Digital intelligence enabling method based on large language model

Abstract

The invention relates to the technical field of AI teaching, and discloses a big language model-based digital intelligence enabling method, which comprises the following steps of: acquiring a current teaching strategy, learner acceptance information and learner learning information; the first plane data is constructed based on the collected current teaching strategy, the first serial data stream and the second plane data are constructed based on the collected learner acceptance information, and the third plane data and the second serial data stream are constructed based on the collected learner acceptance information. According to the method, the psychological counseling strategy and the teaching strategy are combined, the emotion and psychological needs of the learner are better taken into consideration, all-around support and guidance are provided for the learner, the teaching effect is optimized, comprehensive development of the learner is promoted, the humanized intelligent teaching mode is closer to an actual education scene, and the teaching effect is better. Advantages of AI teaching can be brought into full play, and learning of learners and promotion of AI teaching are facilitated.

Description

Technical Field

[0001] The present invention relates to the field of AI teaching, and more specifically, to a digital intelligence empowerment method based on a large language model. Background Art

[0002] The traditional "one-size-fits-all" teaching model can no longer meet the personalized needs of different learners, so AI-powered teaching is a new choice for more learners. AI teaching is a new teaching model that applies artificial intelligence technology to the field of education and optimizes the teaching process, improves teaching efficiency and learning effects in an intelligent way. AI teaching uses machine learning, deep learning, natural language processing and other technologies to provide learners with a personalized and adaptive learning experience.

[0003] Existing AI teaching only arranges learning plans based on some basic needs of learners, and learners are required to strictly implement the learning plans. However, learners may lack trust in AI teaching, or feel disgusted and uncomfortable with the teaching format, which causes psychological frustration to learners, leading them to give up their learning plans, or fail to achieve the expected results after learning, which is not conducive to learners' learning. Summary of the invention

[0004] The present invention provides a digital intelligence empowerment method based on a large language model to solve technical problems in related technologies.

[0005] The present invention provides a digital intelligence empowerment method based on a large language model, comprising the following steps:

[0006] Step 100, obtaining the current teaching strategy, learner acceptance information and learner learning information; learner learning information includes learning preference information, motivation type information, plan execution information and self-discipline evaluation information; learner acceptance information includes feedback information, emotional state information and acceptable teaching time information; the current teaching strategy includes several teaching units, and the teaching units include incentive and reward mechanisms, teaching content presentation and feedback methods, learning plans and schedules, teacher-student interaction methods and image information;

[0007] Step 200, constructing first plane data based on the collected current teaching strategy, constructing a first serial data stream and a second plane data based on the collected learner acceptance information, and constructing a third plane data and a second serial data stream based on the collected learner acceptance information;

[0008] Step 300, input the first plane data and the second plane data into the strategy update model, and output the result representing the updated teaching strategy, wherein the updated teaching strategy includes all updated teaching units, and the updated teaching units include incentive and reward mechanisms, teaching content presentation and feedback methods, learning plans and schedules, teacher-student interaction methods, and image information; input the updated teaching strategy and the second serial data stream into the psychological counseling model, and output the result representing the psychological counseling strategy, wherein the psychological counseling strategy includes emotional acceptance, empathic feedback, guiding a positive attitude, emotional regulation, building a "growth mindset" mentality, and guiding self-dialogue;

[0009] In step 400, the AI ​​teacher performs teaching according to the updated teaching strategy outputted. Meanwhile, when the learner's feedback is poor and the emotional state deviates from the normal state, the AI ​​teacher stops the execution of the teaching strategy according to the output psychological counseling strategy until the feedback and emotional state return to normal, and then executes the teaching strategy again.

[0010] Further, the first plane data includes a first facet for storing data, one first facet indexes a data object, and the first plane data also includes a relationship list for storing association relationships between data objects indexed by the first facet, and the element value of the i-th row and j-th column of the relationship list represents a quantitative value of the association relationship between the data objects indexed by the i-th and j-th first facets;

[0011] The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0012] Data objects include learners, teaching units, AI teachers, teaching subjects, and image types;

[0013] The data objects with high association are defined as: learners and their AI teachers, learners and their teaching subjects, teaching units and their AI teachers, teaching subjects and their AI teachers, and AI teachers and their image types.

