Key task completion evaluation method and device based on generative AI big model

Abstract

The present invention provides a method and device for evaluating the completion of key tasks based on a generative AI large model, which relates to the field of data processing technology. The work statistical data is feature extracted by a pre-built feature extraction model, and the data is analyzed by a generative AI model built based on the Transformer algorithm. Among them, the connection weights and structure of the feature extraction model are adjusted based on the size of the model output, which can adapt to changes in data. The attention weights and inter-layer connection weights of the Transformer algorithm are determined based on an adaptive error compensation strategy, which can improve the performance and stability of the model to effectively evaluate the completion of employees' work.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and device for evaluating the degree of completion of key tasks based on a generative AI large model. Background Art

[0002] In modern enterprise management, effectively evaluating and monitoring employees' work completion is a key factor in improving organizational efficiency and execution. Traditional work evaluation methods usually rely on manual review of employees' work reports, project management system records and performance evaluation results. These methods are not only time-consuming and labor-intensive, but may also be affected by subjective factors, leading to inconsistency and deviation in evaluation results.

[0003] The data used to evaluate and monitor employee work completion usually comes from the enterprise's internal project management system, employee daily report system, and data derived from the performance evaluation system. These data are multi-source, diverse, time-series, and subjective. In terms of extracting enterprise project features, the model weight adjustment of traditional processing methods is usually relatively fixed, which is not flexible enough to cope with changes in feature importance in training data, and may lead to unstable performance of the model on complex data.

[0004] In addition, for tasks such as work completion evaluation, it is often necessary to understand the content of the entire paragraph or document, and the key information may be distributed in different locations. Traditional models are prone to gradient vanishing or gradient exploding problems when processing long sequences, which makes it difficult for them to capture long-distance dependencies in the sequence. If the model cannot effectively capture these long-distance dependencies, the generated evaluation will be inaccurate or incomplete, affecting the performance of the model in automatic text evaluation tasks. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a method and device for evaluating the completion of key tasks based on a generative AI big model to solve the above-mentioned technical problems.

[0006] In a first aspect, an embodiment of the present invention provides a method for evaluating the completion of key tasks based on a generative AI large model, wherein the method includes: obtaining pre-collected work statistical data, and converting the work statistical data into vector data; performing feature extraction on the vector data through a pre-built feature extraction model, and determining target feature parameters; wherein the connection weights and structure of the feature extraction model are adjusted based on the size of the model output; the target feature parameters are input into a pre-built generative AI model, and the target feature parameters are analyzed to obtain analysis results; wherein the generative AI model is constructed based on the Transformer algorithm, and the attention weights and inter-layer connection weights of the Transformer algorithm are determined based on an adaptive error compensation strategy; based on the analysis results, the work statistical data are evaluated for work completion.

[0007] In combination with the first aspect, an embodiment of the present invention also provides a first implementation method of the first aspect, wherein the target feature parameters are input into a pre-built generative AI model, the target feature parameters are analyzed, and the steps of obtaining analysis results include: position encoding the target feature parameters; using a self-attention mechanism to process data through each layer of the Transformer block of the generative AI model, and outputting an intermediate representation of the sequence; based on the intermediate representation, determining the analysis result indicated by the target feature parameters.

[0008] In combination with the first aspect, an embodiment of the present invention also provides a second implementation of the first aspect, wherein the step of determining the attention weight and inter-layer connection weight of the Transformer algorithm based on the adaptive error compensation strategy includes: inputting a preset training sample set into a preset Transformer model for forward propagation; according to the context information of the training sample set, using a mutual information calculation function to calculate the relevance score of each feature of the training sample set; based on the relevance score, activating the features of the training sample set through a gating mechanism; based on the activated features, determining the attention weight and inter-layer connection weight of the Transformer model.

[0009] In combination with the first aspect, an embodiment of the present invention also provides a third implementation of the first aspect, wherein the step of adjusting the connection weights and structure of the feature extraction model based on the size of the model output includes: adaptively adjusting the preset noise based on changes in the loss function of the feature extraction model in the training batch; dynamically adjusting the preset dynamic adjustment factor based on past weight updates of the feature extraction model; and adjusting the node weights of the feature extraction model by simulating the process of raindrop evaporation based on the adjusted preset noise and the adjusted preset dynamic adjustment factor.

[0010] In combination with the first aspect, an embodiment of the present invention also provides a fourth implementation of the first aspect, wherein the method further includes: screening the extracted features based on a preset weight shrinkage threshold; wherein the weight shrinkage threshold is adaptively determined based on the loss reduction rate of the feature extraction model.

[0011] In combination with the first aspect, an embodiment of the present invention also provides a fifth implementation of the first aspect, wherein the method also includes: performing dimensionality reduction processing on the target feature parameters through a pre-constructed autoencoder; wherein the parameters of the autoencoder are determined based on the feature importance vector of the target feature parameters, and the feature importance vector is determined based on a dynamic importance evaluation of the reconstructed data of the autoencoder.

[0012] In combination with the first aspect, an embodiment of the present invention also provides a sixth implementation of the first aspect, wherein a method for constructing an autoencoder comprises: encoding a preset training sample set through a preset autoencoder to obtain a potential representation corresponding to the training sample set; based on the reconstruction error of the potential representation, calculating the loss of the autoencoder and the feature importance vector corresponding to the training sample set; based on the gradient of the loss and the feature importance vector, updating the parameters of the autoencoder through network back propagation.

[0013] In combination with the first aspect, an embodiment of the present invention also provides a seventh implementation method of the first aspect, wherein the above method also includes: obtaining pre-collected work statistical samples, labeling the work statistical samples according to the work completion status represented by the work statistical samples, and constructing an initial sample set; calculating the mutual information of the initial sample set corresponding to the preset latent space vector, and based on the mutual information, performing data expansion on the initial sample set to generate expanded samples; merging the expanded samples and the initial sample set to construct a training sample set.

