Artificial intelligence-based employee automatic business process processing method and system

Abstract

The invention relates to the technical field of business process automation, in particular to an employee automatic business process processing method and system based on artificial intelligence, and the method comprises the steps: firstly obtaining a general teacher model and a target student model used for an automatic business process, and a feedback confirmation result of the operation of the target student model; and performing primary knowledge distillation optimization on the target student model according to the general teacher model. Wherein the feedback confirmation result is used for controlling the learning strength of the target learning model to the general teacher model during the optimization. Compared with the prior art, the target student model is dynamically adjusted by using the feedback confirmation result, so that the target student model is more and more accordant with the personalized operation habit of the user along with the increase of the operation frequency, and the feedback confirmation result is combined with the knowledge distillation technology to accelerate the training speed of the target student model; the application flexibility of the automatic business process technology is improved, and the problem that the business process automation technology in the prior art is lack of individuation is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of business process automation, and in particular to an employee automated business process processing method and system based on artificial intelligence. Background Art

[0002] Business Process Automation (BPA) is a strategy that automates complex and repetitive business processes through software to simplify daily operations and keep the business running smoothly. The core goal of business process automation is to improve operational efficiency, reduce human errors, standardize processes, and enable employees to focus on more strategic tasks. This automation can range from simple task automation to complex process integration, allowing enterprises to perform business operations more flexibly and efficiently.

[0003] The combination of business process automation and artificial intelligence (AI) is the current trend of technological development. This combination can greatly expand the boundaries of automation, making "digital employees" both intelligent and automated. For example, by integrating AI technology, business process automation systems can achieve data collection, recognition, and processing, such as automatically identifying data in invoices and contracts. This integration not only improves the efficiency of data processing, but also enhances the decision-making ability of the system, making the automation process more intelligent and accurate.

[0004] However, although the existing combination method performs well in data collection and other aspects, in scenarios involving decision-making and classification, the results are different due to the different weights and thinking focuses of each employee. When a general model is used for automation, it may lack personalization and cannot fully meet the specific needs of each employee. In existing technologies, personalization requires a large amount of personalized data, which will inevitably increase the cost of enterprises using process automation technology. Therefore, people need a more personalized AI-based employee automation business process processing method. Summary of the invention

[0005] Therefore, the present invention provides an employee automated business process processing method and system based on artificial intelligence to solve the problem of lack of personalization in the business process automation technology in the prior art.

[0006] The present invention provides an employee automated business process processing method based on artificial intelligence, comprising:

[0007] Obtain a generic teacher model and target student model for automating business processes;

[0008] Run the target student model and obtain feedback confirmation results of the target student model operation;

[0009] Based on the feedback confirmation results, the target student model is optimized through knowledge distillation according to the general teacher model;

[0010] Among them, the feedback confirmation result is used to control the learning intensity of the target learning model towards the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model towards the general teacher model.

[0011] In a preferred embodiment: based on the feedback confirmation result, a knowledge distillation optimization is performed on the target student model according to the general teacher model, including:

[0012] Obtaining a preset first distillation temperature, and establishing a first loss function according to a general teacher model and a target student model based on the preset first distillation temperature;

[0013] Obtaining a preset second distillation temperature, and adjusting the preset second distillation temperature according to the feedback confirmation result to obtain a target second distillation temperature;

[0014] Based on the target second distillation temperature, establishing a second loss function according to the general teacher model and the target student model;

[0015] Combine the first loss function and the second loss function to establish a total loss function;

[0016] Perform a knowledge distillation optimization on the target student model based on the total loss function.

[0017] In a preferred embodiment, the first loss function and the second loss function are combined to establish a total loss function, including:

[0018] According to the operation results and feedback confirmation results of the target student model, the actual loss function is established;

[0019] Obtain feedback confirmation results from multiple employees belonging to the same department, and compare multiple feedback confirmation results to obtain feedback consistency level values;

[0020] According to the feedback consistency, the influence weights of the first loss function and the second loss function are obtained, wherein the higher the consistency represented by the feedback consistency level value, the greater the influence weight of the first loss function relative to the influence weight of the second loss function;

[0021] A total loss function is established according to the actual loss function, the first loss function, the second loss function, and the influence weights corresponding to the first loss function and the second loss function.

