A meta-cognition driven heuristic hint question recommendation method

Abstract

This invention discloses a metacognitive-driven heuristic prompt question recommendation method. Its key feature is that this method employs a recommendation system constructed from a knowledge tracking module based on emotion and cognitive style, a personality-enhanced error cause analysis module, and a metacognitive strategy-driven heuristic prompt question recommendation module. The knowledge tracking module uses an LSTM model and emotion and cognitive style auxiliary tasks to obtain a vector representation of the user's current knowledge level. The error cause analysis module obtains error cause representations through an attention mechanism. The heuristic prompt question recommendation module uses a large language model to generate a selection of heuristic prompt questions, ranks them, and recommends the highest-ranked heuristic questions to the user. Compared with existing technologies, this invention provides more accurate and interpretable recommendation results, recommends higher-quality heuristic prompt questions to users, improves users' metacognitive abilities, and thus achieves a synergistic enhancement of human-machine metacognition.

Description

Technical Field

[0001] This invention relates to the field of dialogue recommendation task technology, specifically a metacognitive-driven heuristic question recommendation method. Background Technology

[0002] Conversational recommendation is a crucial task in recommender systems. Its purpose is to enhance the transparency of the recommender system to users and increase user control over the system through real-time, multi-round dialogue interactions. Extracting users' true needs by combining historical behavior and dialogue feedback becomes particularly important. Taking exercise recommendation as an example, a conversational recommendation system can provide exercise recommendations that better match the user's current developmental needs based on user feedback, thereby improving the user's problem-solving experience and ability.

[0003] In early research on recommender systems, researchers used implicit user feedback to learn user preferences, such as clicks and favorites. While implicit feedback data is easier to collect, this feedback may stem from users' fleeting interests or accidental actions, adding noise to user behavior modeling. To increase explicit user feedback to recommender systems, researchers have adopted interactive recommendation methods, such as dialogue-based recommendations, allowing users to engage in real-time conversations with the machine. This enables more personalized and accurate suggestions based on user feedback and further questions.

[0004] On the one hand, dialogue recommendation systems have limitations in understanding and meeting user needs. Traditional dialogue recommendation systems learn user preferences based on the assumption that users always know what they want, but in reality, users sometimes don't even know what they want. Ignoring user uncertainty and the variability of needs can easily lead to poor recommendation performance. Therefore, considering user experience, dialogue recommendation systems also need to guide topics and influence user psychology. Adopting exploratory recommendation strategies and presenting users with diverse heuristic prompts can help users improve their metacognition and achieve personal development.

[0005] On the other hand, in traditional dialogue recommendation, due to a lack of basic common sense and emotion, machine-generated language often differs from human-generated language in some aspects. Furthermore, machines tend to generate safe, universally applicable answers, lacking personalization and authenticity. With the development of large language models, generative dialogue recommendation systems have made significant progress in providing users with a smooth dialogue experience. One advantage of using large language models for dialogue recommendation is that they exhibit naturalness and unpredictability in dialogue; however, when providing heuristic prompts to users, the prompts may lack precision. Therefore, improving the machine's metacognitive level is crucial. Improving the machine's metacognitive level through reinforcement learning based on user feedback can provide better recommendations, achieving a synergistic improvement in human-machine metacognition and resulting in a better dialogue and recommendation experience. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a metacognitive-driven heuristic question recommendation method. This dialogue recommendation method employs human-computer metacognition as its driving force, utilizing the user's current dialogue and historical question-answering records to obtain the user's current knowledge level. Then, based on the current knowledge level, it predicts the knowledge points the user has not yet mastered. Furthermore, it analyzes the reasons for errors in questions answered incorrectly based on the user's personality. Finally, with the goal of improving the user's metacognitive level, it provides heuristic prompts and question recommendations. The entire process uses a reinforcement learning framework to synergistically improve the metacognitive levels of both the user and the system. This method has a stronger dynamic grasp of knowledge level and error reasons, resulting in more accurate and interpretable recommendation results. Experimental results on real datasets demonstrate its effectiveness and higher accuracy.

[0007] The technical solution to achieve the purpose of this invention is: a metacognitive-driven heuristic prompt question recommendation method, characterized by employing a recommendation system constructed from a knowledge tracking module based on emotion and cognitive style, a personality-enhanced error analysis module, and a metacognitive strategy-driven heuristic prompt question recommendation module. This system achieves heuristic prompt question recommendation. The emotion- and cognitive style-based knowledge tracking module extracts the user's cognitive style and knowledge mastery level from the user's historical question-answering records and extracts the user's current emotion from the current dialogue, constructing a knowledge level representation based on emotion and cognitive style. The personality-based error analysis module extracts the user's personality from historical dialogues and uses an attention mechanism to capture the relationship between personality and incorrect questions at the current knowledge level, constructing a personality-based error representation. The metacognitive strategy-driven heuristic prompt question recommendation module, based on the error representation, uses a large language model to generate a series of selectable heuristic prompt questions, uses a graph neural network to rank the generated prompt questions, and continuously adjusts the prompt strategy within a reinforcement learning framework to synergistically improve the human-computer metacognitive level. This method specifically includes the following steps:

[0008] Step 1: Obtain the user's historical test-taking history and conversation interactions.

