Task processing method, computing device and computer readable storage medium

Abstract

The invention discloses a task processing method, computing equipment and a readable storage medium. The task processing method comprises the steps of decomposing a target task into a plurality of sub-tasks based on a deep learning model; evaluating the plurality of sub-tasks, and determining first evaluation values of the plurality of sub-tasks; if the first evaluation value is lower than a first preset value, the target task is decomposed again based on the deep learning model until the first evaluation values of multiple sub-tasks obtained through decomposition of the target task are not lower than the first preset value; and if the first evaluation value is not lower than the first preset value, executing a corresponding processing strategy on the subtask according to the processing difficulty of the subtask. Therefore, the target task is optimized and decomposed by evaluating the task decomposition process, and the optimal processing strategy is matched according to the processing difficulty of each sub-task after decomposition, so that the resource utilization efficiency and the automation and intelligence level of task decomposition are improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a task processing method, a computing device, and a computer-readable storage medium. Background Art

[0002] In recent years, generative large models (such as GPT) have made significant progress in the field of natural language processing, showing excellent versatility and generative capabilities, and can complete a variety of tasks such as text generation, language translation, code generation, etc. However, when faced with complex tasks, these models have exposed certain limitations, such as unreasonable task decomposition, low execution efficiency, or waste of resources. Summary of the invention

[0003] The purpose of this application is to provide a task processing method, a computing device and a computer-readable storage medium, which optimizes the decomposition of target tasks by evaluating the task decomposition process, and matches the optimal processing strategy according to the processing difficulty of each subtask after decomposition, thereby improving resource utilization efficiency and the level of automation and intelligence of task decomposition.

[0004] To achieve the above objectives: In a first aspect, an embodiment of the present application provides a task processing method, comprising the following steps: Decompose the target task into multiple subtasks based on the deep learning model; Evaluate the plurality of subtasks and determine first evaluation values ​​of the plurality of subtasks; If the first evaluation value is lower than the first preset value, decompose the target task again based on the deep learning model until the first evaluation values ​​of the multiple subtasks obtained by decomposing the target task are not lower than the first preset value; If the first evaluation value is not lower than the first preset value, a corresponding processing strategy is executed on the subtask according to the processing difficulty of the subtask.

[0005] In one embodiment, multiple subtasks are evaluated, including: Multiple subtasks are evaluated based on task decomposition evaluation criteria, where the task decomposition evaluation criteria include at least one of logical coherence, executability, rationality of subtask granularity, coverage completeness, input-output consistency, subtask relevance, and workload balance.

[0006] In one embodiment, executing a corresponding task processing strategy on a subtask according to the processing difficulty of the subtask includes: Evaluate the processing difficulty of the subtask and determine a second evaluation value; If the second evaluation value is lower than a second preset value, the subtask is processed based on multiple deep learning models.

[0007] In one embodiment, evaluating the difficulty of processing a subtask includes: The processing difficulty of the subtask is evaluated based on the task difficulty evaluation criteria, which include at least one of the task complexity, resource consumption estimation, data dependency, domain knowledge depth, and result uncertainty.

[0008] In one embodiment, the subtasks are processed based on multiple deep learning models, including: Process subtasks based on multiple deep learning models to obtain multiple deep learning model processing results; Evaluate the processing results of multiple deep learning models based on preset evaluation indicators; Based on the evaluation results, the processing results of multiple deep learning models are integrated to obtain the processing results of the subtasks.

[0009] In one embodiment, before processing subtasks based on multiple deep learning models, the task processing method further includes: Evaluate the input data of the subtask based on the preset evaluation indicators; If the input data of the subtask does not meet the input evaluation criteria, the input data of the subtask is adjusted until it meets the input evaluation criteria.

[0010] In one embodiment, the preset evaluation index includes at least one of description matching, structural integrity, input accuracy, content consistency, and semantic coherence.

[0011] In one embodiment, the evaluation result is an evaluation score of multiple deep learning model processing results, and the multiple deep learning model processing results are fused based on the evaluation result to obtain the processing result of the subtask, including: Extracting the highest scores corresponding to the respective evaluation indicators in the preset evaluation indicators from the evaluation scores of the processing results of the multiple deep learning models respectively as the model processing results corresponding to the evaluation indicators; The model processing results corresponding to each evaluation indicator are integrated to obtain the processing results of the subtask.

[0012] In a second aspect, an embodiment of the present application provides a computing device, specifically comprising: processor; a memory for storing processor-executable instructions; The processor is configured to execute instructions for executing the task processing method as described in the first aspect.

[0013] In a third aspect, an embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored. When the instructions in the computer-readable storage medium are executed by a processor of a computing device, the computing device is enabled to implement the task processing method described in the first aspect.

[0014] The task processing method, computing device and readable storage medium provided by the embodiment of the present application include: decomposing the target task into multiple subtasks based on a deep learning model; evaluating multiple subtasks and determining the first evaluation values ​​of the multiple subtasks; if the first evaluation value is lower than the first preset value, decomposing the target task again based on the deep learning model until the first evaluation values ​​of the multiple subtasks obtained by decomposing the target task are not lower than the first preset value; if the first evaluation value is not lower than the first preset value, executing the corresponding processing strategy for the subtask according to the processing difficulty of the subtask. In this way, the target task is decomposed by evaluating and optimizing the task decomposition process, and the optimal processing strategy is matched according to the processing difficulty of each subtask after decomposition, thereby improving the resource utilization efficiency and the automation and intelligence level of task decomposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a task processing method provided by an embodiment of the present invention.