[0022] In a preferred embodiment: the total loss function is:

[0023] Loss = L hard +λ×(w1×L s1 +w2×L s2 );

[0024] Among them, Loss represents the total loss function, L hard represents the actual loss function, λ represents the preset learning rate, L s1 represents the first loss function, w1 represents the influence weight corresponding to the first loss function, L s2 represents the second loss function, and w2 represents the influence weight corresponding to the second loss function.

[0025] In a preferred embodiment, obtaining a preset second distillation temperature, and adjusting the preset second distillation temperature according to the feedback confirmation result to obtain a target second distillation temperature includes:

[0026] Obtaining a preset second distillation temperature;

[0027] If the feedback confirmation result indicates that the target student model operates accurately, then the value of the preset second distillation temperature is reduced to obtain the target second distillation temperature;

[0028] If the feedback confirmation result indicates that the target student model is not running accurately, the value of the preset second distillation temperature is increased to obtain the target second distillation temperature.

[0029] In a preferred embodiment, lowering the value of the preset second distillation temperature to obtain the target second distillation temperature comprises:

[0030] According to the feedback confirmation result, a hard label is established, and preset key elements and preset non-key elements in the hard label are determined. The preset key elements are elements corresponding to the categories representing the correct results in the feedback confirmation result;

[0031] Get the output vector of the target student model;

[0032] According to the difference between the preset key elements and the preset non-key elements in the output vector and the hard label, the temperature adjustment amplitude is obtained;

[0033] The preset second distillation temperature value is adjusted based on the temperature adjustment range to obtain the target second distillation temperature.

[0034] In a preferred embodiment, the temperature adjustment range is obtained according to the difference between the preset key elements and the preset non-key elements in the output vector and the hard tag, including:

[0035] The temperature adjustment range is obtained by the following formula:

[0036]

[0037] Among them, A represents the temperature adjustment amplitude, d1 represents the difference of preset key elements, Δx1 represents the numerical difference of a preset key element in the output vector and the hard label, n represents the number of preset key elements, d2 represents the difference of preset non-key elements, Δx2 represents the numerical difference of a preset non-key element in the output vector and the hard label, m represents the number of preset non-key elements, a, b and c represent the preset weight adjustment parameters respectively, e represents the natural constant, and ln represents the natural logarithm.

[0038] In a preferred embodiment: based on the target second distillation temperature, a second loss function is established according to the general teacher model and the target student model, including:

[0039] Get preferred categories and non-preferred categories;

[0040] Based on the target second distillation temperature, the initial soft labels of the general teacher model are obtained;

[0041] The initial soft label is normalized twice to obtain a target soft label, wherein the difference between the preferred category and the non-preferred category in the target soft label is greater than the difference between the preferred category and the non-preferred category in the initial soft label;

[0042] According to the target soft label, the second loss function is established.

[0043] In a preferred embodiment, the initial soft label is normalized twice to obtain the target soft label, including:

[0044] The values ​​of the non-preferred categories in the initial soft label are set to zero, and based on the values ​​of the preferred categories in the initial soft label, the elements of the preferred categories in the initial soft label are normalized to obtain the target soft label.

[0045] The present invention also provides an employee automated business process processing system based on artificial intelligence, comprising:

[0046] Initialization module, used to obtain the general teacher model and target student model for automating business processes;

[0047] A trial operation module is used to run the target student model and obtain feedback confirmation results of the target student model operation;

[0048] The personalization module is used to perform a knowledge distillation optimization on the target student model based on the general teacher model based on the feedback confirmation results;

[0049] Among them, the feedback confirmation result is used to control the learning intensity of the target learning model towards the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model towards the general teacher model.

[0050] The beneficial effects of adopting the above embodiment are:

[0051] The present invention provides an employee automated business process processing method based on artificial intelligence, which first obtains a general teacher model and a target student model for automated business processes, then runs the target student model, and obtains the feedback confirmation result of the target student model operation, and finally performs a knowledge distillation optimization on the target student model based on the general teacher model based on the feedback confirmation result. The feedback confirmation result is used to control the learning intensity of the target learning model to the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model to the general teacher model. Compared with the prior art, the present invention uses the feedback confirmation result to dynamically adjust the target student model so that it increasingly conforms to the personalized operation habits of the user as the number of runs increases. At the same time, the feedback confirmation result is combined with the knowledge distillation technology to enable it to converge quickly, accelerate the training speed of the target learning model, reduce the cost of model deployment, and improve the flexibility of the application of automated business process technology, solving the problem of lack of personalization of business process automation technology in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A method flow chart of the method for processing employee automated business processes based on artificial intelligence provided by the present invention;

