Task execution method and system

Abstract

The invention relates to the technical field of data processing, and provides a task execution method and system.The method comprises the steps that a figure portrait is constructed, the figure portrait comprises identity information and skill information, then a to-be-executed task is obtained, a task understanding module is called, the to-be-executed task is analyzed through the task understanding module, and the to-be-executed task is obtained; generating a target list, wherein the target list comprises task items; inputting the target list into a task evaluator so as to decompose the task item into sub-tasks; according to the subtasks, matching the figure portraits to generate recommendation resources; and obtaining a first execution result, wherein the first execution result is a result of executing the subtask by the user according to the recommended resource. According to the method, multi-level task decomposition is realized by constructing a figure portrait and combining a task understanding module and a task evaluator, recommendation resources are matched according to identity and skill information of a user, matching of sub-tasks and user capability is ensured, and the success rate of task execution is improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a task execution method and system. Background Art

[0002] With the continuous advancement of information technology, differences in knowledge levels and skills have emerged among individuals. During task execution, the differences in knowledge and skills among different individuals can be simply matched through task allocation models, but they focus on individual experience and core skills to allocate and assign tasks.

[0003] In some cases, the task allocation model can achieve task allocation, and it is based on simple skill matching or experience background, failing to fully consider the individual's understanding ability and adaptability in specific fields, resulting in low efficiency of task execution.

[0004] There are also intelligent assistance systems based on large-scale pre-trained models (Large Language Model, LLM) and machine learning technologies. These systems optimize task allocation by analyzing individual capabilities and experience, but lack solutions in multi-level task decomposition and personalized guidance, resulting in a low success rate in task execution. Summary of the invention

[0005] The present application provides a task execution method and system to solve the problem of low task execution success rate.

[0006] In a first aspect, the present application provides a task execution method, comprising: Constructing a character portrait, wherein the character portrait includes identity information and skill information; Acquire the tasks to be performed, and call the task understanding module to parse the tasks to be performed through the task understanding module to generate a target list, wherein the target list includes task items; Inputting the goal list into a task evaluator to decompose the task item into subtasks; According to the subtask, matching the character portrait to generate recommended resources; A first execution result is obtained, where the first execution result is a result of the user executing the subtask according to the recommended resource.

[0007] In some feasible embodiments, the obtaining of the tasks to be performed and calling the task understanding module to parse the tasks to be performed through the task understanding module to generate a target list includes: Calling the language model in the task understanding module; Constructing a first prompt item and a second prompt item, wherein the first prompt item is used to guide the language model to output a target list, and the second prompt item is used to guide the language model to output a target focus; Based on the first prompt item, parsing the task to be performed by using the language model to output a task item; Based on the second prompt item, parsing the task item through the language model to output a target focus; The goals are highlighted on the task items to generate a goal list.

[0008] In some feasible embodiments, inputting the target list into a task evaluator to decompose the task item into subtasks includes: Acquire decision information and weight information, wherein the weight information is the weight of the decision information; Calculate first complexity information according to the decision information and the weight information, where the first complexity information is the task complexity of the task item; Based on the first complexity information, calling a knowledge database to extract knowledge point information of the subtask through the knowledge database; Based on the first complexity information, generating a skill requirement matrix through the language model to output requirement information; Generate a subtask list through the first complexity information, knowledge point information and requirement information; The subtask list is decomposed into subtasks.

[0009] In some feasible embodiments, the recommended resource includes a first recommended target and / or a second recommended target; According to the subtask, matching the character portrait to generate recommended resources includes: Converting the knowledge point information into a knowledge point vector, and converting the demand information into a demand vector; Calculating a first similarity and a second similarity by cosine similarity to output a similarity point score, wherein the first similarity is the similarity of the requirement vector, and the second similarity is the similarity of the knowledge point vector; The knowledge point information is sorted according to the similarity score to output a first recommendation target, where the first recommendation target is the knowledge point information whose similarity score is sorted in a preset position.

[0010] In some feasible embodiments, matching the character portrait according to the subtask to generate recommended resources includes: Converting the character portrait into a user vector; Acquire a trained neural network model, wherein the trained neural network model is a neural network model obtained by optimizing target parameter training model parameters; Inputting the user vector, knowledge point vector and demand vector into the trained neural network model to output a prediction score; The knowledge point information is sorted according to the predicted score to output a second recommendation target, where the second recommendation target is the knowledge point information with the predicted score sorted in a preset position.

