A method, device and medium for eloquence training based on automatic weight adjustment

Abstract

The present invention discloses an eloquence training method, device and medium based on automatic weight adjustment, the method comprising: adjusting the weight set of a speech sample to obtain a first weight set; adjusting the weight set of the speech sample according to an immediate reward and a strategy gradient to obtain a second weight set; using an optimal value function to adjust and guide the third weight set to obtain a final weight set, and training the user according to the eloquence training method generated by the final weight set. The present invention proposes an eloquence training method, device and medium based on automatic weight adjustment, which helps to generate a first weight set that meets the actual target needs of the user through a first objective function, and adjusts the weight set through immediate rewards and strategy gradients, which can focus on solving the actual weaknesses of the user's eloquence expression, and then generate a highly personalized eloquence training method, which can solve the problem of difficulty in generating personalized training programs through weight adjustment according to the user's weaknesses in eloquence expression.

Description

Technical Field

[0001] The invention relates to the field of speech training, and in particular to an eloquence training method, device and medium based on automatic weight adjustment. Background Art

[0002] Personalized eloquence training programs play a significant role in improving personal communication skills, expression skills, and shaping personal style. Through systematic and scientific training, it can not only improve personal performance in various occasions, but also build self-confidence, so that the trainees can be more calm in various occasions, laying a solid foundation for achieving personal goals and career development. Existing eloquence training methods cover basic voice exercises, language organization skills, listening comprehension skills, logical thinking training, speech skills improvement, psychological quality training, interactive communication skills and simulation practice applications. Through comprehensive training and practical application, trainees can comprehensively improve their eloquence.

[0003] However, current training methods are usually too templated and single-minded, failing to fully consider the individual differences and characteristics of each trainee. It is difficult to adjust the training content according to the trainee's weak oral expression, and to develop a more personalized oral training plan that meets individual needs, helping trainees to comprehensively improve their oral ability. Summary of the invention

[0004] The present invention provides an eloquence training method, device and medium based on automatic weight adjustment, so as to solve the problem that it is difficult to generate a personalized training program by weight adjustment according to the user's eloquence expression weaknesses.

[0005] In order to solve the above problems, the present invention provides an eloquence training method based on automatic weight adjustment, comprising:

[0006] Get user's speech samples;

[0007] Adjusting the weight set of the speech sample according to the first objective function of the user to obtain a first weight set;

[0008] The weight set of the speech sample is adjusted according to a preset instant reward and a strategy gradient to obtain a second weight set; wherein the strategy gradient is established by measuring the probability of the weight distribution;

[0009] generating a third weight set according to the first weight set and the second weight set;

[0010] The third weight set is adjusted and guided by using a preset optimal value function to obtain a final weight set, and the user is trained according to the eloquence training method generated by the final weight set.

[0011] The first objective function of the present invention can directly reflect the user's expected value. Therefore, adjusting the weight set of the speech sample according to the first objective function can make the obtained first weight set evolve in the adjustment direction desired by the user, which is more in line with the user's actual needs; the immediate reward, as the reward obtained immediately after taking a certain action, is directly related to the planned short-term goal, which helps to make better decisions in subsequent steps, and because the policy gradient is established by measuring the probability of weight distribution, the second weight set is adjusted on the basis of knowing the performance status of each indicator of the user, so that the obtained second weight set is more focused on solving the user's actual eloquence expression weaknesses; finally, the third weight set is adjusted and guided by the optimal value function, which can maximize the reasonable arrangement of the weight distribution in the final weight set based on the first weight set and the second weight set, so that the generated eloquence training method can effectively help users train.

[0012] Compared with the prior art, the present invention is based on the actual eloquence of the user, and uses the first objective function to help generate a first weight set that meets the actual target needs of the user, and adjusts the weight set through instant rewards and strategy gradients. It can focus on solving the actual weaknesses of the user's eloquence expression, and then generate a highly personalized eloquence training method. Therefore, it can solve the problem of difficulty in generating personalized training plans through weight adjustment according to the user's weaknesses in eloquence expression.

[0013] As a preferred solution, the weight set of the speech sample is adjusted according to the first objective function of the user to obtain a first weight set, which is specifically:

[0014] The speech sample is scored according to the initial first eloquence dimension indicator set to obtain a quantitative score; wherein the initial first eloquence dimension indicator set includes a speech recognition indicator, a language logic indicator, and an emotional expression indicator;

[0015] Establishing the first objective function by measuring the difference between the quantized score and the user target score;

[0016] The weight set of the speech sample is updated according to the first objective function to obtain the first weight set; wherein the first weight set is established based on the speech sample and the first eloquence dimension indicator set.

[0017] This preferred solution scores the speech samples according to the initial first eloquence dimension indicator set, and then obtains a quantitative score, so it can reflect the user's actual eloquence status from the dimensions of speech recognition, language logic and emotional expression; because the first objective function is established based on the difference between the quantitative score and the user's target score, the first weight set of the speech sample is updated according to the first objective function, which can make the indicator weights in the speech sample adjusted according to the user's advanced plan, so that the obtained first weight set is more in line with the user's actual needs and can be targeted.

[0018] As a preferred solution, the weight set of the speech sample is adjusted according to the preset instant reward and strategy gradient to obtain a second weight set, which is specifically:

[0019] Calculating the speech sample according to the initial second eloquence dimension indicator set to obtain a second eloquence dimension indicator set;

[0020] Obtaining an initial instant reward corresponding to an initial second weight set in the second eloquence dimension indicator set, and establishing a second objective function according to a discount factor and the initial instant reward;

[0021] Optimizing the initial second weight set according to the second objective function to obtain a final instant reward;

[0022] The second weight set is adjusted according to the final immediate reward and the policy gradient to obtain a second weight set.

[0023] The initial instant reward of this preferred solution is obtained based on the user's second eloquence dimension indicator set, so it can reflect the user's actual current speech effect, which is an objective evaluation of the user's market performance; since the policy gradient is established by measuring the probability of weight distribution, the second objective function is established based on the user's actual performance, based on the knowledge of the user's performance of each indicator, and the second weight set is adjusted on this basis, so that the obtained second weight set is more focused on solving the user's actual eloquence expression weaknesses.

[0024] As a preferred solution, the policy gradient is established by measuring the probability of weight distribution, specifically:

[0025] Obtaining weight strategy parameters of the initial second weight set;

[0026] According to the weight strategy parameters, the probability of weight distribution is obtained by calculating the random strategy;

[0027] The policy gradient is obtained by calculating the probability of the weight distribution based on the criterion of reducing the variance.

[0028] In this preferred solution, the probability of weight distribution is obtained by calculating the random strategy, so that the random strategy can be learned to ensure that the importance of the index arrangement is the optimal choice under different conditions; and the variance can be reduced to make the data more concentrated and reduce the discreteness of the data, which helps the policy gradient to more accurately describe the characteristics of the data and improve the reliability of data information adjustment.

[0029] As a preferred solution, the optimal value function is specifically:

[0030] Calculating the speech sample according to the initial third eloquence dimension indicator set to obtain the third eloquence dimension indicator set;

[0031] Using the preset first formula, a comprehensive reward function is established according to the time series of the third eloquence dimension indicator set and the third weight set;

[0032] The parameters of the value function are updated according to the comprehensive reward function to obtain the optimal value function; wherein the value function is established according to the current language behavior feature vector state of the speech sample and the third weight set.

[0033] In this preferred solution, since time series data can contain more information, it can provide rich data features for the generation of a comprehensive reward function; the comprehensive reward function is used as the basis for the reinforcement learning algorithm to guide the intelligent agent to learn how to make the best decision. It defines the sum of rewards obtained by the user based on a series of actions taken and the observed state sequence in the process of interacting with the environment. Therefore, using the comprehensive reward function to update the parameters of the value function can enable the obtained optimal value function to adjust its evaluation of the state and behavior based on historical experience, and measure the expected value of the long-term cumulative reward that the user can obtain in a given state and the expected value of the long-term cumulative reward that can be obtained after taking a specific behavior.

[0034] As a preferred solution, a preset optimal value function is used to adjust and guide the third weight set to obtain a final weight set, specifically:

[0035] generating a policy network according to the third weight set;

[0036] The optimal value function is used to guide the policy network to update parameters, so that the policy network evolves in a direction of increasing future cumulative rewards, thereby obtaining an optimized final policy network;

[0037] The weight set corresponding to the final strategy network is used as the final weight set.

[0038] In this preferred solution, each decision or action will trigger an immediate reward, and this immediate reward is only a part of the total future reward. The goal of this solution is to learn a strategy that can maximize the sum of rewards in all future steps, that is, the future cumulative reward; by evolving the strategy network in the direction of improving future cumulative rewards, the future cumulative rewards can be maximized. This evolution process can learn how to make long-term optimal decisions in an uncertain environment, not just based on the current immediate rewards. This method can maximize the reasonable arrangement of the final weight concentration and weight distribution.

[0039] As a preferred solution, generating a third weight set according to the first weight set and the second weight set is specifically:

[0040] generating a transition weight set according to the first weight set and the second weight set;

[0041] Using a deep neural network model, the third weight set is generated on a time scale based on the transition weight set and the language expression feature vector in the speech sample.

[0042] The time scale of this preferred solution can help the deep neural network model find a balance between immediate rewards and future potential rewards by adjusting the model's emphasis on future rewards when generating new weights, thereby helping the third weight set achieve better long-term performance.

[0043] The present invention also provides an eloquence training device based on automatic weight adjustment, comprising a sample module, an adjustment module, a reward module, a weight module and a training module;

[0044] Wherein, the sample module is used to obtain the user's speech sample;

[0045] The adjustment module is used to adjust the weight set of the speech sample according to the first objective function of the user to obtain a first weight set;

[0046] The reward module is used to adjust the weight set of the speech sample according to the preset instant reward and the strategy gradient to obtain a second weight set; wherein the strategy gradient is established by measuring the probability of the weight distribution;

[0047] The weight module is used to generate a third weight set according to the first weight set and the second weight set;

[0048] The training module is used to use a preset optimal value function to adjust and guide the third weight set to obtain a final weight set, and train the user according to the eloquence training method generated by the final weight set.