[0014] In combination with the first aspect, an embodiment of the present invention also provides an eighth implementation of the first aspect, wherein the step of calculating the mutual information of the initial sample set corresponding to a preset latent space vector includes: using a preset parameterized model to determine the conditional probability distribution corresponding to the latent space vector; wherein the parameterized model is a neural network function controlled by parameters of a preset data expansion model; based on the conditional probability distribution and the Gaussian distribution of the latent space vector, calculating the mutual information of the initial sample set corresponding to the latent space vector.

[0015] In the second aspect, an embodiment of the present invention provides a key task completion evaluation device based on a generative AI large model, wherein the device includes: a data acquisition module, used to acquire pre-collected work statistical data and convert the work statistical data into vector data; a feature extraction module, used to extract features from the vector data through a pre-built feature extraction model, and determine the target feature parameters; wherein the connection weights and structure of the feature extraction model are adjusted based on the size of the model output; an execution module, used to input the target feature parameters into a pre-built generative AI model, analyze the target feature parameters, and obtain analysis results; wherein the generative AI model is constructed based on the Transformer algorithm, and the attention weights and inter-layer connection weights of the Transformer algorithm are determined based on an adaptive error compensation strategy; an output module, used to evaluate the work completion of the work statistical data based on the analysis results.

[0016] The embodiments of the present invention bring the following beneficial effects: The present invention provides a method and device for evaluating the completion of key tasks based on a generative AI large model, which extracts features from work statistical data through a pre-built feature extraction model, and performs data analysis by a generative AI model built based on the Transformer algorithm. Among them, the connection weights and structure of the feature extraction model are adjusted based on the size of the model output, and can adapt to changes in data. The attention weights and inter-layer connection weights of the Transformer algorithm are determined based on an adaptive error compensation strategy, which can improve model performance and stability to effectively evaluate employee work completion. Moreover, the present invention can not only provide static historical data analysis, but also dynamically handle new business situations, providing enterprises with real-time, data-based decision support, thereby helping enterprises achieve more efficient human resource management and project monitoring.

[0017] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 A flowchart of a key task completion evaluation method based on a generative AI big model provided in an embodiment of the present invention;

[0021] Figure 2 A flowchart of another key task completion evaluation method based on a generative AI big model provided in an embodiment of the present invention;

[0022] Figure 3 A flowchart of a method for constructing a training sample set provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of the structure of a key task completion evaluation device based on a generative AI large model provided in an embodiment of the present invention;

[0024] Figure 5 A schematic diagram of the structure of another key task completion evaluation device based on a generative AI large model provided in an embodiment of the present invention;

[0025] Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the following embodiments of the present disclosure are described by specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0027] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described in the present invention may be embodied in a wide variety of forms, and any specific structure and / or function described in the present invention is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described in the present invention may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described in the present invention may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described in the present invention may be used to implement this device and / or practice this method.

[0028] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The drawings only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. The type, quantity and proportion of each component in actual implementation may be changed at will, and the component layout type may also be more complicated. In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the aspects described can be practiced without these specific details.

[0029] The embodiments of the present invention provide a method and device for evaluating the completion of key tasks based on a generative AI big model, which can effectively evaluate the work of employees.

[0030] For ease of understanding, a key task completion evaluation method based on a generative AI big model provided in an embodiment of the present invention is first described in detail. Figure 1 A flowchart of a key task completion evaluation method based on a generative AI big model provided by an embodiment of the present invention is shown, such as Figure 1 As shown, the method comprises the following steps:

[0031] Step S102, obtaining pre-collected work statistics data, and converting the work statistics data into vector data.

[0032] The data collected in the present invention is derived from the exported data of the enterprise's internal project management system, employee daily report system and performance evaluation system. The collected data is stored in JSON format. In one embodiment, the attributes of the data include: a1 R is the work content; a2 R is the employee ID; a3 For project introduction; R a4 R is the job category; a5 is the start time; R a6 is the end time; R a7 is the working hours; Ra8 R is the job difficulty assessment; a9 For actual effect; R a10 It should be noted that this embodiment is only used to illustrate a data format and type of the present invention. In practical applications, the attributes of data are usually more than 10 attributes, and the number of attributes of data may reach dozens or even hundreds.

[0033] Furthermore, the collected data is converted into a vector format to adapt to the input format of the machine learning model.

[0034] In one embodiment, the present invention uses a pre-trained word embedding model to realize the conversion between words and vectors. Assume that the natural language text is T wb ={t wb1 , t wb2 ,…, t wbi ,…, t wbnw}, where t wbi is the i-th word in the text, nw is the length of the text word vector, and further, through the word embedding model E wb (), each word t wbi is mapped to a vector space of fixed dimension, namely:

[0035]

[0036] In the formula, E wb () is the word embedding model, t wbi is the word in the i-th text, v wbi is the i-th converted word vector.

[0037] Based on this, the text T wb Convert to vector sequence {v wb1 ,v wb2 ,…,v wbi ,…,v wbnw}.

[0038] Step S104, extracting features from the vector data using a pre-built feature extraction model to determine target feature parameters.

[0039] Step S106, input the target feature parameters into a pre-built generative AI model, analyze the target feature parameters, and obtain analysis results.

[0040] Use pre-trained feature extraction models, such as autoencoders, PCA (Principal Component Analysis), LDA (Linear Discriminant Analysis), etc.) to extract features from vector data. According to business needs and model output, it can be determined which feature parameters are the most important, which may include: completion of key tasks, average task time, error rate, resource utilization. Further, use pre-trained generative AI models (such as variational autoencoders VAE, generative adversarial networks GAN, Transformer, etc.) to analyze the target feature parameters. The generative AI model can generate various predictions or evaluation results about the completion of the work, such as: predicting the progress of work in the future, identifying potential risk points, and providing improvement suggestions. In an embodiment of the present invention, the analysis results include an evaluation of the completion of the work, such as, for a piece of data collected, it is marked as: "This work analyzed the sales data of the past week, produced a detailed report, covered all key indicators, had a high degree of completion, accurate data, and was evaluated as an excellent work."