[0011] In some feasible embodiments, after obtaining the first execution result, the following steps are included: Acquiring operation log information, wherein the first execution result and the operation log information are displayed through a tree structure model; If the first execution result is successful, marking the tree structure model; In response to an instruction to obtain a task to be performed, if the task to be performed is extracted from the marked tree structure model, the language model of the task understanding module is controlled to output the tree structure model.

[0012] In some feasible embodiments, after obtaining the first execution result, the method further includes: If the first execution result is failure, obtaining feedback information, the feedback information including the task identifier, the failure reason and context information; inputting the subtask into a task evaluator to output second complexity information; Generate a refinement factor according to the second complexity information, wherein the refinement factor is used to determine the refinement degree of the subtask; A decomposition function is obtained to decompose the subtasks based on the refinement factor and the decomposition function.

[0013] In some feasible embodiments, after decomposing the subtasks, the method further includes: Obtaining a skill derivation function and a knowledge point derivation function, wherein the skill derivation function is used to determine a skill set, and the knowledge point derivation function is used to determine a knowledge point set; Based on the decomposed subtasks and context information, determining a skill set through the skill derivation function; Based on the decomposed subtasks, context information and failure reasons, determining a knowledge point set through the knowledge point derivation function; The skill requirements and knowledge point information of the decomposed subtasks are output through the skill set and knowledge point set.

[0014] In some feasible embodiments, the method further includes: Preset an atomic task level, where the atomic task level is a basic knowledge level; Obtain the decomposition level of the subtask, where the initial level of the subtask is 1; If the decomposition level is the atomic task level, or feedback information is obtained and a second execution result is output, and the feedback information is information that the language model of the task understanding module feedbacks that further decomposition is impossible, the second execution result is failure.

[0015] In a second aspect, the present application provides a task execution system, including: a data processing module, a task understanding module, a decomposition module, an analysis module, and a result acquisition module; The data processing module is used to construct a character portrait, which includes identity information and skill information; and obtain tasks to be performed; The task understanding module is used to parse the to-be-performed task and generate a target list, wherein the target list includes task items; The decomposition module is used to input the target list into the task evaluator to decompose the task item into subtasks; The analysis module is used to match the character portrait according to the subtask to generate recommended resources; The result acquisition module is used to acquire a first execution result, where the first execution result is a result of a user executing a subtask according to the recommended resource.

[0016] It can be seen from the above technical solutions that the present application provides a task execution method and system, the method comprising: constructing a character portrait, the character portrait including identity information and skill information, then obtaining the task to be executed, and calling the task understanding module to parse the task to be executed through the task understanding module to generate a target list, the target list including task items; inputting the target list into the task evaluator to decompose the task items into subtasks; matching the character portrait according to the subtasks to generate recommended resources; obtaining a first execution result, the first execution result being the result of the user executing the subtask according to the recommended resources. The method realizes the decomposition of multi-level tasks by constructing a character portrait, combining the task understanding module and the task evaluator, and matching the recommended resources according to the user's identity and skill information, ensuring that the subtasks match the user's capabilities, and improving the success rate of task execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0018] Figure 1 A flowchart of a task execution method provided in an embodiment of the present application; Figure 2 A schematic diagram of the task execution process provided in an embodiment of the present application; Figure 3 A schematic diagram of a decomposition process after the task execution provided in the embodiment of the present application; Figure 4 A schematic diagram of the first termination condition provided in the embodiment of the present application; Figure 5 A schematic diagram of a second termination condition provided in an embodiment of the present application; Figure 6 A schematic diagram of the task execution system structure provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The following embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following embodiments do not represent all embodiments consistent with the present application.

[0020] The specific process of task allocation in the task allocation model includes demand analysis, candidate assessment, skill matching, priority sorting, task assignment, and feedback and adjustment.

[0021] Demand analysis is used to determine the nature and requirements of the task and identify the skills, knowledge and experience level required for the task. Candidate assessment is to collect personal information of team members, including work experience, professional skills, past project performance, etc. Then, through skill matching, compare the task requirements with the skills and experience of the candidates, and screen suitable candidates through preset matching rules. According to the matching results, the candidates are sorted. The screened candidates are assigned to the corresponding tasks. At this time, the project manager will comprehensively consider factors such as the workload and time arrangement of each candidate to achieve the optimal task allocation. During the execution of the task, feedback from the task executors is collected to evaluate the completion and efficiency of the task. During the execution process, if problems are encountered, corresponding task and personnel adjustments may be made.

