Simulation enhanced cognitive recommendation method and system based on large language model
Through simulation based on large language model, more static simulation data are generated, which solves the sparse problem of knowledge tracking data set, improves the prediction accuracy and robustness of the model, and solves the coldness of the knowledge tracking model Startup problem.
Patent Information
- Application Number
- CN202411619004.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-13
- Publication Date
- 2025-02-07
AI Technical Summary
There is a sparse problem with the knowledge tracking data set, which makes it difficult for the model to learn effective patterns or inference rules from the data, which in turn affects the prediction effect of unseen knowledge points.
Using simulation-enhanced cognitive recommendation method based on large language models, more static simulation data is generated through static simulation data generation module and KT dynamic proxy module, and using adversarial learning and Transformer Encoder architecture to enhance the generalization and robustness of the model. .
By generating more static simulation data, the sparse problem of knowledge tracking data set is solved, the prediction accuracy of the model is improved, and the cold start problem of knowledge tracking model is solved.
Smart Images
Figure CN119398177A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a simulation-enhanced cognitive recommendation method and system based on a large language model. Background Art
[0002] Knowledge tracking is based on modeling of student behavior sequences to predict students' mastery of knowledge. Knowledge tracking is the core and key to building an adaptive education system. In an adaptive education system, whether it is precise push, student learning path planning or knowledge graph construction, the first step is to accurately predict students' mastery of knowledge.
[0003] The sparsity problem of knowledge tracing datasets refers to the imbalance and incompleteness of various knowledge relations in the dataset in the knowledge tracing task. Specifically, this may include the imbalance in the number of samples between certain knowledge relations and the lack of sample support for some potential knowledge relations in the dataset.
[0004] The main reasons for the sparsity problem are as follows: 1. Difficulties in data collection and labeling: Knowledge tracking data needs to be collected from multiple sources and multiple fields, involving a large amount of information extraction and relationship labeling work, which is challenging in terms of manpower and time costs. Therefore, some knowledge relationships may not have sufficient data sample support. 2. Data bias: In the real world, some knowledge relationships may be more common, so relevant data samples are easier to obtain. Relatively speaking, some rarer knowledge relationships may be more difficult to obtain a sufficient number of samples, resulting in sparsity problems in the data set. 3. Incomplete knowledge base: The knowledge tracking task relies on multiple knowledge bases and databases for knowledge extraction and relationship mining, but these knowledge bases may contain missing or erroneous information. Therefore, some real knowledge relationships may be omitted and cannot form complete data support.
[0005] For these reasons, the knowledge relations in knowledge tracking datasets are often sparse, which brings challenges to the training and evaluation of knowledge tracking models. For example, it has a negative impact on the performance of knowledge tracking tasks, or lacks sufficient data support. The model may find it difficult to learn effective patterns or reasoning rules from the data, resulting in poor performance when predicting unseen knowledge points. Current knowledge tracking real-world datasets, such as Assistment2009 and Assistment2017, all have the above problems. Summary of the invention
[0006] Based on the technical problems existing in the background technology, the present invention proposes a simulation enhanced cognitive recommendation method and system based on a large language model to solve the problem of sparse data sets and the cold start problem of the knowledge tracking model.
[0007] The present invention proposes a simulation-enhanced cognitive recommendation method and system based on a large language model, which is used to provide data support for the sparse problem of knowledge tracking data sets. The specific method is as follows:
[0008] Step 1: In the static simulation data generation module, the real data (X, Y) obtained from the real-world data set is input into the large language model to output the generated data (X′, Y′, flag), where flag is the label of whether the generated data is real data. The real data label is set for the intercepted data of the same length directly intercepted from the real-world data set, and the simulation data label is set for the simulation data generated by discretely selecting the question information and sorting it by itself. X is the question number, and Y is the corresponding answer situation;
[0009] Step 2: Input the real data features corresponding to the real data and the generated data features corresponding to the generated data into the discriminator to output the probability of whether the data in the generated data is real data. Through the constructed static simulation loss, first fix the large language model, train the discriminator, and then fix the trained discriminator to train the large language model.
[0010] Step 3: Input the real data (X, Y) and the generated data (X′, Y′, flag) into the KT dynamic agent module respectively, output the answer prediction result at the next moment, and construct the dynamic agent loss based on the answer prediction result and the real answer to train the KT dynamic agent module.
