Task execution method, electronic device, and product

Abstract

The application provides a task execution method, an electronic device and a product. The method determines a task intention corresponding to task information based on the obtained task information and a pre-constructed multi-layer memory network. The multi-layer memory network comprises a short-term memory layer and a long-term memory layer. According to the task intention, an execution task sequence corresponding to the task information is determined, and the execution task sequence comprises at least one subtask related to the task information. The subtask related to the task information is executed according to the execution task sequence. The technical solution of the application can automatically execute the subtask related to the task information, without the need for staff processing, thereby improving the work efficiency of the staff. The task intention recognition of the task information is based on the cooperative recognition of the short-term memory layer and the long-term memory layer, and the historical data in the long-term memory layer is combined for task intention recognition, thereby improving the accuracy of task intention recognition and subtask determination, and further improving the task processing accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a task execution method, electronic device, and product. Background Technology

[0002] As enterprises deepen their digital transformation, the office model in the digital environment needs to handle fragmented task instructions from platforms such as chat tools, emails, and documents. These task instructions originate from different platforms, and employees need to switch between different platforms and jump between tasks during the process, which takes a long time and affects employee work efficiency. In addition, employees need to constantly adjust their thinking framework to deal with different types of tasks, which can easily lead to the omission or misunderstanding of key information and affect the accuracy of task processing. Summary of the Invention

[0003] Based on the above requirements, this application proposes a task execution method, electronic device, and product that can automatically execute sub-tasks related to task information without the need for employee processing, thereby improving employee work efficiency. Furthermore, the task intent recognition of task information is based on the collaborative recognition of short-term memory and long-term memory layers. Combining historical data in the long-term memory layer for task intent recognition can improve the accuracy of task intent recognition and sub-task determination, thereby improving the accuracy of task processing.

[0004] To achieve the above objectives, this application proposes the following technical solution: According to a first aspect of the embodiments of this application, a task execution method is provided, including: Based on the acquired task information and the pre-constructed multi-layer memory network, the task intent corresponding to the task information is determined; the multi-layer memory network includes a short-term memory layer and a long-term memory layer. Based on the task intent, an execution task sequence corresponding to the task information is determined, wherein the execution task sequence includes at least one subtask related to the task information; According to the execution task sequence, subtasks related to the task information are executed.

[0005] Optionally, based on the acquired task information and a pre-constructed multi-layer memory network, the task intent corresponding to the task information is determined, including: The obtained task information is processed by extracting its text content to obtain the task text corresponding to the task information. Based on a pre-constructed multi-layer memory network, the task text is subjected to intent recognition to obtain the task intent corresponding to the task information.

[0006] Optionally, the obtained task information is subjected to text content extraction to obtain the task text corresponding to the task information, including: Based on the modality type of the task information, the text content of the task information is extracted to obtain the task text corresponding to the task information, and the recognition accuracy of the task text is calculated. If the recognition accuracy does not reach the accuracy threshold corresponding to the modality type, the task text is corrected.

[0007] Optionally, based on a pre-constructed multi-layer memory network, intent recognition is performed on the task text to obtain the task intent corresponding to the task information, including: Based on the short-term memory layer, the task text is classified into intents to obtain the initial task intent corresponding to the task information. Based on the long-term memory layer and the initial task intent, determine the intent supplementary data associated with the initial task intent; The initial task intent and the supplementary intent data are fused together, and the task intent corresponding to the task information is generated based on the fused intent information.

[0008] Optionally, the long-term memory layer includes: a contextual memory layer and a semantic memory layer, wherein the contextual memory layer is used to store historical task information and the task intent and execution sequence corresponding to the historical task information, and the semantic memory layer is used to store the domain knowledge and rules corresponding to the task information; Based on the long-term memory layer and the initial task intent, supplementary intent data associated with the initial task intent is determined, including: Based on the context information corresponding to the initial task intent and the task information, historical association information corresponding to the initial task intent is obtained by retrieving from the context memory layer. Based on the initial task intent and the historical association information, the associated knowledge corresponding to the initial task intent is obtained by retrieving from the semantic memory layer. The historical association information and the associated knowledge corresponding to the initial task intent are used as supplementary intent data associated with the initial task intent.