[0022] Although these models can achieve task allocation to a certain extent, they rely on individuals' existing experience and clear core skills, and fail to fully consider individuals' understanding ability and adaptability to specific fields. This skill matching method may lead to inefficiency in the task execution process. Secondly, the task allocation model is ineffective in multi-level task decomposition and personalized guidance. For example, in large projects, tasks often involve multiple links and levels, and a single skill matching may not meet actual needs. This leads to a lack of effective solutions in multi-level task decomposition and personalized guidance, resulting in a low success rate in the final task execution.

[0023] To solve the above problems, see Figure 1 , Figure 2 Some embodiments of the present application provide a task execution method, including: S100: Constructing character portraits.

[0024] A character portrait is a multi-dimensional description of a user, including identity information and skill information. Identity information refers to the user's information, such as basic information such as age, gender, major, and language; skill information refers to the user's skill-related information, such as technical ability, language ability, soft skills, and other characteristic information.

[0025] After obtaining the identity information and skill information, the information may be preprocessed. In some embodiments, the identity information and skill information are cleansed by first checking and deleting redundant data, and then unifying and standardizing the data format. For example, numerical data may be standardized to a distribution with a mean of 0 and a standard deviation of 1, and categorical data may be one-hot encoded.

[0026] After preprocessing the information, the identity information and skill information are stored in a relational database. In some embodiments, the data table structure includes a user basic information table, a dynamic capability feature table, and a historical task performance table. The user basic information table may include basic information such as ID, name, age, and language. The dynamic capability feature table includes feature information such as technical ability, language ability, soft skills, etc. extracted from the identity information and skill information. The historical task performance table includes records of users' historical task execution and scores.

[0027] It is understandable that when a user performs a task for the first time, the historical task performance table is not included.

[0028] In some embodiments, the input information is analyzed by the language model LLM, wherein the input information may include a variety of information, such as identity information, skill information, and other information. Feature information is extracted through an API or a language model interface, and the feature information includes language ability, technical ability, and soft skills. Language ability is extracted by analyzing the complexity and grammatical accuracy of language use, technical ability is obtained by identifying the tools, programming languages, or technology stacks mentioned by the user, and soft skills are identified through sentiment analysis. Communication ability, problem-solving ability, etc.

[0029] If the task to be executed is not the first time the task is executed, in some embodiments, the historical task performance can also be associated. If the task performed by the user in the past is successful, the skills corresponding to the historical task can be extracted, and the historical task performance can be associated with the user's initial portrait. Data association can be performed through tools such as SQL or SQLAlchemy in Python.

[0030] After data association, the extracted feature information is used to generate a dynamic capability portrait, i.e., a character portrait, which can be saved in the form of a json file in the form of: user ID->[language ability, technical ability, soft skills…].

[0031] Among them, LLM is a language model with a large number of parameters after large-scale data training. These models are based on deep learning technology and can perform a variety of natural language processing tasks, such as text generation, question-answering systems, machine translation, summary generation, etc. LLM can be a GLM (Generative Language Model) model. The GLM model combines pre-training and fine-tuning techniques to improve the performance of language models on various tasks. GLM adopts the Transformer architecture and is optimized on this basis, including multi-layer encoders and decoders, which can capture long dependencies in text and generate high-quality text output.

[0032] Since the character portraits of different users are different, in some embodiments, the character portraits are saved in a character database. When the user completes a task, the character portrait in the character database will also be updated. That is, the character database will be updated once each task is completed.

[0033] S200: Obtain tasks to be executed, and call a task understanding module to parse the tasks to be executed through the task understanding module to generate a target list.

[0034] A pending task is a specific work or activity that needs to be performed by the user. A pending task can be a request initiated by the user, a task automatically generated by the system, or a task input by another party that needs to be performed by the user. A pending task can include a variety of information, such as task description, task goal, priority, deadline, etc.

[0035] The task understanding module is used to parse and understand the module of the task to be executed, and convert the task into executable task description information and target list. The task understanding module includes LLM, which performs semantic analysis on the task to be executed, generates task description information first, and then generates target list.

[0036] The target list includes multiple task items. In some embodiments, the language model in the task understanding module is called to construct a first prompt item and a second prompt item. The first prompt item is used to guide the language model to output the target list, and the second prompt item is used to guide the language model to output the target focus. Based on the first prompt item, the task to be executed is parsed through the language model to output the task item. Based on the second prompt item, the task item is parsed through the language model to output the target focus, and the target focus is marked on the task item to generate a target list.

[0037] Among them, the first prompt item and the second prompt item are both prompt engineering. The prompt item is an instruction or information that guides the language model to output content. It can be a sentence, a question, a description, or structured data. The first prompt item and the second prompt item can improve the relevance and accuracy of the output.