[0009] User u's historical problem-solving activities are a sequence, represented as follows: Where, q i Let a represent the i-th question the user answers, expressed as a k-dimensional vector. i The answer to the i-th question is represented by a one-dimensional vector. A list of statements is used. This represents a dialogue interaction, where l k This refers to the utterance in the k-th round of dialogue.

[0010] Step 2: Generate implicit representations of users' emotions and cognitive styles

[0011] Using an n-dimensional vector E = {e1, e2, ..., e} n} and m-dimensional vector C = {c1, c2, ..., c m The user's emotion representation E and cognitive representation C are concatenated to obtain the emotion and cognitive style representation E. * Where n and m both take values ​​between 100 and 200, and e i c is the saliency of the i-th mental lexicon category. j It is the saliency of the j-th cognitive style category.

[0012] The emotion representation E is calculated from the user's current dialogue text information. Specifically, the TF-IDF algorithm is used to extract emotion features. For each user, the occurrence count W of the i-th mental lexicon category word is obtained based on the current dialogue. i And the number of discourses D containing words of this mental lexicon category. i First, use the total number of words in the current dialogue, n w Calculate the word frequency (TF) of category words in a mental lexicon. i Specifically, as shown in equation (ii) below:

[0013]

[0014] Next, the inverse text frequency df is calculated based on the number of utterances n1 in the current dialogue, as shown in equation (iii) below:

[0015]

[0016] Thus, the emotion representation E of user u is obtained, as shown in equation (a) below:

[0017] E = {e1,e2,…,e} n}=tf i ×idf i ×{R1,R2,…,R n} (a);

[0018] Among them, R i It is the correlation factor between the i-th psychological vocabulary category and emotion.

[0019] The cognitive representation C is obtained from the questionnaire results through one-hot encoding. The user's emotional representation and cognitive representation are concatenated to obtain the final implicit representation E of emotion and cognitive style. * Specifically, as shown in equation (b) below:

[0020] E * =concat(E,C) (b).

[0021] Here, concat represents the concatenation operation.

[0022] Step 3: Generate user personality representation

[0023] Using an n-dimensional vector P = {p1, p2, ..., p n As a representation of user personality, n takes a value between 100 and 200, and e i is the saliency of the i-th mental lexicon category. The personality representation P is calculated from the user's historical dialogue text information. The calculation method is the same as that of the emotion representation E in step two. The difference is that the current dialogue pattern reflects short-term emotional characteristics, while the long-term dialogue pattern reflects personality characteristics.

[0024] Specifically, for each user, the number of occurrences W of the i-th mental vocabulary category word is obtained based on the historical dialogue. i And the number of discourses D containing words of this mental lexicon category. i First, use the total number of words N in the historical dialogue. w Calculate the word frequency (TF) of category words in a mental lexicon. i Specifically, as shown in equation (ii) below:

[0025]

[0026] Then, based on the number of words N in the historical dialogue... d Calculate the inverse text frequency (idf) i Specifically, as shown in equation (iii) below:

[0027]

[0028] Thus, the personality representation P of user u is obtained, specifically expressed as follows (i):

[0029] P = {p1, p2, ..., p} n}=tf i ×idf i ×{R1,R2,…,R n} (i).

[0030] Among them, R i It is the correlation factor between the i-th psychological vocabulary category and emotion.

[0031] Step 4: Representation of knowledge level based on emotion and cognitive style

[0032] The first step is to generate a difficulty representation for the question based on the difficulty of the knowledge points it contains and the characteristics of the question itself. Specifically, the objective difficulty of question j is first represented as D. j The calculation is expressed as shown in equation (iv) below:

[0033] D j =∑ i∈K(j) α i +β j (iv)

[0034] Here, K(j) is the set of knowledge points involved in question j, and each knowledge point i has a corresponding difficulty α. i ;β j The difficulty level is inherent to the questions and is unrelated to the specific knowledge points covered.

[0035] Next, considering the current state of user u, the difficulty of problem j for user u is represented as D. u,j The calculation is expressed as shown in equation (c) below:

[0036]

[0037] in, It is based on the implicit representation of user emotion and cognitive style obtained in step two, that is, the user's current state, θ in It is an input integration function that integrates the user's current state into the objective difficulty of the question.

[0038] The second step is to use the problem-solving sequence obtained in step one. The LSTM model is used to obtain the user's current knowledge level. Specifically, firstly, the question q is... t The corresponding question difficulty, represented by D, incorporates the user's current mood and cognitive style. u,j Answer indicates a t By concatenating the vectors, we obtain the input vector x of the LSTM model. t That is, it can be expressed by the following equation (d):

[0039] x t =concat(D u,j ,a t (d).