[0049] As a preferred solution, the adjustment module includes a quantization unit, a difference unit and an update unit;

[0050] The quantification unit is used to evaluate the speech sample according to the initial first eloquence dimension indicator set to obtain a quantitative score; wherein the initial first eloquence dimension indicator set includes a speech recognition indicator, a language logic indicator and an emotional expression indicator;

[0051] The difference unit is used to establish the first objective function by measuring the difference between the quantized score and the user target score;

[0052] The updating unit is used to update the weight set of the speech sample according to the first objective function to obtain the first weight set; wherein the first weight set is established based on the speech sample and the first eloquence dimension indicator set.

[0053] As a preferred solution, the reward module includes a dimension unit, a target unit, an optimization unit and an adjustment unit;

[0054] The dimension unit is used to calculate the speech sample according to the initial second eloquence dimension indicator set to obtain the second eloquence dimension indicator set;

[0055] The target unit is used to obtain an initial instant reward corresponding to an initial second weight set in the second eloquence dimension indicator set, and establish a second target function according to a discount factor and the initial instant reward;

[0056] The optimization unit is used to optimize the initial second weight set according to the second objective function to obtain a final instant reward;

[0057] The adjustment unit is used to adjust the second weight set according to the final immediate reward and the policy gradient to obtain a second weight set.

[0058] As a preferred solution, the reward module includes a strategy unit, a probability unit and a gradient unit;

[0059] Wherein, the strategy unit is used to obtain the weight strategy parameters of the initial second weight set;

[0060] The probability unit is used to obtain the probability of weight distribution by calculating the random strategy according to the weight strategy parameters;

[0061] The gradient unit is used to obtain the policy gradient based on the probability calculation of the weight distribution based on the standard of reducing the variance.

[0062] As a preferred solution, the optimal value function is specifically:

[0063] Calculating the speech sample according to the initial third eloquence dimension indicator set to obtain the third eloquence dimension indicator set;

[0064] Using the preset first formula, a comprehensive reward function is established according to the time series of the third eloquence dimension indicator set and the third weight set;

[0065] The parameters of the value function are updated according to the comprehensive reward function to obtain the optimal value function; wherein the value function is established according to the current language behavior feature vector state of the speech sample and the third weight set.

[0066] As a preferred solution, the training module includes

[0067] Wherein, the generating unit is used to generate a policy network according to the third weight set;

[0068] The evolution unit is used to guide the policy network to update parameters through the optimal value function, so that the policy network evolves in a direction of improving future cumulative rewards, thereby obtaining an optimized final policy network;

[0069] The weight unit is used to use the weight set corresponding to the final strategy network as the final weight set.

[0070] As a preferred solution, the weight module includes a transition unit and a network unit;

[0071] The transition unit is used to generate a transition weight set according to the first weight set and the second weight set;

[0072] The network unit is used to use a deep neural network model to generate the third weight set based on the transition weight set and the language expression feature vector in the speech sample on a time scale.

[0073] The present invention also provides a storage medium, on which a computer program is stored. The computer program is called and executed by a computer to implement the above-mentioned eloquence training method based on automatic weight adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flowchart of an eloquence training method based on automatic weight adjustment provided by an embodiment of the present invention;

[0075] Figure 2 It is a structural schematic diagram of an eloquence training device based on automatic weight adjustment provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0077] In the description of the present application, it should be understood that the terms "first", "second" ... "sixth" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first", "second" ... "sixth" may explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, "several" means two or more.

[0078] The eloquence training method based on automatic weight adjustment provided by the embodiment of the present invention is mainly used in the situation where it is difficult to adjust the corresponding weight of the training content according to the weak eloquence expression of the trainee, highlight the key content, and formulate a personalized eloquence training plan that better meets individual needs, so as to help the trainee comprehensively improve his oratory ability.

[0079] Embodiment 1:

[0080] See also Figure 1 The embodiment of the present invention provides an eloquence training method based on automatic weight adjustment, including S1 to S5, and the specific implementation steps are as follows:

[0081] S1. Obtain the user's speech sample.

[0082] Step S1 of the embodiment of the present invention is specifically:

[0083] Acquire speech samples of the user; wherein the speech samples include voice signals, video streams, and text records.

[0084] S2. Adjust the weight set of the speech sample according to the user's first objective function to obtain a first weight set.

[0085] In step S2 of the embodiment of the present invention, S2 includes S2.1 to S2.5; wherein S2.1 is a process of defining an initial first eloquence dimension indicator set, S2.2 is a process of performing data preprocessing, S2.3 is a process of calculating a quantitative score, S2.4 is a process of constructing a first weight set, and S2.5 is a process of performing weight optimization, specifically:

[0086] S2.1. Define the initial first eloquence dimension indicator set; the initial first eloquence dimension indicator set includes speech recognition indicators, language logic indicators and emotional expression indicators, specifically including: content richness (CR), language clarity (LC), logical rigor (LS), emotional expressiveness (EE), audience appeal (AA), interactive feedback response (IF), voice rhythm (PR), non-verbal expression (NVE), speech rate control (SC), audience reaction (AR) and content novelty (CN).

[0087] S2.2, cleaning the speech samples; wherein the cleaning means include removing noise, invalid information, etc.;

[0088] The speech samples of each dimension d were standardized according to the initial first eloquence dimension indicator set;

[0089] The formula for standardization is:

[0090]

[0091] Among them, Z d,i is the standardized value of the i-th sample on dimension d, X d,i is the original value, μ d is the mean of dimension d, σ d is the standard deviation of dimension d.

[0092] In this embodiment S2.2, the data of each dimension is converted into a standardized value, which is helpful for subsequent mathematical calculations.

[0093] S2.3. Using the deep neural network model F d According to each dimension d of the initial first eloquence dimension indicator set, the speech sample is scored and the quantitative score Score is obtained. d,i ;

[0094] The quantitative scores are:

[0095] Score d,i =F d (Z d,i )

[0096] S2.4. Initialize the weight vector W of each dimension d in the speech sample according to the initial first eloquence dimension indicator set, ensuring that the sum of all dimension weights is 1; and set the initial weight distribution set according to preset historical data or expert experience;

[0097] Quantify the score by measuring d,i The difference between the user's target score and the target score is used to establish the first objective function Loss(W);

[0098] Using gradient descent or a preset optimization algorithm, the initial weight distribution set of the speech sample is updated according to the first objective function Loss(W) to obtain a first weight set;

[0099] Among them, the first objective function is:

[0100]

[0101] For the parameter OverallScore i (W) has:

[0102]

[0103] The formula for updating the initial weight distribution set of speech samples is:

[0104]

[0105] Among them, OverallScore i (W) is the comprehensive score of the i-th sample calculated based on the current weight distribution, TargetScore i is its true performance score, t is the current iteration number, α7 is the learning rate; W d is the weight of dimension d in the weight vector, satisfying

[0106] S2.5. Calculate the new quantization score after each weight update and compare it with the actual result. If the first objective function Loss (W) decreases, keep the updated weight; otherwise, reselect the learning rate or adjust the optimization algorithm.

[0107] The weight allocation scheme is continuously optimized in multiple iterations until a preset convergence condition or target performance is reached, and an optimized first weight set is obtained.

[0108] Overall, this embodiment S2 scores the speech sample according to the initial first eloquence dimension indicator set to obtain a quantitative score, so it can reflect the user's actual eloquence status from the dimensions of speech recognition, language logic and emotional expression; because the first objective function is established based on the difference between the quantitative score and the user's target score, the first weight set of the speech sample is updated according to the first objective function, which can adjust the indicator weights in the speech sample according to the user's advanced plan, so that the obtained first weight set is more in line with the user's actual needs and can be targeted.

[0109] S3. Adjust the weight set of the speech sample according to the preset instant reward and strategy gradient to obtain a second weight set; wherein the strategy gradient is established by measuring the probability of weight distribution.

[0110] In step S3 of the embodiment of the present invention, S3 includes S3.1 to S3.5; wherein S3.1 is a process of defining an initial second eloquence dimension indicator set, S3.2 is a process of calculating a final instant reward, S3.3 is a process of calculating a policy gradient, S3.4 is a process of constructing a second weight set, and S3.5 is a process of weight optimization, specifically:

[0111] S3.1. Define an initial second eloquence dimension indicator set containing m eloquence dimension indicators and a corresponding evaluation score vector S = [S1, S2, ..., S m ]; the initial second eloquence dimension indicator set includes indicators D1, D2, ..., D m ;

[0112] And set the feedback signal F and the historical performance indicator set H.

[0113] S3.2, calculating the speech sample according to the initial second eloquence dimension indicator set to obtain the second eloquence dimension indicator set;

[0114] Initialize the weights corresponding to each indicator in the second eloquence dimension indicator set, and assign a weight value to each dimension i in the second eloquence dimension indicator set Obtaining an initial second weight set;

[0115] Obtain the initial instant reward corresponding to the initial second weight set in the second eloquence dimension indicator set, and establish the second objective function J(W) based on the discount factor and the initial instant reward (t) ); the initial immediate reward can be expressed as the actual effect feedback (such as user satisfaction, market performance, etc.) after the strategy selection (i.e. weight distribution);

[0116] According to the second objective function J(W (t) ) Optimize the initial second weight set to get the final instant reward R k (W (t) );

[0117] Furthermore, the weighted comprehensive score of each dimension i in the second eloquence dimension indicator set is displayed by calculating the comprehensive score;

[0118] Among them, the weights can be expressed as follows after initialization:

[0119]

[0120] The second objective function is:

[0121]

[0122] The comprehensive score is calculated as follows:

[0123]

[0124] in, is the initial weight of the i-th eloquence dimension, W (t) is the weight vector at time step t, R k (W (t) ) is the immediate reward depending on the current weight distribution, γ k is the discount factor.

[0125] S3.3, obtaining weight strategy parameters of the initial second weight set;

[0126] According to the weight strategy parameters, the probability of weight distribution is obtained by calculating the random strategy;

[0127] Taking reducing variance as the criterion, the policy gradient is calculated based on the probability of weight distribution;

[0128] Among them, the update formula of the policy gradient is:

[0129]

[0130] Among them, θ represents the weight strategy parameter, N is the number of sampling times, T j is the duration of the jth trial, a j is in state s j The action taken under (i.e. the weight distribution of the choice), π θ (a j ∣s j ) is the current policy parameter θ in state s j Next select action a j The probability of; b(θ) can be a baseline function used to reduce variance, such as V value estimation.