[0041] In an embodiment of the present invention, the connection weights and structure of the feature extraction model are adjusted based on the size of the model output, so that the model can adapt to the changes of data more flexibly during the training process, and improve its generalization ability and robustness. In one embodiment, the learning rate can be adjusted according to the output size, so that the learning rate of the node with larger output is smaller, and the learning rate of the node with smaller output is larger. Noise can also be added when the weight is updated to increase the robustness of the model. The intensity of the noise can be adjusted according to the output size, and the node with larger output can add smaller noise, and vice versa. It is also possible to remove nodes with long-term small or unimportant outputs by setting a threshold. If the output of a node is lower than the threshold for many consecutive times, it is removed from the network. Alternatively, remove the entire layer, especially when all the node outputs of a certain layer are low. This can be determined by evaluating the importance of each layer. New nodes can also be added in certain important layers to enhance their representation capabilities. New layers can also be added to the existing network structure, especially when it is found that the current network structure is not enough to capture complex features.

[0042] In addition, the generative AI model of the embodiment of the present invention is built based on the Transformer algorithm. The Transformer architecture can effectively capture long-distance dependencies in the input sequence through the self-attention mechanism. This is very important for understanding the specific requirements, background information and contextual connections of the work task, so that the model can more accurately evaluate the completion of the task. The attention weight and inter-layer connection weight of the Transformer algorithm are determined based on the adaptive error compensation strategy. The adaptive error compensation strategy is to dynamically adjust the parameters according to the performance of the model during the training process to improve the performance and stability of the model. Including but not limited to: 1. Learning rate adjustment: adjust the learning rate according to the current error size. 2. Weight decay: adjust the weight decay coefficient according to the error size. 3. Noise injection: adjust the noise intensity added to the weight update according to the error size. 4. Structural adjustment: dynamically adjust the structure of the model according to the error size, such as increasing or decreasing the number of layers, the number of nodes, etc. The adaptive error compensation strategy proposed in the present invention dynamically adjusts the learning rate by considering the error size between the output and the target after each iteration. If the error is large, it means that the prediction accuracy is low, and the model will increase the learning rate to accelerate the learning process; if the error is small, the learning rate is reduced, and the model is fine-tuned to maintain the existing accuracy. By adopting an immediate feedback mechanism based on the current error, it allows the model to respond quickly to errors and compensate itself.

[0043] Step S108: Based on the analysis results, the work completion degree is evaluated on the work statistical data.

[0044] The present invention can not only provide static historical data analysis, but also dynamically process new business situations, providing enterprises with real-time, data-based decision support, thereby helping enterprises achieve more efficient human resource management and project monitoring.

[0045] Furthermore, based on the above embodiment, the embodiment of the present invention also provides another key task completion evaluation method based on the generative AI big model. Figure 2 A flowchart of another key task completion evaluation method based on a generative AI big model provided by an embodiment of the present invention is shown, such as Figure 2 As shown, the method comprises the following steps:

[0046] Step S202, obtaining pre-collected work statistics data, and converting the work statistics data into vector data.

[0047] Step S204: extract features from the vector data using a pre-built feature extraction model to determine target feature parameters.

[0048] The present invention adopts a 3-layer fully connected neural network for feature extraction. In the prior art, some solutions use neural networks for feature extraction. In some neural network structures, problems such as gradient vanishing, gradient exploding, or falling into a local optimal solution may be encountered, affecting the stability of training and the performance of the model.

[0049] The connection weights and structure of the feature extraction model of the embodiment of the present invention are adjusted based on the size of the model output. In a specific implementation, feature extraction is performed by using a neural network based on dynamic correction, and in combination with simulating the evaporation process of raindrops through different paths in nature, the connection weights and structure of the neural network are dynamically adjusted to optimize feature extraction. The weights of unimportant features will gradually decrease until they are completely not considered by the model, thereby achieving adaptive refinement of features.

[0050] First, the parameters of the neural network are initialized. In one embodiment, the initialization method is random initialization, and the initialization interval is [0, 1].

[0051] Furthermore, the input data is fed into the neural network through forward propagation, and the nodes of each layer calculate the output according to the current weights and biases. Specifically, the input layer receives vectorized data , and propagates forward through the multi-layer network to the output layer, expressed as:

[0052]

[0053]

[0054] In the formula, is the data input to the neural network, and are the weights and biases of the lth layer of the neural network, is the Sigmoid activation function, L is the number of network layers, is the hidden layer output of the lth layer of the neural network, is the hidden layer output of the l-1th layer of the neural network, is the weight of the L+1th layer of the neural network, is the bias of the L+1th layer of the neural network.

[0055] Furthermore, the connection weights and structure of the feature extraction model are adjusted by the following steps:

[0056] 1) Adaptively adjust the preset noise based on the change of the loss function of the feature extraction model in the training batch.

[0057] In nature, the evaporation of raindrops is not only affected by climatic conditions, but also by environmental noise (such as wind changes). In neural networks, the effect of this environmental noise can be simulated by dynamically adjusting the intensity of the noise according to the change of the loss function during the training process. The added noise can help the network explore a wider parameter space, prevent the model from converging to the local optimal solution too early, and improve the generalization ability of the model. The present invention adopts an adaptive noise injection mechanism to dynamically adjust the intensity of the noise according to the change of the loss function in each training batch to increase the robustness of the neural network weight parameters. The calculation method is expressed as:

[0058]

[0059] In the formula, is the noise intensity, is the maximum value of the noise intensity, is a parameter that adjusts the steepness of the curve. is the current loss value, is the target loss value. Preferably, Set to 0.95, Set to 0.1, Set to 10.

[0060] 2) Based on the past weight updates of the feature extraction model, the preset dynamic adjustment factor is dynamically adjusted.

[0061] During the evaporation of raindrops, the evaporation rate may be adjusted based on the previous evaporation history (such as the cumulative effect of the evaporation rate). Similarly, the step size of the weight update in the neural network can be dynamically adjusted by considering the previous weight update history. In one embodiment, the dynamic adjustment factor Adjust the step size of weight update to ensure the adaptability and stability of the training process. It is dynamically adjusted based on the exponentially weighted average of past weight updates, expressed as:

[0062]

[0063] In the formula, is the adjustment factor of the previous iteration; and are the last and current weight updates respectively; is the smoothing coefficient; represents the L2 norm. Preferably, Set to 0.9.