[0038] After the task to be executed is input into the language model, the language model will output the task description information about the task to be executed. The task description information is a detailed description of the task, including the task goal, execution process, background, requirements, etc. Then, a target list is generated based on the task description information. The target list refers to a list of one or more task items generated by the task understanding module based on the task description information.

[0039] For example, the task to be performed is to develop a machine learning model and evaluate it. The first prompt is: The following is a task description. Please parse and generate a list of goals to be completed: \nTask description. The second prompt is {Task description}\nGenerate a list of goals and mark the key points of each goal.

[0040] For developing a machine learning model and evaluating it, after passing LLM, the output task description is to collect and prepare data sets, select and develop machine learning models, train models, evaluate models, and finally generate a final report. Through the first prompt, output a list of goals according to the task description. The goal list can be 1. Collect and prepare data sets; 2. Select and develop machine learning models; 3. Train models; 4. Evaluate models; 5. Generate a final report.

[0041] Then, through the second prompt item, the target list is parsed and the target focus is output. For collecting and preparing data sets, the target focus is data preprocessing and cleaning. For selecting and developing machine learning models, the target focus is supervised learning or unsupervised learning. For training models, the target focus is setting hyperparameters and optimizing. For evaluating models, the target focus is using evaluation indicators such as accuracy, recall, etc.

[0042] It is understandable that for a task item in a certain target list, there may be no target focus. For example, for generating a final report, there is no target focus.

[0043] S300: Inputting a target list into a task evaluator to decompose a task item into subtasks.

[0044] The task evaluator is a tool or algorithm for evaluating and refining task items, which can identify the different components in a complex task and reasonably decompose them into smaller, manageable subtasks. In this embodiment, the task evaluator is an AI agent that combines LLM and knowledge database to evaluate the complexity of task items and perform task decomposition.

[0045] To facilitate user execution, in some embodiments, decision information and weight information are obtained, and the decision information may include the number of knowledge points involved in the task, skill proficiency requirements, complexity of dependencies between subtasks, time or resource requirements, etc. The weight information is the weight of the decision information, and it can be understood that the weights of different demand information are adjustable, and the weights of different demand information can be adjusted according to actual needs.

[0046] The first complexity information is then calculated based on the decision information and the weight information. The first complexity information is the task complexity score of the task item. The complexity score can be calculated by the following formula: ; in, is the number of knowledge points involved in the task, for The weight of For skill proficiency requirements, for The weight of is the complexity of the dependencies between subtasks, for The weight of For time or resource requirements, for The weight of .

[0047] Based on the first complexity information, the knowledge database is called to extract the knowledge point information of the subtask through the knowledge database. The knowledge database is a database that stores a large number of knowledge points and related information. The knowledge point information can be extracted through the knowledge database. In order to extract more accurate knowledge point information, the knowledge database can also be updated regularly. Exemplarily, if the subtask is to collect and prepare a data set, data preprocessing and cleaning technology, Python programming, and Pandas library knowledge are required.

[0048] Based on the first complexity information, a skill requirement matrix is ​​generated through a language model to output the demand information. The skill requirement matrix is ​​a two-dimensional table used to represent the various skills required for a task and the corresponding degree of demand. It can express the skill combination required to complete the task and facilitate resource matching.

[0049] A subtask list is generated through the first complexity information, knowledge point information and requirement information, and then the subtask list is decomposed into subtasks and output as a subtask list, which includes subtasks and complexity scores of subtasks, skill requirement information and knowledge point information.

[0050] Exemplarily, the subtask list includes the subtask data collection and cleaning, which has a low complexity score, skill requirement information including Python programming (elementary), data analysis (intermediate), and knowledge point information including Pandas library and data cleaning methods; the subtask list also includes the subtask model development, whose complexity score includes machine learning (intermediate), Scikit-learn (intermediate), and knowledge point information includes model selection and hyperparameter adjustment; the subtask list also includes the subtask model evaluation, whose complexity score includes data analysis (advanced), machine learning (intermediate), and knowledge point information includes accuracy, recall rate, and confusion matrix; the subtask list also includes the subtask report generation, which has a low complexity score, skill requirement information including technical document writing (elementary), and knowledge point information including report format and data visualization.

[0051] S400: Matching character portraits according to subtasks to generate recommended resources.