[0131] In this embodiment S3.3, the probability of weight distribution is obtained by calculating the random strategy, so that the random strategy can be learned to ensure that the importance of the indicator arrangement is the optimal choice under different conditions; and the method of reducing the variance can make the data more concentrated and reduce the discreteness of the data, which helps the policy gradient to more accurately describe the characteristics of the data and improve the reliability of data information adjustment.

[0132] S3.4. According to the final instant reward R k (W (t) ) and the policy gradient to update and adjust the second weight set to obtain a second weight set;

[0133] Among them, the update adjustment formula of the second weight set is:

[0134]

[0135] Among them, α is the learning rate, which is used to control the magnitude of weight change in each iteration; is the second objective function J(W (t) ), namely, the policy gradient.

[0136] S3.5, based on the feedback signal F and the historical performance indicator set H, the weight strategy is continuously adjusted to adapt to the changing environment;

[0137] The iterative optimization process is continued until a preset convergence condition is reached (for example, the weight changes very little after several consecutive iterations), or the maximum number of iterations is reached, to obtain an optimized second weight set.

[0138] It should be noted that on the basis of this weight strategy, if a more complex adaptive algorithm is adopted, such as combining performance within a time window, market trends or other external factors, the complexity and effectiveness of the adjustment mechanism will be further enriched.

[0139] Overall, the initial instant reward of S3 in this embodiment is obtained based on the user's second eloquence dimension indicator set, so it can reflect the user's actual speech effect at the moment, which is an objective evaluation of the user's market performance. Since the policy gradient is established by measuring the probability of weight distribution, the second objective function is established based on the user's actual performance on the basis of knowing the performance of each indicator of the user, so as to adjust the second weight set on this basis, so that the obtained second weight set is more focused on solving the user's actual eloquence expression weaknesses.

[0140] Moreover, the present embodiment S3 is based on an advanced machine learning algorithm, and can calculate the weight values ​​of different eloquence dimension indicators in real time according to factors such as pre-set target scenarios, user characteristics and current performance characteristics; the generation of weights can be dynamically adjusted as the user performance changes, ensuring that the evaluation system always meets actual needs.

[0141] S4. Generate a third weight set according to the first weight set and the second weight set.

[0142] In step S4 of the embodiment of the present invention, S4 includes S4.1 to S4.2, S4.1 is a process of defining an initial third eloquence dimension indicator set, and S4.2 is a process of constructing a third weight set, specifically:

[0143] S4.1. Based on the initial first eloquence dimension indicator set and the initial second eloquence dimension indicator set, the eloquence dimension indicator set is divided into multiple sub-dimensions to obtain the initial third eloquence dimension indicator set C u =[C u1 ,C u2 ,...,Cum ], such as traditional behavioral dimensions such as user click-through rate, dwell time, and content creation quality, as well as eloquence dimensions such as language fluency and opinion influence;

[0144] Convert speech samples into actionable state features S u ** ;

[0145] Initialize the algorithm model parameters that the personalized strategy generation depends on, and set the initial state S u(0) The corresponding eloquence dimension status C u(0) .

[0146] S4.2. Generate a transition weight set according to the first weight set and the second weight set;

[0147] Using the deep neural network model f, a third weight set W is generated based on the transition weight set and the language expression feature vector in the speech sample on a time scale u *(t) ;

[0148] Among them, the language expression feature vector is:

[0149] S u * =[S u1 ,S u2 ,...,S un ,C u ]

[0150] The third weight set is:

[0151]

[0152] Among them, S un Represents the traditional behavior feature vector of user u; C u It is the initial third eloquence dimension indicator set;

[0153] W u *(t) is the personalized weight vector assigned to user u at time step t; function f is a deep neural network model based on the current language behavior feature vector state S u ** and historical information To dynamically generate weights.

[0154] The time scale of S4.2 of this embodiment can help the deep neural network model find a balance between immediate rewards and future potential rewards when generating new weights by adjusting the model's emphasis on future rewards, thereby helping the third weight set achieve better long-term performance.

[0155] S5. Use the preset optimal value function to adjust and guide the third weight set to obtain a final weight set, and train the user according to the eloquence training method generated by the final weight set.

[0156] In step S5 of the embodiment of the present invention, S5 includes S5.1 to S5.6; wherein S5.1 is a process of constructing an optimal value function, S5.2 is a process of adjusting weight distribution, S5.3 is a process of obtaining instant rewards, S5.4 is a process of constructing a final weight set, S5.5 is a process of generating an eloquence training method and training a user, and S5.6 is a process of verifying the effect, specifically:

[0157] S5.1. Calculate the speech sample according to the initial third eloquence dimension indicator set to obtain the third eloquence dimension indicator set;

[0158] Using the preset first formula, a comprehensive reward function J(W) is established based on the time series of the third eloquence dimension indicator set and the third weight set. u *(t) );

[0159] Using the TD error in reinforcement learning, according to the comprehensive reward function J(W u *(t) ) updates the parameters of the value function and obtains the optimal value function Q(S u ** ,W u *(t) ), where the value function is based on the current language behavior feature vector state S of the speech sample u **(t) and the third weight set W u *(t) and established;

[0160] Among them, the comprehensive reward function is:

[0161]

[0162] Among them, the comprehensive reward function J(W u *(t) ) is specifically used to quantitatively evaluate the effect of the personalized strategy at a certain time step t.

[0163] In this embodiment S5.1, since time series data can accommodate more information, it can provide rich data features for the generation of a comprehensive reward function; the comprehensive reward function is used as the basis for the reinforcement learning algorithm to guide the intelligent agent to learn how to make the best decision. It defines the sum of rewards obtained by the user based on a series of actions taken and the observed state sequence in the process of interacting with the environment. Therefore, using the comprehensive reward function to update the parameters of the value function can enable the obtained optimal value function to adjust its evaluation of the state and behavior based on historical experience, and measure the expected value of the long-term cumulative reward that the user can obtain in a given state and the expected value of the long-term cumulative reward that can be obtained after taking a specific behavior.

[0164] S5.2. Conduct cross-impact analysis on different dimensions (traditional behavior and eloquence sub-dimensions) and use the multimodal fusion coefficient To quantify the degree of interaction between them, and adjust the weight distribution in the strategy accordingly; among them, the multimodal fusion coefficient It is used to measure the interaction effects between different dimensions (including traditional behavior and eloquence sub-dimensions) at time step t;

[0165] Among them, the multimodal fusion coefficient is:

[0166]

[0167] in, is the learned weight matrix; σ is the activation function (such as Sigmoid function), ensuring that the fusion coefficient is between [0,1].

[0168] S5.3, according to the third weight set W u *(t) Push corresponding content or services to users, and observe users' responses to the pushed content in real time, so as to obtain immediate feedback signals (such as user satisfaction, conversion rate, etc.), which can be regarded as an immediate reward Rt given by the environment.

[0169] S5.4, generating a policy network according to the third weight set;

[0170] Using the gradient boosting method, the optimal Q value provided by the optimal value function guides the policy network to update the policy network parameters θ, so that the policy network evolves in the direction of improving future cumulative rewards, and obtains the optimized final policy network; where the future cumulative reward is established on the basis of the immediate reward Rt by considering the reward discount and possible future state transitions;

[0171] The weight set corresponding to the final strategy network is used as the final weight set;

[0172] Among them, the formula for updating the parameters of the policy network is:

[0173]

[0174] In this embodiment S5.4, each decision or action will trigger an immediate reward, and this immediate reward is only a part of the total future reward. The goal of this solution is to learn a strategy that can maximize the sum of rewards in all future steps, that is, the future cumulative reward; by making the strategy network evolve in the direction of improving the future cumulative reward, the future cumulative reward can be maximized. This evolution process can learn how to make the best long-term decision in an uncertain environment, not just based on the current immediate reward. This method can maximize the reasonable arrangement of the weight distribution in the final weight concentration;

[0175] Moreover, when updating the policy network, the contribution of the eloquence dimension to the Q-value estimation is taken into account, and the policy network parameters θ can be effectively optimized by gradient ascent.

[0176] S5.5, by repeatedly calculating the third weight set W u *(t) The final weight set is continuously optimized iteratively until the preset number of learning times is reached or the strategy performance index reaches the convergence standard, and the optimized final weight set W is obtained. u *(t+1) ;

[0177] According to the final weight set W u *(t+1) The generated eloquence training method is used to train users.

[0178] S5.6. Conduct short-term and long-term verification of the user's usage results, and visualize the user data, specifically:

[0179] Short-term effect verification: Compare the changes in the scores of various eloquence indicators of users before and after applying the eloquence training method. If a certain indicator is significantly improved, it means that the method is effective. In addition, the improvement of users in simulated situations or actual communication can also be used as direct feedback on short-term effects.

[0180] Long-term effect verification: Track the user's growth curve over a long period of time to observe whether the overall oral ability is continuously and steadily improved after multiple rounds of iterative optimization; conduct satisfaction surveys and case studies regularly to understand the actual acceptance of users' oral training methods and effectiveness evaluation.

[0181] Visual display: Through data analysis tools, the user's progress data can be visualized, such as making line charts, bar charts or radar charts, etc., to intuitively present the user's progress trends in various dimensions.

[0182] The first embodiment of the present invention also includes S6, based on the user's current ability and improvement goals, using artificial intelligence technology to generate personalized training plans and decision support strategies, and presenting the growth path and expected improvement effects in a visual form.

[0183] In this embodiment S6, S6 includes S6.1 to S6.6, specifically:

[0184] S6.1, collect and process data; S6.1 includes S6.11 to S6.15, S6.11 is the process of data source access, S6.12 is the process of capturing real-time data, S6.13 is the process of data preprocessing and cleaning, S6.14 is the process of feature extraction and fusion, S6.15 is the process of format conversion and storage, specifically:

[0185] S6.11. By integrating various sensors and interface technologies, the source of multimodal data is connected to the platform.

[0186] For example, use a highly sensitive microphone to collect the user's audio stream; capture facial expressions and body movements through a camera or a dedicated video analysis system; connect physiological signal monitoring equipment (such as a heart rate belt, skin conductance sensor, etc.) to record the user's physiological responses.

[0187] S6.12, collect real-time voice data, monitor physiological signals and analyze video streams of users to obtain raw data;

[0188] Among them, real-time voice data collection: using advanced speech recognition technology (ASR) to transcribe the user's voice information into text in real time, and extract voice features such as speaking speed, pitch, volume and other parameters.

[0189] Physiological signal monitoring: read and record the user's physiological signal data, including but not limited to heart rate, breathing rate and skin electrical response, etc. These indicators help reflect the user's true emotions and psychological stress.

