Task processing strategy determination method and device, equipment, medium and product

Abstract

The invention discloses a task processing strategy determination method and device, equipment, a medium and a product. The method comprises the steps of obtaining target task data; decomposing the target task data by using a preset hierarchical sub-network to obtain candidate actions corresponding to a plurality of sub-tasks; performing risk prediction on source data in the target task data corresponding to the candidate actions by using a preset risk prediction sub-network to obtain a risk tag corresponding to each candidate action; performing action reward calculation on the candidate actions by utilizing a preset meta-model to obtain action values corresponding to the candidate actions; and determining a target action path based on each candidate action, the risk tag and the action value, so as to determine a task processing strategy according to the target action path. According to the embodiment of the invention, the intelligent and dynamic adaptation of task decomposition can be realized, the risk of determining the task processing strategy is reduced, and the accuracy of determining the task processing strategy is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of big data processing, and particularly relates to a task processing strategy determination method and device, equipment, medium and product. BACKGROUND

[0002] Large language models (LLM) have received extensive attention and application in recent years as an important breakthrough in the field of artificial intelligence. LLM can learn rich language knowledge and world knowledge through pre-training on massive text data, and has achieved remarkable results in natural language processing tasks such as text classification, named entity recognition, and question-answering systems.

[0003] Although LLM performs well in handling relatively simple and common tasks, the accuracy of the determined task processing strategy is low when facing high complexity and long-term tasks, and cannot face the possible complex environment and technical risks. SUMMARY

[0004] The application embodiment provides a task processing strategy determination method, device, equipment, medium and product, which can realize intelligent and dynamic adaptation of task decomposition, reduce the risk of task processing strategy determination, and improve the accuracy of task processing strategy determination.

[0005] In one aspect, the application embodiment provides a task processing strategy determination method, which comprises: obtaining target task data; decomposing and disassembling the target task data using a preset hierarchical subnetwork to obtain candidate actions corresponding to multiple subtasks; predicting the risk of source data in the target task data corresponding to the candidate actions using a preset risk prediction subnetwork to obtain a risk label corresponding to each candidate action; calculating the action reward of the candidate action using a preset meta-model to obtain an action value corresponding to the candidate action; determining a target action path based on each candidate action, risk label and action value to determine a task processing strategy according to the target action path.

[0006] Optionally, the decomposing and disassembling the target task data using a preset hierarchical subnetwork to obtain candidate actions corresponding to multiple subtasks comprises: performing feature analysis on the target task data to obtain multiple data features; performing task level division based on the multiple data features using the preset hierarchical subnetwork to obtain multiple task levels; perform complexity analysis based on the plurality of task levels to obtain a growth degree threshold value, the growth degree threshold value being used to represent a complexity degree of the data feature; perform optimization on the growth degree threshold value to obtain an optimal generation length threshold value, the optimal generation length threshold value being used to represent an optimal complexity degree of the target task data decomposition; segment the target task data according to the optimal generation length threshold value to generate candidate actions corresponding to the plurality of sub-tasks.

[0007] Optionally, before the optimization on the growth degree threshold value to obtain the optimal generation length threshold value, the method further comprises: perform mask sequence conversion on the target task data to obtain a text mask sequence of the target task data, the text mask sequence comprising a plurality of to-be-reconstructed sequences; perform iterative deduplication on the text mask sequence to obtain a reconstructed text; in the process of iterative deduplication, perform quality evaluation on the reconstructed text to obtain an evaluation index of the reconstructed text; in a case where the evaluation index meets a preset optimal threshold value, determine a sequence length corresponding to the current reconstructed text as the optimal generation length threshold value.

[0008] Optionally, before the risk prediction on the source data in the target task data corresponding to the candidate action by using the preset risk prediction sub-network to obtain a risk label corresponding to each candidate action, the method further comprises: obtain risk definition interaction data; perform analysis and feature extraction on the risk definition interaction data to obtain risk features and corresponding risk categories; perform analysis on the risk features and corresponding risk categories based on a preset risk prediction sub-network to obtain a risk determination condition; the risk prediction on the source data in the target task data corresponding to the candidate action by using the preset risk prediction sub-network to obtain a risk label corresponding to each candidate action, comprises: input the source data corresponding to the candidate action in the target task data into the preset risk prediction sub-network, the preset risk prediction sub-network performing feature extraction on the source data to obtain risk features; match the risk features with the risk determination condition to obtain a matching result; based on the matching result, determine a risk label corresponding to the candidate action.