[0040] Then, calculate the unit state c. t Specifically, it is expressed as follows in equation (e):

[0041]

[0042] Among them, f t For the forgetting gate, control c t-1 Used to update c t Features; i t For input gates, control Used to update c t Features; The cell state update value is represented by equations (v) to (vii) below:

[0043] f t =σ(W f [x t ,h t-1 ]+b f (v);

[0044] i t =σ(W i [x t ,h t-1 ]+b i (vi);

[0045]

[0046] Where W and b represent the weight matrix and bias, respectively, and the same applies below.

[0047] Finally, calculate the output gate o. t Use output gate o t and unit state c t Calculate the current knowledge level h t Specifically, as shown in equations (viii) and (f) below:

[0048] o t =σ(W o [x t ,h t-1 ]+b o (viii);

[0049] h t =o t ×tanh(c t (f).

[0050] Step 5: Predict the questions you haven't mastered

[0051] Based on the current knowledge level representation h obtained in step four t Then, the student's level of understanding of the question can be obtained, specifically expressed as formula (g) below:

[0052] y t=σ(W y h t +b y (g).

[0053] Where σ is the sigmoid function, y t The length of the question is equal to the length of the problem.

[0054] The training objective is the negative log-likelihood of the student's answer sequence, using binary cross-entropy loss. Calculate the loss function Specifically, it is expressed as follows (h):

[0055]

[0056] Step Six: Error Analysis Based on User Personality

[0057] Based on the prediction of unmastered questions in step five, questions are recommended to users. When users answer questions incorrectly, error analysis is performed based on the user personality representation P obtained in step three.

[0058] The first step involves preprocessing incorrect questions, personality traits, and reasons for errors. Specifically, on one hand, different types of explainable error reason information are collected to generate pre-defined explainable error reason information. An incorrect question-personality matrix N is used to store the relationship between incorrect questions and personality traits. The matrix is ​​constructed as follows: when a user u with personality trait P answers question q incorrectly, N is incremented. q,p The value of . A personality-fault factor matrix M is used to store the relationship between personality and fault factors, where the elements M p,b The degree of correlation between personality type P and the cause of error b.

[0059] On the other hand, using matrix K key Store knowledge points, where each column represents a knowledge point; use matrix K. ealue Store knowledge states, each column represents K key The degree of correlation between the corresponding knowledge points and the reasons for the errors.

[0060] The second step is to represent the user's current incorrect question as q. t The inner product of the incorrect question-personality matrix N is taken to obtain the personality p′ that is most likely to answer the question incorrectly. Then, the inner product of the obtained personality p′ and the user's own personality P is taken to obtain the matching degree between the two, as specifically expressed by the following formula (m):

[0061] λ=p T p′ (m).

[0062] Among them, a higher matching degree λ indicates a greater contribution of personality to the cause of error; p T Let P be the transpose of the personality representation.

[0063] The third step is to determine the error cause T, which is caused by both the user's knowledge level and personality. The calculation is expressed as follows (1):

[0064] T = λT1 + (1-λ)T2 (1).

[0065] Where T1 and T2 are the error causes caused by the user's personality and knowledge level, respectively, and λ is the degree of personality matching calculated in the second step.

[0066] The error T1 caused by personality is obtained from the personality-error matrix M, and the error T2 caused by knowledge level is calculated by the attention mechanism. Specifically, the question is first represented as q. t With knowledge point matrix K key The inner product of each knowledge point is calculated, and after passing it through an activation function, the relevance between the question and the knowledge point is obtained, i.e., the attention weight w. t Specifically, it is expressed as follows in equation (j):

[0067] w t =Softmax(q t K key (j).

[0068] Therefore, the error T2 that may arise from the question itself is calculated based on the attention weight, as specifically expressed by the following formula (k):

[0069] T2 = w t K value (k).

[0070] Among them, K value K represents key The degree of correlation between the corresponding knowledge points and the reasons for the errors.

[0071] Step 7: Metacognitive-Driven Heuristic Prompt Question Recommendation

[0072] The first step involves using a large language model to generate a series of heuristic hint questions. Specifically, after analyzing the reasons for errors in the questions, language generation constraints are constructed based on the user's individual personality, knowledge level, and historical learning data. Based on these constraints, a series of selectable heuristic hint questions are generated using the large language model.

[0073] The second step involves ranking the generated heuristic hint questions using a graph neural network. Specifically, a hint question graph G = (V, E) is constructed, where the node set V represents the set of hint question nodes, and the edge set E represents the relationships between the hint question nodes. A graph neural network model based on a graph attention mechanism is used for ranking, and this network model is specifically represented by the following equation (n):

[0074]

[0075] in, Represents node v i The hidden state at layer l Represents node v i Aggregated information of the hidden states of neighboring nodes, w l Let represent the parameters of the l-th layer. The final representation h of each hint question node is obtained by iteratively calculating the hidden states. This representation is then concatenated with the user's knowledge level representation ht and sentiment representation E to obtain a comprehensive node feature representation h. * Specifically, it is expressed as follows in equation (o):

[0076] h * =[h,h t ,E] (o).