[0052] The recommended resources include a first recommended target and / or a second recommended target, wherein the first recommended target is knowledge point information and the second recommended target is a link resource. The recommended resources can obtain knowledge point information and link resources respectively through two ways. When providing recommended resources to users, the first recommended target and the second recommended target can be provided at the same time, or a certain recommended target can be provided separately, which is not limited in this embodiment.

[0053] It is understandable that when a user performs a task and fails to perform the task, if the recommended resource is generated as one of the two, when generating the recommended resource again, the other recommended target can be selected, and the recommended target corresponding to the failed execution is marked as an incorrect recommended resource. Exemplarily, if the generated recommended resource is the second recommended target, and the user's execution result is a failure, the subtask will be decomposed again, and the task profile will be matched according to the decomposed subtasks to generate recommended resources. At this time, the recommended resource is the first recommended target, and the second recommended target is marked as an incorrect recommended resource. When other users generate recommended resources based on the same subtask, they will regenerate the recommended resources instead of directly providing the generated second recommended target.

[0054] For the first recommendation target, in some embodiments, the knowledge point information is converted into a knowledge point vector, and the demand information is converted into a demand vector, and then the first similarity and the second similarity are calculated by cosine similarity to output a similarity score, wherein the first similarity is the similarity of the demand vector, and the second similarity is the similarity of the knowledge point vector; according to the similarity score, the knowledge point information is sorted to output the first recommendation target, wherein the first recommendation target is the knowledge point information sorted in a preset number of places by similarity score.

[0055] Natural language processing technology, such as word embedding model, can be used to convert textual knowledge point information into numerical vectors. NLP technology can also be used to convert knowledge point information or demand information into vector form. The cosine similarity between the demand vector and the knowledge point vector can be calculated using the following formula: ; in, is the demand vector, is the knowledge point vector, and is the norm of the vector.

[0056] After sorting by similarity score, the most relevant knowledge points are selected from all candidate knowledge points according to their scores as recommended content. Through vectorization and similarity calculation, knowledge points that match task requirements can be accurately selected to improve the effectiveness and pertinence of recommendations. Considering the uniqueness of each character portrait, the recommendation results can better meet the needs of individual users.

[0057] For the second recommendation target, in some embodiments, the character portrait is converted into a user vector to obtain a trained neural network model; the user vector, the knowledge point vector and the demand vector are input into the trained neural network model to output a predicted score; according to the predicted score, the knowledge point information is sorted to output a second recommendation target, and the second recommendation target is the knowledge point information with a predicted score sorted in a preset number of digits.

[0058] Convert user portraits into vector form , see the following formula: ; in, is the user’s score on the i-th capability dimension, i=1, 2, …, n.

[0059] The trained neural network model is a neural network model obtained by optimizing the target parameter training model parameters. It is a deep learning model that has been trained to achieve a specific performance level and is used to predict the user's matching degree for different resources. The trained neural network model includes an input layer, one or more hidden layers, and an output layer, and each layer is connected by an activation function.

[0060] The value calculated by the trained neural network model based on the input user vector, knowledge point vector and demand vector is used to evaluate the matching degree between resources and users. The model architecture of the trained neural network model can be a neural collaborative filtering or a Transformer-based recommendation model.

[0061] The prediction score is calculated by the following formula: ; in, is the activation function, is the weight matrix, is the user vector, is the knowledge point vector, is the demand vector, is the bias term.

[0062] The activation function can be Sigmoid, ReLU (Rectified Linear Unit), Tanh, etc. The specific choice depends on the model architecture. The bias term is a constant term that allows the model to produce a non-zero output even when all input features are zero.

[0063] The prediction score is a value calculated by the model based on the input user vector, demand vector, and knowledge point vector, and is used to evaluate the matching degree between resources and users. The model not only considers the importance of each input feature, but also considers the ability to provide reasonable output even when some features are missing. The model can more accurately predict the matching degree between resources and users, thereby achieving more personalized recommendations.

[0064] After sorting by prediction score, the most relevant link resources are selected from all candidate link resources according to the prediction score, and the external resource links that are most likely to help users complete their tasks are provided to them. Link resources refer to external resources that can be accessed through links, such as web pages, videos, audio, etc. Link resources can be captured and organized from the Internet through search engines, crawlers and other technologies.

[0065] S500: Obtain a first execution result.

[0066] After generating the recommended resources, the user performs the subtask according to the recommended resources. During the process of the user performing the subtask, the user's execution results, i.e., the first execution results, are collected in real time through dynamic knowledge mapping. Dynamic knowledge mapping is to track the process of the user performing the subtask in real time and generate new knowledge units according to the user's successful operations. The first execution results include success and failure.