[0009] Optionally, after determining the intent supplementary data associated with the initial task intent based on the long-term memory layer and the initial task intent, the method further includes: Based on the initial task intent and the supplementary intent data, the intent data of relevant personnel associated with the task information is determined using a pre-constructed other-person model network corresponding to the relevant personnel of the task information; wherein, the other-person model network corresponding to the relevant personnel is a model constructed based on the behavioral data of the relevant personnel that can predict the intent data of the relevant personnel. Based on the initial task intent and the intent supplementary data, public information associated with the task information is queried from a pre-built shared group memory; wherein, the shared group memory is used to store group public information; Correspondingly, the initial task intent and the supplementary intent data are fused, and based on the fused intent information, a task intent corresponding to the task information is generated, including: The initial task intent, the supplementary intent data, the relevant personnel intent data, and the public information are merged, and the task intent corresponding to the task information is generated based on the merged intent information.

[0010] Optionally, based on the task intent, determining the execution task sequence corresponding to the task information includes: Extract task elements from the task intent to obtain the task elements corresponding to the task intent; Based on the task elements, determine the sub-tasks corresponding to the task information; The execution order of the subtasks corresponding to the task information is planned to obtain the execution task sequence corresponding to the task information; wherein, the execution task sequence contains subtasks arranged in the planned execution order.

[0011] Optionally, based on the task elements, determining the sub-tasks corresponding to the task information includes: Based on pre-set complex task determination rules, the task constructed from the task elements is determined to be a complex task, and the determination result is obtained. If the determination result indicates that the task constructed from the task elements is a complex task, then the task constructed from the task elements is decomposed to obtain the sub-tasks corresponding to the task information.

[0012] Optionally, according to the execution task sequence, subtasks related to the task information are executed, including: Construct a virtual environment isolated from the user's current work environment; In accordance with the execution task sequence, subtasks related to the task information are executed in the virtual environment.

[0013] Optionally, the task execution method also includes: Obtain the task execution result and verify the task execution result to obtain the verification result; If the verification result indicates that the task execution has failed, then the task will be executed again and / or a prompt message indicating the task execution failure will be output. If the verification result indicates that the task was executed successfully, the task information, the task intent corresponding to the task information, and the execution task sequence corresponding to the task information are written into the short-term memory layer and the long-term memory layer, respectively.

[0014] According to a second aspect of the embodiments of this application, an electronic device is provided, including: a memory and a processor; The memory is connected to the processor and is used to store programs; The processor is used to implement the above-described task execution method by running the program in the memory.

[0015] According to a third aspect of the embodiments of this application, a computer program product is provided, including computer program instructions, which, when executed by a processor, cause the processor to perform the above-described task execution method.

[0016] The task execution method proposed in this application determines the task intent corresponding to the task information based on the acquired task information and a pre-constructed multi-layer memory network. The multi-layer memory network includes a short-term memory layer and a long-term memory layer. Based on the task intent, an execution task sequence corresponding to the task information is determined, and the execution task sequence includes at least one subtask related to the task information. The subtasks related to the task information are then executed according to the execution task sequence. Using the technical solution of this application, the automatic execution of subtasks related to task information can be achieved without employee intervention, improving employee work efficiency. Furthermore, the task intent recognition of the task information is based on the collaborative recognition of the short-term and long-term memory layers, combined with historical data in the long-term memory layer, which improves the accuracy of task intent recognition and subtask determination, thereby improving task processing accuracy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a task execution method provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the process for determining task intent provided in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram of the process for determining the intention to supplement data provided in the embodiments of this application.

[0021] Figure 4 This is a schematic diagram of the process for determining task intent provided in an embodiment of this application.

[0022] Figure 5 This is a flowchart illustrating another task execution method provided in an embodiment of this application.

[0023] Figure 6 This is a schematic diagram of the structure of a task execution device provided in an embodiment of this application.

[0024] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of this application are applicable to scenarios involving automated task processing. By employing the technical solutions of this application, the automatic execution of sub-tasks related to task information can be achieved without human intervention, thus improving employee work efficiency. Furthermore, the task intent recognition of task information is based on the collaborative recognition of short-term and long-term memory layers. Combining historical data in the long-term memory layer for task intent recognition improves the accuracy of task intent recognition and sub-task determination, thereby enhancing the accuracy of task processing.

[0026] As enterprises continue their digital transformation, more and more organizations are migrating their daily office processes online. Work content that was originally concentrated in a single office system is gradually being distributed across different digital platforms such as instant messaging tools, corporate email, online collaborative documents, project management platforms, and approval systems. The tasks and instructions employees receive daily have also broken away from the traditional "top-down, unified deployment" model, evolving into fragmented information from different channels, senders, and levels of urgency. For example, a project manager might simultaneously receive meeting notifications via WeChat Work, obtain report requirements via email, receive approval requests via DingTalk, and also handle automatically pushed schedule update notifications from the office system.