[0011] Furthermore, in step 2, the static simulation loss L DG as follows:
[0012] min G max D V(D,G)=E z～pdata (x)[logD(x)]
[0013] +E z～pz (z)[log(1-D(G(z)))]
[0014] Among them, G is the generator, corresponding to the large language model, D is the discriminator, x represents the data sampled from the real data distribution, represents the real data; z represents the noise variable, from the noise distribution p z The sample is used as the input of the generator G, and the generator G(z) generates simulation data; E z～pdata (·) represents the expectation of the sample x obtained by sampling the real data distribution pdata; E z～pz (·) represents the noise distribution p zThe expected value of is used to measure the performance of the generator under different noise inputs; V(·) represents the target loss function, which is used to describe the game process between the discriminator D and the generator G, and optimizes min G max D V(D,G) is used to train the generator and discriminator.
[0015] Furthermore, in step 2, the real data features corresponding to the real data and the generated data features corresponding to the generated data are specifically:
[0016] Feat=f((Y+1)*X)
[0017] Feat′=f((Y′+1)*X′)
[0018] Among them, Feat is the real data feature, Feat′ is the generated data feature, and * is the dot product.
[0019] Furthermore, in step three, the KT dynamic agent module includes a two-layer Transformer Encoder architecture, a prediction layer, a reward layer, a weighted output layer, and a discriminant loss;
[0020] At time t, the question sequence and answer sequence in the real data and generated data are encoded by the first layer encoder of the two-layer Transformer Encoder architecture to obtain the question embedding q t , interactive embedded qa t ;
[0021] The second layer encoder of the two-layer Transformer Encoder architecture embeds the topic into q t , interactive embedded qa t Encode and obtain the latent variable h of the degree of knowledge mastery at time t t ;
[0022] The hidden variable h t and the topic embedding q at time t+1 t+1 After splicing, it is sent to the prediction layer to predict the probability of the student answering the question correctly at time t+1
[0023] The hidden variable h t , the latent variable h of the knowledge mastery at time t-1 t-1 and interactive embedded qa t The state change variable h obtained after splicing * Input to the reward layer to output the reward score at time t The final reward layer outputs the reward scores at different times and arranges them in descending order, and recommends the question numbers corresponding to the top s reward scores to the real-world dataset of the static simulation data generation module;
[0024] Based on probability and bonus points Add together to get the answer prediction result at time t+1
[0025] Furthermore, the answer prediction results The details are as follows:
[0026]
[0027] Among them, γ represents a constant parameter used to balance the reward score The predicted probability of the next moment The weight of .
[0028] Furthermore, in step three, based on the dynamic agent loss KT dynamic agent module, for real data, the standard loss is used to evaluate the prediction accuracy; for generated data, sequence consistency or adversarial loss is introduced to enhance the generalization and robustness of the model.
[0029] A simulation-enhanced cognitive recommendation system based on a large language model, used to provide data support for the sparsity problem of knowledge tracking datasets, including a static simulation data generation module and a KT dynamic agent module;
[0030] In the static simulation data generation module, the real data (X, Y) obtained from the real-world data set is input into the large language model to output the generated data (X′, Y′, flag), where flag is the label of whether the generated data is real data. The real data label is set for the intercepted data of the same length directly intercepted from the real-world data set, and the simulated data label is set for the simulated data generated by discretely selecting the question information and sorting it by itself, where X is the question number and Y is the corresponding answer situation; the real data features corresponding to the real data and the generated data features corresponding to the generated data are input into the discriminator to output the probability of whether the data in the generated data is real data. Through the constructed static simulation loss, the large language model is first fixed to train the discriminator, and then the trained discriminator is fixed to train the large language model;
[0031] The real data (X, Y) and generated data (X′, Y′, flag) are respectively input into the KT dynamic agent module, and the answer prediction result at the next moment is output. Based on the answer prediction result and the real answer, the dynamic agent loss is constructed to train the KT dynamic agent module.