[0009] Optionally, before the action reward calculation on the candidate action by using the preset meta-model to obtain an action value corresponding to the candidate action, the method further comprises: Obtain a training set, which includes action descriptions, task data, and historical interaction data; The action description is input into an initial recurrent neural network model, which generates a dynamic reward action based on the action description and a reward function. Based on the dynamic reward action and the task data, a criticism signal is generated; Based on the criticism signal and the dynamic reward action, a preset voting mechanism is used to vote on the candidate reward functions to obtain the dynamic reward function of the preset meta-model; Based on the dynamic reward function, the preset meta-model is constructed.

[0010] Optionally, the step of calculating the action reward for the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action includes: A decision tree is generated based on the candidate actions; Based on the decision tree, a preset tree search algorithm is used to generate multiple candidate action paths; For each candidate action path, a reward action is performed using a preset environment model and a preset meta-model to obtain the reward value for each candidate path; The summation of the reward values ​​yields the action value of the decision tree.

[0011] Optionally, before determining a target action path based on each candidate action, risk label, and action value, and then determining a task processing strategy based on the target action path, the method further includes: For each candidate action path, the path value of the current candidate action path is determined using a preset path value formula based on the action value of the candidate action path and the reward value of the current candidate action. The process of determining a target action path based on each candidate action, risk label, and action value, and then determining a task processing strategy based on the target action path, includes: Based on each candidate action, risk label, action value, and path value, a target action path is determined, and a task processing strategy is determined based on the target action path.

[0012] Optionally, determining the target action path based on each of the candidate actions, risk tags, action values, and path values ​​includes: Risk simulation data is generated based on the candidate actions, risk tags, action values, and path values. Under each risk simulation data set, determine the simulated risk value for each candidate action path; The risk indicators for each candidate action path are determined using a pre-set comprehensive evaluation formula. The candidate action path corresponding to the highest risk indicator is determined as the target action path.

[0013] On the other hand, embodiments of this application provide a task processing strategy determination apparatus, the apparatus comprising: The acquisition module is used to acquire target task data; The decomposition module is used to decompose the target task data using a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks; The risk prediction module is used to perform risk prediction on the source data in the target task data corresponding to the candidate action using a preset risk prediction sub-network, and obtain the risk label corresponding to each candidate action. The calculation module is used to calculate the action reward of the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action; The determination module is used to determine the target action path based on each of the candidate actions, risk tags, and action values, so as to determine the task processing strategy according to the target action path.

[0014] In another aspect, embodiments of this application provide an electronic device, the device comprising: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the task processing strategy determination method as described in the first aspect.

[0015] In another aspect, embodiments of this application provide a computer storage medium on which computer program instructions are stored, and when executed by a processor, the computer program instructions implement the task processing strategy determination method as described in the first aspect.

[0016] In another aspect, embodiments of this application provide a computer program product in which instructions, when executed by a processor of an electronic device, cause the electronic device to perform the task processing strategy determination method as described in the first aspect.

[0017] The task processing strategy determination method, apparatus, device, medium, and product of this application embodiment can decompose target task data through a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks. Then, risk labels and action values ​​are determined through the candidate actions. Based on the candidate actions, risk labels, and action values, risk perception is added to the task processing strategy, which effectively reduces the risk of the task processing strategy. At the same time, it can effectively ensure the accuracy of the task processing strategy when the scenario and technical risks change, and realize the intelligent determination of task processing strategy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a task processing strategy determination method provided in one embodiment of this application; Figure 2 This is a flowchart illustrating a task processing strategy determination method provided in another embodiment of this application; Figure 3 This is a flowchart illustrating a task processing strategy determination method provided in another embodiment of this application; Figure 4 This is a schematic diagram of the structure of a task processing strategy determination device provided in another embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0020] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0021] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0022] To address the problems of existing technologies, embodiments of this application provide a method, apparatus, device, medium, and product for determining task processing strategies. In these embodiments, target task data is decomposed using a pre-defined hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks. Risk labels and action values ​​are then determined based on these candidate actions. Finally, risk perception is incorporated into the task processing strategy based on the candidate actions, risk labels, and action values. This effectively reduces the risk of the task processing strategy while ensuring its accuracy even when scenario and technical risks change, thus achieving intelligent determination of the task processing strategy.

[0023] The task processing strategy determination method provided in the embodiments of this application will be introduced first below.