[0064] 3) Based on the adjusted preset noise and the adjusted preset dynamic adjustment factor, the node weights of the feature extraction model are adjusted to simulate the process of raindrop evaporation.

[0065] In each layer of the neural network, the process of simulating the evaporation of raindrops dynamically adjusts the weights of the nodes in that layer according to the size of the output. In nature, the evaporation rate of raindrops depends on many factors, such as temperature, humidity, wind speed, etc. In the neural network, the output of each node can be regarded as its "environmental condition". Adjusting the weight of each node according to the output size is equivalent to adjusting the evaporation rate according to the different conditions of the environment in which the raindrops are located. The weights of nodes with larger outputs (i.e., more important features in the current training data and tasks) are reduced less, similar to the fact that raindrops are more intact under lower evaporation conditions. Conversely, the weights of nodes with smaller outputs are reduced more, similar to the fact that raindrops disappear quickly under high evaporation conditions. That is, the larger the output of the neural network node, the less its corresponding weight is reduced in the next iteration, and vice versa. The calculation method of the dynamically adjusted weight decay strategy is expressed as:

[0066]

[0067]

[0068] In the formula, is the learning rate, is the parameter update operation, is the loss function of the neural network, is the dynamic adjustment factor, is the adaptive noise, is the noise intensity, is the identity matrix, is the L1 regularization parameter, is the L1 norm. Preferably, Set to 0.01.

[0069] Furthermore, the embodiment of the present invention also screens the extracted features based on a preset weight shrinkage threshold; wherein the weight shrinkage threshold is adaptively determined based on the loss reduction rate of the feature extraction model. Specifically, by evaluating the contribution of all features and deciding whether to continue to retain these features based on the set threshold, the weights of features with low contribution gradually shrink until they disappear, and the feature screening process is expressed as:

[0070]

[0071]

[0072] In the formula, is the threshold for weight shrinkage, which is a feature screening function.

[0073] In one embodiment, the adaptive shrink threshold Based on the dynamic adjustment of the loss reduction rate during training, the calculation method is expressed as:

[0074]

[0075] In the formula, is the initial threshold; is to adjust the sensitivity parameter; represents the reduction of the loss function between two consecutive iterations. Preferably, Set to 1, Set to 2.

[0076] Furthermore, the embodiment of the present invention also calculates the error between the output and the target, and adjusts the weights and biases in the network through a back propagation algorithm to minimize the error between the output and the target. The calculation method is expressed as:

[0077]

[0078] In the formula, is the loss function of the neural network, is the mean square error loss function, is the L1 norm, is the L2 norm; is the target output; is the output of the neural network calculated by the preset Softmax function. and are the weights of L1 and L2 regularization respectively. Preferably, and Set to 0.3 and 0.7 respectively.

[0079] Repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0080] In summary, the embodiment of the present invention uses a neural network based on dynamic correction for feature extraction. Traditional neural networks may encounter inflexible weight adjustment during feature extraction, thereby affecting the adaptability of the model to feature importance. By simulating the natural evaporation process and dynamically adjusting the weights, the network is allowed to automatically abandon unimportant features during training, thereby improving the accuracy of feature extraction and the training efficiency of the network.

[0081] Furthermore, the data being evaluated also has high-dimensional characteristics. The existing technology uses autoencoder networks to process high-dimensional complex data. However, traditional autoencoders usually have fixed network structures and weights, which means that they cannot be dynamically adjusted during the training process to adapt to the changes in the importance of different features, reducing the accuracy and reliability of the model in practical applications.

[0082] Correspondingly, the present invention performs dimensionality reduction processing on the target feature parameters through a pre-constructed autoencoder; wherein, the parameters of the autoencoder are determined based on the feature importance vector of the target feature parameters, and the feature importance vector is determined based on a dynamic importance evaluation of the reconstructed data of the autoencoder. In a specific implementation, an autoencoder neural network based on dynamic feature evaluation is used to perform feature dimensionality reduction, and the autoencoder neural network includes two parts: an encoder and a decoder. The encoder is responsible for mapping high-dimensional input data to a low-dimensional hidden layer, and the decoder reconstructs the original data from this low-dimensional hidden state. The present invention enhances the model's ability to process complex data and improves the accuracy of feature compression and data reconstruction by adopting a dynamic feature importance evaluation strategy to evaluate and adjust the importance of each input feature. Specifically, the method for constructing an autoencoder includes:

[0083] 1) The preset training sample set is encoded by the preset autoencoder to obtain the potential representation corresponding to the training sample set.

[0084] First, initialize the parameters of the autoencoder neural network, including the weight parameters and bias parameters of the autoencoder neural network. In one embodiment, the weight of the autoencoder is , the bias is , the initialization method is random initialization, and the initialization interval is [0,1].

[0085] In the feature dimensionality reduction stage, the data is processed by the autoencoder algorithm. During the encoding process, the input feature is extracted and the data is is converted into a compressed latent representation , and then reconstructed back through the decoding process , expressed as:

[0086]

[0087] In the formula, is the weight matrix of the encoder; is the bias vector of the encoder; is the ReLU activation function; The latent representation compressed by the encoder.

[0088] And, the decoding process of the decoder is expressed as:

[0089]

[0090] In the formula, is the weight matrix of the decoder; is the bias vector of the decoder; Reconstructed data for the decoder.

[0091] 2) Based on the reconstruction error of the latent representation, calculate the loss of the autoencoder and the feature importance vector corresponding to the training sample set.

[0092] The difference between the reconstructed data and the original data is calculated. The loss function used in the present invention optimizes the performance of the model and prevents overfitting by combining the regularization term reconstruction error and the regularization term. The calculation method is expressed as:

[0093]

[0094] In the formula, is the loss function of the autoencoder, is the L1 norm, is the L2 norm; is the L1 regularization coefficient; represents the L1 norm of the weight. Preferably, Set to 0.3.