[0032] The advantages of the simulation-enhanced cognitive recommendation method and system based on a large language model provided by the present invention are: the simulation-enhanced cognitive recommendation method and system based on a large language model provided in the structure of the present invention is a method for solving the sparsity problem of knowledge tracking data sets based on adversarial learning of a large language model. By combining the powerful generation ability of the large language model with adversarial learning, more static simulation data is generated for knowledge tracking through sparse, long-tailed knowledge tracking real-world data sets. This method enables the knowledge tracking model to obtain a large amount of data, solving the sparsity problem of the data set and the cold start problem of the knowledge tracking model. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a schematic diagram of the structure of the present invention. DETAILED DESCRIPTION
[0034] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.
[0035] This embodiment proposes a method for solving the sparsity problem of knowledge tracking datasets based on adversarial learning of large language models in the online learning scenario of teenagers. Using sparse, long-tail real-world datasets for knowledge tracking, the strong generation ability of large language models and adversarial learning of GAN networks, and combining the dynamic information of KT models during training, the large language model generates appropriate, dense knowledge tracking data, as follows.
[0036] like Figure 1 As shown, the present invention proposes a simulation-enhanced cognitive recommendation method based on a large language model, which is used to provide data support for the sparsity problem of knowledge tracking data sets. The specific method is as follows:
[0037] Step 1: In the static simulation data generation module, the real data (X, Y) obtained from the real-world data set is input into the large language model to output the generated data (X′, Y′, flag), where flag is the label of whether the generated data is real data. The real data label is set for the intercepted data of the same length directly intercepted from the real-world data set, and the simulation data label is set for the simulation data generated by discretely selecting the question information and sorting it by itself. X is the question number, and Y is the corresponding answer situation;
[0038] Get real data (X, Y) from the real-world dataset or from the recommendation part of the KT dynamic agent module. Its size is (batchsize, seqlen) and its data format is (X1, Y1)(X2, Y2), ···, (X i ,Y i ),···,(X seqlen ,Y seqlen ), X i is the question number, Y i ∈{0,1} is the corresponding correct or incorrect answer. Batch size is the amount of data input to the model for training each time, and seq_len refers to the length of each input sequence, that is, the number of time steps contained in each input sequence.
[0039] In knowledge tracing adversarial learning, a pair of data contains (X, Y), (X, Y) ∈ R, where R is a knowledge tracing real-world dataset, and its feature space is defined as: F = {f: D → R}, which is modified and maintained by the knowledge tracing process. The following operations are taken to obtain adversarial sample features:
[0040] Feat = f((Y+1)*X);
[0041] Considering that parallel processing of information can improve efficiency, assuming that a batch of sequences are obtained at one time, the above features are concatenated, that is, Feats = concat (Featall, dim = 0), and a feature matrix with a shape of (batchsize, dimemb) is obtained as the real data feature, that is, the adversarial sample feature.
[0042] The real data (X, Y) is input into the large language model, which aims to simulate the student interaction data of the data set through the student interaction information in the real-world data set, thereby improving the sparsity problem and a large number of long-tail distributions of the real-world data set in the knowledge tracking direction. During the fine-tuning process, it is necessary to set a reasonable prompt to understand the input of the data, regularize the output, and finally adjust the model through the output and label.
[0043] In this embodiment, the prompt template is designed as:
[0044] (Identity confirmation): Suppose you are a teacher of subject xx with rich teaching experience.
[0045] (Task determination): Now you have a set of samples (X, Y), and you need to filter or extract the same number of questions X′ and corresponding answer situations Y′ from the real-world dataset and sort them.
[0046] (Task requirements): It is required to deceive the discriminant model as much as possible to think that the generated data (X′, Y′)
[0047] It is from real-world datasets.
[0048] (Enter OK): Now we pass you data from the real world in the following format:
[0049] <(X1,Y1)(X2,Y2),···,(X i ,Y i ),···,(X seqlen ,Y seqlen ), X i is the question number, Y i ∈{0,1} is the corresponding correct or incorrect answer>;
[0050] Please complete the above tasks.
[0051] (Output determination): If the question information is discretely selected and sorted by itself, then the output is (X′, Y′, flags = 1). If a real-world data set of the same length as the sample is randomly intercepted from the real-world data set, then the output is (X′, Y′, flag = 0). That is, when generating data, there are two types of data generated from the real data (X, Y). One is to randomly shuffle the real data and set the flag to 1, and the other is simulated data generated by the generator (large language model) and the flag is set to 0. These three types of data (real data, generated data with flag = 1, and generated data with flag = 0) are used as inputs in the KT dynamic agent module.