[0024] Figure 1 A flowchart illustrating a task processing strategy determination method according to an embodiment of this application is shown. Figure 1 As shown, the task processing strategy determination method may include S101-S105: S101, Obtain target task data.

[0025] In some embodiments, the target task data can be text input by the user describing the task to be processed, which may include text and operators. That is, the user can input the target task data into the electronic device in advance through an interactive method to predict the risk of the target task data and obtain the task processing strategy of the target task data, thereby improving the task processing strategy.

[0026] S102, the target task data is decomposed and broken down using a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks.

[0027] In some embodiments, since each target task data includes multiple subtasks and these multiple subtasks are hierarchical, it is necessary to decompose and break down the target task data using a preset hierarchical subnetwork to obtain candidate actions corresponding to multiple subtasks. Furthermore, the candidate action corresponding to each child node inherits the candidate action of the parent node, and the candidate action can be the execution action for processing each subtask.

[0028] In this embodiment, the preset hierarchical subnetwork can be a large language model.

[0029] S103, using a preset risk prediction sub-network to perform risk prediction on the source data in the target task data corresponding to the candidate action, and obtain the risk label corresponding to each candidate action.

[0030] In this embodiment, after decomposing the target task data using a preset hierarchical sub-network to obtain candidate actions corresponding to each sub-task, it is necessary to analyze and predict the execution of each candidate action to obtain the risk label corresponding to each candidate action, so as to introduce risk perception for subsequent determination of task processing strategy and improve the accuracy of task processing strategy.

[0031] It is worth noting that the preset risk prediction subnetwork can be a large language model.

[0032] S104. Calculate the action reward for the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action.

[0033] In this embodiment, the preset meta-model can be a reward function generator that can generate a customized reward function in real time based on multi-level candidate actions that protect context information, so as to reduce the optimization deviation caused by fixed rules.

[0034] The preset meta-model can be formed by embedding a task encoder in a recurrent neural network, where the recurrent neural network establishes task sequence dependencies and the task encoder maps task features to the reward function space.

[0035] S105. Based on each candidate action, risk label, and action value, determine the target action path, and determine the task processing strategy according to the target action path.

[0036] In this embodiment, the target task data is decomposed by a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks. Then, risk labels and action values ​​are determined by the candidate actions. Based on the candidate actions, risk labels, and action values, risk perception is added to the task processing strategy, which effectively reduces the risk of the task processing strategy. At the same time, it can effectively ensure the accuracy of the task processing strategy when the scenario and technical risks change, thus realizing the intelligent determination of the task processing strategy.

[0037] Reference Figure 2 In some embodiments, S102 may include: S1021, Perform feature analysis on the target task data to obtain multiple data features; S1022, Using a pre-defined hierarchical sub-network, task levels are divided based on multiple data features to obtain multiple task levels; S1023, complexity analysis is performed based on multiple task levels to obtain the growth degree threshold, which is used to characterize the complexity of data features; S1024, optimize the growth degree threshold to obtain the optimal generation length threshold; S1025, the target task data is segmented according to the optimal generation length threshold to generate candidate actions corresponding to multiple subtasks.

[0038] In this embodiment, since the target task data may include multiple subtasks, during the decomposition process, it is necessary to generate corresponding candidate actions for each subtask. It is worth noting that the preset hierarchical subnetwork adapts to the text length limit of the target task data during the subtask decomposition process. Based on the candidate actions corresponding to each level of subtask, a lightweight subnetwork can be used to perform problem complexity analysis on the target task data, output growth degree threshold, and then the target task data is segmented according to the growth degree threshold to generate candidate actions corresponding to the subtasks.

[0039] Specifically, the target task data can first be analyzed to obtain multiple data features, such as accuracy and logical chains. During the task hierarchy division using a pre-defined hierarchical sub-network, dynamic hierarchy division can be performed. For example, a smaller growth threshold can be output for highly complex paragraphs such as scientific arguments and multi-threaded narratives, such as 200 words / paragraph, to ensure local logical rigor. For less complex paragraphs such as background descriptions and data listings, a larger threshold, such as 800 words / paragraph, can be used to improve generation efficiency. Then, the conventional optimization algorithm within the pre-defined hierarchical sub-network is used to optimize and obtain the optimal generation length threshold. The optimal generation length threshold is used to characterize the optimal complexity of the target task data decomposition. Finally, the target task data is segmented according to the optimal generation length threshold, and candidate actions corresponding to the sub-tasks are generated.