[0053] Figure 2 for Figure 1 Specific step diagram of step S103;

[0054] Figure 3 This is a specific method step diagram for lowering the preset second distillation temperature in the present invention;

[0055] Figure 4 for Figure 2 Specific step diagram of step S203;

[0056] Figure 5 This is a system structure diagram of the artificial intelligence-based employee automated business process processing system provided by the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] Combination Figure 1As shown, a specific embodiment of the present invention discloses an employee automated business process processing method based on artificial intelligence, including:

[0059] S101, obtaining a general teacher model and a target student model for automating a business process;

[0060] S102, running the target student model, and obtaining feedback confirmation results of the target student model running;

[0061] S103, based on the feedback confirmation result, perform a knowledge distillation optimization on the target student model according to the general teacher model;

[0062] Among them, the feedback confirmation result is used to control the learning intensity of the target learning model towards the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model towards the general teacher model.

[0063] It should be noted that knowledge distillation technology is a method used for model compression and knowledge transfer in the field of deep learning. Its core idea is to transfer the knowledge of a large and complex model (teacher model) to a small and simple model (student model) to optimize the performance of the student model and speed up the model training.

[0064] In the present invention, the general teacher model is a pre-trained artificial intelligence model with good performance, more complex and able to adapt to most work needs, which serves as a teacher model for transferring knowledge in the knowledge distillation process. The target student model is an artificial intelligence model that is actually used by employees and is to be optimized and personalized, that is, a student model for transferring knowledge in the knowledge distillation process. Both are used to assist in business process automation, decision-making, classification and other types of work, such as fault diagnosis, resume screening, administrative approval, etc. The feedback confirmation result is the feedback information input by the employee who actually runs the target student model, such as the actual fault cause of the fault diagnosis, whether a resume is qualified by the employee, whether a certain material is approved, etc.

[0065] Knowledge distillation and traditional neural network training are essentially repeated processes of forward propagation and back propagation. In this embodiment, step S103 refers to a back propagation process. In practice, the target student model can be initially trained with a small amount of data and then given to employees for trial use. During operation, the employees continuously confirm and modify (i.e., repeat the above steps) until it is personalized, after which the business process automation can be completely realized.

[0066] Compared with the prior art, the present invention utilizes feedback confirmation results to dynamically adjust the target student model so that it becomes more and more in line with the user's personalized operating habits as the number of runs increases. At the same time, the feedback confirmation results are combined with knowledge distillation technology to enable it to converge quickly, thereby accelerating the training speed of the target learning model, reducing the cost of model deployment, and improving the flexibility of the application of automated business process technology, thus solving the problem of lack of personalization in business process automation technology in the prior art.

[0067] Further, combined with Figure 2 As shown, in a preferred embodiment, the above step S103, based on the feedback confirmation result, performs a knowledge distillation optimization on the target student model according to the general teacher model, specifically including:

[0068] S201, obtaining a preset first distillation temperature, and establishing a first loss function according to a general teacher model and a target student model based on the preset first distillation temperature;

[0069] S202, obtaining a preset second distillation temperature, and adjusting the preset second distillation temperature according to the feedback confirmation result to obtain a target second distillation temperature;

[0070] S203, based on the target second distillation temperature, establishing a second loss function according to the general teacher model and the target student model;

[0071] S204, combining the first loss function and the second loss function to establish a total loss function;

[0072] S205. Perform a knowledge distillation optimization on the target student model based on the total loss function.

[0073] Similarly, performing a knowledge distillation optimization in this embodiment also refers to performing a back propagation. In the above process, the preset first distillation temperature is a pre-set hyperparameter obtained based on experience or experiments, and the preset second distillation temperature is only a pre-set hyperparameter obtained based on experience or experiments in the first iteration. In the subsequent iteration process, the preset second distillation temperature refers to the distillation temperature used to calculate the loss function during the previous back propagation.

[0074] In the prior art, the total loss function in the knowledge distillation technology generally consists of two parts: one is the actual loss function that represents the gap between the student model and the actual situation, and the other is the distillation loss function used to reflect the difference between the student model and the teacher model. A common expression in the prior art is:

[0075] Loss = L hard +λ×L soft

[0076] Among them, Loss represents the total loss function, Lhard Represents the actual loss function, L soft represents the distillation loss function, and λ represents the preset learning rate.

