149] 3. During model training, minimizing the total loss function achieves synergistic optimization of prediction accuracy and capability rationality, ensuring that the goals of each module in the system are consistent and the performance is optimal.

[0150] Based on the aforementioned multi-task training mechanism, the ability prediction module, question prediction module, and knowledge point prediction module can output three independent types of answer probability prediction results. However, without clear rules for integrating the results, the three types of prediction data cannot be effectively merged—relying solely on one type of result will result in dimensional bias, while simple splicing will fail to leverage synergistic value, ultimately affecting the accuracy of the cognitive analysis results. Therefore, it is necessary to establish a fair and reproducible result integration method to fully leverage the three-dimensional prediction advantages of "subject ability + question characteristics + knowledge point adaptation" to output comprehensive and accurate final probabilities.

[0151] In view of this, this embodiment further clarifies the prediction and integration method of the multi-task fusion module: based on the probability of answering the next exercise correctly under the ability dimension, the probability of answering the next exercise correctly under the question feature dimension, and the probability of answering the next exercise correctly under the knowledge point adaptation dimension, the average value method is used for integration to obtain the final probability of answering the next exercise correctly.

[0152] ,

[0153] in, It is the predicted final probability of answering correctly in the next practice session. These are the probabilities of answering correctly in the next practice session, estimated based on ability, the question, and the knowledge points.

[0154] This integration approach treats all three types of prediction results equally, avoiding the bias of a single dimension from dominating the final conclusion. At the same time, it fully integrates the combined influence of the subject's own characteristics, the question, and the knowledge point adaptation. The final output probability provides core data support for the cognitive analysis results, ensuring the comprehensiveness and credibility of the results.

[0155] The results of using this cognitive analysis system include:

[0156] 1. Ability change curves of subjects for each knowledge point: Based on the ability values ​​of each practice session calculated by the ability acquisition module, the dynamic improvement or decline trajectory of subjects in different knowledge points is intuitively displayed, clearly presenting the learning effect;

[0157] 2. Analysis of weaknesses in knowledge points: By combining the knowledge point fit probability and ability change curve, the knowledge points that the subjects have not mastered well are accurately located (e.g., "the ability value of subtraction knowledge point has been lower than 0.4 for a long time").

[0158] 3. The final probability prediction result of answering correctly in the next practice session from the multi-task fusion module: provides a direct basis for question recommendation and learning difficulty adaptation;

[0159] 4. Personalized Question Recommendation Suggestions: Based on the ability change curve, weakness analysis, and final answer probability, questions suitable for the subject's current level are recommended (e.g., basic consolidation questions are recommended for weak knowledge points, and advanced questions are recommended for knowledge points that are well mastered), realizing a closed-loop application of "precise diagnosis + targeted guidance".

[0160] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A cognitive analysis system that focuses on changes in ability and the characteristics of questions, characterized in that, include: The module includes an interactive coding network module, a capability acquisition module, a capability prediction module, a capability rationality module, a question prediction module, a knowledge point prediction module, and a multi-task fusion module. The interactive coding network module is used to encode the question-answer sequence containing question-knowledge point-answer to obtain historical interactive feature coding, which includes ability-based historical interactive feature coding, question-based historical interactive feature coding and knowledge point-based historical interactive feature coding. The ability acquisition module is used to obtain the ability value sequence of the subject after t practice sessions based on the historical interaction feature encoding of the subject's ability class during the t-th practice session. Among them, the ability value sequence This represents the probability of answering a question correctly in each practice session; The ability prediction module is used to combine the ability value sequence after each historical practice session. And the knowledge points involved in the next question The probability of answering the next exercise correctly under the predictive ability dimension; The ability rationality module is used to constrain the rationality of ability value changes by comparing the changes in ability values ​​between different practice sessions and whether the answers are correct. The question prediction module is used to predict the probability of answering the question correctly in the next exercise based on the historical interaction feature encoding of the question type and the question-knowledge point embedding vector of the next exercise. The knowledge point prediction module is used to predict the probability of answering the next exercise correctly under the knowledge point adaptation dimension based on the historical interaction feature encoding of the knowledge point class and the question-knowledge point embedding vector of the next exercise. The multi-task fusion module is used to balance the prediction probabilities obtained by the ability prediction module, question prediction module, and knowledge point prediction module through fusion training to obtain the final probability of answering correctly in the next exercise. The final probability is used for the cognitive analysis of the subject. The total loss function of the multi-task fusion module in the fusion training stage consists of cross-entropy loss based on ability dimension prediction, cross-entropy loss based on question feature dimension prediction, cross-entropy loss based on knowledge point adaptation dimension prediction, and ability rationality loss. The weight parameters of each type of loss are constants between 0 and 1, and their sum is equal to 1. ， in, It is the total loss during the training process. , These are cross-entropy loss methods based on prediction probabilities, namely, ability dimension, question feature dimension, and knowledge point adaptation dimension. It is a loss of reasonableness. , , , They are , , , The corresponding weight parameters are all constants between 0 and 1, and satisfy the following conditions: ; Based on the probabilities of answering correctly in the next practice session under the ability dimension, the probability of answering correctly in the next practice session under the question feature dimension, and the probability of answering correctly in the next practice session under the knowledge point adaptation dimension, the final probability of answering correctly in the next practice session is obtained by integrating the probabilities using the average method: ， in, It is the predicted final probability of answering correctly in the next practice session. These are the probabilities of answering correctly in the next practice session, estimated based on ability, the question, and the knowledge points.