[0040] It is worth noting that the preset hierarchical subnetworks can be small Transformers or graph neural networks.

[0041] In other embodiments, in order to meet the real needs of different target tasks, the method may further include the following steps before S1024: The target task data is transformed into a mask sequence to obtain a text mask sequence of the target task data, which includes multiple segments of the sequence to be reconstructed. Iterative deduplication of the text mask sequence yields the reconstructed text; During the iterative deduplication process, the quality of the reconstructed text is evaluated to obtain evaluation metrics for the reconstructed text. If the evaluation metrics meet the preset optimal threshold, the sequence length corresponding to the current reconstructed text is determined as the optimal generation length threshold.

[0042] In this embodiment, the text can also be reconstructed gradually through multiple iterations using a mask diffusion strategy based on the global context of the complete sequence, and the optimal generation length threshold can be determined based on the reconstructed text, so that the determined candidate actions are more in line with the target task data.

[0043] Specifically, taking the pre-defined hierarchical sub-network Transformer as an example, the Transformer model has a bidirectional attention mechanism. In any discussion iteration, the pre-defined hierarchical sub-network can see all positions of the entire sequence, including the determined text and the mask to be filled, thereby ensuring that the generation of each word is based on complete contextual information and effectively capturing long-range dependencies.

[0044] In this embodiment, we can start with a sequence of masks or random noise, which contains only prompt words or a few anchor words for initialization. Then, in each iteration, the preset hierarchical subnetwork can predict the probability distribution of all masks that are children based on the current complete mask sequence, select the part of the mask with the highest confidence to fill or replace it according to a certain proportion, and then use the updated sequence as the input for the next iteration to perform iterative text reconstruction and obtain the reconstructed text.

[0045] After each round or a preset number of iterations, the quality of the reconstructed text can be evaluated to obtain evaluation metrics. These metrics can include reconstruction confidence, change convergence, and semantic integrity. Reconstruction confidence can be the reconstruction probability or perplexity of the generated parts by a preset hierarchical subnetwork. Change convergence can characterize whether the changes in text content tend to level off between consecutive iterations. Semantic integrity can be determined by an auxiliary classifier to determine whether the current text has answered the question, completed the instruction, or formed a complete narrative.

[0046] After obtaining the evaluation metrics for the reconstructed text, the evaluation metrics can be compared with the preset optimal threshold. If the reconstruction confidence is greater than the preset threshold or the change convergence is lower than the preset threshold, it can be determined that convergence has been achieved and the iteration can be stopped. The sequence length at this point is the optimal generation length threshold.

[0047] In some embodiments, the loss function of the optimal generation length threshold configured in the preset hierarchical subnetwork can be expressed by the following formula (1):

[0048] Among them, R acc For accuracy, R coh For semantic coherence scoring, L / L opt To penalize length deviation In some other embodiments, prior to S103, the method may further include: Obtain risk definition interaction data; The risk definition interaction data is parsed and features are extracted to obtain risk features and corresponding risk types; Based on the pre-set risk prediction sub-network, the risk characteristics and corresponding risk types are analyzed to obtain the risk judgment conditions; S104 may include: The source data corresponding to the candidate actions in the target task data is input into the preset risk prediction sub-network. The preset risk sub-network extracts features from the source data to obtain risk features. The risk characteristics are matched with the risk assessment criteria to obtain the matching results; Based on the matching results, the risk labels corresponding to the candidate actions are determined.

[0049] In this embodiment, before generating risk labels using the preset risk prediction sub-network, it is necessary to define risk judgment conditions. Risk labels are determined by different risk judgment conditions. Specifically, when determining risk judgment conditions, in order to generate risk labels quickly, user-input risk definition interaction data can be obtained. This risk definition interaction data can be text directly entered by the user, a text file uploaded by the user, or text extracted from a link provided by the user. No limitation is made here.

[0050] At this point, a pre-defined risk prediction sub-network can be used to parse and extract features from the risk definition interaction data to obtain risk features and corresponding risk types. Then, the risk judgment conditions can be determined based on the risk features and risk types.

[0051] In other embodiments, when the user does not input risk definition interaction data, risk definition standard data that matches the target task data can be determined based on similarity, and risk judgment conditions that match the risk definition standard data can be obtained.