[0077] Here, [] denotes the vector concatenation operation.

[0078] Then, the ranking score s of the nodes is predicted using the fully connected layer. i Specifically, it is expressed as follows (p):

[0079] s i =σ(W s h * +b s (p).

[0080] Where σ is the activation function, W s and b s These are the weight matrix and bias vector of the fully connected layer, respectively.

[0081] Finally, the nodes are sorted according to the obtained ranking scores, and the heuristic hint questions with the highest rankings are recommended to the user.

[0082] The third step is to enhance human-machine metacognitive collaboration based on reinforcement learning. Specifically, using the DDPG deep deterministic policy gradient algorithm, the dialogue content, user features, and system state are first represented as a state vector s. Based on the current state s, an action a is generated using a policy network, as shown in equation (q) below:

[0083] a = π(s) (q).

[0084] Where π(s) represents the policy network.

[0085] Then, the user's answer after receiving the prompt is used as a feedback signal r. When the user answers correctly, a reward signal R is calculated based on the feedback signal r, as shown in the following formula (r):

[0086] R = r + γ*R′ (r).

[0087] Where γ is a discount factor that controls the importance of future rewards to the current decision, and its value ranges from [0,1]. R′ is the reward signal at the next time step.

[0088] Finally, the parameters of the policy network and the value function network are updated according to the reward signal R. The specific process is as follows: The policy network parameters and the target policy network parameters are represented as θ and θ', respectively, and the value function network parameters and the target value function network parameters are represented as... and

[0089] Based on the target policy network parameters θ', the target value function Q' is calculated using the target value function network; based on the policy network parameters θ, the current value function Q is calculated using the value function network, as specifically expressed in equations (s) and (t) below:

[0090] Q′(s′,a′) (s);

[0091] Q(s,a) (t).

[0092] Where s′ and a′ are the next state and the next action, respectively, and s and a are the current state and the current action, respectively.

[0093] The time difference error δ is calculated using the reward signal R and the target value function Q′, as specifically expressed by the following equation (u):

[0094] δ=R+γ*Q′(s′,a′)-Q(s,a) (u).

[0095] Where γ is the discount factor.

[0096] To minimize the square of the time difference error δ, the parameters are updated by backpropagating the gradient of the policy network parameters θ, as expressed in equation (v) below:

[0097]

[0098] Where α is the learning rate of the policy network.

[0099] The value function network parameters are updated in batches using sample data from the replay buffer, minimizing the squared temporal difference error δ, by adjusting the value network parameters. The gradient is backpropagated to update the parameters, as shown in the following equation (w):

[0100]

[0101] Where β is the learning rate of the value function network.

[0102] Compared with existing technologies, this invention offers more accurate, interpretable, and effective recommendations. It acquires a user's knowledge level representation based on their emotions and cognitive styles, and analyzes the causes of incorrect answers based on user personality, fully considering the combined influence of long-term personality and short-term emotions. Aiming to improve users' metacognitive level, it generates a series of heuristic prompts based on the causes of errors using a large language model. These prompts are then ranked and recommended to the user. Based on user feedback, reinforcement learning is used to improve the machine's metacognitive level, while the machine can recommend higher-quality heuristic prompts to enhance the user's metacognitive ability, thus achieving a synergistic improvement in human-machine metacognition. Attached Figure Description

[0103] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0104] The present invention will be further described in detail below with reference to the specific embodiments and accompanying drawings. Except for the contents specifically mentioned below, the processes, conditions, and experimental methods for implementing the present invention are all common knowledge and general knowledge in the art, and the present invention does not have any particular limitations.

[0105] Example 1

[0106] See Figure 1 To implement heuristic suggestion questions, follow these steps:

[0107] Step 1: Obtain the user's historical test-taking history and conversation interactions.

[0108] User answer information and dialogue interaction information can be obtained from public datasets such as ASSIST2012, Persona-Chat, and DailyDialog. User questions are represented by vectors with a dimension of 100, and correct answers are represented by vectors with a dimension of 1, with 1 representing a correct answer and 0 representing an incorrect answer.

[0109] Step 2: Generate implicit representations of users' emotions and cognitive styles

[0110] The psychological lexicon dictionary obtained from LIWC2007 or LIWC2015 categorizes users' cognitive styles into {intuitive-perceptual, concrete-abstract, sequential-synthetic}, which can be obtained through questionnaires.

[0111] Both the user's emotion representation vector and cognitive representation vector are 100-dimensional, using an n-dimensional vector E = {e1, e2, ..., e}. n} and m-dimensional vector C = {c1, c2, ..., c m The user's emotion representation E and cognitive representation C are concatenated to obtain the emotion and cognitive style representation E. *Among them, e i c is the saliency of the i-th mental lexicon category. j It is the saliency of the j-th cognitive style category.