[0095] A dynamic feature importance evaluation strategy is used to evaluate and adjust the importance of each input feature. The calculation method is expressed as:

[0096]

[0097] In the formula, is the feature importance vector, is the Softmax function, are the encoder weights, is the ith input sample, is the i-th reconstruction sample, The number of batches of data input for the current iteration, is a normalization factor. Preferably, Set to the i-th input sample The adjustment process of the normalization factor shows that the model is considering how to adjust the contribution of each feature in each training step to achieve the best prediction performance. This method of dynamically adjusting parameters based on each training step embodies a dynamic feature importance evaluation strategy that can capture and respond to the impact of different features on the model's prediction ability in real time.

[0098] 3) Based on the gradient of the loss and the feature importance vector, the parameters of the autoencoder are updated through network back propagation.

[0099] In the embodiment of the present invention, the gradient of the loss function is back-propagated through the network to update the weights and biases in the model to reduce the prediction error and ensure the minimization of the loss function. The parameter update method of the autoencoder is expressed as:

[0100]

[0101]

[0102]

[0103]

[0104] In the formula, is the learning rate; and is the partial derivative of the loss function with respect to the encoder weights and biases; and is the partial derivative of the loss function with respect to the decoder weights and biases; is the feature importance vector; Represents element-wise multiplication.

[0105] In one embodiment, the learning rate It is set by adaptive adjustment to dynamically adjust according to different stages of network training, expressed as:

[0106]

[0107] In the formula, is the learning rate for the t+1th iteration, is the learning rate of the tth iteration, is the sensitivity parameter of learning rate adjustment, is the loss function About parameters Preferably, Set to 0.5.

[0108] Further, the above steps are repeated until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0109] In summary, the embodiment of the present invention uses an autoencoder network based on dynamic feature evaluation to perform feature dimensionality reduction. When processing high-dimensional data, the traditional autoencoder network may not be sufficient to effectively capture the complex structure of the data. By adopting a dynamic feature importance evaluation strategy to evaluate and adjust the importance of each input feature, the model's ability to process complex data is enhanced, and the accuracy of feature compression and data reconstruction is improved.

[0110] Step S206: Position encoding of the target feature parameters.

[0111] Step S208, using the self-attention mechanism to process data through each layer of the Transformer block of the generative AI model, and outputting the intermediate representation of the sequence.

[0112] Step S210, determining the analysis result indicated by the target feature parameter based on the intermediate representation.

[0113] In combination with the above embodiments, the present invention adopts a Transformer algorithm based on correlation scoring as a generative AI model. The generative AI model has the ability to encode the text of the work completion evaluation paragraph. The Transformer algorithm based on correlation scoring includes multiple levels of Transformer blocks, each of which is composed of a self-attention mechanism and a position encoding. The self-attention mechanism enables the model to focus on the correlation of different parts of the sequence feature vector after the input feature dimension reduction, and the position encoding provides the position information of each element in the sequence, enhancing the model's ability to understand the sequence structure.

[0114] In order to improve the accuracy of prediction and the response speed of the model, the present invention adopts an adaptive error compensation strategy based on the Transformer model, and automatically adjusts the model's attention weight and inter-layer connection weight according to the error of the previous prediction, so that the model pays more attention to those parts that are difficult to predict during the training process. Correspondingly, the attention weight and inter-layer connection weight of the Transformer algorithm are determined by the following steps:

[0115] 1) Perform forward propagation by inputting the preset training sample set into the preset Transformer model.

[0116] First, the parameters of the Transformer model are initialized. In one embodiment, the initialization method is random initialization, and the initialization interval is [0, 1].

[0117] Furthermore, the input data is first positionally encoded and then input into a multi-layer Transformer block. Each layer of Transformer blocks processes the data through a self-attention mechanism and outputs an intermediate representation of the sequence, which is then passed to the next layer. The forward propagation process is expressed as:

[0118]

[0119]

[0120] In the formula, is the i-th attention head; is the multi-head self-attention function; For cascade operation; is the data input into the Transformer model, that is, the feature vector after dimensionality reduction; are the weight matrices corresponding to query, key, and value respectively; is the weight matrix of the output linear transformation; is the attention module function.

[0121] In one embodiment, the attention module function allows the model to focus on information in different representation subspaces in parallel, and the calculation method is expressed as:

[0122]

[0123] In the formula, is the dimension of the key vector.

[0124] 2) Based on the context information of the training sample set, the mutual information calculation function is used to calculate the relevance score of each feature of the training sample set.

[0125] The embodiment of the present invention adopts a context-sensitive dynamic feature selection mechanism, which enables the model to dynamically select the most relevant features according to the context when generating predictions, thereby improving the accuracy of predictions and the explanatory power of the model. The relevance score of each feature to the current prediction task is calculated. The score is based on the mutual information between the feature and the context, and the calculation method is expressed as:

[0126]

[0127] In the formula, is the relevance score of the ith feature; is the number of eigenvalues ​​of the input sample; is the i-th data input into the Transformer model; represents the mutual information calculation function; Represents the current context information.

[0128] 3) Based on the relevance score, the features of the training sample set are activated through a gating mechanism.

[0129] According to the calculated relevance score, a gating mechanism is used to decide whether to activate a feature. The calculation method is expressed as:

[0130]

[0131] In the formula, is the gated output of the i-th feature; and are learnable parameters used to adjust the sensitivity and activation threshold of the gate; is the Sigmoid activation function.

[0132] Further, By accumulating It is calculated by:

[0133]

[0134] 4) Based on the activated features, determine the attention weights and inter-layer connection weights of the Transformer model.

[0135] In each iteration, the attention weights and inter-layer connection weights are adjusted according to the error between the output and the target sequence. The adjustment method is expressed as:

[0136]

[0137]

[0138] In the formula, is the weight of the updated Transformer model; is the initial learning rate; is the gated output vector of all features; is the learning rate adjusted by the adaptive error compensation mechanism; is the mean square error function; is the gradient of the loss function with respect to the weights; is a parameter that controls the sensitivity of learning rate adjustment. Preferably, Set to 3.