[0014] Obtaining a first serial data stream based on the learner learning information sorting, the first serial data stream comprising n data units sorted by time, the tth data unit storing the second plane data generated by the learner learning information collected at the tth moment;

[0015] The second plane data includes a second facet for storing data, each second facet indexes a data object, and the second plane data also includes a relationship list for storing association relationships between the data objects indexed by the second facet, wherein the element value in the i-th row and j-th column of the relationship list represents a quantitative value of the association relationship between the data objects indexed by the i-th and j-th second facets;

[0016] The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0017] Data objects include learners, preference types, motivation types, execution status, and self-discipline assessment scores;

[0018] The data objects with high correlation are defined as: the learner’s preference type in the teaching, the learner’s motivation type in learning, the learner’s implementation of his / her learning plan, and the learner’s self-discipline evaluation score;

[0019] Generating third plane data based on the learner's acceptance information includes a third facet for storing data, a third facet indexing a second serial data stream of a data object, the third plane data also includes a relationship list for storing association relationships between data objects indexed by the third facet, the element value of the i-th row and j-th column of the relationship list representing the quantitative value of the association relationship between the data objects indexed by the i-th and j-th third facets;

[0020] The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0021] Data objects include learners, AI teachers, opinion types, and emotional states;

[0022] The data objects with high correlation are defined as: learners and the types of opinions they express, learners and the emotional states they give back when receiving teaching, learners and the AI ​​teachers they teach, and opinion types and their corresponding AI teachers.

[0023] The second serial data stream of a data object includes n data units sorted by time, and the t-th data unit represents the learner's acceptance information collected at the t-th moment of the data object.

[0024] Furthermore, the strategy update model includes a first trunk network and a first branch network, the first trunk network includes a first encoder, a third encoder, a first feature fusion layer and a first FC layer, the first branch network includes a second encoder, wherein the first encoder inputs first plane data and outputs a first recognition feature, while the second encoder inputs second plane data and outputs a second recognition feature, the first recognition feature and the second recognition feature are input into a first fusion feature layer, the first fusion feature layer outputs a first fusion state, the first fusion state is input into the third encoder, the third encoder outputs a third recognition feature to the first FC layer, and the first FC layer outputs a result representing the updated teaching strategy.

[0025] Furthermore, the psychological counseling model includes a second trunk network and a second branch network, the second trunk network includes a fourth encoder, a fifth encoder, a second feature fusion layer and a second FC layer, the second branch network includes a sixth encoder, wherein the fourth encoder inputs a second serial data stream and outputs a fourth recognition feature, the fourth recognition feature is input into the fifth encoder, the fifth encoder outputs a fifth recognition feature, and at the same time the sixth encoder inputs an updated teaching strategy to output a sixth recognition feature, the sixth recognition feature and the fifth recognition feature are input into the second fusion feature layer, the second fusion feature layer outputs a second fusion state, the second fusion state is input into the second FC layer, and the second FC layer outputs a result representing the psychological counseling strategy.

[0026] Furthermore, the calculation formula of the first encoder is as follows:

[0027] ;

[0028] in represents the panel feature of the vth first panel of the lth layer, represents the set of first facets with high correlation with facet v, represents the first plane recognition weight matrix of layer l, is the attention weight of the first surfel u of layer l to the first surfel v, which is calculated by the following formula:

[0029] ;

[0030] in is a learnable attention vector, represents the vector concatenation operation, represents the second plane recognition weight matrix of layer l, represents transpose; and They represent the panel features of the vth first panel of the l-1th layer, , E represents the total number of layers, when l=1 , when l=E is equal to the identification feature of the first facet v; and The feature representations of the objects representing the vth and uth first facet indices respectively.

[0031] Furthermore, the calculation formula of the second encoder is as follows:

[0032] ;

[0033] represents the panel feature of the cth second panel of the lth layer, represents the panel feature of the d-th second panel of the l-1th layer, and represent the set of second facets with high correlation with the second facet c and the second facet d respectively, represents the cardinality of a set, represents the third plane recognition weight matrix of layer l, , E represents the total number of layers, when l=1 , The feature representation of the object with the dth second facet index, when l=E is equal to the second identification feature of the second facet c, is the sigmoid function.

[0034] Furthermore, the calculation formula of the first fusion feature layer is as follows:

[0035] ;

[0036] in, represents the first combination feature, represents the feature combination function (concatenation function or summation function), represents the panel feature of the vth first panel of the Eth layer of the first encoder, represents the set of panel features of the first panel, represents the panel feature of the c-th second panel of the E-th layer of the second encoder, a set of panel features representing a second panel;

[0037] The calculation formula of the third encoder is as follows:

[0038] ;

[0039] in represents the tth third identification feature, represents the t-1th third identification feature, , represents the tth data unit of the first combined feature, and is the serial recognition weight matrix of the third encoder, is the serial recognition bias parameter of the third encoder, ReLU is the ReLU activation function;

[0040] The calculation formula of the first FC layer is as follows:

[0041] ;

[0042] in represents the first output vector, and its cth component value represents the probability value of the cth strategy. The strategy with the largest probability value is selected as the output. The strategy group contains all executable strategies. A strategy includes incentive and reward mechanism, teaching content presentation and feedback method, learning plan, progress schedule, teacher-student interaction method, and image information;

[0043] For more complex strategies, vector representation can be used: a strategy is represented by a vector, and the jth component of the vector represents the jth teaching information of the updated teaching strategy;

[0044] represents the last third identification feature output by the third encoder, Indicates full connection.