[0112] The emotion representation E is calculated from the user's current dialogue text information. Specifically, the TF-IDF algorithm is used to extract emotion features. For each user, the occurrence count W of the i-th mental lexicon category word is obtained based on the current dialogue. i And the number of discourses D containing words of this mental lexicon category. i First, use the total number of words in the current dialogue, n w Calculate the word frequency (TF) of category words in a mental lexicon. i Specifically, as shown in equation (ii) below:

[0113]

[0114] Next, based on the number of utterances n in the current dialogue... d Calculate the reverse file frequency (idf) i Specifically, as shown in equation (iii) below:

[0115]

[0116] Thus, the emotion representation E of user u is obtained, specifically as shown in equation (a) below:

[0117] E = {e1,e2,…,e} n}=tf i ×idf i ×{R1,R2,…,R n} (a).

[0118] Among them, R i It is the correlation factor between the i-th psychological vocabulary category and emotion.

[0119] The cognitive representation C is obtained by one-hot encoding, and the user's emotion representation and cognitive style representation are concatenated to obtain the final implicit representation of emotion and cognitive style E. * Specifically, it is expressed as in equation (b) below:

[0120] E * =concat(E,C) (b).

[0121] Here, concat represents the concatenation operation.

[0122] Step 3: Generate user personality representation

[0123] The psychological lexicon is obtained from LIWC2007 or LIWC2015, and the user's personality representation vector has a dimension of 100. An n-dimensional vector P = {p1, p2, ..., p...} is used. n As a representation of user personality, e i This represents the saliency of the i-th mental vocabulary category. The personality representation P is calculated from the user's historical dialogue text information, using the same calculation method as the emotion representation E in step two. The difference lies in that the current dialogue pattern reflects short-term emotional characteristics, while the long-term dialogue pattern reflects personality characteristics.

[0124] Specifically, for each user, the number of occurrences W of the i-th mental vocabulary category word is obtained based on the historical dialogue. i And the number of discourses D containing words of this mental lexicon category. i First, use the total number of words N in the historical dialogue. w Calculate the word frequency (TF) of category words in a mental lexicon. i Specifically, it is expressed as follows in equation (ii):

[0125]

[0126] Then, based on the number of words N in the historical dialogue... d Calculate the inverse text frequency (idf) i Specifically, as shown in equation (iii) below:

[0127]

[0128] Thus, the personality representation P of user u is obtained, specifically expressed as follows (i):

[0129] P = {p1, p2, ..., p} n}=tf i ×idf i ×{R1,R2,…,R n} (i).

[0130] Among them, R i It is the correlation factor between the i-th psychological vocabulary category and emotion.

[0131] Step 4: Representation of knowledge level based on emotion and cognitive style

[0132] The first step is to generate a difficulty representation for the question based on the difficulty of the knowledge points it contains and the characteristics of the question itself. Specifically, the objective difficulty of question j is first represented by the following equation (iv):

[0133] D j =∑ i∈K(j) α i +β j (iv)

[0134] Where K(j) is the set of knowledge points involved in question j, and each knowledge point i has a corresponding difficulty αi; β j The difficulty level is inherent to the questions and is unrelated to the specific knowledge points covered.

[0135] Next, considering the current state of user u, the difficulty of question j for user u is represented by the following equation (c):

[0136]

[0137] in, This is based on the implicit representation of user emotion and cognitive style obtained in step two, i.e., the user's current state. θ in It is an input integration function that integrates the user's current state into the objective difficulty of the question.

[0138] The second step is to use the problem-solving sequence obtained in step one. The LSTM model is used to obtain the user's current knowledge level. Specifically, firstly, q... t a t and The input vector x of the LSTM model is obtained by concatenating the vectors. t That is, it can be expressed by the following equation (d):

[0139] x t =concat(D u,j a t (d).

[0140] Then, calculate the unit state c. t Specifically, it is expressed as follows in equation (e):

[0141]

[0142] Among them, f t For the forgetting gate, control c t-1 Used to update c t Features; i t For input gates, control Used to update c t Features; The cell state update value is represented by equations (v) to (vii) below:

[0143] f t =σ(W f [x t ,h t-1 ]+b f (v);

[0144] i t =σ(Wi [x t ,h t-1 ]+b i (vi);

[0145]

[0146] Where W and b represent the weight matrix and bias, respectively, and the same applies below.

[0147] Finally, calculate the output gate o. t Use output gate o t and unit state c t Calculate the current knowledge level h t Specifically, as shown in equations (viii) and (f) below:

[0148] o t =σ(W o [x t ,h t-1 ]+b o (viii);

[0149] h t =o t ×tanh(c t (f).

[0150] Step 5: Predict the questions you haven't mastered

[0151] Based on the current knowledge level representation h obtained in step four t Then, the student's level of understanding of the question can be obtained, specifically expressed as formula (g) below:

[0152] y t =σ(W y h t +b y (g).

[0153] Where σ is the sigmoid function, y t The length of the question is equal to the length of the problem; in this embodiment, the length of the problem is 1000.