[0077] It is conceivable that adjusting the learning intensity of the target learning model toward the general teacher model based on the feedback confirmation result can be achieved by adjusting the above-mentioned learning rate, for example, making the preset learning rate positively correlated with the accuracy of the feedback confirmation result representation.

[0078] A key parameter in the distillation loss function is the distillation temperature. The direct effect of the distillation temperature is to adjust the smoothness of the output labels of the neural network model. It can be simply understood as adjusting the difference between the numerical values ​​of each category output by the neural network model. This difference will directly affect the fluctuation of the loss function, and then adjust the modification range of the back propagation. The specific meaning and principle of the distillation temperature are all existing technologies, so this article will not explain it in detail. In general, the distillation loss function can be regarded as a tool for the teacher model to transfer knowledge to the student model, "distilling" the knowledge of the teacher model and handing it over to the student model. The amount of content "distilled" by this tool and whether these contents are useful components or useless impurities all depend on the distillation temperature.

[0079] In this embodiment, the distillation loss function in the prior art is innovatively further split into a first loss function and a second loss function. The first loss function is obtained based on a preset first distillation temperature, and the second loss function is obtained based on a preset second distillation temperature. The key is that the second distillation temperature changes dynamically according to the feedback confirmation result, which makes the ability of the knowledge transferred by the second loss function dynamically adjusted with the feedback confirmation result, which is equivalent to adjusting the intensity of the target student model learning from the general teacher model with the feedback, so as to achieve personalization.

[0080] For example, if an employee makes mistakes repeatedly when using the target student model, it means that the thinking logic of the general teacher model is not consistent with that of the employee. At this time, the second distillation temperature can be increased to make the target student model ignore the details of the general teacher model and only learn the general part that does not participate in decision-making, while the specific details are trained by the actual loss function. It is conceivable that if the second distillation temperature is high enough, the probabilities of each category output by the neural network model tend to be the same, and the second loss function tends to zero. At this time, it is equivalent to the student model not learning from the teacher model at all, achieving extreme personalization.

[0081] Obviously, the above extreme situation makes the teacher model useless, and the process of knowledge distillation completely degenerates into an ordinary back-propagation process. In practice, this process requires a large amount of data and time. In addition, the implementation of automated business processes in practice must comply with the overall decision-making direction of the enterprise. Too personalized and arbitrary decisions are obviously not allowed. Therefore, this embodiment also retains the first loss function, and maintains the minimum learning limit of the target student model from the general teacher model by setting the unchanged preset first distillation temperature, so that the target student model meets the actual needs of the enterprise and also guarantees the convergence rate of the target student model. The two loss functions work together to enable this embodiment to take into account both personalization and learning rate when optimizing the target student model, greatly improving practicality.

[0082] Furthermore, in a preferred embodiment, the above step S204 combines the first loss function and the second loss function to establish a total loss function, which specifically includes:

[0083] According to the operation results and feedback confirmation results of the target student model, the actual loss function is established;

[0084] Obtain feedback confirmation results from multiple employees belonging to the same department, and compare multiple feedback confirmation results to obtain feedback consistency level values;

[0085] According to the feedback consistency, the influence weights of the first loss function and the second loss function are obtained, wherein the higher the consistency represented by the feedback consistency level value, the greater the influence weight of the first loss function relative to the influence weight of the second loss function;

[0086] A total loss function is established according to the actual loss function, the first loss function, the second loss function, and the influence weights corresponding to the first loss function and the second loss function.

[0087] In the above process, "multiple employees belonging to the same department" refers to multiple employees in a target group who perform the same function through the target student model. It can refer to colleagues in the same group or all employees in the enterprise. The feedback consistency level value indicates the consistency of the feedback confirmation results of multiple employees in the same department, such as the average accuracy rate, or the similarity of the distribution of each person's predicted correct classification results. It can be flexibly adjusted according to actual needs.

[0088] The key point is that this embodiment adjusts the weights of the first loss function and the second loss function according to the feedback consistency, that is, further adjusts the learning rate and personalized preference of the target student model. When the consistency represented by the feedback consistency level value is higher, it means that the preference of classification decision in the group is more strict. At this time, the proportion of personalized learning can be appropriately reduced, further improving the flexibility and practicality of this method.