[0190] Video stream analysis: Use computer vision algorithms to analyze video streams to parse out the user's emotional state (such as happiness, surprise, and nervousness) and body language characteristics, such as changes in gestures and head movement trajectories.

[0191] S6.13. Preprocess the raw data, including removing noise interference, standardizing signal strength, and correcting time synchronization problems, to ensure that the data quality is reliable and can be used for subsequent analysis, so as to obtain modal data;

[0192] Among them, for unstructured video and image data, a series of image processing operations such as image segmentation, feature point positioning and expression recognition are performed.

[0193] S6.14, extract key features of modal data, such as MFCC features of audio, facial key point change sequence of video, trend change of physiological signals, etc.;

[0194] Fuse data from different modalities and establish multimodal feature vectors to better characterize the user's comprehensive expressive ability.

[0195] S6.15. Convert the processed multimodal data into a unified format to facilitate subsequent analysis, mining and visualization;

[0196] Storing this data in the database provides a real-time updated basic data source for upper-level decision support and analysis.

[0197] In this embodiment S6.1, data is collected from different dimensions, and preliminarily processed and stored, so as to prepare for subsequent analysis and presentation.

[0198] S6.2, perform deep learning and machine learning on the collected data; wherein S6.2 includes S6.21 to S6.25, S6.21 is the process of data collection and preprocessing, S6.22 is the process of feature engineering and extraction, S6.23 is the process of machine learning and deep learning modeling, S6.24 is the process of real-time analysis and feedback, and S6.25 is the process of result interpretation and visualization, specifically:

[0199] S6.21. Obtaining raw data from users during speech or communication from different sources (such as audio, video, and text records);

[0200] Preprocessing is performed on different types of data, including but not limited to noise reduction and standardization of speech signals, cleaning, word segmentation and removal of irrelevant information of text content, frame extraction of video images, facial expression recognition, etc.

[0201] S6.22. For the language logic dimension, use NLP technology to construct syntax trees, semantic networks or dependency graphs, and calculate sentence complexity, coherence and rationality indicators;

[0202] The emotional expressiveness dimension uses sentiment analysis algorithms to identify and quantify the emotional color in speech and text. For example, deep learning models are used to analyze the pitch and intensity changes of speech to capture emotional fluctuations.

[0203] In terms of timbre control, the clarity, loudness, and stability of the sound are evaluated through spectrum analysis, Mel-frequency cepstral coefficients (MFCCs) or other acoustic features;

[0204] The rhythm and prosody dimension uses rhythm analysis algorithms, such as counting parameters such as pronunciation duration, pause time, and speech speed changes, to determine the speaker's ability to control rhythm.

[0205] S6.23. Based on the extracted features related to the eloquence dimension, design and train specific machine learning models and deep learning networks, such as logistic regression, support vector machine, neural network, convolutional neural network (CNN), recurrent neural network (RNN) or Transformer structures;

[0206] These models map input features to corresponding quantitative evaluation indicators and optimize model parameters through training sets to ensure that they can accurately predict and evaluate the performance of each eloquence dimension.

[0207] S6.24, when data continues to flow in, the intelligent analysis engine runs the above algorithm in real time, performs online analysis on the new data stream, and outputs the corresponding quantitative evaluation results;

[0208] Combining multi-dimensional quantitative indicators, the system comprehensively considers the weight of each factor to form a comprehensive score that fully reflects the individual's oral ability.

[0209] S6.25. Transform complex mathematical model outputs into easily understandable and interpretable forms, such as charts, dashboards, or reports, to provide intuitive feedback to users;

[0210] The scores and development trends of each eloquence dimension are displayed on the visual decision support platform to help users quickly identify their own strengths and improvement directions.

[0211] This embodiment S6.2 includes reinforcement learning algorithms and data analysis tools for dynamic weight adjustment, strategy generation and dimensional indicator analysis, which can effectively analyze data.

[0212] S6.3. Establish a knowledge graph of eloquence to associate the relationships between indicators of different dimensions, and provide functions such as data comparison and trend analysis. S6.3 includes S6.31 to S6.36. S6.31 is the process of data storage and management, S6.32 is the process of building a knowledge graph, S6.33 is the process of real-time data update and synchronization, S6.34 is the process of multi-dimensional association analysis, S6.35 is the process of making personalized recommendations and optimization suggestions, and S6.36 is the process of comparison and trend analysis. Specifically:

[0213] S6.31. Build a dynamically updated database system to store users’ historical speech and communication performance data;

[0214] Among them, historical speech and communication performance data include but are not limited to voice recordings, text content, video clips, and various eloquence dimension indicator results generated by the intelligent analysis engine.

[0215] S6.32. The platform uses relational database and knowledge graph technology to structure the oral ability data of different users into nodes and edges, where each node represents a specific evaluation indicator or event, and the edge represents the association or causal relationship between different indicators;

[0216] Each user's data is organized into an individual knowledge sub-graph, and the rules of mutual influence between different eloquence dimensions are revealed through attribute association and logical reasoning.

[0217] S6.33. As new speech data is continuously input and analysis results are produced, the database and knowledge graph will be updated in real time to ensure that the information is always kept up to date, reflecting the changes and development trends of the user's oral ability.

[0218] S6.34, based on the knowledge graph, the system can perform complex multi-dimensional correlation analysis;

[0219] For example, by querying and traversing the graph, we can discover deeper connections such as whether improving emotional expression helps improve language logic, or how tone control affects the audience's acceptance of users.

[0220] S6.35. Based on the user's historical data and pattern mining in the knowledge graph, the platform can provide personalized training suggestions and improvement plans to help users improve their eloquence skills in one or more aspects in a targeted manner.

[0221] S6.36. Users can use the functions of this layer to compare their performance in different periods, view the development trend of indicators in each dimension, and intuitively display the reasons and paths for progress or regression with the help of knowledge graphs;

[0222] At the same time, users can also compare themselves with other users, learn from others' successful experiences, and continuously optimize their personal speaking skills.

[0223] This embodiment S6.3 includes user data and domain knowledge, and can support user interaction and generate personalized suggestions.

[0224] S6.4: Display the analysis results of the data and provide targeted exercises; S6.4 includes S6.41 to S6.46, S6.41 is the process of data visualization design, S6.42 is the process of building a personalized dashboard, S6.43 is the process of realizing dynamic interactive functions, S6.44 is the process of making intelligent suggestions and feedback, S6.45 is the process of real-time monitoring and updating, and S6.46 is the process of tracking and responding to user behavior, specifically:

[0225] S6.41, using advanced data visualization technology, presents the complex analysis results of S6.3 (i.e., knowledge graph and database layer) to users in an intuitive and easy-to-understand way;

[0226] For example, through various graphical means such as charts, dashboards, heat maps and timelines, the user's scores, rankings and changing trends in various dimensions of oral ability can be clearly displayed.

[0227] S6.42. Provide each user with a customized dashboard interface, summarize their key indicators, and configure different views according to user needs, such as speech content quality, speech speed control, volume adjustment, body language effect, etc., so that users can quickly obtain their own advantages and areas for improvement.

[0228] S6.43. Users can interact with the visualization interface in real time by clicking, sliding or filtering, and gain in-depth understanding of the specific information behind each data, such as viewing detailed records of performance in a certain period of time, comparing performance differences in different occasions, etc.

[0229] S6.44. Based on the system's in-depth analysis of the user's historical data, targeted practice suggestions and optimization strategies are provided directly on the visual interface. When a user views a certain indicator, the system can provide an action guide to improve the indicator or recommend related learning resources based on the relevance of the knowledge graph.

[0230] S6.45. The visual presentation content is connected to the backend database in real time, which means that every time a user uploads new data, it will be immediately reflected in the corresponding visual chart, ensuring that the user sees the latest and most comprehensive performance evaluation results.

[0231] S6.46. Capture user operation behaviors, dynamically adjust and optimize the visual interface layout and content push according to the user's focus and usage habits, and form a good human-computer interaction experience.

[0232] This embodiment S6.4 provides a user-friendly interface that can intuitively and effectively present quantitative evaluation results, weight distribution and personalized recommendations; by providing a highly intuitive visualization interface, complex quantitative results are displayed in the form of graphics, charts, etc., helping decision makers to clearly understand the correlation between various indicators, quickly identify key factors and potential risks, simplify the decision-making process, and improve decision-making speed.

[0233] S6.5: Based on the user's current capabilities and improvement goals, the platform uses artificial intelligence technology to generate personalized training plans and decision support strategies, and displays the growth path and expected improvement effects in a visual form; S6.5 includes S6.51 to S6.56, S6.51 is the process of analyzing user capabilities, S6.52 is the process of intelligent algorithm matching, S6.53 is the process of generating personalized training plans, S6.54 is the process of formulating decision support strategies, S6.55 is the process of constructing a visual presentation of the growth path, and S6.56 is the process of simulating and displaying the expected improvement effects, specifically:

[0234] S6.51. By collecting and analyzing users' historical behavior data, evaluation results, and clear improvement goals, a user capability model is formed, including in-depth exploration and quantitative evaluation of the user's current skill level, strengths, and weaknesses.

[0235] S6.52. Based on artificial intelligence technologies such as machine learning or deep learning, compare and analyze the user capability model with the huge knowledge base and successful case database to find the growth path and training method that best suits the user's characteristics.

[0236] S6.53. Based on the analysis results, the system automatically generates a personalized training plan, which includes specific steps such as learning content, practice tasks and actual combat simulations at different stages, and ensures that these contents are targeted at the specific needs of users and help them improve their capabilities in all dimensions.

[0237] S6.54, the system generates relevant decision support strategies based on the user's goals and capability development;

[0238] For example, it can provide suggestions for adjusting training focus at key points, or predict the expected improvement effects under different decisions.

[0239] S6.55. Using visualization tools and techniques, the system will display the user's growth path in the form of charts, timelines, etc., clearly depicting the gradual progress from the current state to the ideal goal, allowing users to intuitively understand and track their own progress.

[0240] S6.56. Enable the visual interface to simulate and display the expected results after implementing the recommended strategy;

[0241] For example, the expected change trend chart of various capability indicators can help users predict the results after adopting suggestions and enhance their confidence and motivation to implement personalized recommendation plans.

[0242] This embodiment S6.5 uses user data and engine analysis results to generate personalized improvement suggestions and optimization strategies, which can effectively train users and improve their eloquence and expression skills.