[0045] Furthermore, the calculation formula of the fourth encoder is as follows:

[0046] ;

[0047] ;

[0048] ;

[0049] in represents the t-th data unit of the second serial data stream corresponding to the v-th third plane element of the third plane data, represents the t-1th fourth recognition feature of the vth third facet, and denote the first and second intermediate features of the vth third facet, respectively. represents the tth data unit of the second serial data stream of the vth third bin, and is the serial recognition weight matrix in the fourth encoder, is the serial recognition bias parameter in the fourth encoder, ReLU is the ReLU activation function, Represents the Sigmoid function;

[0050] The calculation formula of the fifth encoder is as follows:

[0051] ;

[0052] represents the panel feature of the third panel of the fth layer, represents the panel feature of the third panel of the rth panel in the l-1th layer, and They represent the set of facets with high correlation with facet f and facet r, respectively. represents the cardinality of a set, represents the plane recognition weight matrix of the lth layer in the fifth encoder, , E represents the total number of layers, when l=1 , The feature representation of the object that represents the index of the third facet of the rth facet, when l=E is equal to the fifth identification feature of the third panel element f, is the sigmoid function;

[0053] The calculation formula of the sixth encoder is as follows:

[0054] ;

[0055] in represents the sixth identification feature, represents the updated teaching strategy of the input, Represents the convolution function.

[0056] The present invention provides a digital intelligence empowerment system based on a large language model, comprising:

[0057] Data acquisition module: responsible for collecting current teaching strategies, learners’ learning information and learners’ acceptance information, and converting the collected data into structured plane data and serial data streams;

[0058] Strategy update model: including a first trunk network and a first branch network, inputting first plane data and second plane data, and outputting an updated teaching strategy;

[0059] Psychological counseling model: including a second trunk network and a second branch network, inputting the updated teaching strategy and the second serial data stream, and outputting the psychological counseling strategy;

[0060] Large language model module: Based on the pre-trained large language model, it generates corresponding teaching content and interactive responses according to the updated teaching strategies and psychological counseling strategies;

[0061] AI teaching module: responsible for implementing updated teaching strategies, interacting with learners, monitoring learners' feedback types and emotional states, and providing corresponding psychological counseling based on psychological counseling strategies.

[0062] The present invention provides a storage medium storing non-temporary computer-readable instructions for executing the steps in the aforementioned digital intelligence empowerment method based on a large language model.

[0063] The beneficial effects of the present invention are:

[0064] The present invention combines psychological counseling strategies with teaching strategies in parallel to better take care of learners' emotions and psychological needs, provide them with all-round support and guidance, thereby optimizing teaching effects and promoting learners' all-round development. This humanized intelligent teaching model is closer to actual educational scenarios, which is conducive to giving full play to the advantages of AI teaching, and is conducive to learners' learning and the promotion of AI teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the digital intelligence empowerment method based on a large language model of the present invention. DETAILED DESCRIPTION

[0066] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0067] like Figure 1 As shown, a digital intelligence empowerment method based on a large language model includes the following steps:

[0068] Step 100, obtaining current teaching strategy, learner acceptance information and learner learning information;

[0069] Learners’ learning information includes learning preference information, motivation type information, plan execution information and self-discipline evaluation information;

[0070] The information learners receive includes feedback, emotional state, and acceptable teaching time.

[0071] The current teaching strategy includes several teaching units, which include incentive and reward mechanisms, teaching content presentation and feedback methods, learning plans and schedules, teacher-student interaction methods, and image information;

[0072] Step 200, generating first plane data based on the collected current teaching strategy, the first plane data including a first facet for storing data, one first facet indexing a data object, and the first plane data also including a relationship list for storing association relationships between data objects indexed by the first facet, the element value of the i-th row and j-th column of the relationship list representing the quantitative value of the association relationship between the data objects indexed by the i-th and j-th first facets;

[0073] The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0074] Data objects include learners, teaching units, AI teachers, teaching subjects, and image types;

[0075] The data objects with high association are defined as: learners and their AI teachers, learners and their teaching subjects, teaching units and their AI teachers, teaching subjects and their AI teachers, and AI teachers and their image types.