[0052] In this embodiment, as an example, risk assessment criteria may include: brute-force attack risk, data leakage risk, etc. The brute-force attack risk can be further specifically defined as: more than 15 failed login attempts from the same IP address to the same username within one minute, thus being deemed an illegal login attack risk; the data leakage risk can be further specifically defined as: internal employee accounts downloading more than 1GB of data outside of working hours and attempting to send data outwards through non-standard ports (such as non-80 / 443), thus being deemed an employee account illegally downloading data and risking data leakage.

[0053] In this embodiment, when generating risk labels for candidate actions corresponding to target task data, the source data corresponding to the candidate actions can be located from the target task data. The source data is then injected into a preset risk prediction sub-network. The risk prediction sub-network extracts the data features of the source data and matches them with risk judgment conditions. Based on the matching results, the risk label for a candidate action is determined. One or more risk judgment conditions correspond to one risk label.

[0054] In this embodiment, the preset risk prediction subnetwork can be a large language model.

[0055] In some other embodiments, prior to S104, the method may further include: Obtain the training set; The action description is input into the initial recurrent neural network model; Based on dynamic reward actions and task data, criticism signals are generated. Based on criticism signals and dynamic reward actions, a preset voting mechanism is used to vote on candidate reward functions to obtain the dynamic reward function of the preset meta-model. A pre-defined meta-model is constructed based on a dynamic reward function.

[0056] In this embodiment, the preset meta-reward model can be a reward function generator that can generate a customized reward function in real time based on multi-level candidate actions of context information, thus avoiding optimization deviations caused by fixed rules.

[0057] Specifically, when training a pre-defined meta-model, a training set can be obtained. The training set can include action descriptions, task data, and historical interaction data. The task descriptions can be candidate actions, and the historical interaction data can include action paths and user evaluation information of the action paths. The sample labels can be inferred by using the action paths and user evaluation information of the action paths.

[0058] In the process of generating the preset meta-reward model, action descriptions are input into an initial recurrent neural network model. This initial model generates dynamic reward actions based on the action descriptions, then generates dynamic reward actions from candidate actions. Subsequently, it generates a criticism signal based on the dynamic reward actions and the target task data. Based on this criticism signal and the dynamic reward actions, the candidate reward functions are voted on to obtain the final dynamic reward function. Finally, the preset meta-model is constructed based on this dynamic reward function. The criticism signal characterizes the criticism function on the target task data, preventing overfitting. Furthermore, using a title-based approach can improve the robustness of the reward function.

[0059] In this embodiment, during the training of the preset meta-model, the loss function of the preset meta-model is as shown in formula (2):

[0060] in, The optimal dynamic reward function is derived by back-calculating the target action path and the user's evaluation of the path. This is the current dynamic reward function.

[0061] In some other embodiments, after the preset meta-model has been trained, S105 may include: Generate a decision tree based on candidate actions; Multiple candidate action paths are generated based on a decision tree using a pre-defined tree search algorithm. For each candidate action path, a reward action is performed using a preset environment model and a preset meta-model to obtain the reward value for each candidate path; The summation of reward values ​​yields the action value of the decision tree.

[0062] In this embodiment, a decision tree is first generated using multi-level candidate actions. Each node of the decision tree carries a risk label. Then, a Monte Carlo tree can be used to search the decision tree to generate multiple candidate action paths, which include multiple candidate actions for executing subtasks.

[0063] Specifically, in determining the value of an action, a pre-judgment environment model can be used to predict state transitions for each candidate action path. For each candidate action path [a1, a2, ..., aT], starting from the initial state s0, actions at are applied step by step. The environment model yields a new state s_{t+1}, and the dynamic reward function R_dynamic(s_t, a_t) is called to obtain the reward value r_t for each candidate action. Finally, the total reward value of the candidate action path is the sum of the reward values ​​(discounts) of all candidate actions: G = Σr_t. That is, the total reward value of the candidate action path can be considered the action value of the decision tree.

[0064] In some other embodiments, in order to improve the accuracy of task processing strategy determination, the method may further include the following steps before S105: For each candidate action path, the path value of the current candidate action path is determined using a preset path value formula based on the action value of the candidate action path and the reward value of the current candidate action. S105 may include: Based on each candidate action, risk label, action value, and path value, the target action path is determined, and the task processing strategy is determined according to the target action path.

[0065] In this embodiment, for a candidate action path, the discount factor corresponding to each candidate action can be calculated using the action value or the aforementioned reward value, and then the path value of the candidate action path can be calculated using the following preset path value formula (3):

[0066] Where T represents the path depth, t represents the current candidate action, and r t Q represents the discount factor. t R represents the value of an action. t This represents the reward value.