[0243] S6.6. Verify the user experience, including:

[0244] The platform effect can be verified through comparative experiments, such as whether the eloquence improvement of the user group under the guidance of the application platform after a certain period of time is significantly different from that of the control group;

[0245] Through regular user satisfaction surveys and feedback collection, we can understand the actual effectiveness of platform functions and whether user needs are met;

[0246] Combined with specific application scenarios, such as public speaking competitions, business negotiations, etc., track and record users' actual performance improvements as direct evidence of the effectiveness of the platform's functions.

[0247] Generally speaking, this embodiment has the following beneficial effects:

[0248] The first objective function of the embodiment of the present invention can directly reflect the user's expected value. Therefore, by adjusting the weight set of the speech sample according to the first objective function, the obtained first weight set can evolve in the adjustment direction desired by the user, which is more in line with the actual needs of the user. The immediate reward is a reward obtained immediately after taking a certain action. It is directly related to the short-term goal of the plan, which helps to make better decisions in subsequent steps. Since the policy gradient is established by measuring the probability of weight distribution, the second weight set is adjusted on the basis of knowing the performance status of each indicator of the user, so that the obtained second weight set is more focused on solving the actual eloquence expression weaknesses of the user. Finally, the third weight set is adjusted and guided by the optimal value function, which can maximize the reasonable arrangement of the weight distribution in the final weight set based on the first weight set and the second weight set, so that the generated eloquence training method can effectively help users to train.

[0249] Furthermore, through real-time data updates and automatic adjustment of weight strategies, this embodiment can flexibly change the decision basis according to environmental changes and new information inputs, ensuring continuous provision of optimal decision support in a dynamic environment, and can continuously learn lessons from historical decision results, iteratively optimize algorithms and related parameters, thereby achieving the evolution of training programs and performance improvements.

[0250] Embodiment 2:

[0251] See also Figure 2 , an embodiment of the present invention provides a modal analysis device for harmonic resonance of a power distribution system, comprising a sample module 10, an adjustment module 20, a reward module 30, a weight module 40 and a training module 50;

[0252] Wherein, the sample module 10 is used to obtain the user's speech sample;

[0253] An adjustment module 20, configured to adjust the weight set of the speech sample according to the user's first objective function to obtain a first weight set;

[0254] A reward module 30, configured to adjust the weight set of the speech sample according to a preset instant reward and a policy gradient to obtain a second weight set; wherein the policy gradient is established by measuring the probability of the weight distribution;

[0255] A weight module 40, configured to generate a third weight set according to the first weight set and the second weight set;

[0256] The training module 50 is used to use a preset optimal value function to adjust and guide the third weight set to obtain a final weight set, and train the user according to the eloquence training method generated by the final weight set.

[0257] In one embodiment, the sample module 10 is specifically:

[0258] Acquire speech samples of the user; wherein the speech samples include voice signals, video streams, and text records.

[0259] In one embodiment, the adjustment module 20 includes a first definition unit, a data processing unit, a quantification unit, a difference unit, an update unit and a first optimization unit; wherein the first definition unit is a process of defining an initial first eloquence dimension indicator set, the data processing unit is a process of performing data preprocessing, the quantification unit is a process of calculating a quantitative score, the difference unit and the update unit are processes of constructing a first weight set, and the first optimization unit is a process of performing weight optimization, specifically:

[0260] The first definition unit is used to define an initial first eloquence dimension indicator set; wherein the initial first eloquence dimension indicator set includes speech recognition indicators, language logic indicators and emotional expression indicators, specifically including: content richness (CR), language clarity (LC), logical rigor (LS), emotional expressiveness (EE), audience appeal (AA), interactive feedback response (IF), voice rhythm (PR), non-verbal expression (NVE), speech rate control (SC), audience reaction (AR) and content novelty (CN).

[0261] A data processing unit, used for cleaning the speech samples; wherein the cleaning means include removing noise, invalid information, etc.;

[0262] The speech samples of each dimension d were standardized according to the initial first eloquence dimension indicator set;

[0263] The formula for standardization is:

[0264]

[0265] Among them, Z d,i is the standardized value of the i-th sample on dimension d, X d,i is the original value, μ d is the mean of dimension d, σ d is the standard deviation of dimension d.

[0266] The data processing unit of this embodiment converts the data of each dimension into a standardized value, which is helpful for subsequent mathematical calculations.

[0267] Quantization unit for using deep neural network model F d According to each dimension d of the initial first eloquence dimension indicator set, the speech sample is scored and the quantitative score Score is obtained. d,i ;

[0268] The quantitative scores are:

[0269] Score d,i =F d (Z d,i )

[0270] The difference unit is used to initialize the weight vector W of each dimension d in the speech sample according to the initial first eloquence dimension indicator set, ensuring that the sum of all dimension weights is 1; and setting the initial weight distribution set according to preset historical data or expert experience;

[0271] The difference unit is also used to quantify the score by measuring Score d,i The difference between the user's target score and the target score is used to establish the first objective function Loss(W);

[0272] An updating unit, configured to update an initial weight distribution set of the speech sample according to a first objective function Loss(W) using a gradient descent or a preset optimization algorithm to obtain a first weight set;

[0273] Among them, the first objective function is:

[0274]

[0275] For the parameter OverallScore i (W) has:

[0276]

[0277] The formula for updating the initial weight distribution set of speech samples is:

[0278]

[0279] Among them, OverallScore i(W) is the comprehensive score of the i-th sample calculated based on the current weight distribution, TargetScore i is its true performance score, t is the current iteration number, α7 is the learning rate; W d is the weight of dimension d in the weight vector, satisfying

[0280] The first optimization unit is used to calculate the new quantization score after each weight update and compare and verify it with the actual result. If the first objective function Loss (W) decreases, the updated weight is retained; otherwise, the learning rate needs to be reselected or the optimization algorithm needs to be adjusted;

[0281] The weight allocation scheme is continuously optimized in multiple iterations until a preset convergence condition or target performance is reached, and an optimized first weight set is obtained.

[0282] Overall, the adjustment module 20 of this embodiment scores the speech sample according to the initial first eloquence dimension indicator set, and then obtains a quantitative score, so it can reflect the user's actual eloquence status from the dimensions of speech recognition, language logic and emotional expression; because the first objective function is established based on the difference between the quantitative score and the user's target score, the first weight set of the speech sample is updated according to the first objective function, which can adjust the indicator weights in the speech sample according to the user's advanced plan, so that the obtained first weight set is more in line with the user's actual needs and can be targeted.

[0283] In one embodiment, the reward module 30 includes a second definition unit, a dimension unit, a target unit, an optimization unit, a strategy unit, a probability unit, a gradient unit, an adjustment unit, and a second optimization unit; wherein the second definition unit is a process of defining an initial second eloquence dimension indicator set, the dimension unit, the target unit, and the optimization unit are processes of calculating the final instant reward, the strategy unit, the probability unit, and the gradient unit are processes of calculating the policy gradient, the adjustment unit is a process of constructing a second weight set, and the second optimization unit is a process of weight optimization, specifically:

[0284] The second definition unit is used to define an initial second eloquence dimension indicator set including m eloquence dimension indicators and a corresponding evaluation score vector S=[S1, S2, ..., S m ]; the initial second eloquence dimension indicator set includes indicators D1, D2, ..., D m ;

[0285] And set the feedback signal F and the historical performance indicator set H.

[0286] A dimension unit, used to calculate the speech sample according to the initial second eloquence dimension indicator set to obtain a second eloquence dimension indicator set;

[0287] The dimension unit is also used to initialize the weights corresponding to each indicator in the second eloquence dimension indicator set, and assign a weight value to each dimension i in the second eloquence dimension indicator set Obtaining an initial second weight set;

[0288] The target unit is used to obtain the initial instant reward corresponding to the initial second weight set in the second eloquence dimension indicator set, and establish the second target function J(W) according to the discount factor and the initial instant reward. (t) ); the initial immediate reward can be expressed as the actual effect feedback (such as user satisfaction, market performance, etc.) after the strategy selection (i.e. weight distribution);

[0289] The optimization unit is used to optimize the second objective function J(W (t) ) Optimize the initial second weight set to get the final instant reward R k (W (t) );

[0290] Furthermore, the weighted comprehensive score of each dimension i in the second eloquence dimension indicator set is displayed by calculating the comprehensive score;

[0291] Among them, the weights can be expressed as follows after initialization:

[0292]

[0293] The second objective function is:

[0294]

[0295] The comprehensive score is calculated as follows:

[0296]

[0297] in, is the initial weight of the i-th eloquence dimension, W (t) is the weight vector at time step t, R k (W (t) ) is the immediate reward depending on the current weight distribution, γ k is the discount factor.

[0298] A strategy unit, used to obtain weight strategy parameters of an initial second weight set;

[0299] The probability unit is used to obtain the probability of weight distribution by calculating the random strategy according to the weight strategy parameters;

[0300] Gradient unit, which is used to calculate the policy gradient based on the probability of weight distribution with the reduction of variance as the criterion;

[0301] Among them, the update formula of the policy gradient is:

[0302]

[0303] Among them, θ represents the weight strategy parameter, N is the number of sampling times, T j is the duration of the jth trial, a j is in state s j The action taken under (i.e. the weight distribution of the choice), π θ (a j ∣s j ) is the current policy parameter θ in state s j Next select action a The probability of j; b(θ) can be a baseline function used to reduce variance, such as V value estimation.

[0304] In the strategy unit, probability unit and gradient unit of the present embodiment, the probability of weight distribution is obtained by calculating the random strategy, so that the random strategy can be learned to ensure that the importance arrangement of the indicators is the optimal choice under different conditions; and the variance can be reduced to make the data more concentrated and reduce the discreteness of the data, which helps the policy gradient to more accurately describe the characteristics of the data and improve the reliability of data information adjustment.

[0305] The adjustment unit is used to adjust the final instant reward R k (W (t) ) and the policy gradient to update and adjust the second weight set to obtain a second weight set;

[0306] Among them, the update adjustment formula of the second weight set is:

[0307]

[0308] Among them, α is the learning rate, which is used to control the magnitude of weight change in each iteration; For the weight W i The second objective function J(W (t) ), namely, the policy gradient.