[0076] Obtaining a first serial data stream based on the learner learning information sorting, the first serial data stream comprising n data units sorted by time, the tth data unit storing the second plane data generated by the learner learning information collected at the tth moment;

[0077] The second plane data includes a second facet for storing data, each second facet indexes a data object, and the second plane data also includes a relationship list for storing association relationships between the data objects indexed by the second facet, wherein the element value in the i-th row and j-th column of the relationship list represents a quantitative value of the association relationship between the data objects indexed by the i-th and j-th second facets;

[0078] The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0079] Data objects include learners, preference types, motivation types, execution status, and self-discipline assessment scores;

[0080] The data objects with high correlation are defined as: the learner’s preference type in the teaching, the learner’s motivation type in learning, the learner’s implementation of his / her learning plan, and the learner’s self-discipline evaluation score;

[0081] Generating third plane data based on the learner's acceptance information includes a third facet for storing data, a third facet indexing a second serial data stream of a data object, the third plane data also includes a relationship list for storing association relationships between data objects indexed by the third facet, the element value of the i-th row and j-th column of the relationship list representing the quantitative value of the association relationship between the data objects indexed by the i-th and j-th third facets;

[0082] The association relationship includes high and low, corresponding to two quantitative values ​​respectively.

[0083] Data objects include learners, AI teachers, opinion types, and emotional states;

[0084] The data objects with high correlation are defined as: learners and the types of opinions they express, learners and the emotional states they give back when receiving teaching, learners and the AI ​​teachers they teach, and opinion types and their corresponding AI teachers.

[0085] The second serial data stream of a data object includes n data units ordered by time, and the t-th data unit represents the learner acceptance information collected at the t-th moment of the data object;

[0086] Step 300, inputting the first plane data and the second plane data into the strategy update model, the strategy update model includes a first trunk network and a first branch network, the first trunk network includes a first encoder, a third encoder, a first feature fusion layer and a first FC (fully connected layer) layer, the first branch network includes a second encoder, wherein the first encoder inputs the first plane data and outputs a first recognition feature, while the second encoder inputs the second plane data and outputs a second recognition feature, the first recognition feature and the second recognition feature are input into a first fusion feature layer, the first fusion feature layer outputs a first fusion state, the first fusion state is input into a third encoder, the third encoder outputs a third recognition feature to a first FC layer, and the first FC layer outputs a result representing the updated teaching strategy;

[0087] The updated teaching strategy includes all updated teaching units, and the updated teaching units include but are not limited to incentive and reward mechanisms, teaching content presentation and feedback methods, learning plans and schedules, teacher-student interaction methods and image information;

[0088] Input the updated teaching strategy and the second serial data stream into the psychological counseling model, the psychological counseling model includes a second trunk network and a second branch network, the second trunk network includes a fourth encoder, a fifth encoder, a second feature fusion layer and a second FC layer, the second branch network includes a sixth encoder, wherein the fourth encoder inputs the second serial data stream and outputs a fourth recognition feature, the fourth recognition feature is input into the fifth encoder, the fifth encoder outputs a fifth recognition feature, and at the same time the sixth encoder inputs the updated teaching strategy and outputs a sixth recognition feature, the sixth recognition feature and the fifth recognition feature are input into the second fusion feature layer, the second fusion feature layer outputs a second fusion state, the second fusion state is input into the second FC layer, and the second FC layer outputs a result representing the psychological counseling strategy;

[0089] Psychological counseling strategies include but are not limited to emotional acceptance, empathetic feedback, guiding a positive mindset, emotional regulation, building a "growth mindset", and guiding self-dialogue;

[0090] In step 400, the AI ​​teacher performs teaching according to the updated teaching strategy outputted. Meanwhile, when the learner's feedback type or emotional state deviates from the normal threshold of acceptance of learning level, the AI ​​teacher stops the execution of the teaching strategy according to the output psychological counseling strategy until the feedback type or emotional state meets the normal threshold of acceptance of learning level, and then executes the teaching strategy.

[0091] In one embodiment of the present invention, the calculation formula of the first encoder is as follows:

[0092] ;

[0093] in represents the panel feature of the vth first panel of the lth layer, represents the set of first facets with high correlation with facet v, represents the first plane recognition weight matrix of layer l, is the attention weight of the first surfel u of layer l to the first surfel v, which is calculated by the following formula:

[0094] ;

[0095] in is a learnable attention vector, represents the vector concatenation operation, represents the second plane recognition weight matrix of layer l, represents transpose; and They represent the panel features of the vth first panel of the l-1th layer, , E represents the total number of layers, when l=1 , when l=E is equal to the identification feature of the first facet v; and The feature representations of the objects representing the vth and uth first facet indices respectively.