[0067] In this embodiment, the determined path value can be used to optimize the task determination strategy from multiple candidate action paths, and find the best task strategy to perform the target task.

[0068] In this embodiment, for each candidate action path, the risk labels of each candidate action on the candidate action path are first deduplicated, then classified and graded according to the deduplicated risk labels, and finally the risk labels of the highest grade in each category are weighted and summed to obtain risk assessment data.

[0069] In some other embodiments, S105 may include: Risk simulation data is generated by simulating candidate actions, risk labels, action value, and path value. Under each risk simulation data set, determine the simulated risk value for each candidate action path; The risk indicators for each candidate action path are determined using a pre-set comprehensive evaluation formula. The candidate action path corresponding to the highest risk indicator is determined as the target action path.

[0070] In this embodiment, during the process of determining the task processing strategy, risk data simulation can be performed based on candidate actions, risk labels, action value, and path value to generate various technical risk data and various market risk data. Then, the technical risk data and market risk data are combined to obtain various risk simulation data. Under each risk simulation data, the risk simulation value of each candidate action path is determined, and the comprehensive evaluation index of each candidate action path is calculated using the following preset comprehensive evaluation formula (4):

[0071] Where a, λ, μ and As weights, they can be determined through model training during implementation. At this point, the candidate action path with the highest comprehensive evaluation index can be taken as the target action path, and then the task processing strategy can be determined based on the candidate actions on the target action path.

[0072] Reference Figure 3 As a specific example, methods for determining task processing strategies may include: S301, Obtain target task data; S302, multi-level task decomposition and processing, generating multi-level candidate actions; S303, Construct a decision tree that carries risk labels and action values; S304, Tree search generates candidate action paths; S305, Generate reward value and path value based on candidate action paths, determine risk assessment data for candidate action paths, simulate risk data, and generate risk simulation data; S306 integrates path value, risk simulation data, reward value, and risk assessment data to determine a comprehensive evaluation index for candidate action paths; S307, Filter target action paths; S308, Generate task processing strategy.

[0073] Based on the task processing strategy determination method provided in the above embodiments, this application also provides specific implementation methods of the task processing strategy determination device. Please refer to the following embodiments.

[0074] First see Figure 4 The task processing strategy determination device 400 provided in this application embodiment may include: Module 401 is used to acquire target task data; The decomposition module 402 is used to decompose the target task data using a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks. Risk prediction module 403 is used to perform risk prediction on the source data in the target task data corresponding to the candidate action using a preset risk prediction sub-network, and obtain the risk label corresponding to each candidate action; The calculation module 404 is used to calculate the action reward of the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action. The determination module 405 is used to determine the target action path based on each candidate action, risk label, and action value, so as to determine the task processing strategy according to the target action path.

[0075] As an optional implementation, the disassembly module 402 can be specifically used for: Feature analysis is performed on the target task data to obtain multiple data features; By using a pre-defined hierarchical subnetwork based on multiple data features, task levels are divided into multiple levels. Complexity analysis is performed based on multiple task levels to obtain growth degree thresholds, which are used to characterize the complexity of data features. The optimal generation length threshold is obtained by optimizing the growth degree threshold. The optimal generation length threshold is used to characterize the optimal complexity of the target task data decomposition. The target task data is segmented based on the optimal generation length threshold to generate candidate actions for multiple subtasks.

[0076] As an optional implementation, the disassembly module 402 can also be specifically used for: The target task data is transformed into a mask sequence to obtain a text mask sequence of the target task data, which includes multiple segments of the sequence to be reconstructed. Iterative deduplication of the text mask sequence yields the reconstructed text; During the iterative deduplication process, the quality of the reconstructed text is evaluated to obtain evaluation metrics for the reconstructed text. If the evaluation metrics meet the preset optimal threshold, the sequence length corresponding to the current reconstructed text is determined as the optimal generation length threshold.

[0077] As an optional implementation, the risk prediction module 403 can also be used for: Obtain risk definition interaction data; The risk definition interaction data is parsed and features are extracted to obtain risk features and corresponding risk types; Based on the pre-set risk prediction sub-network, the risk characteristics and corresponding risk types are analyzed to obtain the risk judgment conditions; Risk prediction module 403 can also be used for: The source data corresponding to the candidate actions in the target task data is input into the preset risk prediction sub-network. The preset risk sub-network extracts features from the source data to obtain risk features. The risk characteristics are matched with the risk assessment criteria to obtain the matching results; Based on the matching results, the risk labels corresponding to the candidate actions are determined.