[0154] The training objective is the negative log-likelihood of the student's answer sequence, using binary cross-entropy loss. Calculate the loss function Specifically, it is expressed as follows (h):

[0155]

[0156] Step Six: Error Analysis Based on User Personality

[0157] Based on the prediction of unmastered questions in step five, questions are recommended to users. When users answer questions incorrectly, error analysis is performed based on the user personality representation P obtained in step three.

[0158] The first step involves preprocessing incorrect questions, personality traits, and reasons for errors. Specifically, on one hand, different types of explainable error reason information are collected to generate pre-defined explainable error reason information. An incorrect question-personality matrix N is used to store the relationship between incorrect questions and personality traits. The matrix is ​​constructed as follows: when a user u with personality trait P answers question q incorrectly, N is incremented. q,p The value of . A personality-fault factor matrix M is used to store the relationship between personality and fault factors, where the elements M p,b The degree of correlation between personality type P and the cause of error b.

[0159] On the other hand, using matrix K key Store knowledge points, where each column represents a knowledge point; use matrix K. ealue Store knowledge states, each column represents K key The degree of correlation between the corresponding knowledge points and the reasons for the errors.

[0160] The second step is to represent the user's current incorrect question as q. t The inner product of the incorrect question-personality matrix N is taken to obtain the personality p′ that is most likely to answer the question incorrectly. Then, the inner product of the obtained personality p′ and the user's own personality P is taken to obtain the matching degree between the two, as specifically expressed by the following formula (m):

[0161] λ=p T p′ (m).

[0162] Among them, a higher matching degree λ indicates a greater contribution of personality to the cause of error; p T It is the transpose of the personality representation P.

[0163] The third step is to determine the error cause T, which is caused by both the user's knowledge level and personality. The calculation is expressed as follows (1):

[0164] T = λT1 + (1-λ)T2 (1).

[0165] Where T1 and T2 are the error causes caused by the user's personality and knowledge level, respectively, and λ is the degree of personality matching calculated in the second step.

[0166] The error T1 caused by personality is obtained from the personality-error matrix M, and the error T2 caused by knowledge level is calculated by the attention mechanism. Specifically, the question is first represented as q. t and the knowledge point matrix K key The inner product of each knowledge point is calculated, and after passing it through an activation function, the relevance between the question and the knowledge point is obtained, i.e., the attention weight w. tSpecifically, it is expressed as follows in equation (j):

[0167] w t =Softmax(q t K key (j).

[0168] Therefore, the error T2 that may arise from the question itself is calculated based on the attention weight, as specifically expressed by the following formula (k):

[0169] T2 = w t K value (k).

[0170] Among them, K value K represents key The degree of correlation between the corresponding knowledge points and the reasons for the errors.

[0171] Step 7: Metacognitive-Driven Heuristic Prompt Question Recommendation

[0172] ChatGPT was chosen as the large language model. The first step involved generating a series of heuristic hint questions using the large language model. Specifically, after analyzing the reasons for errors in the incorrect questions, language generation constraints were constructed based on the user's individual personality, knowledge level, and historical learning data. Based on these constraints, the large language model was then used to generate a series of selectable heuristic hint questions.

[0173] The second step involves ranking the generated heuristic hint questions based on a graph neural network. Specifically, a hint question graph G = (V, E) is constructed, where the node set V represents the set of hint question nodes, and the edge set E represents the relationships between the hint question nodes. A graph neural network model based on a graph attention mechanism is used for ranking, and this network model is specifically represented by the following equation (n):

[0174]

[0175] in, Represents node v i The hidden state at layer l Represents node v i Aggregated information of the hidden states of neighboring nodes, w l Let represent the parameters of the l-th layer. The final representation h of each hint question node is obtained by iteratively calculating the hidden states, and then compared with the user's knowledge level representation h. t By connecting the emotion representation E, a comprehensive node feature representation h is obtained. * Specifically, it is expressed as follows in equation (o):

[0176] h * =[h,h t ,E] (o).

[0177] Here, [] denotes the vector concatenation operation.

[0178] Then, the ranking score s of the nodes is predicted using the fully connected layer. i Specifically, it is expressed as follows (p):

[0179] s i =σ(W s h * +b s (p).

[0180] Where σ is the activation function, W s and b s These are the weight matrix and bias vector of the fully connected layer, respectively.

[0181] Finally, the nodes are sorted according to the obtained ranking scores, and the heuristic hint questions with the highest rankings are recommended to the user.

[0182] The third step is to enhance human-machine metacognitive collaboration based on reinforcement learning. Specifically, on the one hand, the DDPG deep deterministic policy gradient algorithm is used to improve the machine's metacognitive level to improve recommendation quality; on the other hand, higher-quality heuristic prompts are recommended to users to improve their metacognitive abilities. First, the dialogue content, user features, and system state are represented as a state vector s. Based on the current state s, a policy network is used to generate action a, as shown in equation (q) below:

[0183] a = π(s) (q).

[0184] Where π(s) represents the policy network.