[0016] Figure 2 A schematic diagram of a specific process flow in a task processing method provided in an embodiment of the present invention.

[0017] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0019] It should be noted that, in this article, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element. In addition, components, features, and elements with the same name in different embodiments of the present application may have the same meaning or different meanings, and their specific meanings need to be determined by their explanation in the specific embodiment or further combined with the context of the specific embodiment.

[0020] It should be understood that, although the terms first, second, third, etc. may be used to describe various information in this article, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this article, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein can be interpreted as "at the time of" or "when" or "in response to determination". Furthermore, as used in this article, the singular forms "one", "one" and "the" are intended to also include plural forms, unless there is an opposite indication in the context. It should be further understood that the terms "comprising", "including" indicate that there are described features, steps, operations, elements, components, projects, kinds, and / or groups, but do not exclude the existence, occurrence or addition of one or more other features, steps, operations, elements, components, projects, kinds, and / or groups. The terms "or" and "and / or" used herein are interpreted as inclusive, or mean any one or any combination. Thus, “A, B, or C” or “A, B and / or C” means “any of the following: A; B; C; A and B; A and C; B and C; A, B, and C.” An exception to this definition will occur only when a combination of elements, functions, steps, or operations are inherently mutually exclusive in some manner.

[0021] It should be understood that, although the various steps in the flowchart in the embodiment of the present application are displayed in sequence according to the indication of the arrows, these steps are not necessarily performed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and it can be performed in other orders. Moreover, at least a portion of the steps in the figure may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and their execution order is not necessarily performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0022] It should be noted that in this article, step codes such as S1 and S2 are used for the purpose of expressing the corresponding content more clearly and concisely, and do not constitute a substantial limitation on the order. When implementing the step, those skilled in the art may execute S2 first and then S1, etc., but these should all be within the scope of protection of this application.

[0023] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0024] In the subsequent description, the suffixes such as "module", "component" or "unit" used to represent elements are only used to facilitate the description of the present application, and have no specific meanings. Therefore, "module", "component" or "unit" can be used in a mixed manner.

[0025] Most existing task decomposition methods are static and fail to dynamically adjust according to the real-time status of subtasks. At the same time, the quality of input data and output integration strategy are crucial to the performance of tasks, but traditional methods often ignore the evaluation and optimization of input integrity and accuracy, and lack the ability to flexibly adjust when faced with input data that does not meet expected requirements, which affects the accuracy and consistency of output results.

[0026] In addition, for the fusion of different answers in complex tasks, most existing methods use a single algorithm, which cannot fully utilize the advantages of each answer, limiting the potential of generative models in handling complex tasks. Therefore, in order to better solve complex tasks, it is necessary to develop a more intelligent task decomposition mechanism, optimize the input data processing process, and improve the output fusion strategy to enhance the ability of large generative models to cope with complex tasks.

[0027] See also Figure 1 In order to cope with the above challenges, an embodiment of the present application provides a task processing method. In this embodiment, the task processing method is applied to a computing device as an example. The task processing method provided by this embodiment includes: Step S1: Decompose the target task into multiple subtasks based on the deep learning model.

[0028] Among them, the target task is composed of multiple subtasks with dependencies and logic. By decomposing the target into multiple subtasks, the processing difficulty of the target task can be reduced and the execution efficiency can be improved. When decomposing the target task, the target task is decomposed into multiple subtasks with logical associations by calling a generative large model based on deep learning. You can use a self-developed deep learning model or call an existing model, such as the GPT series (Generative Pre-trained Transformer), BERT (Bidirectional Encoder Representations from Transformers), language model Qwen, etc.

[0029] Step S2: Evaluate multiple subtasks and determine first evaluation values ​​of the multiple subtasks.

[0030] Among them, multiple subtasks are evaluated through task decomposition evaluation criteria, and the evaluation content includes but is not limited to the content of the subtasks and the relationship between the subtasks. The decomposition results of the target task are optimized through evaluation, so that complex target tasks can be scientifically and reasonably decomposed into multiple subtasks, improving the efficiency and accuracy of target task execution.

[0031] In one embodiment, multiple subtasks are evaluated, including: Multiple subtasks are evaluated based on task decomposition evaluation criteria, where the task decomposition evaluation criteria include at least one of logical coherence, executability, rationality of subtask granularity, coverage completeness, input-output consistency, subtask relevance, and workload balance.