2. The system according to claim 1, characterized in that: The method for obtaining historical interaction feature encoding in the interactive coding network module includes: Calculate the knowledge point vector for the t-th practice session; Based on the knowledge point vector and the corresponding question vector at the t-th practice session, establish the question-knowledge point-answer interaction vector at the t-th practice session; Using an LSTM network, the question-knowledge point-answer interaction vector is encoded to obtain the historical interaction feature encoding.

3. The system according to claim 2, characterized in that: Methods for calculating the knowledge point vector in the t-th exercise include: The sum of all knowledge point embedding vectors corresponding to the question in the t-th practice session is obtained by dividing by the number of knowledge points; wherein, the number of knowledge points is obtained by statistical analysis of the knowledge point index contained in the question.

4. The system according to claim 2, characterized in that: The encoding format of the question-knowledge point-answer interaction vector is as follows: When the answer is correct, the question vector, the knowledge point vector, and a 2d vector of all zeros are concatenated in sequence. When an answer is incorrect, the vector consisting of all zeros, the question vector, and the knowledge point vector are concatenated in sequence. Where d is the vector length of the question and the knowledge point.

5. The system according to any one of claims 2-4, characterized in that: The interactive coding network module contains three LSTM networks with identical structures but independent parameters. The input to each of the three LSTM networks is a question-knowledge point-answer interaction vector, and each network independently calculates its feature encoding using its previous output feature value and gate mechanism parameters. The first LSTM network outputs historical interaction feature encodings for ability categories, the second LSTM network outputs historical interaction feature encodings for question categories, and the third LSTM network outputs historical interaction feature encodings for knowledge point categories.

6. The system according to claim 1, characterized in that: The capability acquisition module contains several fully connected layers, which implement a capability value sequence through progressive processing. Calculation; The ability acquisition module obtains the subject's ability value sequence after t training sessions. The methods include: First, a linear rectified activation operation is performed on the output of the previous layer, then the learnable weight matrix is ​​transformed and a bias term is added, and then a second linear rectified activation operation is performed. Then, the result after the second activation is multiplied element-wise with the feature transformation weight vector, and the result is summed and pooled. Finally, the activation function outputs a normalized sequence of ability values ​​after t training sessions. .

7. The system according to claim 6, characterized in that: The method by which the ability prediction module predicts the probability of answering correctly in the next exercise includes: First, based on the subject's ability value sequence after t training sessions. By performing dot product operations, the ability values ​​corresponding to the knowledge points involved in the next exercise are extracted, and a historical sequence of ability values ​​is constructed. Based on the historical sequence of the ability values, an LSTM network is used, combined with a network gate mechanism to assist in the calculation, to obtain the historical ability feature vector corresponding to the t-th practice session. Based on the historical ability feature vector, after passing through several fully connected layers, and finally through an activation function, the probability of answering correctly in the next exercise under the ability dimension is obtained.

8. The system according to claim 6, characterized in that: The method for judging changes in capability values ​​by the capability rationality module includes: Based on the ability values ​​of two consecutive practice sessions, after eliminating the influence of numerical normalization through the inverse operation of the activation function, the difference between the next practice session and the previous practice session is calculated as the actual ability change value. An actual ability change value greater than 0 corresponds to an increase in ability, and a value less than 0 corresponds to a decrease in ability.

9. The system according to claim 8, characterized in that: In the ability rationality module, when the answer is correct, the ideal ability change value is the product of the one-hot vector of the knowledge point involved in the question and 0.5, superimposed with a vector of the same dimension and all elements of 0.5; when the answer is incorrect, the ideal ability change value is the product of the one-hot vector of the knowledge point involved in the question and -0.5, superimposed with a vector of the same dimension and all elements of -0.

5.

10. The system according to claim 9, characterized in that: The method for constraining changes in capability values ​​by the capability rationality module includes: Based on the actual ability change sequence from the second practice session to the tth practice session, and the corresponding ideal ability change sequence, a mean squared error loss function is used. By minimizing the calculation result of this loss function, the actual ability change value is forced to approach the ideal ability change value, thus ultimately constraining the ability change value.

11. The system according to claim 1, characterized in that: The method used by the question prediction module to predict the probability of answering the next question correctly includes: The question-knowledge point embedding vector of the next practice session is concatenated with the historical interaction feature encoding of the question class to form the question class prediction feature vector. The predicted feature vectors of the question type are input into several fully connected layers, and the probability of answering correctly in the next practice session is output through the activation function.

12. The system according to claim 1, characterized in that: The knowledge point prediction module uses the following methods to predict the probability of answering the next question correctly: The knowledge point embedding vector of the next practice question is concatenated with the historical interaction feature encoding of the knowledge point class to form the knowledge point class prediction feature vector. The prediction feature vectors of knowledge points are input into several fully connected layers, and the prediction accuracy of the subjects for each knowledge point is output through the activation function. Calculate the average of all prediction accuracy rates to obtain the probability of answering the next exercise correctly under the knowledge point adaptation dimension.

Images