[0185] Then, the user's answer after receiving the prompt is used as a feedback signal r. When the user answers correctly, a reward signal R is calculated based on the feedback signal r, as shown in the following formula (r):

[0186] R = r + γ*R′ (r).

[0187] Where γ is a discount factor that controls the importance of future rewards to the current decision, and its value ranges from [0,1]. R′ is the reward signal at the next time step.

[0188] Finally, the parameters of the policy network and the value function network are updated according to the reward signal R. The specific process is as follows: The policy network parameters and the target policy network parameters are represented as θ and θ', respectively, and the value function network parameters and the target value function network parameters are represented as... and

[0189] Based on the target policy network parameters θ', the target value function Q' is calculated using the target value function network; based on the policy network parameters θ, the current value function Q is calculated using the value function network. Specifically, this is expressed in equations (s) and (t) below:

[0190] Q′(s′,a′) (s);

[0191] Q(s,a) (t).

[0192] Where s′ and a′ are the next state and the next action, respectively, and s and a are the current state and the current action, respectively.

[0193] The time difference error δ is calculated using the reward signal R and the target value function Q′, as specifically expressed by the following equation (u):

[0194] δ=R+γ*Q′(s′,a′)-Q(s,a) (u).

[0195] Where γ is the discount factor.

[0196] To minimize the square of the time difference error δ, the parameters are updated by backpropagating the gradient of the policy network parameters θ, as expressed in equation (v) below:

[0197]

[0198] Where α is the learning rate of the policy network.

[0199] The value function network parameters are updated in batches using sample data from the replay buffer, minimizing the squared temporal difference error δ, by adjusting the value network parameters. The gradient is backpropagated to update the parameters, as shown in the following equation (w):

[0200]

[0201] Where β is the learning rate of the value function network.

[0202] This invention is interpretable, acquiring a user's knowledge level representation based on user emotions and cognitive styles, and analyzing the causes of incorrect answers based on user personality, fully considering the combined influence of long-term personality and short-term emotions. Aiming to improve users' metacognitive level, it uses a large language model to generate a series of heuristic prompts based on the causes of errors. These prompts are then ranked and recommended to the user. Based on user feedback, reinforcement learning is used to improve the machine's metacognitive level, while the machine can recommend higher-quality heuristic prompts to enhance the user's metacognitive ability, thus achieving a synergistic improvement in human-machine metacognition. The method of this invention is effective; experiments on real datasets show that this invention has higher accuracy compared to existing technologies.

[0203] The scope of protection of this invention is not limited to the above embodiments. Any variations and advantages that can be conceived by those skilled in the art without departing from the spirit and scope of the inventive concept are included in this invention and are protected by the appended claims.