[0096] In one embodiment of the present invention, the calculation formula of the second encoder is as follows:

[0097] ;

[0098] represents the panel feature of the cth second panel of the lth layer, represents the panel feature of the d-th second panel of the l-1th layer, and represent the set of second facets with high correlation with the second facet c and the second facet d respectively, represents the cardinality of a set, represents the third plane recognition weight matrix of layer l, , E represents the total number of layers, when l=1 , The feature representation of the object with the dth second facet index, when l=E is equal to the second identification feature of the second facet c, is the sigmoid function;

[0099] In one embodiment of the present invention, the calculation formula of the first fusion feature layer is as follows:

[0100] ;

[0101] in, represents the first combination feature, represents the feature combination function (concatenation function or summation function), represents the panel feature of the vth first panel of the Eth layer of the first encoder, represents the set of panel features of the first panel, represents the panel feature of the c-th second panel of the E-th layer of the second encoder, A collection of surfel features representing the second surfel.

[0102] In one embodiment of the present invention, the calculation formula of the third encoder is as follows:

[0103] ;

[0104] in represents the tth third identification feature, represents the t-1th third identification feature, , represents the tth data unit of the first combined feature, and is the serial recognition weight matrix of the third encoder, is the serial recognition bias parameter of the third encoder, and ReLU is the ReLU activation function.

[0105] In one embodiment of the present invention, the calculation formula of the first FC layer is as follows:

[0106] ;

[0107] in represents the first output vector, and its cth component value represents the probability value of the cth strategy. The strategy with the largest probability value is selected as the output. The strategy group contains all executable strategies. A strategy includes incentive and reward mechanism, teaching content presentation and feedback method, learning plan, progress schedule, teacher-student interaction method, and image information;

[0108] For more complex strategies, vector representation can be used: a strategy is represented by a vector, and the jth component of the vector represents the jth teaching information of the updated teaching strategy;

[0109] represents the last third identification feature output by the third encoder, Indicates full connection.

[0110] In one embodiment of the present invention, the calculation formula of the fourth encoder is as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] in represents the t-th data unit of the second serial data stream corresponding to the v-th third plane element of the third plane data, represents the t-1th fourth recognition feature of the vth third facet, and denote the first and second intermediate features of the vth third facet, respectively. represents the tth data unit of the second serial data stream of the vth third bin, and is the serial recognition weight matrix in the fourth encoder, is the serial recognition bias parameter in the fourth encoder, ReLU is the ReLU activation function, Represents the Sigmoid function.

[0115] In one embodiment of the present invention, the calculation formula of the fifth encoder is as follows:

[0116] ;

[0117] represents the panel feature of the third panel of the fth layer, represents the panel feature of the third panel of the rth panel in the l-1th layer, and They represent the set of facets with high correlation with facet f and facet r, respectively. represents the cardinality of a set, represents the plane recognition weight matrix of the lth layer in the fifth encoder, , E represents the total number of layers, when l=1 , The feature representation of the object that represents the index of the third facet of the rth facet, when l=E is equal to the fifth identification feature of the third panel element f, is the sigmoid function;

[0118] In one embodiment of the present invention, the calculation formula of the sixth encoder is as follows:

[0119] ;

[0120] in represents the sixth identification feature, represents the updated teaching strategy of the input, Represents the convolution function.

[0121] In one embodiment of the present invention, the calculation formula of the second fusion feature layer is as follows:

[0122] ;

[0123] in, represents the second combined feature, represents the feature combination function (concatenation function or summation function), represents the bin feature of the f-th third bin of the E-th layer of the fifth encoder, A collection of panel features representing a third panel.

[0124] In one embodiment of the present invention, the calculation formula of the second FC layer is as follows:

[0125]

[0126] in represents the second output vector, and its e-th component value represents the probability value of the e-th strategy. The strategy with the largest probability value is selected as the output. The strategy group contains all executable strategies. A strategy includes emotional acceptance, empathic feedback, guiding positive mentality, emotional regulation, building a "growth mindset" mentality, and guiding self-dialogue;

[0127] For more complex strategies, vector representation can be used: one strategy is represented by a vector, and the j-th component of the vector represents the j-th psychological counseling method of the psychological counseling strategy; FC means fully connected.

[0128] In one embodiment of the present invention, the training steps of the policy update model are as follows:

[0129] ;

[0130] in represents the loss value, represents the total number of teaching unit classification categories, represents the total number of training samples, It is The symbol value of the training sample, if the The actual teaching unit classification of the training sample is the t-th classification category and its value is 1, otherwise it is 0. is the first prediction of the policy update model The probability that a training sample belongs to the tth classification category, Represents the logarithmic function with the natural constant e as the base.