[0032] Optionally, the task decomposition evaluation criteria are used to evaluate the rationality and effectiveness of the generative large model in decomposing complex tasks into subtasks. By setting scientific evaluation indicators, it is ensured that the decomposed subtasks meet high standards in terms of logical coherence, executability and completeness, thereby laying a solid foundation for the smooth execution of subsequent tasks. The task decomposition evaluation criteria include multiple evaluation indicators, each of which is only for a specific aspect of the task decomposition performance, that is, it is used to evaluate a subtask or a certain characteristic between subtasks to ensure the interpretability and operability of the evaluation indicators. The accuracy of the evaluation can be improved by decomposing complex evaluation tasks into clear and easy-to-execute independent sub-evaluation tasks through specific evaluation indicators of multiple different dimensions. On this basis, the scoring range of each evaluation indicator is unified and standardized to facilitate the comparison and comprehensive processing of different evaluation indicators.

[0033] The following is a description of each evaluation indicator in the task decomposition evaluation criteria.

[0034] Optionally, the logical coherence indicator is used to evaluate whether the logical relationship between subtasks is reasonable to ensure that they meet the overall task objectives. This indicator checks whether there are clear dependencies between subtasks and whether their sequence follows a reasonable logical flow to avoid logical conflicts or inconsistencies.

[0035] The evaluation content of the logical coherence index includes whether the dependencies between subtasks are clear, for example, the completion of one subtask is the prerequisite for the start of another subtask, and whether the processing order of each subtask conforms to the logical process, for example, data collection is carried out before data analysis.

[0036] The scoring range of logical coherence is 0~1, where 1 means that the logic between the subtasks after the target task is decomposed is completely coherent, and no logical loopholes or conflicts are found; 0 means that there are obvious logical problems, which seriously affect the coherence and executability of the task.

[0037] Optionally, the executability metric is used to evaluate whether each subtask has clear inputs and outputs to ensure that they can actually be performed. This metric checks whether the description of the subtask is clear and specific enough, and whether there are clear input requirements and expected outputs to ensure the actual feasibility of the task.

[0038] The evaluation of the executable indicators includes: checking the clarity and specificity of the subtask descriptions to confirm whether they provide sufficient details to guide execution; and confirming whether the input requirements and expected outputs of each subtask are clear, including the required resources, tools, data formats, etc.

[0039] The score range of the executability index is 0~1, where 1 means it is fully executable, meaning that all subtasks have clear operating instructions and expected results; 0 means there are obvious execution barriers, making the task difficult to implement.

[0040] Optionally, the subtask granularity rationality index is used to evaluate whether the subtask decomposition is appropriate to avoid being too detailed or too coarse. This index ensures that the task decomposition will neither increase unnecessary complexity and management burden due to being too detailed, nor affect the efficiency and quality of task execution due to being too coarse.

[0041] The evaluation of the rationality of subtask granularity includes comparing the number of subtasks with the overall complexity of the task to determine whether the decomposition is reasonable; and judging the complexity of individual subtasks to ensure that they are neither too simple nor too complex to be completed independently.

[0042] The scoring range of the rationality of subtask granularity is 0~1, where 1 means that the subtask granularity is completely appropriate and can effectively support task execution; 0 means that the granularity is too fine or too coarse, which has a negative impact on task execution.

[0043] Optionally, the coverage completeness indicator is used to evaluate whether the decomposed subtasks fully cover all requirements and objectives of the original task, avoid missing task steps, and ensure the completeness of task completion.

[0044] The evaluation of the coverage completeness indicator includes listing all the requirements of the original task and comparing them with the decomposed subtasks to check whether there are any key tasks or steps that have not been decomposed. The scoring range of the coverage completeness indicator is 0~1, where 1 means complete coverage.

[0045] Optionally, the input-output consistency index is used to evaluate whether the input and output between subtasks match and whether the information transmission is smooth, so as to ensure smooth connection of subtasks, avoid information discontinuity or repeated processing, and improve task processing efficiency.

[0046] The evaluation of the input-output consistency index includes checking whether the output of the previous subtask meets the input requirements of the next subtask, and confirming the accuracy of data and information transferred between subtasks. The input-output consistency index is scored from 0 to 1, where 1 means complete consistency.

[0047] Optionally, a subtask relevance indicator is used to assess whether the subtask directly contributes to the completion of the overall task goal, avoiding irrelevant or secondary tasks and ensuring that each subtask actually contributes to the completion of the overall task.

[0048] The evaluation of the subtask relevance index includes judging the relevance of the subtask goal to the overall task, and identifying whether there are subtasks that are not related to the task goal. The subtask relevance index is scored from 0 to 1, where 1 indicates a high relevance.

[0049] Optionally, the workload balance index is used to evaluate whether the complexity between subtasks is balanced, to avoid some subtasks being too complex or simple, to ensure reasonable task allocation, to avoid uneven resource allocation, to ensure efficient task execution and reasonable resource utilization.

[0050] The workload balance indicator is evaluated by comparing the expected workload and difficulty of each subtask, and analyzing and confirming whether there is a significant complexity imbalance between the subtasks. The workload balance indicator is scored on a scale of 0 to 1, where 1 indicates a good complexity balance.

[0051] For example, suppose the target task is to develop a new mobile application. This task can be broken down into the following subtasks: Requirements analysis: Communicate with customers to determine the functional requirements and user experience requirements of the application.

[0052] Deployment and launch: Release the final version of the app to the app store for users to download.

[0053] Design Prototyping: Based on the results of the requirements analysis, create a preliminary design and user interface prototype of the application.