[0027] Task instructions originating from different platforms have their own scenario attributes and information formats. When employees process task instructions from different platforms, they need to switch between different platforms and jump between tasks, which makes the task execution time longer and affects the work efficiency of employees.

[0028] Furthermore, switching between different platforms and navigating between tasks continuously consumes mental energy. Each time an employee transitions from a task on one platform to one on another, they need to readjust their thought process. For example, they might shift from handling instant messaging in a group chat to studying formal business logic in an email, and then to understanding detailed modification requirements in document annotations. This constant adjustment of thought framework affects an employee's task processing approach and concentration, potentially leading to the omission or misunderstanding of key information, thus impacting task accuracy.

[0029] Based on this, this application proposes a task execution method. This technical solution can automatically execute sub-tasks related to task information without the need for employee processing. Furthermore, the task intent recognition of task information is based on the collaborative recognition of short-term memory and long-term memory layers. Combining historical data in the long-term memory layer for task intent recognition can improve the accuracy of task intent recognition and sub-task determination, thereby solving the problems of low employee work efficiency and low task processing accuracy in the prior art.

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Exemplary Implementation Environment The implementation environment of this application comprises three main parts: a user terminal, a task execution server, and several external business systems. The user terminal can be a desktop computer, laptop, tablet, or smart mobile device, running an operating system and applications for daily office use, such as instant messaging clients, email clients, document editors, and browsers. The user terminal is equipped with input and output devices. Input devices include a keyboard, mouse, touchscreen, microphone, and screen capture component for acquiring task information in various modalities, such as highlighted text, screenshots, and voice. Output devices include a monitor and speakers for presenting processing results and interactive prompts to the user.

[0032] The task execution server can be deployed on a local computing device, a server within a local area network, or a remote cloud server. At the hardware level, the task execution server includes a processor, memory, communication interfaces, and a bus. The processor is coupled to the memory and communication interfaces via the bus. The memory stores program instructions that can be loaded and executed by the processor, as well as the data required during execution. At the software level, the task execution server includes an operating system, a runtime environment, and virtualization components. The virtualization components are used to partition an independent running space within the server, isolated from the user's current working environment, for background tasks to run within.

[0033] External business systems include email systems, calendar and meeting room booking systems, instant messaging systems, document management systems, and approval systems, each providing access interfaces. The task execution server establishes data connections with user terminals and various external business systems through communication interfaces. The connection method can be one or more of wired networks, wireless LANs, or mobile communication networks.

[0034] In the data flow direction, the user terminal sends the collected multimodal task information to the task execution server via the communication link; the task execution server processes the task information internally and calls the corresponding external business system through the interface to complete the specific action based on the processing result; the external business system returns the execution feedback to the task execution server via the communication link; when the task execution server needs to confirm with the user or prompt the user, it sends the corresponding information back to the user terminal and presents it to the user through the output device.

[0035] The task execution server also has an internal storage area for persistently storing data generated during operation and reference data required for subsequent processing. This storage area can be provided by the server's local disk or by an external database or distributed storage system. The aforementioned hardware resources, software modules, and external systems together constitute the operating platform of this embodiment, and the specific processing flows described in subsequent embodiments will all be carried out within this implementation environment.

[0036] Exemplary methods See Figure 1 As shown in the figure, this application proposes a task execution method. The method includes: S101. Based on the acquired task information and the pre-constructed multi-layer memory network, determine the task intent corresponding to the task information.

[0037] The user terminal can collect raw input data expressing pending tasks through an input device, and use this raw input data as task information. That is, task information broadly refers to any form of input content or instruction that carries the user's intention to complete a pending task. This embodiment acquires task information transmitted by the user terminal. Specifically, task information can be multimodal; for example, task information may include at least one of text, image, or voice information. This embodiment supports multimodal task information input, enabling the task execution method to adapt to task information from different sources and in different formats, avoiding the need for manual secondary processing by users due to platform differences, thereby expanding the applicable scenarios of the task execution method.

[0038] For example, text information can be text fragments extracted by the user from platforms such as chat tools, documents, or browsers through word selection operations, where word selection is performed using a mouse, keyboard, or other means to select text on the screen. This method directly retrieves task information from the user's current work interface, reducing the need for manual input of task instructions. Image information can be images obtained by the user taking screenshots of chat interfaces or other application interfaces. Screenshots allow for the simultaneous capture of task content including multi-person conversations and multiple messages, suitable for group chat scenarios and other situations requiring the extraction of task information from complex interfaces. Voice information can be voice commands input by the user through voice input. Supporting voice input allows users to issue task commands even in scenarios where text input is inconvenient, improving the convenience of task input.