[0139] Calculate the loss function of the Transformer model and update the model parameters through back propagation. The calculation method of the loss function is expressed as:

[0140]

[0141] In the formula, is the target sequence; is the output sequence of the model; is the coefficient of the control error compensation effect. Preferably, Set to 0.3.

[0142] Further, the above steps are repeated until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000 times.

[0143] In summary, the embodiment of the present invention uses the Transformer model based on relevance scoring to evaluate the completion degree. For the automatic evaluation of work completion, the model needs to be able to understand and generate complex text data. The Transformer model optimizes the processing ability of long sequence data through self-attention mechanism and position encoding, enabling the model to achieve higher accuracy and response speed in evaluation tasks.

[0144] Step S212: Based on the analysis results, the work completion degree is evaluated on the work statistical data.

[0145] Furthermore, based on the above embodiments, the embodiments of the present invention also designs a training sample set. Correspondingly, the embodiments of the present invention also provide another key task completion evaluation method based on the generative AI big model. This embodiment is to illustrate the method of constructing the training sample set. Figure 3 A flow chart of a method for constructing a training sample set provided by an embodiment of the present invention is shown. Figure 3 As shown, the method comprises the following steps:

[0146] Step S10, obtaining pre-collected work statistical samples, labeling the work statistical samples according to the work completion status represented by the work statistical samples, and constructing an initial sample set.

[0147] The work statistical sample includes various work-related data, such as task completion time, task type, employee performance, etc. In one embodiment, the data source of the embodiment of the present invention refers to the above-mentioned work statistical data. Furthermore, the collected data can be labeled by defining labeling rules for work completion. For example, the task completion can be divided into several levels (such as 0-20% for low completion, 20%-50% for medium completion, 50%-80% for high completion, and 80%-100% for very high completion). In the specific implementation, the labeling method of the present invention is manual labeling. Among them, the labeling content of the embodiment of the present invention is the completion evaluation of the work. For example, for a piece of collected data, it is labeled: "This work analyzed the sales data of the past week and made a detailed report covering all key indicators. It has a high degree of completion, accurate data, and is evaluated as excellent work."

[0148] Step S11, calculating the mutual information of the initial sample set corresponding to the preset latent space vector, and performing data expansion on the initial sample set based on the mutual information.

[0149] It is understandable that in the task of the present invention, the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor generalization of the model and affect the accuracy of the model. In terms of expanding enterprise project data, the existing generative adversarial network may fail to ensure the consistency of the statistical characteristics of the generated data with the real data, limiting the effectiveness and accuracy of the model in practical applications.

[0150] Correspondingly, the embodiment of the present invention uses mutual information to measure the correlation between samples and latent space vectors to increase the diversity of the data set.

[0151] In specific implementation, the embodiment of the present invention uses a preset parameterized model to determine the conditional probability distribution corresponding to the latent space vector. Based on the conditional probability distribution and the Gaussian distribution of the latent space vector, the mutual information of the initial sample set corresponding to the latent space vector is calculated.

[0152] In one embodiment, the mutual information term of the generator The calculation method is expressed as:

[0153]

[0154] In the formula, is a Gaussian weighted conditional probability distribution.

[0155] Furthermore, the Gaussian weighted conditional probability distribution The calculation method is expressed as:

[0156]

[0157] In the formula, is a Gaussian distribution.

[0158] Furthermore, we define the parameterized model , then the conditional probability distribution is based on the numerical model To map arrive Calculate and express it as:

[0159]

[0160] In the formula, It is a parameterized model, specifically the parameters of the model are expanded by the preset data (such as parameters ) controlled by the neural network function. Characterized by the parameters From hidden space Generate data ; is the variance of the generated data. Preferably, Set to 0.01.

[0161] Specifically, the embodiment of the present invention adopts a generative adversarial network for data expansion. The traditional generative adversarial network relies on adversarial learning between the generator and the discriminator to generate new data. The generator tends to generate only a few samples instead of covering the entire data distribution, resulting in insufficient diversity of the generated data and failure to fully represent the statistical characteristics of the real data. Moreover, the evaluation indicators of the existing generative adversarial network mainly focus on the quality and diversity of the generated data, and do not directly measure the consistency of the statistical characteristics of the generated data with the real data, resulting in some limitations in the consistency of the statistical characteristics of the generated data with the real data. In this regard, the embodiment of the present invention adopts a generative adversarial network algorithm based on mutual information for sample generation, and maximizes the mutual information to ensure that the generated data has a higher similarity with the real data in terms of statistical characteristics, thereby improving the diversity and quality of the generated data.

[0162] Combined with the above steps, the parameters of the data augmentation model indicated by the above parameterized model are the parameters of the generator Correspondingly, the training steps of the generative adversarial network of the embodiment of the present invention are as follows:

[0163] 1) Initialize the network parameters of the generator and the discriminator. The generator maps the transformation from the latent space to the data space, and the discriminator tries to distinguish the generated data from the real data. In one embodiment, let the generator be , the discriminator is , the parameters of the generator are , the parameters of the discriminator are , the initialization method is expressed as:

[0164]

[0165]

[0166] In the formula, is the initial weight of the generator; is the initial weight of the discriminator; To obey a specific distribution; represents a normal distribution with mean 0 and covariance the identity matrix.

[0167] 2) In the iterative training stage, the generator is updated by combining the mutual information maximization strategy with the adversarial loss. The goal is to maximize the mutual information while reducing the adversarial loss of the generative adversarial network. The training goal of the generator is , the calculation method is expressed as:

[0168]

[0169]

[0170]

[0171] In the formula, is the training target of the generator; is the minimization function; The coefficient for weighing mutual information and generating adversarial loss; is the mutual information function; represents the generated data sample; represents the latent space vector; Adversarial loss for generative adversarial networks; is the generator function; is the discriminator function; To obey a specific distribution; represents the conditional probability distribution, calculated above. Represents the marginal distribution of real data; The marginal distribution of vectors in the implicit space; Indicates the desired operation. Preferably, Set to 0.3.