[0067] In some embodiments, operation log information and a first execution result are obtained, and an automatic evaluation function is used to verify whether the user's execution result is successful. If the first execution result is successful, the user's operation log information is converted into a knowledge unit. A knowledge unit refers to the smallest knowledge fragment extracted and structured from the user's operation log information, and the knowledge unit is stored in a knowledge database to update the knowledge database.

[0068] In the process of storing knowledge units, the structured storage of knowledge units requires the definition of data models, which may include subtasks. , Execution steps , Execution Results , knowledge point information , Recommended Resources The fields are stored in the following format: ; Among them, the first execution result and operation log information are displayed through a tree structure model. The tree structure model is a data structure that organizes information in a hierarchical manner. Each node can have zero or more child nodes, but only one parent node (except the root node). In this embodiment, according to the logical flow of the subtasks and the user's operation log information, the relevant information is organized into a tree structure, in which each node represents an operation or decision point.

[0069] The tree structure model can be expressed as follows: ; in, is the subtask obtained by decomposition.

[0070] If the first execution result is successful, the successful path is marked in the tree structure model, for example, by color change, icon addition, etc. In response to the instruction to obtain the task to be executed, if the task to be executed is extracted in the marked tree structure model, the language model is controlled to output the tree structure model. When other users execute this task again, they first check whether the marked tree structure model contains content that matches the new task. If it does, the language model can be directly triggered to generate the operation log information corresponding to the marked successful path, including subtask list and recommended resources and other information. By providing successful cases as templates, the user's execution time can be reduced and the task execution efficiency can be improved.

[0071] In some embodiments, if the first execution result is failure, feedback information is obtained, and the feedback information includes a task identifier, a failure reason, and context information, wherein the task identifier is an identifier of the subtask that was not successfully executed, the failure reason is a specific failure reason, for example, insufficient recommended resources, lack of knowledge points, etc., and the context information is the environment and conditions involved in the task execution process.

[0072] The feedback data can be expressed as:

[0073] in: A collection of user feedback. For the Unsuccessful tasks, and For the error reason, For context information.

[0074] In the first task failure or within the preset number of executions, if the execution result is a failure, the subtasks that failed to execute are decomposed. In some embodiments, the atomic task level can be preset. The atomic task level is the basic knowledge level, such as a formula. The atomic task level is the most fine-grained task level in the task decomposition process, representing the basic operations or knowledge points that cannot be further decomposed. Then obtain the decomposition level of the subtask. The initial level of the subtask is 1. Each time a larger task is decomposed into smaller subtasks, the decomposition level is increased by 1.

[0075] If the decomposition level is the atomic task level, or, feedback information is obtained, the second execution result is output, and the feedback information is the information that the language model feedback cannot be further decomposed, and the second execution result is failure. In other words, when the language model determines that a subtask cannot be decomposed any further, or the decomposition level reaches the atomic task level, the second execution result is output as failure, and the current task processing flow is terminated. The upper limit of the user's understanding ability and the lower limit of the task can be inferred and updated to the user profile.

[0076] If within the preset atomic task level, or if no information is received from the language model indicating that further decomposition cannot be continued, or if the result of continued decomposition by the language model is different from the result of the previous decomposition, the task identifier, failure reason and context information are obtained, the subtask is input into the task evaluator, the second complexity information is output, and a refinement factor is generated based on the second complexity information. The refinement factor is used to determine the degree of refinement of the subtask; the decomposition function is obtained to decompose the subtask based on the refinement factor and the decomposition function.

[0077] According to the failure cause and context information, the complexity score of the subtask is re-determined. Different from the first complexity information, the second complexity information is a new indicator reflecting the complexity of the subtask. The subtask is re-input into the task evaluator and calculated according to the failure cause and context information.

[0078] In order to improve the success rate of execution, by introducing refinement factors , decompose the subtasks into more basic tasks, where the refinement factor is a granularity adjustment factor, which is used to adjust the fineness of task decomposition. By adjusting the refinement factor, the granularity of the task decomposition can be changed, thereby improving the task execution efficiency.

[0079] And by decomposing the function , see the following formula to output the decomposed subtasks: ; in, For the reasons of failure, For subtasks.

[0080] For the decomposed subtasks, in some embodiments, a skill derivation function and a knowledge point derivation function are obtained, the skill derivation function is used to determine the skill set, and the knowledge point derivation function is used to determine the knowledge point set. Based on the decomposed subtasks and context information, the skill set is determined through the skill derivation function; based on the decomposed subtasks, context information and failure reasons, the knowledge point set is determined through the knowledge point derivation function; the skill requirements and knowledge point information of the decomposed subtasks are output through the skill set and the knowledge point set.