[0131] In one embodiment of the present invention, the training steps of the psychological counseling model are as follows:

[0132] Step 301, copy the psychological counseling model to generate an Actor network and a TargetActor network, use a feedforward neural network as a Critic network, copy the Critic network to obtain a TargetCritic network, and randomly initialize the parameters of the Actor network, the TargetActor network, the Critic network, and the TargetCritic network;

[0133] Step 302, input The learners received information collected Go to the Actor network to get the Psychological counseling strategies implemented ;

[0134] No. Psychological counseling strategies implemented Add random noise to get Psychological counseling strategies for adding noise ;

[0135] The first Psychological counseling strategies implemented Input into the environment and get immediate rewards from feedback and The learners received information collected ; Experience Deposit into experience pool;

[0136] Step 303, randomly obtain N experiences from the experience pool;

[0137] enter Go to the TargetActor network to get the Psychological counseling strategies implemented ;

[0138] Step 304, calculate the first loss value and the second loss value , gradient descent method updates the Actor network and Critic network;

[0139] ;

[0140] ;

[0141] ;

[0142] in Represents input and The output of the Critic network is Indicates that and Input the output of the Critic network, Indicates execution The comprehensive Q value when express and The output of the TargetCritic network;

[0143]

[0144] in, , are the first weight parameter and the second weight parameter in the instant reward, As a result of emotional stability, For teaching acceptance;

[0145] ;

[0146] ;

[0147] in, is the learner’s emotional state score at time step t, such as the emotional score extracted from the learner’s language expression by a sentiment analysis algorithm, is the psychological state at time step t, including the learner’s current emotional state, historical interaction records, and other information. is the action taken by the counseling model at time step t (i.e., the counseling response given), is the new psychological state at time step t+1, reflecting the learner's Changes in emotional state after represents the action taken by the learner at time step t;

[0148] If the learner's emotional state improves after taking action (i.e. > ),but is positive, otherwise it is negative; by maximizing the cumulative reward, the model will learn which types of responses can effectively improve the learner's emotional state.

[0149] ;

[0150] in, Indicates direct feedback from learners (such as satisfaction ratings), Indicates interaction indicators (such as learner response speed, interaction duration), , They refer to the weight coefficients of direct feedback and interaction indicators respectively.

[0151] is the discount factor, which is a value between 0 and 1, used to balance the weights of immediate rewards and future rewards. The default value is 0.6;

[0152] pass Update the Actor network by Update the Critic network;

[0153] Step 305, every fixed number of network updates, update the TargetActor network so that its parameters are the same as those of the current Actor network, and update the TargetCritic network so that its parameters are the same as those of the current Critic network;

[0154] Step 306, until the Actor network and the Critic network converge, the step is terminated, or after the number of loops of steps 303-305 reaches a set value, the Actor network replaces the psychological counseling model, and the step is terminated.

[0155] Based on the above digital intelligence empowerment method, a digital intelligence empowerment system based on a large language model is also proposed, including:

[0156] Data acquisition module: responsible for collecting current teaching strategies, learners’ learning information and learners’ acceptance information, and converting the collected data into structured plane data and serial data streams;

[0157] Strategy update model: includes the first backbone network and the first branch network, inputs the first plane data and the second plane data, and outputs the updated teaching strategy

[0158] Psychological counseling model: including the second backbone network and the second branch network, inputting the updated teaching strategy and the second serial data stream, and outputting the psychological counseling strategy

[0159] Large language model module: Based on the pre-trained large language model (such as GPT, BERT, etc.), generate corresponding teaching content and interactive responses according to the updated teaching strategies and psychological counseling strategies

[0160] AI teaching module: responsible for implementing updated teaching strategies, interacting with learners, monitoring learners' feedback types and emotional states, and providing corresponding psychological counseling based on psychological counseling strategies.

[0161] At least one embodiment of the present disclosure provides a storage medium storing non-temporary computer-readable instructions for executing one or more steps in the aforementioned large language model-based digital intelligence empowerment method.

[0162] The computer program may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. Any reference signs in the claims should not be construed as limiting the scope.

[0163] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are protected by the present embodiment.

Claims

1. A digital intelligence empowerment method based on a large language model, characterized in that: The following steps are involved: Step 100, obtaining current teaching strategy, learner acceptance information and learner learning information; Learners' learning information includes learning preference information, motivation type information, plan execution information and self-discipline evaluation information; learners' acceptance information includes feedback information, emotional state information and acceptable teaching time information; current teaching strategies include several teaching units, which include incentive and reward mechanisms, teaching content presentation and feedback methods, learning plans and progress arrangements, teacher-student interaction methods and image information; Step 200, constructing first plane data based on the collected current teaching strategy, constructing a first serial data stream and a second plane data based on the collected learner acceptance information, and constructing a third plane data and a second serial data stream based on the collected learner acceptance information; Step 300, input the first plane data and the second plane data into the strategy update model, and output the result representing the updated teaching strategy, wherein the updated teaching strategy includes all updated teaching units, and the updated teaching units include incentive and reward mechanisms, teaching content presentation and feedback methods, learning plans and schedules, teacher-student interaction methods, and image information; input the updated teaching strategy and the second serial data stream into the psychological counseling model, and output the result representing the psychological counseling strategy, wherein the psychological counseling strategy includes emotional acceptance, empathic feedback, guiding a positive attitude, emotional regulation, building a "growth mindset" mentality, and guiding self-dialogue; In step 400, the AI ​​teacher performs teaching according to the updated teaching strategy outputted. Meanwhile, when the learner's feedback is poor and the emotional state deviates from the normal state, the AI ​​teacher stops the execution of the teaching strategy according to the output psychological counseling strategy until the feedback and emotional state return to normal, and then executes the teaching strategy again.