[0309] The second optimization unit is used to continuously adjust the weight strategy to adapt to the changing environment according to the feedback signal F and the historical performance indicator set H;

[0310] The iterative optimization process is continued until a preset convergence condition is reached (for example, the weight changes very little after several consecutive iterations), or the maximum number of iterations is reached, and the optimized second weight set is obtained.

[0311] It should be noted that on the basis of this weight strategy, if a more complex adaptive algorithm is adopted, such as combining performance within a time window, market trends or other external factors, the complexity and effectiveness of the adjustment mechanism will be further enriched.

[0312] From a holistic perspective, the initial instant reward of the reward module 30 of this embodiment is obtained based on the user's second eloquence dimension indicator set, so it can reflect the user's actual speech effect at the moment, which is an objective evaluation of the user's market performance. Since the policy gradient is established by measuring the probability of weight distribution, the second objective function is established based on the user's actual performance on the basis of knowing the performance of each indicator of the user, so as to adjust the second weight set on this basis, so that the obtained second weight set is more focused on solving the user's actual eloquence expression weaknesses.

[0313] In addition, the reward module 30 of this embodiment is based on advanced machine learning algorithms, and can calculate the weight values ​​of different eloquence dimension indicators in real time according to pre-set target scenarios, user characteristics, current performance characteristics and other factors; the generation of weights can be dynamically adjusted as user performance changes, ensuring that the evaluation system always meets actual needs.

[0314] In one embodiment, the weight module 40 includes a third definition unit, a transition unit and a network unit. The third definition unit is a process of defining an initial third eloquence dimension indicator set, and the transition unit and the network unit are a process of constructing a third weight set, specifically:

[0315] The third definition unit is used to refine and divide the eloquence dimension indicator set into multiple sub-dimensions based on the initial first eloquence dimension indicator set and the initial second eloquence dimension indicator set, and obtain the initial third eloquence dimension indicator set Cu=[Cu1, Cu2,..., Cum], such as multiple traditional behavior dimensions such as user click rate, dwell time and content creation quality, as well as eloquence dimensions such as language expression fluency and opinion influence;

[0316] Convert speech samples into actionable state features S u ** ;

[0317] Initialize the algorithm model parameters that the personalized strategy generation depends on, and set the initial state S u(0) The corresponding eloquence dimension status C u(0) .

[0318] A transition unit, configured to generate a transition weight set according to the first weight set and the second weight set;

[0319] A network unit is used to generate a third weight set W based on the transition weight set and the language expression feature vector in the speech sample on a time scale using a deep neural network model f. u *(t) ;

[0320] Among them, the language expression feature vector is:

[0321] S u * =[S u1 ,S u2 ,...,S un ,C u ]

[0322] The third weight set is:

[0323]

[0324] Among them, S un Represents the traditional behavior feature vector of user u; C u It is the initial third eloquence dimension indicator set;

[0325] W u *(t) is the personalized weight vector assigned to user u at time step t; function f is a deep neural network model based on the current language behavior feature vector state S u ** and historical information To dynamically generate weights.

[0326] The time scales of the transition units and network units in this embodiment can help the deep neural network model find a balance between immediate rewards and future potential rewards by adjusting the model's emphasis on future rewards when generating new weights, thereby helping the third weight set achieve better long-term performance.

[0327] In one embodiment, the training module 50 includes a function construction unit, a strategy adjustment unit, a reward unit, a generation unit, an evolution unit, a weight unit, an actual training unit and an effect verification unit; wherein the function construction unit is a process of constructing an optimal value function, the strategy adjustment unit is a process of adjusting weight distribution, the reward unit is a process of obtaining instant rewards, the generation unit, the evolution unit and the weight unit are processes of constructing a final weight set, the actual training unit is a process of generating an eloquence training method and training a user, and the effect verification unit is a process of performing effect verification, specifically:

[0328] A function construction unit, used for calculating the speech sample according to the initial third eloquence dimension indicator set to obtain the third eloquence dimension indicator set;

[0329] The function building unit is further used to use the preset first formula to establish a comprehensive reward function J(W) according to the time series of the third eloquence dimension indicator set and the third weight set. u *(t) );

[0330] Function building unit, also used to use TD error in reinforcement learning, according to the comprehensive reward function J(W u *(t) ) updates the parameters of the value function and obtains the optimal value function Q(S u ** ,W u *(t) ), where the value function is based on the current language behavior feature vector state S of the speech sample u **(t) and the third weight set W u *(t) and established;

[0331] Among them, the comprehensive reward function is:

[0332]

[0333] Among them, the comprehensive reward function J(W u *(t) ) is specifically used to quantitatively evaluate the effect of the personalized strategy at a certain time step t.

[0334] In the function construction unit of this embodiment, since time series data can accommodate more information, it can provide rich data features for the generation of a comprehensive reward function; the comprehensive reward function is the basis for the reinforcement learning algorithm to guide the intelligent agent to learn how to make the best decision. It defines the sum of rewards obtained by the user based on a series of actions taken and the observed state sequence in the process of interacting with the environment. Therefore, using the comprehensive reward function to update the parameters of the value function can enable the obtained optimal value function to adjust its evaluation of the state and behavior based on historical experience, and measure the expected value of the long-term cumulative reward that the user can obtain in a given state and the expected value of the long-term cumulative reward that can be obtained after taking a specific behavior.

[0335] Strategy adjustment unit, used to conduct cross-impact analysis on different dimensions (traditional behavior and eloquence sub-dimensions), through multimodal fusion coefficient To quantify the degree of interaction between them, and adjust the weight distribution in the strategy accordingly; among them, the multimodal fusion coefficient It is used to measure the interaction effects between different dimensions (including traditional behavior and eloquence sub-dimensions) at time step t;

[0336] Among them, the multimodal fusion coefficient is:

[0337]

[0338] in, is the learned weight matrix; σ is the activation function (such as Sigmoid function), ensuring that the fusion coefficient is between [0,1].

[0339] Reward unit, for rewarding u *(t) Push corresponding content or services to users, and observe users' responses to the pushed content in real time, so as to obtain immediate feedback signals (such as user satisfaction, conversion rate, etc.), which can be regarded as an immediate reward Rt given by the environment.

[0340] A generating unit, configured to generate a policy network according to a third weight set;

[0341] The evolution unit is used to use the gradient boosting method to guide the policy network to update the policy network parameters θ through the optimal Q value provided by the optimal value function, so that the policy network evolves in the direction of improving the future cumulative reward, and obtains the optimized final policy network; wherein the future cumulative reward is established on the basis of the immediate reward Rt by considering the reward discount and the possible state transfer in the future;

[0342] A weight unit, used to take the weight set corresponding to the final strategy network as the final weight set;

[0343] Among them, the formula for updating the parameters of the policy network is:

[0344]

[0345] In the generation unit, evolution unit and weight unit of this embodiment, each decision or action will trigger an immediate reward, and this immediate reward is only a part of the future total reward. The goal of this solution is to learn a strategy that can maximize the sum of rewards in all future steps, that is, the future cumulative reward; by making the strategy network evolve in the direction of improving the future cumulative reward, the future cumulative reward can be maximized. This evolution process can learn how to make the best long-term decision in an uncertain environment, not just based on the current immediate reward. This method can maximize the reasonable arrangement of the weight distribution in the final weight concentration;

[0346] Moreover, when updating the policy network, the contribution of the eloquence dimension to the Q-value estimation is taken into account, and the policy network parameters θ can be effectively optimized by gradient ascent.

[0347] The actual training unit is used to repeatedly calculate the third set of weights W u *(t)The final weight set is continuously optimized iteratively until the preset number of learning times is reached or the strategy performance index reaches the convergence standard, and the optimized final weight set W is obtained. u *(t+1) ;

[0348] According to the final weight set W u *(t+1) The generated eloquence training method is used to train users.

[0349] The effect verification unit is used to verify the short-term and long-term effects of the user's use effect and visualize the user data, specifically:

[0350] Short-term effect verification: Compare the changes in the scores of various eloquence indicators of users before and after applying the eloquence training method. If a certain indicator is significantly improved, it means that the method is effective. In addition, the improvement of users in simulated situations or actual communication can also be used as direct feedback on short-term effects.

[0351] Long-term effect verification: Track the user's growth curve over a long period of time to observe whether the overall oral ability is continuously and steadily improved after multiple rounds of iterative optimization; conduct satisfaction surveys and case studies regularly to understand the actual acceptance of users' oral training methods and effectiveness evaluation.

[0352] Visual display: Through data analysis tools, the user's progress data can be visualized, such as making line charts, bar charts or radar charts, etc., to intuitively present the user's progress trends in various dimensions.

[0353] The second embodiment also includes a decision-making module;

[0354] Among them, the decision module 60 is used to generate personalized training plans and decision support strategies based on the user's current capabilities and improvement goals using artificial intelligence technology, and to present the growth path and expected improvement effects in a visual form.

[0355] In one embodiment, the decision module 60 includes a data unit, a learning unit, a graph unit, a display unit, a strategy unit and a verification unit, specifically:

[0356] The data unit is used to collect and process data; wherein the data unit includes a first data sub-unit, a second data sub-unit, a third data sub-unit, a fourth data sub-unit and a fifth data sub-unit, wherein the first data sub-unit is a process of data source access, the second data sub-unit is a process of capturing real-time data, the third data sub-unit is a process of data preprocessing and cleaning, the fourth data sub-unit is a process of feature extraction and fusion, and the fifth data sub-unit is a process of format conversion and storage, specifically:

[0357] The first data subunit is used to connect the source of multimodal data to the platform by integrating various sensors and interface technologies.

[0358] For example, use a highly sensitive microphone to collect the user's audio stream; capture facial expressions and body movements through a camera or a dedicated video analysis system; connect physiological signal monitoring equipment (such as a heart rate belt, skin conductance sensor, etc.) to record the user's physiological responses.

[0359] The second data subunit is used to collect real-time voice data, monitor physiological signals and analyze video streams of users to obtain raw data;

[0360] Among them, real-time voice data collection: using advanced speech recognition technology (ASR) to transcribe the user's voice information into text in real time, and extract voice features such as speaking speed, pitch, volume and other parameters.

[0361] Physiological signal monitoring: read and record the user's physiological signal data, including but not limited to heart rate, breathing rate and skin electrical response, etc. These indicators help reflect the user's true emotions and psychological stress.

[0362] Video stream analysis: Use computer vision algorithms to analyze video streams to parse out the user's emotional state (such as happiness, surprise, and nervousness) and body language characteristics, such as changes in gestures and head movement trajectories.