[0081] The skill set is a set of specific skills output by the skill derivation function, which represents all the capabilities required to complete the decomposed subtasks. The skill set is calculated by the following formula: ; in, is the context information, Derive functions for skill sets.

[0082] The knowledge point set is a set of specific knowledge points output by the knowledge point derivation function, which represents all the knowledge background required to complete the decomposed subtasks. The knowledge point set is calculated by the following formula: ; in, Derive functions for knowledge points.

[0083] Integrate skill sets and knowledge point sets to generate a detailed skill requirement list and knowledge point information for users to use when performing the decomposed subtasks.

[0084] See also Figure 3,Exemplarily, the task to be performed is obtained. The task is decomposed into subtask 1, subtask 2, and subtask 3. Each subtask has its knowledge requirements, execution process and ,related data links. The user performs subtask 1. If it succeeds, the user continues with subtask 2. The user performs subtask 2. If it succeeds, the user continues with subtask 3. The user performs subtask 3. If it fails, the task is further decomposed into more ,fine-grained subtasks 3.1, subtask 3.2, subtask 3.3, and subtask 3.4.

[0085] The user performs subtask 3.1. If successful, the user continues with subtask 3.2. The user performs subtask 3.2. If successful, the user continues with subtask 3.3. The user performs subtask 3.3. If successful, the user continues with subtask 3.4. When all subtasks are successfully executed, the task reaches the termination condition and ends. By gradually breaking down the task into smaller and more specific subtasks, each subtask can be successfully executed by the user, thereby improving the overall success rate of the task.

[0086] In this embodiment, see Figure 4 , when all subtasks are successfully executed, that is, when the execution results are all unsuccessful, the task ends. Figure 5 ,If a subtask reaches the atomic task level after multiple decompositions, the task is also defined as ,completed.

[0087] In some embodiments, the user's operation log information and execution results, updated character portraits and other data will all be saved in the knowledge database as knowledge units.

[0088] Based on the above task execution method, see Figure 6 ,Some embodiments of the present application also provide a task execution system, including: a data processing module, a task understanding module, a decomposition module, an analysis module, and a result acquisition module; The data processing module is used to construct a character portrait, which includes identity information and skill information; and obtain tasks to be performed; The task understanding module is used to parse the to-be-performed task and generate a target list, wherein the target list includes task items; The decomposition module is used to input the target list into the task evaluator to decompose the task item into subtasks; The analysis module is used to match the character portrait according to the subtask to generate recommended resources; The result acquisition module is used to acquire a first execution result, where the first execution result is a result of a user executing a subtask according to the recommended resource.

[0089] The effects of the above system embodiment when running can refer to the effects of the above method embodiment, which will not be described in detail here.

[0090] The present application provides a task execution method and system, the method comprising: constructing a character portrait, the character portrait comprising identity information and skill information, then obtaining a task to be executed, and calling a task understanding module to parse the task to be executed through the task understanding module to generate a target list, the target list comprising task items; inputting the target list into a task evaluator to decompose the task items into subtasks; matching the character portrait according to the subtasks to generate recommended resources; obtaining a first execution result, the first execution result being the result of the user executing the subtask according to the recommended resources. The method realizes the decomposition of multi-level tasks by constructing a character portrait, combining a task understanding module and a task evaluator, and matching recommended resources according to the user's identity and skill information, ensuring that the subtasks match the user's capabilities, and improving the success rate of task execution.

[0091] Similar parts between the embodiments provided in this application can be referenced to each other. The specific implementation methods provided above are only a few examples under the general concept of this application and do not constitute a limitation on the protection scope of this application. For those skilled in the art, any other implementation methods expanded based on the scheme of this application without creative work belong to the protection scope of this application.

Claims

1. A task execution method, characterized in that: include: Constructing a character portrait, wherein the character portrait includes identity information and skill information; Acquire the tasks to be performed, and call the task understanding module to parse the tasks to be performed through the task understanding module to generate a target list, wherein the target list includes task items; Inputting the goal list into a task evaluator to decompose the task item into subtasks; According to the subtask, matching the character portrait to generate recommended resources; A first execution result is obtained, where the first execution result is a result of the user executing the subtask according to the recommended resource.