Claims

1. A meta-cognition driven heuristic hint question recommendation method, characterized in that, The method adopts a recommendation system constructed by a knowledge tracking module based on emotion and cognitive style, a personality-enhanced error cause analysis module, and a meta-cognitive strategy-driven heuristic hint question recommendation module to realize heuristic hint question recommendation. The knowledge tracking module based on emotion and cognitive style extracts the cognitive style and knowledge mastery level of the user from the user's historical test records, extracts the current emotion of the user from the current dialogue, and constructs a knowledge level representation based on emotion and cognitive style. The error cause analysis module extracts the user's personality from the historical dialogue, uses an attention mechanism to capture the relationship between personality and wrong questions at the current knowledge level, and constructs a personality-based error cause representation. The meta-cognitive strategy-driven heuristic hint question recommendation module generates a series of heuristic hint questions for selection based on the error cause representation using a large language model, sorts the generated heuristic hint questions using a graph neural network, and continuously adjusts the hint strategy in a reinforcement learning framework to collaboratively improve human-computer meta-cognition. The error cause analysis module specifically includes the following two parts: The personality is calculated from the historical dialogue text information of the user by the following (i) formula : （i）； wherein, is the term frequency of the th psychological lexicon category word in the historical dialogue; is the inverse document frequency of the th psychological lexicon category word in the historical dialogue; is the correlation factor of the th psychological lexicon category and the emotion; 1) Generating user personality representation part 2-2-1: According to the personality-mistake matrix , get the personality caused by the mistake ; According to the mistake-personality matrix , get the personality that is most likely to make mistakes in this question 2-2-2: Representing the question with each knowledge point in the knowledge point matrix The inner product is taken with each knowledge point in the knowledge point matrix, and the relevance of the question to the knowledge point is obtained by the activation function, i.e., the attention weight represented by the following (j) formula : （j）； 2-2-3: According to the attention weight The error factor caused by the knowledge level is calculated by the following (k) formula : （k）； wherein, represents the degree of relevance of the corresponding knowledge point and the error cause in the 2-2-4: According to the error factor and the error factor The error factor is calculated from the following (l) formula : （l）； wherein, is the personality most likely to make a mistake on the question with the user's personality an inner product is obtained, indicating the degree of matching, which is calculated as shown in the following (m) formula: （m）； wherein is a personality representation is the transpose of. 2.The meta-cognition driven heuristic prompting question recommendation method of claim 1, wherein, 2) Error cause analysis based on user personality part The knowledge tracking module specifically includes the following three parts: 1-1-1: Extracting sentiment features using a TF-IDF algorithm based on the user's current conversation text information to calculate the user's sentiment representation , is specifically represented by the following (a) formula: （a）； wherein, is the word frequency of the current dialogue for the th psychological lexicon category word; is the inverse text frequency of the current dialogue for the th psychological lexicon category word; is the correlation factor of the th psychological lexicon category and the emotion; 1-1-2: User Emotional expression With cognitive representation By splicing the components together, the implicit representation of emotion and cognitive style is obtained from equation (b) below. : （b）； wherein represents a concatenation operation; The cognitive representation Obtained by one-hot encoding of the questionnaire results; 1) Generating implicit representation of user emotion and cognitive style part According to the question sequence The current knowledge level of the user is obtained using the LSTM model, as follows: 1-2-1: Implicit representation based on user emotion and cognitive style The title is obtained from the following (c) formula The difficulty of the user is represented : （c）； wherein, is the difficulty representation of the question based on the knowledge point difficulty contained in the question and the characteristics of the question; is an input integration function that integrates the user's current state into the objective difficulty of the question; 1-2-2: The title The corresponding title difficulty expression fused with the current emotion and cognitive style of the user And the answer expression Splicing, the input vector of the LSTM model is obtained by the following (d) formula : （d）； 1-2-3: Using a forget gate and an input gate The cell state is calculated by the following (e) formula : （e）； wherein, is a cell state update value; forget gate control in the cell for updating a feature; input gate control in the cell for updating a feature; 1-2-4: Using an output gate and cell state Current knowledge level is calculated by the following (f) formula : （f）； 2) Knowledge level representation based on emotion and cognitive style part 1-3-1: Representing the current knowledge level The degree of mastery of the student over the question is calculated by the following (g) formula : （g）； wherein, is a sigmoid function; the length of the is equal to the length of the question. 1-3-2: The training objective is the negative log-likelihood of the student answer sequence, and the loss is computed using binary cross-entropy (BCE) as follows (h) : （h）； wherein, is the binary cross-entropy loss. 3.The meta-cognition driven heuristics prompting question recommendation method of claim 1, wherein, 3) Predicting unmastered questions part The heuristic hint question recommendation module specifically includes the following three parts: 1) Heuristic hint question generation part According to the individual personality, knowledge level and historical learning data information of the user, construct language generation constraints, and generate a series of selectable heuristic hint questions using a large language model; 3-2-1 : Constructing the hint question graph where the set of nodes represents the set of hint question nodes, the set of edges represents the relationships between hint question nodes; 2) Heuristic hint question recommendation part （n）； wherein, is a node In a first hidden state of a layer; is a node aggregated information of hidden states of neighbor nodes of the node is a parameter of a first layer; 3-2-3: Obtain the final representation of each prompt question node by iteratively calculating the hidden state and then connect with the user's knowledge level representation and emotion representation to obtain the integrated node feature representation by the following (o) formula : （o）； wherein denotes a concatenation operation of vectors; 3-2-4: Predicting the ranking score of a node from the following (p) using a fully connected layer : （p）； wherein, is an activation function; and are the weight matrix and bias vector of the fully connected layer, respectively; 3-2-2: Sort using a graph neural network model based on graph attention mechanism, which is represented by the following (n) formula: 3-2-5: Sort the nodes according to the obtained ranking scores, and recommend the top-ranked heuristic hint question to the user; 3) Reinforcement learning-based human-computer meta-cognition collaborative improvement part 3-3-1: Representing the contents of the dialogue, the user characteristics, and the system state as a state vector , according to the current state , using a policy network to generate an action from the following (q) formula : （q）； wherein denotes a policy network; 3-3-2: The answer of the user after accepting the prompt question is used as the feedback signal When the user answers the question correctly, the feedback signal is used to calculate the reward signal The reward signal is calculated by the following (r) formula : （r）； wherein, is a discount factor, controlling the importance of future rewards for current decisions, which takes values in the range ; is the reward signal at the next time instant; 3-3-3: Let the policy network parameters and the target policy network parameters be denoted as and Let the value function network parameters and the target value function network parameters be denoted as and ; 3-3-4: Target policy network parameters , the target value function is calculated using a target value function network and is represented by the following (s) formula; （s）； wherein, next state and next action, respectively; 3-3-5: Strategy network parameter-based , the current value function is calculated using the value function network , and is represented by the following (t) formula: （t）； wherein, and are the current state and the current action, respectively. 3-3-6: Using a reward signal and an objective value function The time-difference error is calculated by the following (u) formula : （u）； wherein is a discount factor; 3-3-7: Update the parameters by backpropagating the gradients of the policy network parameters as represented by the following equation (v): （v）； wherein, is the learning rate of the policy network; 3-3-8: Update parameters by backpropagating gradients through the value network parameters as represented by the following equation (w): （w）； wherein, is the learning rate for the value function network.

Images