[0089] Specifically, in the above process, the total loss function is reflected as follows:

[0090] Loss = L hard +λ×(w1×L s1 +w2×L s2 );

[0091] Among them, Loss represents the total loss function, L hard represents the actual loss function, λ represents the preset learning rate, L s1 represents the first loss function, w1 represents the influence weight corresponding to the first loss function, L s2 represents the second loss function, and w2 represents the influence weight corresponding to the second loss function.

[0092] Further, in a preferred embodiment, the above step S202, obtaining the preset second distillation temperature, and adjusting the preset second distillation temperature according to the feedback confirmation result to obtain the target second distillation temperature, specifically includes:

[0093] Obtaining a preset second distillation temperature;

[0094] If the feedback confirmation result indicates that the target student model operates accurately, then the value of the preset second distillation temperature is reduced to obtain the target second distillation temperature;

[0095] If the feedback confirmation result indicates that the target student model is not running accurately, the value of the preset second distillation temperature is increased to obtain the target second distillation temperature.

[0096] The specific adjustment method in the above process can be flexibly determined according to actual needs. For example, each time the target student model is fed back to run correctly, the value of the preset second distillation temperature is lowered a little, or the adjustment range is calculated by a preset linear function based on the cumulative accuracy.

[0097] Combination Figure 3 As shown, the present invention provides a preferred embodiment. In this embodiment, the above step of reducing the value of the preset second distillation temperature to obtain the target second distillation temperature specifically includes:

[0098] S301, establishing a hard tag according to the feedback confirmation result, and determining a preset key element and a preset non-key element in the hard tag, wherein the preset key element is an element corresponding to a category representing a correct result in the feedback confirmation result;

[0099] S302, obtaining an output vector of a target student model;

[0100] S303, obtaining a temperature adjustment range according to a difference between a preset key element and a preset non-key element in the output vector and the hard label;

[0101] S304: Adjust the preset second distillation temperature value based on the temperature adjustment range to obtain a target second distillation temperature.

[0102] In the above process, the preset key elements refer to the more critical categories among the option categories that the target student model needs to classify or make decisions. The remaining categories are preset non-key elements, which can be obtained according to manual designation or statistics based on feedback confirmation results, and the categories with higher classification proportions are selected as preset key elements. For example, in a defect detection scenario, if the proportion of a certain defect exceeds 90%, then this defect is the preset key category.

[0103] The preset key elements are naturally the primary consideration. The significance of this embodiment is that the temperature adjustment amplitude is jointly decided based on the difference between the preset key elements in the hard label and the output vector and the difference between the preset non-key elements, so that the target student model can adjust the learning intensity according to the specific situation.

[0104] Specifically, in a preferred embodiment, in the above process, the specific formula for calculating the temperature adjustment range is:

[0105]

[0106] Among them, A represents the temperature adjustment amplitude, d1 represents the difference of preset key elements, Δx1 represents the numerical difference of a preset key element in the output vector and the hard label, n represents the number of preset key elements, d2 represents the difference of preset non-key elements, Δx2 represents the numerical difference of a preset non-key element in the output vector and the hard label, m represents the number of preset non-key elements, a, b and c represent the preset weight adjustment parameters respectively, e represents the natural constant, and ln represents the natural logarithm.

[0107] The significance of the above formula is that the average value of the difference between a preset key element or a preset non-key element is used as the difference between the two. The difference between the preset key elements has the greatest impact on the adjustment range, followed by the similarity between the difference between the preset key elements and the difference between the preset non-key elements, and finally the difference between the preset non-key elements. In this way, the target student model can be adjusted based on the preset key elements first, so that it can reach the standard of actual use more quickly and improve practicality.

[0108] Further, combined with Figure 4 As shown, in a preferred embodiment, the above step S203, based on the target second distillation temperature, establishes a second loss function according to the general teacher model and the target student model, specifically including:

[0109] S401, obtaining preferred categories and non-preferred categories;

[0110] S402, obtaining an initial soft label of the general teacher model based on the target second distillation temperature;

[0111] S403, performing secondary normalization on the initial soft label to obtain a target soft label, wherein the difference between the preferred category and the non-preferred category in the target soft label is greater than the difference between the preferred category and the non-preferred category in the initial soft label;

[0112] S404: Establish a second loss function according to the target soft label.

[0113] Similar to preset key elements and preset non-key elements, preference categories and non-preference categories also refer to one of the option categories that the target student model needs to classify or make decisions. However, preset key elements and preset non-key elements are more inclined to the overall decision trend, while preference categories and non-preference categories indicate a more personalized decision preference of an individual. For example, when screening resumes, an employee is more inclined to let the candidate pass and screen through actual interviews. At this time, the option of "pass" is the preference category, and the option of "fail" is the non-preference category. It can be understood that preference categories and non-preference categories can also be obtained based on statistical analysis of feedback confirmation results.