[0052] As can be seen above, the large language model LLM will generate a set of samples (X′, Y′) of the same size as the adversarial sample (X, Y), and project them into the feature space in the same way: Feat′ = f((Y′+1)*X′). The last flag in its output represents whether the sample is real data, 1 for real data and 0 for simulated data, that is:
[0053] Real data features: Feat = f((Y+1)*X);
[0054] Generate data features: Feat′=f((Y′+1)*X′).
[0055] Since Y∈{0,1} is the information of whether the answer is right or wrong, 0 represents an incorrect answer and 1 represents a correct answer; X is the question number information of the question. Through the above Feat operation, it is assumed that the wrong questions retain the original question number, while the correct questions have twice the original question number. After projecting to the feature space, different features of the same question can be obtained for correct and incorrect answers.
[0056] Step 2: Input the real data features corresponding to the real data (X, Y) and the generated data features corresponding to the generated data (X′, Y′, flags) into the discriminator to output the probability of whether the data in the generated data is real data. Through the constructed static simulation loss, first fix the large language model, train the discriminator, and then fix the trained discriminator to train the large language model;
[0057] The discriminator is a binary classifier that determines whether the data generated by the large language model is real data. The generated data feature Feat′ output by the large language model and the real data feature Feat of the sample are input into the discriminator. The final output of the discriminator is:
[0058] p=sigmoid(Discriminator(concat(Feat,Feat′)))
[0059] Among them, p∈[0,1] is the probability of whether the data in the generated data is real data, concat is the concatenation layer, Discriminator is the discriminator, and sigmoid is the activation function.
[0060] Finally, in the static simulation data generation module, the static simulation loss L DG for:
[0061] min G max D V(D,G)=E z～pdata (x)[logD(x)]+E z～pz (z)[log(1-D(G(z)))]; (1)
[0062] Among them, G is the generator, corresponding to the large language model, D is the discriminator, x represents the data sampled from the real data distribution, represents the real data; z represents the noise variable, from the noise distribution p z The sample is used as the input of the generator G, and the generator G(z) generates simulation data; E z～pdata (·) represents the expectation of the sample x obtained by sampling the real data distribution pdata; E z～pz (·) represents the noise distribution p z The expected value of is used to measure the performance of the generator under different noise inputs; V(·) represents the target loss function, which is used to describe the game process between the discriminator D and the generator G, and optimizes min G max D V(D,G) is used to train the generator and discriminator.
[0063] In the training process of the static simulation data generation module, the discriminator is trained first to ensure that the discriminator can distinguish false samples. When the discriminator is trained to a certain level, the generator is trained. At this time, the discriminator is fixed. If the generator is stronger, the discriminator will make a wrong judgment, and the value of the entire formula (1) will become smaller.
[0064] Step 3: Input the real data (X, Y) and generated data (X′, Y′, flag) into the KT dynamic agent module respectively, and output the prediction result of the answer at the next moment. Based on the prediction result and the actual answer situation, construct a dynamic agent loss for training to train the KT dynamic agent module. For real data, standard loss is used to evaluate the prediction accuracy; for generated data, sequence consistency or adversarial loss is introduced to enhance the generalization and robustness of the model.
[0065] For real data, flag = None, then follow the conventional knowledge tracking process, and then use the predicted value and label to calculate the cross entropy loss; for generated data, flag ≠ None, then flag has two forms, one is flag = 1, at this time the data is generated by shuffling the real data, and the other is flag = 0, at this time the data is generated by the generator (large language model).
[0066] In the KT dynamic agent module, the question recommendation model is used as the agent, the student's ability level is used as the reward, the knowledge tracking process is the environment, and the recommendation of questions is the agent's behavior. In reinforcement learning, the agent recommends questions in order to continuously improve the student's ability level.
[0067] The KT dynamic agent module has two major purposes. One is to train the knowledge tracking model to have a higher prediction accuracy. The other is to train the reinforcement learning agent, namely the question recommendation model, to guide the large language model to generate simulation questions.