[0054] Technology selection: Choose the appropriate technology stack and tools to implement the application's functionality.

[0055] Coding implementation: Write code based on design and technology selection to implement application functions.

[0056] Testing and Optimization: Perform comprehensive testing on the app and make necessary optimization adjustments based on feedback.

[0057] For the logical coherence index, the above four subtasks have clear logical dependencies (such as design prototype and technology selection), but there are logical conflicts in two subtasks (for example, deployment and demand analysis should not be directly related). Therefore, the logical coherence score = 4 / 6 = 0.667.

[0058] For the executable index, we evaluate whether each subtask has clear input and output. We found that the input and output of the five subtasks (such as demand analysis, design prototype, technology selection, coding implementation, testing and optimization) are clear and executable, while some details of deployment are too abstract and lack clear operation guidelines. Therefore, the executable score = 5 / 6 = 0.833.

[0059] As for the subtask granularity rationality index, among the six subtasks, it was found that the two tasks of "demand analysis" and "design prototype" were relatively simple and could be merged or further refined; while the workload and complexity of "coding implementation" were much higher than other tasks and seemed too complicated; the remaining three tasks (technology selection, testing and optimization, and deployment and launch) had appropriate granularity. Therefore, the subtask granularity rationality score = 3 / 6 = 0.5.

[0060] For the coverage completeness indicator, the original task has 10 key goals, including user experience design, performance optimization, safety measures, etc. After inspection, it is found that the decomposed subtasks cover 9 goals and miss a key step about user training. Therefore, the coverage completeness score = 9 / 10 = 0.9.

[0061] For the input-output consistency indicator, we checked the input-output matching between each pair of subtasks and found that the information transmission between the four pairs of subtasks (such as design prototype and technology selection) was smooth, and only the deployment and the previous demand analysis phase had information gaps. Therefore, the input-output consistency score = 4 / 5 = 0.8.

[0062] For the subtask relevance indicator, we analyze whether each subtask directly contributes to the overall goal. We find that although the subtask of "deployment and launch" is important, it is only related to the secondary goal in the current decomposition. The remaining five tasks are closely related to the core goal of application development. Therefore, the subtask relevance score = 5 / 6 = 0.833.

[0063] For the workload balance indicator, we evaluated the complexity distribution of the subtasks and found that the "coding implementation" subtask was too complex and took up a lot of resources; "requirements analysis" and "design prototype" were relatively simple; the remaining three tasks were balanced in complexity. Therefore, the workload balance score = 3 / 6 = 0.5.

[0064] Based on the above indicator scores (assuming that each indicator has equal weight), the comprehensive score is calculated as follows: The first evaluation value = (4 / 6+5 / 6+3 / 6+9 / 10+4 / 5+5 / 6+3 / 6) / 7 = 0.719.

[0065] Of course, in other embodiments, different weights may be set according to the indicators that need to be evaluated.

[0066] Step S3: If the first evaluation value is lower than the first preset value, the target task is decomposed again based on the deep learning model until the first evaluation values ​​of the multiple subtasks obtained by decomposing the target task are not lower than the first preset value.

[0067] In the above case, assuming that the first preset value is 0.8, that is, the first evaluation value is lower than the first preset value, the target task is decomposed again based on the deep learning model, and the multiple subtasks obtained by the decomposition are evaluated again until the first evaluation values ​​of the multiple subtasks are not lower than the first preset value, so as to improve the rationality and scientificity of the task decomposition.

[0068] Step S4: If the first evaluation value is not lower than the first preset value, a corresponding processing strategy is executed on the subtask according to the processing difficulty of the subtask.

[0069] Among them, if the first evaluation value is not lower than the first preset value, it indicates that the target task decomposition is completed, the decomposition process meets the task decomposition evaluation criteria, and the multiple subtasks obtained by decomposition can continue to be processed. The dynamic evaluation mechanism of task difficulty enables the system to adjust the strategy according to the real-time status of the task, thereby improving the flexibility of complex task processing.

[0070] In one embodiment, executing a corresponding task processing strategy on a subtask according to the processing difficulty of the subtask includes: Evaluate the processing difficulty of the subtask and determine a second evaluation value; If the second evaluation value is lower than a second preset value, the subtask is processed based on multiple deep learning models.

[0071] Optionally, the processing difficulty of the subtask is first evaluated to determine a quantitative second evaluation value. If the second evaluation value is lower than the second preset value, it indicates that the subtask is relatively simple and will be directly processed in a conventional manner, such as using a single model to output the processing result. On the contrary, if the second evaluation value is higher than or equal to the second preset value, multiple processing models are called to cope with more complex task requirements. The second preset value can be determined based on historical data and dynamically adjusted during the evaluation process. In this way, not only can computing resources be reasonably allocated according to the actual difficulty of the task, but also the efficiency and accuracy of task processing can be significantly improved, ensuring that each subtask can be processed in the most optimized way, thereby improving the overall completion quality and success rate of complex tasks.

[0072] In one embodiment, evaluating the difficulty of processing a subtask includes: The processing difficulty of the subtask is evaluated based on the task difficulty evaluation criteria, which include at least one of the task complexity, resource consumption estimation, data dependency, domain knowledge depth, and result uncertainty.