[0363] The third data subunit is used to pre-process the raw data, including removing noise interference, standardizing signal strength, and correcting time synchronization problems, etc. It is necessary to ensure that the data quality is reliable and can be used for subsequent analysis, so as to obtain modal data;

[0364] Among them, for unstructured video and image data, a series of image processing operations such as image segmentation, feature point positioning and expression recognition are performed.

[0365] The fourth data subunit is used to extract key features of modal data, such as MFCC features of audio, facial key point change sequence of video, trend change of physiological signals, etc.;

[0366] Fuse data from different modalities and establish multimodal feature vectors to better characterize the user's comprehensive expressive ability.

[0367] The fifth data subunit is used to convert the processed multimodal data into a unified format to facilitate subsequent analysis, mining and visualization;

[0368] Storing this data in the database provides a real-time updated basic data source for upper-level decision support and analysis.

[0369] The data unit of this embodiment collects data from different dimensions, and performs preliminary processing and storage on the data, so as to prepare for subsequent analysis and presentation.

[0370] A learning unit, used for performing deep learning and machine learning processing on the collected data;

[0371] Among them, the learning unit includes the first learning subunit, the second learning subunit, the third learning subunit, the fourth learning subunit and the fifth learning subunit. The first learning subunit is the process of data collection and preprocessing, the second learning subunit is the process of feature engineering and extraction, the third learning subunit is the process of machine learning and deep learning modeling, the fourth learning subunit is the process of real-time analysis and feedback, and the fifth learning subunit is the process of result interpretation and visualization, specifically:

[0372] The first learning subunit is used to obtain the original data of the user in speech or communication from different sources (such as audio, video and text records);

[0373] Preprocessing is performed on different types of data, including but not limited to noise reduction and standardization of speech signals, cleaning, word segmentation and removal of irrelevant information of text content, frame extraction of video images, facial expression recognition, etc.

[0374] The second learning sub-unit is used to construct syntax trees, semantic networks or dependency graphs using NLP technology for the language logic dimension, and calculate sentence complexity, coherence and rationality indicators;

[0375] The emotional expressiveness dimension uses sentiment analysis algorithms to identify and quantify the emotional color in speech and text. For example, deep learning models are used to analyze the pitch and intensity changes of speech to capture emotional fluctuations.

[0376] In terms of timbre control, the clarity, loudness, and stability of the sound are evaluated through spectrum analysis, Mel-frequency cepstral coefficients (MFCCs) or other acoustic features;

[0377] The rhythm and prosody dimension uses rhythm analysis algorithms, such as counting parameters such as pronunciation duration, pause time, and speech speed changes, to determine the speaker's ability to control rhythm.

[0378] The third learning subunit is used to design and train specific machine learning models and deep learning networks, such as logistic regression, support vector machine, neural network, convolutional neural network (CNN), recurrent neural network (RNN) or Transformer structures, based on the extracted eloquence dimension related features;

[0379] These models map input features to corresponding quantitative evaluation indicators and optimize model parameters through training sets to ensure that they can accurately predict and evaluate the performance of each eloquence dimension.

[0380] The fourth learning subunit is used for the intelligent analysis engine to run the above algorithm in real time when data is continuously flowing in, to perform online analysis on the new data stream, and to output corresponding quantitative evaluation results;

[0381] Combining multi-dimensional quantitative indicators, the system comprehensively considers the weight of each factor to form a comprehensive score that fully reflects the individual's oral ability.

[0382] The fifth learning subunit is used to transform complex mathematical model outputs into easily understandable and interpretable forms, such as charts, dashboards, or reports, providing intuitive feedback to users;

[0383] The scores and development trends of each eloquence dimension are displayed on the visual decision support platform to help users quickly identify their own strengths and improvement directions.

[0384] The learning unit of this embodiment includes a reinforcement learning algorithm and a data analysis tool for dynamic weight adjustment, strategy generation, and dimensional indicator analysis, and can effectively analyze the data.

[0385] The graph unit is used to establish a knowledge graph of eloquence to associate the relationships between indicators of different dimensions and provide functions such as data comparison and trend analysis;

[0386] Among them, the graph unit includes the first graph subunit, the second graph subunit, the third graph subunit, the fourth graph subunit, the fifth graph subunit and the sixth graph subunit. The first graph subunit is the process of data storage and management, the second graph subunit is the process of building a knowledge graph, the third graph subunit is the process of real-time data update and synchronization, the fourth graph subunit is the process of multi-dimensional association analysis, the fifth graph subunit is the process of personalized recommendation and optimization suggestion, and the sixth graph subunit is the process of comparison and trend analysis, specifically:

[0387] The first graph subunit is used to construct a dynamically updated database system for storing the user's historical speech and communication performance data;

[0388] Among them, historical speech and communication performance data include but are not limited to voice recordings, text content, video clips, and various eloquence dimension indicator results generated by the intelligent analysis engine.

[0389] The second graph sub-unit is used by the platform to structure the oral ability data of different users into nodes and edges using relational database and knowledge graph technology. Each node represents a specific evaluation indicator or event, and the edge represents the association or causal relationship between different indicators.

[0390] Each user's data is organized into an individual knowledge sub-graph, and the rules of mutual influence between different eloquence dimensions are revealed through attribute association and logical reasoning.

[0391] The third graph sub-unit is used to set that as new speech data is continuously input and analysis results are output, the database and knowledge graph will be updated in real time, and ensure that the information is always kept up to date to reflect the changes and development trends of the user's oral ability.

[0392] The fourth graph subunit is used to enable the system to perform complex multi-dimensional association analysis based on the knowledge graph;

[0393] For example, by querying and traversing the graph, we can discover deeper connections such as whether improving emotional expression helps improve language logic, or how tone control affects the audience's acceptance of users.

[0394] The fifth graph sub-unit is used to mine patterns in the user's historical data and knowledge graph. The platform can provide personalized training suggestions and improvement plans to help users improve their eloquence skills in one or more aspects in a targeted manner.

[0395] The sixth graph sub-unit is used for users to compare their performance in different periods through the functions of this layer, view the development trend of indicators in each dimension, and intuitively display the reasons and paths for progress or regression with the help of knowledge graphs;

[0396] At the same time, users can also compare themselves with other users, learn from others' successful experiences, and continuously optimize their personal speaking skills.

[0397] The graph unit of this embodiment includes user data and domain knowledge, and can support user interaction and generate personalized suggestions.

[0398] The presentation unit is used to display the analysis results of the data and provide targeted exercises;

[0399] Among them, the display unit includes the first display sub-unit, the second display sub-unit, the third display sub-unit, the fourth display sub-unit and the fifth display sub-unit; the first display sub-unit is the process of data visualization design, the second display sub-unit is the process of building a personalized dashboard, the third display sub-unit is the process of realizing dynamic interactive functions, the first and fourth display sub-units are the processes of making intelligent suggestions and feedback, the fifth display sub-unit is the process of real-time monitoring and updating, and the sixth display sub-unit is the process of tracking and responding to user behaviors, specifically:

[0400] The first display subunit is used to use advanced data visualization technology to present the complex analysis results of S6.3 (i.e., the knowledge graph and database layer) to users in an intuitive and easy-to-understand way;

[0401] For example, through various graphical means such as charts, dashboards, heat maps and timelines, the user's scores, rankings and changing trends in various dimensions of oral ability can be clearly displayed.

[0402] The second display sub-unit is used to provide each user with a customized dashboard interface, summarize their key indicators, and configure different views according to user needs, such as speech content quality, speech speed control, volume adjustment, body language effects, etc., to help users quickly obtain their own advantages and areas for improvement.

[0403] The third display sub-unit is used for users to interact with the visualization interface in real time by clicking, sliding or filtering, and to gain an in-depth understanding of the specific information behind each piece of data, such as viewing detailed records of performance within a certain period of time, comparing performance differences in different occasions, etc.

[0404] The fourth display sub-unit is used to provide targeted practice suggestions and optimization strategies directly on the visual interface based on the system's in-depth analysis of the user's historical data. When a user views a certain indicator, the system can provide an action guide to improve the indicator or recommend related learning resources based on the relevance of the knowledge graph.

[0405] The fifth display sub-unit, which is used for visual presentation of content, is connected to the backend database in real time. This means that every time a user uploads new data, it will be immediately reflected in the corresponding visual chart, ensuring that the user sees the latest and most comprehensive performance evaluation results.

[0406] The sixth display sub-unit is used to capture the user's operating behavior, dynamically adjust and optimize the visual interface layout and content push according to the user's focus and usage habits, and form a good human-computer interaction experience.

[0407] The display unit of this embodiment provides a user-friendly interface that can intuitively and effectively present quantitative evaluation results, weight distribution and personalized suggestions; by providing a highly intuitive visualization interface, complex quantitative results are displayed in the form of graphics, charts, etc., helping decision makers to clearly understand the correlation between various indicators, quickly identify key factors and potential risks, simplify the decision-making process, and improve decision-making speed.

[0408] Strategy unit: Based on the user's current capabilities and improvement goals, the platform uses artificial intelligence technology to generate personalized training plans and decision support strategies, and displays the growth path and expected improvement effects in a visual form;

[0409] Among them, the strategy unit includes the first strategy subunit, the second strategy subunit, the third strategy subunit, the fourth strategy subunit, the fifth strategy subunit and the sixth strategy subunit. The first strategy subunit is the process of analyzing user capabilities, the second strategy subunit is the process of intelligent algorithm matching, the third strategy subunit is the process of generating personalized training plans, the fourth strategy subunit is the process of formulating decision support strategies, the fifth strategy subunit is the process of constructing a visual presentation of the growth path, and the sixth strategy subunit is the process of simulating and displaying the expected improvement effect, specifically:

[0410] The first strategy sub-unit is used to form a user capability model by collecting and analyzing the user's historical behavior data, evaluation results, and clear improvement goals, including in-depth exploration and quantitative evaluation of the user's current skill level, strengths, and weaknesses.

[0411] The second strategy sub-unit is used to compare and analyze the user capability model with a huge knowledge base and successful case database based on artificial intelligence technologies such as machine learning or deep learning, so as to find the growth path and training method that best suits the user's characteristics.

[0412] The third strategy sub-unit is used to automatically generate a personalized training plan based on the analysis results. The plan includes specific steps such as learning content, practice tasks, and actual combat simulations at different stages, and ensures that these contents are targeted at the specific needs of users and help them improve their capabilities in various dimensions.