[0068] The KT dynamic agent module includes a two-layer Transformer Encoder architecture, a prediction layer, a reward layer, a weighted output layer, and a discriminant loss;
[0069] (1) At time t, the input question sequence and answer sequence in the real data and generated data are encoded by the first layer encoder of the two-layer Transformer Encoder architecture to obtain the question embedding q t , interactive embedded qa t ;
[0070] The real data (X, Y) and generated data (X′, Y′, flag) are input into the KT dynamic proxy module. Taking time t as an example, the title embedding q corresponding to the real data is obtained through the first layer encoder. t , interactive embedded qa t, and the topic embedding q′ corresponding to the generated data t , interactive embedding qa′ t .
[0071] The KT dynamic agent module embeds the question into q t , interactive embedded qa t , and the topic embedding q′ t , interactive embedding qa′ t The data processing process is consistent with that of t , interactive embedded qa t The output processing process of the question is explained, and the question is embedded in q′ t , interactive embedding qa′ t The data processing process can be directly referenced.
[0072] (2) The second layer encoder of the two-layer Transformer Encoder architecture embeds the question into q t , interactive embedded qa t Encode and obtain the latent variable h of the degree of knowledge mastery at time t t ;
[0073] The second-level encoder processes the two types of personalized codes and finally obtains the students' personalized ability levels.
[0074] (3) The hidden variable h t and the topic embedding q at time t+1 t+1 After splicing, it is sent to the prediction layer to predict the probability of the student answering the question correctly at time t+1
[0075] (4) The hidden variable h t , the latent variable h of the knowledge mastery at time t-1 t-1 and interactive embedded qa t The state change variable h obtained after splicing * Input to the reward layer to output the reward score at time t The final reward layer outputs the reward scores at different times and arranges them in descending order, and recommends the question numbers corresponding to the top s reward scores to the real-world dataset of the static simulation data generation module;
[0076] The reward function of the reward layer in this embodiment is different from the commonly used predefined reward function. The instant reward function layer is designed as follows: R(h*)=Sigmoid(W r (h * )+b r), the final reward layer outputs the reward scores at different times. Recommended topics will be recommended based on the reward scores, and topics with high reward scores will be recommended to the static simulation data generation module.
[0077] (5) Based on probability and bonus points Add together to get the answer prediction result at time t+1 Will Compared with the true label, the loss is set to BCE binary cross entropy During the training process, L is continuously optimized;
[0078]
[0079] L=αL DG +βL R ;
[0080] Among them, α and β are balance coefficients, L DG is the static simulation loss, and γ represents a constant parameter used to balance the reward score The predicted probability of the next moment The weight of γ is usually adjusted through experiments to achieve the best balance. According to the needs of actual tasks, the common value range is between [0,1]. The specific value can be verified and adjusted according to the model performance.
[0081] Through steps one to three, this embodiment provides a method for solving the sparsity problem of knowledge tracking datasets based on adversarial learning of large language models in the online learning scenario of teenagers. By combining the powerful generation ability of large language models with adversarial learning, more static simulation data is generated for knowledge tracking through sparse, long-tailed knowledge tracking real-world datasets. This method enables the knowledge tracking model to obtain a large amount of data, solving the sparsity problem of datasets and the cold start problem of the knowledge tracking model.
[0082] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A simulation-enhanced cognitive recommendation method based on a large language model, characterized in that: It is used to provide data support for the sparse problem of knowledge tracking datasets. The specific method is as follows: Step 1: In the static simulation data generation module, the real data (X, Y) obtained from the real-world data set is input into the large language model to output the generated data (X′, Y′, flag), where flag is the label of whether the generated data is real data. The real data label is set for the intercepted data of the same length directly intercepted from the real-world data set, and the simulation data label is set for the simulation data generated by discretely selecting the question information and sorting it by itself. X is the question number, and Y is the corresponding answer situation; Step 2: Input the real data features corresponding to the real data and the generated data features corresponding to the generated data into the discriminator to output the probability of whether the data in the generated data is real data. Through the constructed static simulation loss, first fix the large language model, train the discriminator, and then fix the trained discriminator to train the large language model. Step 3: Input the real data (X, Y) and the generated data (X′, Y′, flag) into the KT dynamic agent module respectively, output the answer prediction result at the next moment, and construct the dynamic agent loss based on the answer prediction result and the real answer to train the KT dynamic agent module.