[0073] Optionally, the task difficulty evaluation standard is used to comprehensively evaluate the difficulty of the generative large model in handling complex tasks, so as to adjust the execution strategy, reasonably allocate resources and optimize the task process. Through scientific indicators and systematic evaluation methods, this standard can help the model accurately identify the complexity, high-risk links and potential execution bottlenecks in the task, thereby significantly improving the efficiency and success rate of task completion.

[0074] The following is a description of each evaluation indicator in the task difficulty evaluation standard.

[0075] Optionally, the task complexity metric is used to evaluate the logical and semantic complexity of the task that requires a large generative model to process. The evaluation content includes but is not limited to whether a multi-step answer is required, multi-level semantic understanding, and the complexity of the input content (such as text length, multi-domain fusion, etc.).

[0076] Optionally, resource consumption estimation indicators are used to evaluate the demand for computing resources (CPU / GPU time), memory usage, and processing time when the task is completed. The resource consumption ratio can be compared with the system resource upper limit to score.

[0077] Optionally, the data dependency index is used to evaluate the degree of dependence of a task on external data or the output of a preceding subtask, and the evaluation method is to analyze the number of data sources required by the task and the quality requirements of the dependent data.

[0078] Optionally, the domain knowledge depth indicator is used to evaluate the depth of the task's demand for expertise or background knowledge in a specific domain. The evaluation content includes the number of domains involved in the subtask and the knowledge modules or model capabilities that need to be called.

[0079] Optionally, the result uncertainty index is used to evaluate the uncertainty or risk of the task completion result, including the ambiguity of the task goal and the reliability of external data, and the score can be calculated by analyzing and weighting the unclear factors in the subtasks.

[0080] For example, suppose the subtask is to analyze the background, process and impact of a specific historical event and provide relevant data references. This task requires a comprehensive and multi-angle discussion of the specific historical event, involving multi-domain knowledge and data references.

[0081] For task complexity indicators, the multi-step ratio and semantic complexity ratio can be included. The subtask requires the model to generate answers to three main parts (background, process, impact), which is a multi-step task. Assuming the highest complexity is set to 5 steps, the actual number of steps required is 3, and the multi-step complexity ratio = 3 / 5 = 0.6. The task involves a high-level semantic understanding, which is set to level 4, the highest level is level 5, and the semantic complexity score = 4 / 5 = 0.8. Taking the average of the above two scores, assuming that the weights of the two are equal, the task complexity score = (0.6+0.8) / 2=1.4 / 2=0.7 can be obtained.

[0082] Resource consumption estimation indicators can include input length ratio and processing time ratio. Assuming that the task input text length is 1500 words and the system's maximum processing capacity is 2000 words, the input length ratio = 1500 / 2000 = 0.75. It is estimated to take 60 seconds, and the maximum processing time allowed by the system is 100 seconds, then the processing time ratio = 60 / 100 = 0.6. Taking the average of the above two scores, assuming that the weights of the two are equal, the resource consumption estimation score can be obtained = (0.75+0.6) / 2=1.35 / 2=0.675.

[0083] For the data dependency index, assume that the current subtask needs to reference 5 external data sources, the maximum number of data sources set by the system is 10, and there is no previous subtask dependency. Then the data dependency score = 5 / 10 = 0.5.

[0084] For the domain knowledge depth indicator, assuming that the subtasks involve multiple fields such as history, politics, and economics, 3 knowledge modules need to be enabled, and the maximum number of knowledge modules available in the system is 5, then the knowledge domain depth score = 3 / 5 = 0.6.

[0085] For the result uncertainty indicator, the subtask description is clear, the goal is clear, and the external data source is highly reliable, but there may be some missing data. The estimated uncertainty is 20%, so the result uncertainty score = 20% = 20 / 100 = 0.2.

[0086] Summarize the scores of each indicator, assuming that each indicator has equal weight, and calculate the comprehensive difficulty score (i.e. the second evaluation value) as follows: The second evaluation value = (task complexity score + resource consumption score + data dependence score + domain knowledge depth score + result uncertainty score) / 5 = (0.7+0.675+0.5+0.6+0.2) / 5 = 0.535.

[0087] Assuming that the second preset value is 0.5, that is, the second evaluation value is higher than the second preset value, the subtask is processed based on multiple deep learning models.

[0088] In one embodiment, before processing subtasks based on multiple deep learning models, the task processing method further includes: Evaluate the input data of the subtask based on the preset evaluation indicators; If the input data of the subtask does not meet the input evaluation criteria, the input data of the subtask is adjusted until it meets the input evaluation criteria.

[0089] Optionally, before processing subtasks based on multiple deep learning models, the input data of the subtasks can be evaluated according to preset evaluation indicators (such as input judgment criteria). If the input data does not meet the input judgment criteria, the input data is adjusted until it meets the criteria. The pre-evaluation and adjustment mechanism ensures that the input data of each subtask has been optimized before entering the deep learning model for processing, thereby significantly improving the accuracy and efficiency of subsequent processing, reducing errors or deviations caused by input data quality issues, and ultimately improving the quality and reliability of the overall task completion. Through the input inspection and adjustment process, the accuracy and completeness of the input data quality are guaranteed, fundamentally improving the reliability and efficiency of task execution.