2. The task execution method according to claim 1, characterized in that: The obtaining of the tasks to be executed and calling the task understanding module to parse the tasks to be executed through the task understanding module and generate a target list includes: Calling the language model in the task understanding module; Constructing a first prompt item and a second prompt item, wherein the first prompt item is used to guide the language model to output a target list, and the second prompt item is used to guide the language model to output a target focus; Based on the first prompt item, parsing the task to be performed by using the language model to output a task item; Based on the second prompt item, parsing the task item through the language model to output a target focus; The goals are highlighted on the task items to generate a goal list.

3. The task execution method according to claim 2, characterized in that: The step of inputting the target list into a task evaluator to decompose the task item into subtasks comprises: Acquire decision information and weight information, wherein the weight information is the weight of the decision information; Calculate first complexity information according to the decision information and the weight information, where the first complexity information is the task complexity of the task item; Based on the first complexity information, calling a knowledge database to extract knowledge point information of the subtask through the knowledge database; Based on the first complexity information, generating a skill requirement matrix through the language model to output requirement information; Generate a subtask list through the first complexity information, knowledge point information and requirement information; The subtask list is decomposed into subtasks.

4. The task execution method according to claim 3, characterized in that: The recommended resource includes a first recommended target and / or a second recommended target; According to the subtask, matching the character portrait to generate recommended resources includes: Converting the knowledge point information into a knowledge point vector, and converting the demand information into a demand vector; Calculating a first similarity and a second similarity by cosine similarity to output a similarity point score, wherein the first similarity is the similarity of the requirement vector, and the second similarity is the similarity of the knowledge point vector; The knowledge point information is sorted according to the similarity score to output a first recommendation target, where the first recommendation target is the knowledge point information whose similarity score is sorted in a preset position.

5. The task execution method according to claim 4, characterized in that: According to the subtask, matching the character portrait to generate recommended resources includes: Converting the character portrait into a user vector; Acquire a trained neural network model, wherein the trained neural network model is a neural network model obtained by optimizing target parameter training model parameters; Inputting the user vector, knowledge point vector and demand vector into the trained neural network model to output a prediction score; The knowledge point information is sorted according to the predicted score to output a second recommendation target, where the second recommendation target is the knowledge point information with the predicted score sorted in a preset position.

6. The task execution method according to claim 1, characterized in that: After obtaining the first execution result, the method includes: Acquiring operation log information, wherein the first execution result and the operation log information are displayed through a tree structure model; If the first execution result is successful, marking the tree structure model; In response to an instruction to obtain a task to be performed, if the task to be performed is extracted from the marked tree structure model, the language model of the task understanding module is controlled to output the tree structure model.

7. The task execution method according to claim 1, characterized in that: After obtaining the first execution result, the method further includes: If the first execution result is failure, obtaining feedback information, the feedback information including the task identifier, the failure reason and context information; inputting the subtask into a task evaluator to output second complexity information; Generate a refinement factor according to the second complexity information, wherein the refinement factor is used to determine the refinement degree of the subtask; A decomposition function is obtained to decompose the subtasks based on the refinement factor and the decomposition function.

8. The task execution method according to claim 7, characterized in that: After decomposing the subtasks, the following steps are further included: Obtaining a skill derivation function and a knowledge point derivation function, wherein the skill derivation function is used to determine a skill set, and the knowledge point derivation function is used to determine a knowledge point set; Based on the decomposed subtasks and context information, determining a skill set through the skill derivation function; Based on the decomposed subtasks, context information and failure reasons, determining a knowledge point set through the knowledge point derivation function; The skill requirements and knowledge point information of the decomposed subtasks are output through the skill set and knowledge point set.

9. The task execution method according to claim 7, characterized in that: The method further comprises: Preset an atomic task level, where the atomic task level is a basic knowledge level; Obtain the decomposition level of the subtask, where the initial level of the subtask is 1; If the decomposition level is the atomic task level, or feedback information is obtained and a second execution result is output, and the feedback information is information fed back by the language model of the task understanding module that further decomposition is impossible, the second execution result is failure.

10. A task execution system, characterized in that: include: Data processing module, task understanding module, decomposition module, analysis module and result acquisition module; The data processing module is used to construct a character portrait, which includes identity information and skill information; and obtain tasks to be performed; The task understanding module is used to parse the to-be-performed task and generate a target list, wherein the target list includes task items; The decomposition module is used to input the target list into the task evaluator to decompose the task item into subtasks; The analysis module is used to match the character portrait according to the subtask to generate recommended resources; The result acquisition module is used to acquire a first execution result, where the first execution result is a result of a user executing a subtask according to the recommended resource.

Images