[0078] As an optional implementation, the computing module 404 can also be used for: Obtain the training set, which includes action descriptions, task data, and historical interaction data; The action description is input into the initial recurrent neural network model, which generates a dynamic reward action based on the action description and the reward function. Based on dynamic reward actions and task data, criticism signals are generated. Based on criticism signals and dynamic reward actions, a preset voting mechanism is used to vote on candidate reward functions to obtain the dynamic reward function of the preset meta-model. A pre-defined meta-model is constructed based on a dynamic reward function.

[0079] As an optional implementation, the computing module 404 can also be used for: Generate a decision tree based on candidate actions; Multiple candidate action paths are generated based on a decision tree using a pre-defined tree search algorithm. For each candidate action path, a reward action is performed using a preset environment model and a preset meta-model to obtain the reward value for each candidate path; The summation of reward values ​​yields the action value of the decision tree.

[0080] As an optional implementation, the determining module 405 can also be used for: For each candidate action path, the path value of the current candidate action path is determined using a preset path value formula based on the action value of the candidate action path and the reward value of the current candidate action. The determination module 405 can also be used for: Based on each candidate action, risk label, action value, and path value, the target action path is determined, and the task processing strategy is determined according to the target action path.

[0081] As an optional implementation, the determining module 405 can also be used for: Risk simulation data is generated by simulating candidate actions, risk labels, action value, and path value. Under each risk simulation data set, determine the simulated risk value for each candidate action path; The risk indicators for each candidate action path are determined using a pre-set comprehensive evaluation formula. The candidate action path corresponding to the highest risk indicator is determined as the target action path.

[0082] Figure 5 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0083] An electronic device may include a processor 501 and a memory 502 storing computer program instructions.

[0084] Specifically, the processor 501 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0085] Memory 502 may include mass storage for data or instructions. For example, and not limitingly, memory 502 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 502 may include removable or non-removable (or fixed) media, or memory 502 may be non-volatile solid-state storage. Memory 502 may be internal or external to the integrated gateway disaster recovery device.

[0086] In one instance, memory 502 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0087] Memory 502 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the task processing strategy determination method according to the first aspect of this disclosure.

[0088] The processor 501 reads and executes computer program instructions stored in the memory 502 to achieve... Figure 1 A method for determining a task processing strategy in the illustrated embodiment.

[0089] In one example, the electronic device may also include a communication interface 503 and a bus 504. For example, Figure 5 As shown, the processor 501, memory 502, and communication interface 503 are connected through bus 504 and complete communication with each other.

[0090] The communication interface 503 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0091] Bus 504 includes hardware, software, or both, that couples components of an electronic device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 504 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0092] The electronic device can execute the task processing strategy determination method in the embodiments of this application, thereby achieving a combination Figures 1-4 The described method and apparatus for determining task processing strategies.

[0093] Furthermore, in conjunction with the task processing strategy determination method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the task processing strategy determination methods in the above embodiments.

[0094] In an optional embodiment, in conjunction with the task processing strategy determination method in the above embodiments, this application embodiment can provide a computer program product to implement it. The instructions in the computer program product are executed by the processor of the electronic device, enabling the electronic device to implement any of the task processing strategy determination methods in the above embodiments.

[0095] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0096] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0097] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0098] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0099] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for determining a task processing strategy, characterized in that, include: Obtain target task data; The target task data is decomposed and broken down using a pre-defined hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks; A risk prediction subnetwork is used to perform risk prediction on the source data in the target task data corresponding to the candidate actions, and a risk label corresponding to each candidate action is obtained. The action reward is calculated for the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action; Based on each candidate action, risk label, and action value, a target action path is determined, and a task processing strategy is determined according to the target action path.

2. The method according to claim 1, characterized in that, The step of decomposing and breaking down the target task data using a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks includes: The target task data is analyzed to obtain multiple data features; The task hierarchy is divided based on the multiple data features using the preset hierarchical subnetwork to obtain multiple task levels; Complexity analysis is performed based on multiple task levels to obtain a growth degree threshold, which is used to characterize the complexity of the data features. The growth degree threshold is optimized to obtain the optimal generation length threshold, which is used to characterize the optimal complexity of the target task data decomposition. The target task data is segmented according to the optimal generation length threshold to generate candidate actions corresponding to the multiple sub-tasks.