[0090] In one embodiment, the subtasks are processed based on multiple deep learning models, including: Process subtasks based on multiple deep learning models to obtain multiple deep learning model processing results; Evaluate the processing results of multiple deep learning models based on preset evaluation indicators; Based on the evaluation results, the processing results of multiple deep learning models are integrated to obtain the processing results of the subtasks.

[0091] Optionally, multiple deep learning models are first used to process the same subtask in parallel to generate multiple different processing results. Next, the output of each model is independently evaluated according to preset evaluation indicators (such as description matching, input structure integrity, input accuracy, content consistency, semantic coherence, etc.) to ensure the quality and reliability of each result. Finally, based on these evaluation results, an intelligent fusion algorithm is used to integrate the outputs of multiple models to form the final subtask processing results. In this way, not only can the advantages of different models be fully utilized to improve the accuracy and robustness of the processing results, but also the possible limitations of a single model can be effectively avoided through the evaluation and fusion mechanism, thereby significantly improving the efficiency and quality of the overall task execution.

[0092] In one embodiment, the preset evaluation index includes at least one of description matching, structural integrity, input accuracy, content consistency, and semantic coherence.

[0093] Optionally, the description matching index is used to evaluate whether the input meets the requirements of the task description, including whether the data content is relevant to the target task. The evaluation method can use semantic similarity to calculate the matching degree between the input content and the task description, and determine whether the input covers the key topics or information points required by the subtask.

[0094] Optionally, the structural integrity indicator is used to evaluate whether the structure of the input data is complete and whether it contains all fields or parts of the content required for task execution. The evaluation method includes but is not limited to checking whether the input contains the fields and formats required for each subtask, and whether the input has missing or redundant information.

[0095] Optionally, the input accuracy indicator is used to evaluate whether the input data content is correct and has no obvious logical or factual errors. The evaluation method includes comparing the degree of fit between the input content and the expected rules, knowledge base or context, and checking whether the input content contains obvious contradictions or errors.

[0096] Optionally, the content consistency index is used to evaluate whether the input content is logically self-consistent and whether it meets the dependency requirements of the subtask on the input data. The evaluation method includes analyzing whether the input content has internal contradictions or logical conflicts that are inconsistent with the task context, and checking whether the input data meets the dependency chain of the subtask.

[0097] Optionally, a semantic coherence metric is used to evaluate whether the expression of the input content is semantically fluent and coherent, and does not cause confusion or semantic jumps in the generated model. The evaluation method includes checking whether there are semantic contradictions or jumps in the input content, and measuring the semantic consistency score of the input paragraph through the natural language model.

[0098] For example, assume that the task is to generate a complete financial report based on the financial data input by the user. The input is required to provide necessary financial indicators (such as revenue, cost, profit, etc.) and meet the task objectives.

[0099] For the description matching index, the ratio of the number of matching information points to the total number of information points in the task description is calculated. The user input provides 5 financial indicators (revenue, cost, profit, tax rate and cash flow), and the task requires 6 indicators (including additional return on investment). The description matching score = 5 / 6 = 0.833.

[0100] For the structural integrity index, the ratio of the number of required fields that actually exist to the total number of required fields is calculated. The table provided by the user contains 8 fields, of which 6 fields are required fields for the task and the remaining 2 fields are redundant fields. The structural integrity score = 6 / 8 = 0.75.

[0101] For the input accuracy metric, calculate the ratio of the number of correct data points to the total number of data points. Of the 6 required fields, 5 fields have correct data, and the data in 1 field (profit) contains logical errors (such as the result of revenue minus cost does not match), the input accuracy score = 5 / 6 = 0.833.

[0102] For the content consistency index, the ratio of the number of consistent data points to the total number of data points is calculated. Among the five numerical dependencies involved in the input (such as revenue-cost=profit), four are logically consistent and one is wrong. The content consistency score = 4 / 5 = 0.8.

[0103] For the semantic coherence index, the ratio of the number of semantically fluent paragraphs to the total number of paragraphs is calculated. The input text contains 3 main paragraphs, 2 of which are semantically fluent, and the other paragraph has semantic jumps that make it difficult to understand. The semantic coherence score = 2 / 3 = 0.667.

[0104] Take the average of the scores of each indicator and calculate the comprehensive input quality score = (description matching score + structural completeness score + input accuracy score + content consistency score + semantic coherence score) / 5=(0.833+0.75+0.833+0.8+0.667) / 5=0.7766.

[0105] In one embodiment, the evaluation result is an evaluation score of multiple deep learning model processing results, and the multiple deep learning model processing results are fused based on the evaluation result to obtain the processing result of the subtask, including: Extracting the highest scores corresponding to the respective evaluation indicators in the preset evaluation indicators from the evaluation scores of the processing results of the multiple deep learning models respectively as the model processing results corresponding to the evaluation indicators; The model processing results corresponding to each evaluation indicator are integrated to obtain the processing results of the subtask.