2. The method for simulation-enhanced cognitive recommendation based on a large language model according to claim 1, characterized in that: In step 2, the static simulation loss L DG as follows: min G max D V(D,G)=E z～pdata (x)[logD(x)]+E z～pz (z)[log(1-D(G(z)))] Among them, G is the generator, corresponding to the large language model, D is the discriminator, x represents the data sampled from the real data distribution, represents the real data; z represents the noise variable, from the noise distribution p z The sample is used as the input of the generator G, and the generator G(z) generates simulation data; E z～pdata (·) represents the expectation of the sample x obtained by sampling the real data distribution pdata; E z～pz (·) represents the noise distribution p z The expected value of is used to measure the performance of the generator under different noise inputs; V(·) represents the target loss function, which is used to describe the game process between the discriminator D and the generator G, and optimizes min G max D V(D,G) is used to train the generator and discriminator.
3. The method for simulation-enhanced cognitive recommendation based on a large language model according to claim 1, characterized in that: In step 2, the real data features corresponding to the real data and the generated data features corresponding to the generated data are specifically: Feat=f((Y+1)*X) Feat′=f((Y′+1)*X′) Among them, Feat is the real data feature, Feat′ is the generated data feature, and * is the dot product.
4. The method for simulation-enhanced cognitive recommendation based on a large language model according to claim 1, characterized in that: In step 3, the KT dynamic agent module includes a two-layer Transformer Encoder architecture, a prediction layer, a reward layer, a weighted output layer, and a discriminant loss; At time t, the question sequence and answer sequence in the real data and generated data are encoded by the first layer encoder of the two-layer TransformerEncoder architecture to obtain the question embedding q t , interactive embedded qa t ; The second layer encoder of the two-layer Transformer Encoder architecture embeds the topic into q t , interactive embedded qa t Encode and obtain the latent variable h of the degree of knowledge mastery at time t t ; The hidden variable h t and the topic embedding q at time t+1 t+1 After splicing, it is sent to the prediction layer to predict the probability of the student answering the question correctly at time t+1 The hidden variable h t , the latent variable h of the knowledge mastery at time t-1 t-1 and interactive embedded qa t The state change variable h obtained after splicing * Input to the reward layer to output the reward score at time t The final reward layer outputs the reward scores at different times and arranges them in descending order, and recommends the question numbers corresponding to the top s reward scores to the real-world dataset of the static simulation data generation module; Based on probability and bonus points Add together to get the answer prediction result at time t+1 5. The method for simulation-enhanced cognitive recommendation based on a large language model according to claim 4, characterized in that: Answer prediction results The details are as follows: Among them, γ represents a constant parameter used to balance the reward score The predicted probability of the next moment The weight of .
6. The method for simulation-enhanced cognitive recommendation based on a large language model according to claim 4, characterized in that: In step three, based on the dynamic agent loss KT dynamic agent module, for real data, the standard loss is used to evaluate the prediction accuracy; for generated data, sequence consistency or adversarial loss is introduced to enhance the generalization and robustness of the model.
7. A simulation-enhanced cognitive recommendation system based on a large language model, characterized in that: Used to provide data support for the sparse problem of knowledge tracking data sets, including static simulation data generation module and KT dynamic agent module; In the static simulation data generation module, the real data (X, Y) obtained from the real-world data set is input into the large language model to output the generated data (X′, Y′, flag), where flag is the label of whether the generated data is real data. The real data label is set for the intercepted data of the same length directly intercepted from the real-world data set, and the simulated data label is set for the simulated data generated by discretely selecting the question information and sorting it by itself, where X is the question number and Y is the corresponding answer situation; the real data features corresponding to the real data and the generated data features corresponding to the generated data are input into the discriminator to output the probability of whether the data in the generated data is real data. Through the constructed static simulation loss, the large language model is first fixed to train the discriminator, and then the trained discriminator is fixed to train the large language model; The real data (X, Y) and generated data (X′, Y′, flag) are respectively input into the KT dynamic agent module, and the answer prediction result at the next moment is output. Based on the answer prediction result and the real answer, the dynamic agent loss is constructed to train the KT dynamic agent module.