[0172] 3) In the discriminator update phase, the generator parameters are fixed to update the discriminator. The goal of the discriminator is to accurately distinguish between real samples and generated samples. Suppose the training goal of the generator is , the calculation method is expressed as:

[0173]

[0174]

[0175] In the formula, is the discriminator loss; is the weight of the discriminator; Represents data sampled from a real dataset; is the probability distribution of the real data.

[0176] 4) Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0177] After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples. In one embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0178] Step S12: merging the expanded sample and the initial sample set to construct a training sample set.

[0179] In summary, the training sample set of the embodiment of the present invention can be used to train feature extraction models, generative AI models, and autoencoders. In one embodiment, the training sample set can be subjected to feature extraction by the trained feature extraction model and then the autoencoder can be trained. Furthermore, the generative AI model can be trained after the training sample set is subjected to dimensionality reduction processing by the autoencoder.

[0180] In summary, an embodiment of the present invention provides another method for evaluating the completion of key tasks based on a generative AI large model, which uses a generative adversarial network based on mutual information for data expansion. Traditional generative adversarial networks may not be able to ensure consistency with the statistical characteristics of real data when generating new samples. The generative adversarial network based on mutual information maximizes the mutual information to ensure that the generated data is highly similar to the real data in statistical characteristics, thereby solving the problems of insufficient samples and weak model generalization ability.

[0181] Furthermore, based on the above embodiments, the embodiments of the present invention also provide a key task completion evaluation device based on a generative AI big model. Figure 4 A structural diagram of a key task completion evaluation method based on a generative AI big model provided by an embodiment of the present invention is shown, Figure 4 As shown, the device includes: a data acquisition module 100, used to acquire pre-collected work statistical data and convert the work statistical data into vector data; a feature extraction module 200, used to extract features from the vector data through a pre-built feature extraction model and determine target feature parameters; wherein the connection weights and structure of the feature extraction model are adjusted based on the size of the model output; an execution module 300, used to input the target feature parameters into a pre-built generative AI model, analyze the target feature parameters, and obtain analysis results; wherein the generative AI model is built based on the Transformer algorithm, and the attention weights and inter-layer connection weights of the Transformer algorithm are determined based on an adaptive error compensation strategy; an output module 400, used to evaluate the work completion of the work statistical data based on the analysis results.

[0182] An embodiment of the present invention provides a key task completion evaluation device based on a generative AI big model, which has the same technical features as a key task completion evaluation method based on a generative AI big model provided in the above embodiment, so it can also solve the same technical problems and achieve the same technical effects.

[0183] Furthermore, based on the above embodiment, the embodiment of the present invention also provides another key task completion evaluation device based on the generative AI big model. Figure 5The schematic diagram of the structure of another key task completion evaluation device based on a generative AI large model provided by an embodiment of the present invention is shown. The execution module 300 is also used to positionally encode the target feature parameters; the self-attention mechanism is used to process data through each layer of the Transformer block of the generative AI model, and the intermediate representation of the output sequence is output; based on the intermediate representation, the analysis result indicated by the target feature parameter is determined.

[0184] The above-mentioned execution module 300 is also used to input the preset training sample set into the preset Transformer model for forward propagation; according to the context information of the training sample set, the mutual information calculation function is used to calculate the relevance score of each feature of the training sample set; based on the relevance score, the features of the training sample set are activated through a gating mechanism; based on the activated features, the attention weight and inter-layer connection weight of the Transformer model are determined.

[0185] The above-mentioned feature extraction module 200 is also used to adaptively adjust the preset noise based on the change of the loss function of the feature extraction model in the training batch; dynamically adjust the preset dynamic adjustment factor based on the past weight update of the feature extraction model; and adjust the node weights of the feature extraction model based on the adjusted preset noise and the adjusted preset dynamic adjustment factor to simulate the process of raindrop evaporation.

[0186] The feature extraction module 200 is also used to screen the extracted features based on a preset weight shrinkage threshold; wherein the weight shrinkage threshold is adaptively determined based on the loss reduction rate of the feature extraction model.

[0187] The above-mentioned execution module 300 is also used to perform dimensionality reduction processing on the target feature parameters through a pre-built autoencoder; wherein the parameters of the autoencoder are determined based on the feature importance vector of the target feature parameters, and the feature importance vector is determined based on a dynamic importance evaluation of the reconstructed data of the autoencoder.

[0188] The above-mentioned execution module 300 is also used to encode a preset training sample set through a preset autoencoder to obtain a potential representation corresponding to the training sample set; based on the reconstruction error of the potential representation, calculate the loss of the autoencoder and the feature importance vector corresponding to the training sample set; based on the gradient of the loss and the feature importance vector, update the parameters of the autoencoder through network back propagation.

[0189] The device also includes a construction module 500, which is used to obtain pre-collected work statistical samples, label the work statistical samples according to the work completion status represented by the work statistical samples, and construct an initial sample set; calculate the mutual information of the initial sample set corresponding to the preset latent space vector, and based on the mutual information, perform data expansion on the initial sample set to generate expanded samples; merge the expanded samples and the initial sample set to construct a training sample set.

[0190] The construction module 500 is also used to determine the conditional probability distribution corresponding to the latent space vector using a preset parameterized model; wherein the parameterized model is a neural network function controlled by the parameters of a preset data expansion model; based on the conditional probability distribution and the Gaussian distribution of the latent space vector, the mutual information of the initial sample set corresponding to the latent space vector is calculated.

[0191] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above Figures 1 to 3 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to perform the above Figures 1 to 3 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 6 FIG. 1 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 61 and a memory 60, the memory 60 stores computer executable instructions that can be executed by the processor 61, and the processor 61 executes the computer executable instructions to implement the above Figures 1 to 3 The method shown.