[0106] Optionally, when processing subtasks, the processing results of multiple deep learning models are first scored based on preset evaluation indicators to generate evaluation scores for each model. Subsequently, the highest scores corresponding to each evaluation indicator in the preset evaluation indicators are extracted as the optimal model processing results. Finally, the model processing results corresponding to these highest scores are fused to form the final subtask processing results. This ensures that the optimal performance under each evaluation indicator is integrated into the final answer, thereby significantly improving the comprehensiveness, accuracy and consistency of the subtask processing results, and effectively avoiding the deviations and limitations that may exist in a single model.

[0107] Optionally, this embodiment uses a multi-model-based answer fusion mechanism to obtain the processing results of the subtasks, ensuring that the fused answers draw on the advantages of each generated answer, avoiding the deviation and limitations of a single answer, and improving the comprehensiveness, accuracy and consistency of the final output. The evaluation of the answers generated by each model can use the same indicators as the input data evaluation, or different indicators. Taking the same input judgment criteria as an example, each generated answer is evaluated to obtain scores for indicators such as description matching, structural integrity, input accuracy, content consistency, and semantic coherence. For each indicator, the answer content with the highest score is selected from all answers. For example, description matching: answer A has the highest score, and its related content is selected. Structural integrity: answer B has the highest score, and its related content is selected. Input accuracy: answer C has the highest score, and its related content is selected. Then, according to the selected high-scoring indicators, the corresponding content is extracted. For example: extract the part related to the task description match from answer A. Extract the part with the best structural integrity from answer B. Integrate the extracted high-scoring content in a logical order and semantic coherence to avoid semantic jumps or logical conflicts. Recheck the logical consistency and coherence of the integrated content, output the final integrated answer, and retain the source of each part of the content and the corresponding high-score evaluation indicator score.

[0108] For example, suppose the subtask is to explain the background, technical implementation and application impact of a certain scientific breakthrough based on a deep learning model. Three answers A, B and C are generated based on different models. Among them, answer A explains the background in detail, but the technical implementation part is omitted. The technical implementation part of answer B is detailed, but the semantic expression is not smooth enough. The overall content of answer C is comprehensive, but the background part lacks details. The input evaluation criteria are used to evaluate each answer, and the evaluation results are shown in Table 1: Table 1

[0109] Next, select the high-scoring content corresponding to each indicator and merge them. For description matching, select the content related to answer A (the highest score is 0.8). For input structure completeness, select the content related to answer B (the highest score is 0.9). For input accuracy, select the content related to answer C (the highest score is 0.9). For content consistency, select the content related to answer A (the highest score is 0.9). For semantic coherence, select the content related to answer A (the highest score is 0.8). Then, by integrating the high-scoring content according to the logical order of the task, the final answer may be as follows: Background: Based on answer A, provide detailed background information about the technological breakthrough, making sure it matches the task description and is logically consistent.

[0110] Technical implementation: Based on answer B, add technical details and complete the content structure.

[0111] Application impact: Based on answer C, ensure accuracy and logic, and expand on the application impact.

[0112] Through the above design, the answer fusion mechanism can not only effectively utilize the advantages of multiple answers generated by the generative large model, but also ensure that the final output answers reach a high level in all aspects, thereby improving the quality and efficiency of task completion. This systematic fusion method provides strong support for handling complex tasks and ensures that the generated answers are both comprehensive and accurate.

[0113] The task decomposition method of the embodiment of the present application can be applied to various complex task execution scenarios, including text generation, code generation, and logical reasoning, and has a wide range of practical value. Figure 2 , is a specific embodiment of the task processing method provided in the embodiment of the present application. Figure 2 As shown, a decomposition operation is performed on the target task based on a preset deep learning model, and the decomposition result is evaluated. It is determined whether the decomposition evaluation result is lower than the first threshold. If so, the target task is decomposed again; if not, the subtasks obtained by decomposition are dynamically evaluated in turn. It is determined whether the difficulty evaluation result of the current subtask exceeds the second threshold. If so, the input data of the subtask is optimized and the input adjustment strategy is performed, and then multiple answers are generated for the subtask based on the multi-model parameters, and the advantages of each answer are evaluated and analyzed and integrated to obtain the processing result of the subtask. If not, a conventional single model is used to process the subtask. After the processing is completed, the next subtask is processed until all subtasks obtained by decomposing the target task are processed according to the above process.

[0114] In summary, in the task processing method provided by the above embodiment, the target task is decomposed into multiple subtasks based on the deep learning model; multiple subtasks are evaluated to determine the first evaluation values ​​of the multiple subtasks; if the first evaluation value is lower than the first preset value, the target task is decomposed again based on the deep learning model until the first evaluation values ​​of the multiple subtasks obtained by decomposing the target task are not lower than the first preset value; if the first evaluation value is not lower than the first preset value, the corresponding processing strategy is executed on the subtask according to the processing difficulty of the subtask. In this way, the target task is decomposed by evaluating and optimizing the task decomposition process, and the optimal processing strategy is matched according to the processing difficulty of each subtask after decomposition, thereby improving the resource utilization efficiency and the automation and intelligence level of task decomposition.