2. According to claim 1, a digital intelligence empowerment method based on a large language model is characterized in that: The first plane data includes a first facet for storing data, wherein each first facet indexes a data object, and the first plane data also includes a relationship list for storing association relationships between the data objects indexed by the first facet, wherein the element value in the i-th row and j-th column of the relationship list represents a quantitative value of the association relationship between the data objects indexed by the i-th and j-th first facets; The association relationships include high and low, corresponding to two quantitative values; Data objects include learners, teaching units, AI teachers, teaching subjects, and image types; The data objects with high association are defined as: learners and their AI teachers, learners and their teaching subjects, teaching units and their AI teachers, teaching subjects and their AI teachers, and AI teachers and their image types. Obtaining a first serial data stream based on the learner learning information sorting, the first serial data stream comprising n data units sorted by time, the tth data unit storing the second plane data generated by the learner learning information collected at the tth moment; The second plane data includes a second facet for storing data, each second facet indexes a data object, and the second plane data also includes a relationship list for storing association relationships between the data objects indexed by the second facet, wherein the element value in the i-th row and j-th column of the relationship list represents a quantitative value of the association relationship between the data objects indexed by the i-th and j-th second facets; The association relationships include high and low, corresponding to two quantitative values; Data objects include learners, preference types, motivation types, execution status, and self-discipline assessment scores; The data objects with high correlation are defined as: the learner’s preference type in the teaching, the learner’s motivation type in learning, the learner’s implementation of his / her learning plan, and the learner’s self-discipline evaluation score; Generating third plane data based on the learner's acceptance information includes a third facet for storing data, a third facet indexing a second serial data stream of a data object, the third plane data also includes a relationship list for storing association relationships between data objects indexed by the third facet, the element value of the i-th row and j-th column of the relationship list representing the quantitative value of the association relationship between the data objects indexed by the i-th and j-th third facets; The association relationships include high and low, corresponding to two quantitative values; Data objects include learners, AI teachers, opinion types, and emotional states; The data objects with high correlation are defined as: learners and the types of opinions they express, learners and the emotional states they give back when receiving teaching, learners and the AI ​​teachers they teach, and opinion types and their corresponding AI teachers. The second serial data stream of a data object includes n data units sorted by time, and the t-th data unit represents the learner's acceptance information collected at the t-th moment of the data object.

3. According to claim 2, a digital intelligence empowerment method based on a large language model is characterized in that: The strategy update model includes a first trunk network and a first branch network. The first trunk network includes a first encoder, a third encoder, a first feature fusion layer and a first FC layer. The first branch network includes a second encoder. The first encoder inputs first plane data and outputs a first recognition feature. At the same time, the second encoder inputs second plane data and outputs a second recognition feature. The first recognition feature and the second recognition feature are input into a first fusion feature layer. The first fusion feature layer outputs a first fusion state. The first fusion state is input into the third encoder. The third encoder outputs a third recognition feature to the first FC layer. The first FC layer outputs a result representing the updated teaching strategy.

4. According to claim 3, a digital intelligence empowerment method based on a large language model is characterized in that: The psychological counseling model includes a second trunk network and a second branch network. The second trunk network includes a fourth encoder, a fifth encoder, a second feature fusion layer and a second FC layer. The second branch network includes a sixth encoder. The fourth encoder inputs a second serial data stream and outputs a fourth recognition feature. The fourth recognition feature is input into the fifth encoder. The fifth encoder outputs a fifth recognition feature. At the same time, the sixth encoder inputs an updated teaching strategy and outputs a sixth recognition feature. The sixth recognition feature and the fifth recognition feature are input into the second fusion feature layer. The second fusion feature layer outputs a second fusion state. The second fusion state is input into the second FC layer. The second FC layer outputs the result representing the psychological counseling strategy.