[0192] exist Figure 6In the illustrated embodiment, the electronic device further includes a bus 62 and a communication interface 63, wherein the processor 61, the communication interface 63 and the memory 60 are connected via the bus 62. The memory 60 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 63 (which may be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. may be used. The bus 62 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. It may also be an AMBA (Advanced Microcontroller Bus Architecture) bus, where AMBA defines three types of buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus, and an AXI (Advanced Xtensible Interface) bus. The bus 62 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6Only one bidirectional arrow is used in the diagram, but it does not mean that there is only one bus or one type of bus. The processor 61 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit in the processor 61 or the instructions in the form of software. The above-mentioned processor 61 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to be executed, or a combination of hardware and software modules in the decoding processor can be executed. The software module can be located in a random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, register or other mature storage media in the art. The storage medium is located in the memory, and the processor 61 reads the information in the memory and completes the above-mentioned Figures 1 to 3 Any of the methods shown.

[0193] A computer program product of a key task completion evaluation method and device based on a generative AI large model provided in an embodiment of the present invention includes a computer-readable storage medium storing a program code, and the instructions included in the program code can be used to execute the method described in the previous method embodiment. The specific implementation can refer to the method embodiment, which will not be repeated here. A person skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here. In addition, in the description of the embodiment of the present invention, unless otherwise clearly specified and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0194] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes. Finally, it should be noted that the above embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions recorded in the above embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A key task completion evaluation method based on a generative AI big model, characterized in that: The method comprises: Acquire pre-collected work statistics data, and convert the work statistics data into vector data; Extracting features from the vector data using a pre-built feature extraction model to determine target feature parameters; wherein the connection weights and structure of the feature extraction model are adjusted based on the size of the model output; Inputting the target feature parameters into a pre-built generative AI model, analyzing the target feature parameters, and obtaining analysis results; wherein the generative AI model is built based on a Transformer algorithm, and the attention weight and inter-layer connection weight of the Transformer algorithm are determined based on an adaptive error compensation strategy; Based on the analysis results, evaluating the work completion of the work statistical data; The step of determining the attention weight and inter-layer connection weight of the Transformer algorithm based on the adaptive error compensation strategy includes: Input the preset training sample set into the preset Transformer model for forward propagation; Calculating the relevance score of each feature of the training sample set using a mutual information calculation function according to the context information of the training sample set; Based on the relevance score, activating features of the training sample set through a gating mechanism; Based on the activated features, the attention weights and inter-layer connection weights of the Transformer model are determined.

2. The method according to claim 1, characterized in that The step of inputting the target feature parameters into a pre-built generative AI model, analyzing the target feature parameters, and obtaining analysis results comprises: Position encoding the target feature parameters; Each layer of the Transformer block of the generative AI model uses a self-attention mechanism to process data and output an intermediate representation of the sequence; Based on the intermediate representation, an analysis result indicated by the target characteristic parameter is determined.

3. The method according to claim 1, characterized in that: The steps of adjusting the connection weights and structure of the feature extraction model based on the size of the model output include: Adaptively adjusting the preset noise based on changes in the loss function of the feature extraction model in the training batch; Dynamically adjusting a preset dynamic adjustment factor based on past weight updates of the feature extraction model; Based on the adjusted preset noise and the adjusted preset dynamic adjustment factor, the node weights of the feature extraction model are adjusted in a process simulating raindrop evaporation.

4. The method according to claim 3, characterized in that The method further comprises: The extracted features are screened based on a preset weight shrinkage threshold; wherein the weight shrinkage threshold is adaptively determined based on the loss reduction rate of the feature extraction model.

5. The method according to claim 1, characterized in that The method further comprises: The target feature parameters are subjected to dimensionality reduction processing by a pre-constructed autoencoder; wherein the parameters of the autoencoder are determined based on a feature importance vector of the target feature parameters, and the feature importance vector is determined based on a dynamic importance evaluation of the reconstructed data of the autoencoder.

6. The method according to claim 5, characterized in that The method for constructing the autoencoder comprises: Encoding a preset training sample set by a preset autoencoder to obtain a potential representation corresponding to the training sample set; Calculating the loss of the autoencoder and the feature importance vector corresponding to the training sample set based on the reconstruction error of the latent representation; Based on the gradient of the loss and the feature importance vector, the parameters of the autoencoder are updated through network back propagation.

7. The method according to claim 1, characterized in that The method further comprises: Acquire pre-collected work statistical samples, label the work statistical samples according to the work completion status represented by the work statistical samples, and construct an initial sample set; Calculating the mutual information of the initial sample set corresponding to the preset latent space vector, and based on the mutual information, performing data expansion on the initial sample set to generate expanded samples; The expanded samples and the initial sample set are combined to construct a training sample set.

8. The method according to claim 7, characterized in that The step of calculating the mutual information of the initial sample set corresponding to the preset latent space vector comprises: A preset parameterized model is used to determine the conditional probability distribution corresponding to the latent space vector; wherein the parameterized model is a neural network function controlled by parameters of a preset data expansion model; Based on the conditional probability distribution and the Gaussian distribution of the latent space vector, the mutual information of the initial sample set corresponding to the latent space vector is calculated.

9. A key task completion evaluation device based on a generative AI large model, characterized in that: The device comprises: A data acquisition module, used to acquire pre-collected work statistical data and convert the work statistical data into vector data; A feature extraction module, used to extract features from the vector data using a pre-built feature extraction model to determine target feature parameters; wherein the connection weights and structure of the feature extraction model are adjusted based on the size of the model output; An execution module, used for inputting the target feature parameters into a pre-built generative AI model, analyzing the target feature parameters, and obtaining analysis results; wherein the generative AI model is built based on a Transformer algorithm, and the attention weight and inter-layer connection weight of the Transformer algorithm are determined based on an adaptive error compensation strategy; An output module, used for evaluating the work completion degree of the work statistical data based on the analysis result; Among them, the step of determining the attention weight and inter-layer connection weight of the Transformer algorithm based on the adaptive error compensation strategy includes: inputting a preset training sample set into a preset Transformer model for forward propagation; according to the context information of the training sample set, using a mutual information calculation function to calculate the relevance score of each feature of the training sample set; based on the relevance score, activating the features of the training sample set through a gating mechanism; based on the activated features, determining the attention weight and inter-layer connection weight of the Transformer model.

Images