3. The method according to claim 2, characterized in that, Before optimizing the growth degree threshold to obtain the optimal generation length threshold, the method further includes: The target task data is subjected to mask sequence conversion to obtain a text mask sequence of the target task data, the text mask sequence including multiple segments of the sequence to be reconstructed; The text mask sequence is iteratively deduplicated to obtain the reconstructed text; During the iterative deduplication process, the quality of the reconstructed text is evaluated to obtain the evaluation index of the reconstructed text; If the evaluation index meets the preset optimal threshold, the sequence length corresponding to the current reconstructed text is determined as the optimal generation length threshold.

4. The method according to claim 1, characterized in that, Before performing risk prediction on the source data in the target task data corresponding to the candidate actions using a preset risk prediction sub-network to obtain the risk label corresponding to each candidate action, the method further includes: Obtain risk definition interaction data; The risk definition interaction data is parsed and features are extracted to obtain risk features and corresponding risk types; The risk characteristics and corresponding risk types are analyzed based on a preset risk prediction subnetwork to obtain risk determination conditions; The step of using a preset risk prediction sub-network to perform risk prediction on the source data in the target task data corresponding to the candidate actions, and obtaining a risk label corresponding to each candidate action, includes: The source data corresponding to the candidate action in the target task data is input into the preset risk prediction sub-network, and the preset risk sub-network extracts features from the source data to obtain risk features; The risk characteristics are matched with the risk determination conditions to obtain the matching results; Based on the matching results, the risk label corresponding to the candidate action is determined.

5. The method according to claim 1, characterized in that, Before calculating the action reward for the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action, the method further includes: Obtain a training set, which includes action descriptions, task data, and historical interaction data; The action description is input into an initial recurrent neural network model, which generates a dynamic reward action based on the action description and a reward function. Based on the dynamic reward action and the task data, a criticism signal is generated; Based on the criticism signal and the dynamic reward action, a preset voting mechanism is used to vote on the candidate reward functions to obtain the dynamic reward function of the preset meta-model; Based on the dynamic reward function, the preset meta-model is constructed.

6. The method according to claim 5, characterized in that, The step of calculating the action reward for the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action includes: A decision tree is generated based on the candidate actions; Based on the decision tree, a preset tree search algorithm is used to generate multiple candidate action paths; For each candidate action path, a reward action is performed using a preset environment model and a preset meta-model to obtain the reward value for each candidate path; The summation of the reward values ​​yields the action value of the decision tree.

7. The method according to claim 6, characterized in that, Before determining a target action path based on each candidate action, risk label, and action value, and then determining a task processing strategy based on the target action path, the method further includes: For each candidate action path, the path value of the current candidate action path is determined using a preset path value formula based on the action value of the candidate action path and the reward value of the current candidate action. The process of determining a target action path based on each candidate action, risk label, and action value, and then determining a task processing strategy based on the target action path, includes: Based on each candidate action, risk label, action value, and path value, a target action path is determined, and a task processing strategy is determined based on the target action path.

8. The method according to claim 7, characterized in that, The determination of the target action path based on each candidate action, risk label, action value, and path value includes: Risk simulation data is generated based on the candidate actions, risk tags, action values, and path values. Under each risk simulation data set, determine the simulated risk value for each candidate action path; The risk indicators for each candidate action path are determined using a pre-set comprehensive evaluation formula. The candidate action path corresponding to the highest risk indicator is determined as the target action path.

9. A task processing strategy determination device, characterized in that, The device includes: The acquisition module is used to acquire target task data; The decomposition module is used to decompose the target task data using a preset hierarchical sub-network to obtain candidate actions corresponding to multiple sub-tasks; The risk prediction module is used to perform risk prediction on the source data in the target task data corresponding to the candidate action using a preset risk prediction sub-network, and obtain the risk label corresponding to each candidate action. The calculation module is used to calculate the action reward of the candidate action using a preset meta-model to obtain the action value corresponding to the candidate action; The determination module is used to determine the target action path based on each of the candidate actions, risk tags, and action values, so as to determine the task processing strategy according to the target action path.

10. An electronic device, characterized in that, The device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the task processing strategy determination method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the task processing strategy determination method as described in any one of claims 1-8.

12. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device performs the task processing strategy determination method as described in any one of claims 1-8.

Images