5. According to claim 4, a digital intelligence empowerment method based on a large language model is characterized in that: The calculation formula of the first encoder is as follows: ； in represents the panel feature of the vth first panel of the lth layer, represents the set of first facets with high correlation with facet v, represents the first plane recognition weight matrix of layer l, is the attention weight of the first surfel u of layer l to the first surfel v, which is calculated by the following formula: ； in is a learnable attention vector, represents the vector concatenation operation, represents the second plane recognition weight matrix of layer l, represents transpose; and They represent the panel features of the vth first panel of the l-1th layer, , E represents the total number of layers, when l=1 , when l=E is equal to the identification feature of the first facet v; and The feature representations of the objects representing the vth and uth first facet indices respectively.

6. A digital intelligence empowerment method based on a large language model according to claim 5, characterized in that: The calculation formula of the second encoder is as follows: ； represents the panel feature of the cth second panel of the lth layer, represents the panel feature of the d-th second panel of the l-1th layer, and represent the set of second facets with high correlation with the second facet c and the second facet d respectively, represents the cardinality of a set, represents the third plane recognition weight matrix of layer l, , E represents the total number of layers, when l=1 , The feature representation of the object with the dth second facet index, when l=E is equal to the second identification feature of the second facet c, is the sigmoid function.

7. A digital intelligence empowerment method based on a large language model according to claim 6, characterized in that: The calculation formula of the first fusion feature layer is as follows: ； in, represents the first combination feature, represents the feature combination function (concatenation function or summation function), represents the panel feature of the vth first panel of the Eth layer of the first encoder, represents the set of panel features of the first panel, represents the panel feature of the c-th second panel of the E-th layer of the second encoder, a set of panel features representing a second panel; The calculation formula of the third encoder is as follows: ； in represents the tth third identification feature, represents the t-1th third identification feature, , represents the tth data unit of the first combined feature, and is the serial recognition weight matrix of the third encoder, is the serial recognition bias parameter of the third encoder, ReLU is the ReLU activation function; The calculation formula of the first FC layer is as follows: ； in represents the first output vector, and its cth component value represents the probability value of the cth strategy. The strategy with the largest probability value is selected as the output. The strategy group contains all executable strategies. A strategy includes incentive and reward mechanism, teaching content presentation and feedback method, learning plan, progress schedule, teacher-student interaction method, and image information; For more complex strategies, vector representation can be used: a strategy is represented by a vector, and the jth component of the vector represents the jth teaching information of the updated teaching strategy; represents the last third identification feature output by the third encoder, Indicates full connection.

8. The digital intelligence empowerment method based on a large language model according to claim 7 is characterized in that: The calculation formula of the fourth encoder is as follows: ； ； ； in represents the t-th data unit of the second serial data stream corresponding to the v-th third plane element of the third plane data, represents the t-1th fourth recognition feature of the vth third facet, and denote the first and second intermediate features of the vth third facet, respectively. represents the tth data unit of the second serial data stream of the vth third bin, and is the serial recognition weight matrix in the fourth encoder, is the serial recognition bias parameter in the fourth encoder, ReLU is the ReLU activation function, Represents the Sigmoid function; The calculation formula of the fifth encoder is as follows: ； represents the panel feature of the third panel of the fth layer, represents the panel feature of the third panel of the rth panel in the l-1th layer, and They represent the set of facets with high correlation with facet f and facet r, respectively. represents the cardinality of a set, represents the plane recognition weight matrix of the lth layer in the fifth encoder, , E represents the total number of layers, when l=1 , The feature representation of the object that represents the index of the third facet of the rth facet, when l=E is equal to the fifth identification feature of the third panel f, is the sigmoid function; The calculation formula of the sixth encoder is as follows: ； in represents the sixth identification feature, represents the updated teaching strategy of the input, Represents the convolution function.

9. A digital intelligence empowerment system based on a large language model, used to execute the steps in the digital intelligence empowerment method based on a large language model as described in any one of claims 1 to 8, characterized in that: include: Data acquisition module: responsible for collecting current teaching strategies, learners’ learning information and learners’ acceptance information, and converting the collected data into structured plane data and serial data streams; Strategy update model: including a first trunk network and a first branch network, inputting first plane data and second plane data, and outputting an updated teaching strategy; Psychological counseling model: including a second trunk network and a second branch network, inputting the updated teaching strategy and the second serial data stream, and outputting the psychological counseling strategy; Large language model module: Based on the pre-trained large language model, it generates corresponding teaching content and interactive responses according to the updated teaching strategies and psychological counseling strategies; AI teaching module: responsible for implementing updated teaching strategies, interacting with learners, monitoring learners' feedback types and emotional states, and providing corresponding psychological counseling based on psychological counseling strategies.

10. A storage medium, characterized in that: Non-temporary computer-readable instructions are stored for executing the steps in the digital intelligence empowerment method based on a large language model as described in any one of claims 1-8.

Images