[0115] Based on the same inventive concept as the above embodiments, an embodiment of the present invention provides a computing device, such as Figure 3 As shown, the computing device includes: a processor 410 and a memory 411 storing a computer program; wherein, Figure 3The processor 410 shown in the figure is not used to indicate that the number of the processor 410 is one, but is only used to indicate the positional relationship of the processor 410 relative to other devices. In actual applications, the number of the processor 410 may be one or more; similarly, Figure 3 The memory 411 shown in the figure has the same meaning, that is, it is only used to refer to the position relationship of the memory 411 relative to other devices. In practical applications, the number of memories 411 can be one or more. When the processor 410 runs the computer program, the above task processing method is implemented.

[0116] The computing device may also include: at least one network interface 412. The various components in the computing device are coupled together via a bus system 413. It is understood that the bus system 413 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 413 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 413 is not used in the example above. Figure 3 Various buses are labeled as bus system 413.

[0117] The memory 411 may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory may be a disk memory or a tape memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM, SyncLink Dynamic Random Access Memory), and direct RAMbus random access memory (DRRAM, Direct Rambus Random Access Memory).The memory 411 described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memories.

[0118] The memory 411 in the embodiment of the present invention is used to store various types of data to support the operation of the computing device. Examples of these data include: any computer program used to operate on the computing device, such as an operating system and an application; contact data; phone book data; messages; pictures; videos, etc. Among them, the operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic services and process hardware-based tasks. The application program may include various applications, such as a media player, a browser, etc., which are used to implement various application services. Here, the program that implements the method of the embodiment of the present invention may be included in the application program.

[0119] Based on the same inventive concept as the above-mentioned embodiment, this embodiment further provides a computer-readable storage medium, in which a computer program is stored. The computer-readable storage medium may be a ferromagnetic random access memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); it may also be various devices including one or any combination of the above memories, such as a mobile phone, a computer, a tablet device, a personal digital assistant, etc. When the computer program stored in the computer-readable storage medium is executed by the processor, the task processing method applied to the above-mentioned computing device is implemented. For the specific steps implemented when the computer program is executed by the processor, please refer to Figure 1 The description of the illustrated embodiment will not be repeated here.

[0120] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0121] In this document, the terms "comprises," "comprising," or any other variations thereof, are intended to cover a non-exclusive inclusion of elements other than those listed and may also include additional elements not expressly listed.

[0122] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A task processing method, characterized in that: The following steps are involved: Decompose the target task into multiple subtasks based on the deep learning model; evaluating the plurality of subtasks and determining first evaluation values ​​of the plurality of subtasks; If the first evaluation value is lower than a first preset value, decomposing the target task again based on the deep learning model until the first evaluation values ​​of the plurality of subtasks obtained by decomposing the target task are not lower than the first preset value; If the first evaluation value is not lower than a first preset value, a corresponding processing strategy is executed on the subtask according to the processing difficulty of the subtask.

2. The method according to claim 1, characterized in that The evaluating the plurality of subtasks comprises: The multiple subtasks are evaluated based on task decomposition evaluation criteria, where the task decomposition evaluation criteria include at least one of logical coherence, executability, rationality of subtask granularity, coverage completeness, input-output consistency, subtask relevance, and workload balance.

3. The method according to claim 1, characterized in that The executing a corresponding task processing strategy on the subtask according to the processing difficulty of the subtask includes: Evaluate the processing difficulty of the subtask and determine a second evaluation value; If the second evaluation value is lower than a second preset value, the subtask is processed based on multiple deep learning models.

4. The method according to claim 3, characterized in that The step of evaluating the processing difficulty of the subtask includes: The processing difficulty of the subtask is evaluated based on a task difficulty evaluation standard, where the task difficulty evaluation standard includes at least one of task complexity, resource consumption estimation, data dependency, domain knowledge depth, and result uncertainty.

5. The method according to claim 3, characterized in that: The processing of the subtasks based on multiple deep learning models includes: Processing the subtasks based on multiple deep learning models to obtain multiple deep learning model processing results; Evaluate the processing results of the multiple deep learning models based on preset evaluation indicators respectively; Based on the evaluation results, the processing results of the multiple deep learning models are integrated to obtain the processing results of the subtasks.

6. The method according to claim 4, characterized in that Before processing the subtasks based on multiple deep learning models, the method further includes: Evaluate the input data of the subtask based on preset evaluation indicators; If the input data of the subtask does not meet the input evaluation criteria, the input data of the subtask is adjusted until it meets the input evaluation criteria.

7. The method according to claim 5 or 6, characterized in that: The preset evaluation index includes at least one of description matching, structural integrity, input accuracy, content consistency, and semantic coherence.

8. The method according to claim 7, characterized in that The evaluation result is an evaluation score of the processing results of the multiple deep learning models, and fusing the processing results of the multiple deep learning models based on the evaluation result to obtain the processing result of the subtask includes: Respectively extracting the highest score corresponding to each evaluation indicator in the preset evaluation indicators from the evaluation scores of the multiple deep learning model processing results as the model processing result corresponding to the evaluation indicator; The model processing results corresponding to each evaluation index are integrated to obtain the processing results of the subtasks.

9. A computing device, characterized in that include: A processor and a memory for storing executable instructions; wherein the processor is configured to execute the instructions to implement the task processing method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor, the task processing method according to any one of claims 1 to 8 is implemented.

Images