[0114] Soft labels are an existing definition in knowledge distillation technology, which can be understood as the probability of each category predicted by the general teacher model or the target student model, and the result after normalization, except that the difference between the probability values ​​of different categories in the result will scale with the target second distillation temperature. The significance of this embodiment is that a secondary normalization is performed on the initial soft labels, and the difference between the preferred categories and the non-preferred categories in the initial soft labels is amplified secondary, so that there is more information related to the preferred categories in the second loss function. Similar to the preset key elements and the preset non-key elements, the above design enables the target student model to prioritize learning based on the personalized preferences of employees, and to meet the personalized needs of matching employees more quickly.

[0115] It is understandable that the secondary normalization can be performed in any existing manner, such as calculating the sum of the preference category and the non-preference category in the initial soft label, and then using different temperatures for normalization for the two preferences. The present invention provides a simpler method:

[0116] In a preferred embodiment, the above step S403, performing secondary normalization on the initial soft label to obtain the target soft label, specifically includes:

[0117] The values ​​of the non-preferred categories in the initial soft label are set to zero, and based on the values ​​of the preferred categories in the initial soft label, the elements of the preferred categories in the initial soft label are normalized to obtain the target soft label.

[0118] This embodiment directly sets the values ​​of non-preferred categories in the initial soft labels to zero, thereby reducing the amount of calculation for secondary normalization and further improving the running speed of the method.

[0119] Combination Figure 5 As shown, the present invention also provides an employee automated business process processing system based on artificial intelligence, comprising:

[0120] Initialization module 510, for obtaining a general teacher model and a target student model for automating a business process;

[0121] A trial operation module 520 is used to run the target student model and obtain feedback confirmation results of the target student model operation;

[0122] A personalization module 530 is used to perform a knowledge distillation optimization on the target student model according to the general teacher model based on the feedback confirmation result;

[0123] Among them, the feedback confirmation result is used to control the learning intensity of the target learning model towards the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model towards the general teacher model.

[0124] It should be noted here that the corresponding system provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding contents in the above method embodiments, which will not be repeated here.

[0125] The present invention provides an employee automated business process processing method based on artificial intelligence, which first obtains a general teacher model and a target student model for automated business processes, then runs the target student model, and obtains the feedback confirmation result of the target student model operation, and finally performs a knowledge distillation optimization on the target student model based on the general teacher model based on the feedback confirmation result. The feedback confirmation result is used to control the learning intensity of the target learning model to the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model to the general teacher model. Compared with the prior art, the present invention uses the feedback confirmation result to dynamically adjust the target student model so that it increasingly conforms to the personalized operation habits of the user as the number of runs increases. At the same time, the feedback confirmation result is combined with the knowledge distillation technology to enable it to converge quickly, accelerate the training speed of the target learning model, reduce the cost of model deployment, and improve the flexibility of the application of automated business process technology, solving the problem of lack of personalization of business process automation technology in the prior art.

[0126] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0127] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An employee automated business process processing method based on artificial intelligence, characterized in that: include: Obtain a generic teacher model and target student model for automating business processes; Run the target student model and obtain feedback confirmation results of the target student model operation; Based on the feedback confirmation results, the target student model is optimized through knowledge distillation according to the general teacher model; Among them, the feedback confirmation result is used to control the learning intensity of the target learning model towards the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model towards the general teacher model.

2. The method for processing employee automated business processes based on artificial intelligence according to claim 1, characterized in that: Based on the feedback confirmation results, the target student model is optimized through knowledge distillation according to the general teacher model, including: Obtaining a preset first distillation temperature, and establishing a first loss function according to a general teacher model and a target student model based on the preset first distillation temperature; Obtaining a preset second distillation temperature, and adjusting the preset second distillation temperature according to the feedback confirmation result to obtain a target second distillation temperature; Based on the target second distillation temperature, establishing a second loss function according to the general teacher model and the target student model; Combine the first loss function and the second loss function to establish a total loss function; Perform a knowledge distillation optimization on the target student model based on the total loss function.