[0413] The fourth strategy subunit is used for the system to generate relevant decision support strategies based on the user's goals and capability development;

[0414] For example, it can provide suggestions for adjusting training focus at key points, or predict the expected improvement effects under different decisions.

[0415] The fifth strategy sub-unit is used to utilize visualization tools and techniques. The system will present the user's growth path in the form of charts, timelines, etc., clearly depicting the gradual progress from the current state to the ideal goal, allowing users to intuitively understand and track their own progress.

[0416] The sixth strategy subunit is used to enable the visualization interface to simulate and display the expected effect after implementing the recommended strategy;

[0417] For example, the expected change trend chart of various capability indicators can help users predict the results after adopting suggestions and enhance their confidence and motivation to implement personalized recommendation plans.

[0418] The strategy unit of this embodiment uses user data and engine analysis results to generate personalized improvement suggestions and optimization strategies, which can effectively train users and improve their eloquence and expression skills.

[0419] The verification unit is used to verify the user's use effect, including:

[0420] The platform effect can be verified through comparative experiments, such as whether the eloquence improvement of the user group under the guidance of the application platform after a certain period of time is significantly different from that of the control group;

[0421] Through regular user satisfaction surveys and feedback collection, we can understand the actual effectiveness of platform functions and whether user needs are met;

[0422] Combined with specific application scenarios, such as public speaking competitions, business negotiations, etc., track and record users' actual performance improvements as direct evidence of the effectiveness of the platform's functions.

[0423] Generally speaking, this embodiment has the following beneficial effects:

[0424] The first objective function of the embodiment of the present invention can directly reflect the user's expected value. Therefore, by adjusting the weight set of the speech sample according to the first objective function, the obtained first weight set can evolve in the adjustment direction desired by the user, which is more in line with the actual needs of the user. The immediate reward is a reward obtained immediately after taking a certain action. It is directly related to the short-term goal of the plan, which helps to make better decisions in subsequent steps. Since the policy gradient is established by measuring the probability of weight distribution, the second weight set is adjusted on the basis of knowing the performance status of each indicator of the user, so that the obtained second weight set is more focused on solving the actual eloquence expression weaknesses of the user. Finally, the third weight set is adjusted and guided by the optimal value function, which can maximize the reasonable arrangement of the weight distribution in the final weight set based on the first weight set and the second weight set, so that the generated eloquence training method can effectively help users to train.

[0425] Furthermore, through real-time data updates and automatic adjustment of weight strategies, this embodiment can flexibly change the decision basis according to environmental changes and new information inputs, ensuring continuous provision of optimal decision support in a dynamic environment, and can continuously learn lessons from historical decision results, iteratively optimize algorithms and related parameters, thereby achieving the evolution of training programs and performance improvements.

[0426] Embodiment three:

[0427] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the eloquence training method based on automatic weight adjustment;

[0428] Wherein, the method for training eloquence based on automatic weight adjustment, if implemented in the form of a software functional unit and used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0429] The above are preferred embodiments of the present invention. It should be noted that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for training eloquence based on automatic weight adjustment, characterized in that: include: Get user's speech samples; Adjusting the original weight set of the speech sample according to the first objective function of the user to obtain a first weight set; The original weight set of the speech sample is adjusted according to a preset instant reward and a strategy gradient to obtain a second weight set; wherein the strategy gradient is established by measuring the probability of the weight distribution; generating a third weight set according to the first weight set and the second weight set; The third weight set is adjusted and guided by using a preset optimal value function to obtain a final weight set, and the user is trained according to the eloquence training method generated by the final weight set.

2. The method for training eloquence based on automatic weight adjustment as claimed in claim 1, characterized in that: The original weight set of the speech sample is adjusted according to the first objective function of the user to obtain a first weight set, which is specifically: The speech sample is scored according to the initial first eloquence dimension indicator set to obtain a quantitative score; wherein the initial first eloquence dimension indicator set includes a speech recognition indicator, a language logic indicator, and an emotional expression indicator; Establishing the first objective function by measuring the difference between the quantitative score and the user target score; The original weight set of the speech sample is updated according to the first objective function to obtain the first weight set; wherein the first weight set is established based on the speech sample and the first eloquence dimension indicator set.

3. The method for training eloquence based on automatic weight adjustment as claimed in claim 1, characterized in that: The original weight set of the speech sample is adjusted according to the preset instant reward and strategy gradient to obtain a second weight set, which is specifically: Calculating the speech sample according to the initial second eloquence dimension indicator set to obtain a second eloquence dimension indicator set; Obtaining an initial instant reward corresponding to an initial second weight set in the second eloquence dimension indicator set, and establishing a second objective function according to a discount factor and the initial instant reward; Optimizing the initial second weight set according to the second objective function to obtain a final instant reward; The initial second weight set is adjusted according to the final immediate reward and the policy gradient to obtain a second weight set.

4. The method for training eloquence based on automatic weight adjustment as claimed in claim 3, characterized in that: The policy gradient is established by measuring the probability of weight distribution, specifically: Obtaining weight strategy parameters of the initial second weight set; According to the weight strategy parameters, the probability of weight distribution is obtained by calculating the random strategy; The policy gradient is obtained by calculating the probability of the weight distribution based on the criterion of reducing the variance.

5. The method for training eloquence based on automatic weight adjustment according to claim 1, characterized in that: The optimal value function is specifically: Calculating the speech sample according to the initial third eloquence dimension indicator set to obtain the third eloquence dimension indicator set; Using the preset first formula, a comprehensive reward function is established according to the time series of the third eloquence dimension indicator set and the third weight set; The parameters of the value function are updated according to the comprehensive reward function to obtain the optimal value function; wherein the value function is established according to the current language behavior feature vector state of the speech sample and the third weight set.

6. The method for training eloquence based on automatic weight adjustment according to claim 1, characterized in that: The third weight set is adjusted and guided using a preset optimal value function to obtain a final weight set, specifically: generating a policy network according to the third weight set; The optimal value function is used to guide the policy network to update parameters, so that the policy network evolves in a direction of increasing future cumulative rewards, thereby obtaining an optimized final policy network; The weight set corresponding to the final strategy network is used as the final weight set.

7. The method for training eloquence based on automatic weight adjustment according to claim 1, characterized in that: Generating a third weight set according to the first weight set and the second weight set is specifically: generating a transition weight set according to the first weight set and the second weight set; Using a deep neural network model, the third weight set is generated on a time scale based on the transition weight set and the language expression feature vector in the speech sample.

8. An eloquence training device based on automatic weight adjustment, characterized in that: Includes sample module, adjustment module, reward module, weight module and training module; Wherein, the sample module is used to obtain the user's speech sample; The adjustment module is used to adjust the original weight set of the speech sample according to the first objective function of the user to obtain a first weight set; The reward module is used to adjust the original weight set of the speech sample according to the preset instant reward and strategy gradient to obtain a second weight set; wherein the strategy gradient is established by measuring the probability of weight distribution; The weight module is used to generate a third weight set according to the first weight set and the second weight set; The training module is used to use a preset optimal value function to adjust and guide the third weight set to obtain a final weight set, and train the user according to the eloquence training method generated by the final weight set.

9. The eloquence training device based on automatic weight adjustment as claimed in claim 8, characterized in that: The adjustment module includes a quantization unit, a difference unit and an update unit; The quantification unit is used to evaluate the speech sample according to the initial first eloquence dimension indicator set to obtain a quantitative score; wherein the initial first eloquence dimension indicator set includes a speech recognition indicator, a language logic indicator and an emotional expression indicator; The difference unit is used to establish the first objective function by measuring the difference between the quantized score and the user target score; The updating unit is used to update the original weight set of the speech sample according to the first objective function to obtain the first weight set; wherein the first weight set is established based on the speech sample and the first eloquence dimension indicator set.

10. The eloquence training device based on automatic weight adjustment according to claim 8, characterized in that: The reward module includes a dimension unit, a target unit, an optimization unit and an adjustment unit; The dimension unit is used to calculate the speech sample according to the initial second eloquence dimension indicator set to obtain the second eloquence dimension indicator set; The target unit is used to obtain an initial instant reward corresponding to an initial second weight set in the second eloquence dimension indicator set, and establish a second target function according to a discount factor and the initial instant reward; The optimization unit is used to optimize the initial second weight set according to the second objective function to obtain a final instant reward; The adjustment unit is used to adjust the initial second weight set according to the final immediate reward and the policy gradient to obtain a second weight set.

11. The eloquence training device based on automatic weight adjustment according to claim 10, characterized in that: The reward module includes a strategy unit, a probability unit and a gradient unit; Wherein, the strategy unit is used to obtain the weight strategy parameters of the initial second weight set; The probability unit is used to obtain the probability of weight distribution by calculating the random strategy according to the weight strategy parameters; The gradient unit is used to obtain the policy gradient based on the probability calculation of the weight distribution based on the standard of reducing the variance.

12. The eloquence training device based on automatic weight adjustment according to claim 8, characterized in that: The optimal value function is specifically: Calculating the speech sample according to the initial third eloquence dimension indicator set to obtain the third eloquence dimension indicator set; Using the preset first formula, a comprehensive reward function is established according to the time series of the third eloquence dimension indicator set and the third weight set; The parameters of the value function are updated according to the comprehensive reward function to obtain the optimal value function; wherein the value function is established according to the current language behavior feature vector state of the speech sample and the third weight set.

13. The eloquence training device based on automatic weight adjustment according to claim 8, wherein the training module comprises a generation unit, an evolution unit and a weight unit; in, The generating unit is used to generate a policy network according to the third weight set; The evolution unit is used to guide the policy network to update parameters through the optimal value function, so that the policy network evolves in a direction of improving future cumulative rewards, thereby obtaining an optimized final policy network; The weight unit is used to use the weight set corresponding to the final strategy network as the final weight set.

14. The eloquence training device based on automatic weight adjustment according to claim 8, characterized in that: The weight module includes a transition unit and a network unit; The transition unit is used to generate a transition weight set according to the first weight set and the second weight set; The network unit is used to use a deep neural network model to generate the third weight set based on the transition weight set and the language expression feature vector in the speech sample on a time scale.

15. A storage medium, characterized in that: The storage medium stores a computer program, which is called and executed by a computer to implement any one of the eloquence training methods based on automatic weight adjustment as claimed in claims 1 to 7.

Images