3. The method for processing employee automated business processes based on artificial intelligence according to claim 2, characterized in that: Combining the first loss function and the second loss function, a total loss function is established, including: According to the operation results and feedback confirmation results of the target student model, the actual loss function is established; Obtain feedback confirmation results from multiple employees belonging to the same department, and compare multiple feedback confirmation results to obtain feedback consistency level values; According to the feedback consistency, the influence weights of the first loss function and the second loss function are obtained, wherein the higher the consistency represented by the feedback consistency level value, the greater the influence weight of the first loss function relative to the influence weight of the second loss function; A total loss function is established according to the actual loss function, the first loss function, the second loss function, and the influence weights corresponding to the first loss function and the second loss function.

4. The method for processing employee automated business processes based on artificial intelligence according to claim 3 is characterized in that: The total loss function is: Loss=L hard +λ×(w1×L s1 +e2×L s2 ); Among them, Loss represents the total loss function, L hard represents the actual loss function, λ represents the preset learning rate, L s1 represents the first loss function, w1 represents the influence weight corresponding to the first loss function, L s2 represents the second loss function, and w2 represents the influence weight corresponding to the second loss function.

5. The method for processing employee automated business processes based on artificial intelligence according to claim 2, characterized in that: Obtaining a preset second distillation temperature, and adjusting the preset second distillation temperature according to the feedback confirmation result to obtain a target second distillation temperature, including: Obtaining a preset second distillation temperature; If the feedback confirmation result indicates that the target student model operates accurately, then the value of the preset second distillation temperature is reduced to obtain the target second distillation temperature; If the feedback confirmation result indicates that the target student model is not running accurately, the value of the preset second distillation temperature is increased to obtain the target second distillation temperature.

6. The method for processing employee automated business processes based on artificial intelligence according to claim 5, characterized in that: Lowering the value of the preset second distillation temperature to obtain the target second distillation temperature includes: According to the feedback confirmation result, a hard label is established, and preset key elements and preset non-key elements in the hard label are determined. The preset key elements are elements corresponding to the categories representing the correct results in the feedback confirmation result; Get the output vector of the target student model; According to the difference between the preset key elements and the preset non-key elements in the output vector and the hard label, the temperature adjustment amplitude is obtained; The preset second distillation temperature value is adjusted based on the temperature adjustment range to obtain the target second distillation temperature.

7. The method for processing employee automated business processes based on artificial intelligence according to claim 6, characterized in that: According to the difference between the preset key elements and the preset non-key elements in the output vector and the hard label, the temperature adjustment range is obtained, including: The temperature adjustment range is obtained by the following formula: Among them, A represents the temperature adjustment amplitude, d1 represents the difference of preset key elements, Δx1 represents the numerical difference of a preset key element in the output vector and the hard label, n represents the number of preset key elements, d2 represents the difference of preset non-key elements, Δx2 represents the numerical difference of a preset non-key element in the output vector and the hard label, m represents the number of preset non-key elements, a, b and c represent the preset weight adjustment parameters respectively, e represents the natural constant, and ln represents the natural logarithm.

8. The method for processing employee automated business processes based on artificial intelligence according to claim 2, characterized in that: Based on the target second distillation temperature, a second loss function is established according to the general teacher model and the target student model, including: Get preferred categories and non-preferred categories; Based on the target second distillation temperature, the initial soft labels of the general teacher model are obtained; The initial soft label is normalized twice to obtain a target soft label, wherein the difference between the preferred category and the non-preferred category in the target soft label is greater than the difference between the preferred category and the non-preferred category in the initial soft label; According to the target soft label, the second loss function is established.

9. The method for processing employee automated business processes based on artificial intelligence according to claim 8, characterized in that: The initial soft label is normalized twice to obtain the target soft label, including: The values ​​of the non-preferred categories in the initial soft label are set to zero, and based on the values ​​of the preferred categories in the initial soft label, the elements of the preferred categories in the initial soft label are normalized to obtain the target soft label.

10. An employee automated business process processing system based on artificial intelligence, characterized in that: include: Initialization module, used to obtain the general teacher model and target student model for automating business processes; A trial operation module is used to run the target student model and obtain feedback confirmation results of the target student model operation; The personalization module is used to perform a knowledge distillation optimization on the target student model based on the general teacher model based on the feedback confirmation results; Among them, the feedback confirmation result is used to control the learning intensity of the target learning model towards the general teacher model during this optimization, and the accuracy of the feedback confirmation result representation is positively correlated with the learning intensity of the target learning model towards the general teacher model.

Images