[0039] In this embodiment, after acquiring task information, the task intent corresponding to the task information is determined based on the task information and a pre-constructed multi-layer memory network. The multi-layer memory network is a network structure composed of at least two memory layers with different time scales and abstraction levels, used to integrate current input with historical information for intent recognition and task execution.

[0040] A multi-layered memory network includes at least a short-term memory layer and a long-term memory layer. The short-term memory layer, also known as working memory, is used to store and process raw information from the current interaction round or short session, and is the primary input buffer for intent recognition. As a real-time cache of current interaction information, the short-term memory layer can provide the latest contextual information during intent recognition, thus providing accurate raw evidence for subsequent deep intent understanding based on the long-term memory layer.

[0041] For example, the information stored in the short-term memory layer may include at least one of the following: current task information, the context of the current dialogue (e.g., recent rounds of dialogue), the current immediate state of the environment (e.g., location, time), and the subtask being executed. The short-term memory layer is characterized by small capacity, fast updates, and high activity. Small capacity means retaining only the most recent and relevant information; fast updates mean it is refreshed or updated after each round of interaction; and high activity means it directly participates in real-time reasoning.

[0042] The long-term memory layer stores information with a long lifespan, such as historical experience, domain knowledge, and user profiles, providing deep contextual support for intent recognition. The long-term memory layer provides historical experience and structured domain knowledge spanning multiple interactions, enabling the system to not only rely on the current input task information during intent recognition but also to perform deep reasoning by combining historical patterns and professional knowledge. This supplements and corrects the initial results of intent recognition based on the short-term memory layer, thereby improving the accuracy and personalization of intent recognition.

[0043] At the hardware level, the storage medium relied upon by a multi-layered memory network can be the local disk of the task execution server, or it can be an external database or distributed storage system. Its physical location can be in a local computing device, a server within a local area network, or a remote cloud server. At the software or logical level, information within a multi-layered memory network can be organized in various forms such as vectors, key-value pairs, structured tables, knowledge graphs, or neural network parameters, and inter-layer interaction can be achieved through read / write, filtering, and consolidation mechanisms.

[0044] Specifically, after obtaining task information, this embodiment performs a comprehensive analysis based on a multi-layered memory network. That is, it combines the task information with the memory data stored in the multi-layered memory network to determine a structured intent description reflecting the user's true needs, i.e., the task intent. In some embodiments, the task intent includes not only surface-level intents (e.g., "set a reminder"), but also refined intents derived from historical experience and domain knowledge (e.g., "set a reminder for the project review meeting at 2:45 PM"). Through the collaborative analysis of the multi-layered memory network, the system can deepen its understanding from surface-level intents to refined intents, avoiding situations where a superficial understanding of intent leads to subsequent task execution deviating from the user's true needs and resulting in low task processing accuracy. For example, when the user inputs a simple instruction like "Remind me this afternoon," the specific event to be reminded cannot be determined solely by the literal meaning. However, by combining it with historical experience and domain knowledge stored in the multi-layered memory network, it can be deduced that the reminder is actually related to a meeting already scheduled in the user's schedule, thus outputting a more refined task intent.

[0045] This embodiment first performs intent recognition on the task information to obtain the initial task intent, and then uses a pre-built multi-layer memory network to supplement and correct the initial task intent to obtain the final task intent. Intent recognition of the task information can be performed using rule-based methods, employing a predefined rule base (such as keyword matching); machine learning-based methods, using open-source large models or pre-trained intent classification models; or a hybrid method combining rule-based and machine learning approaches, assigning corresponding weights to each approach and combining the results of both to determine the initial task intent. Supplementing and correcting the initial task intent using the pre-built multi-layer memory network involves retrieving data associated with the initial task intent and task information from the long-term memory layer of the network as supplementary data. This supplementary data is used to correct the initial task intent, thereby obtaining the final task intent.

[0046] S102. Based on the task intent, determine the sequence of tasks to be executed corresponding to the task information.

[0047] In this embodiment, after determining the task intent corresponding to the task information, an execution task sequence corresponding to the task information is determined based on the task intent. This execution task sequence includes at least one subtask related to the task information. Furthermore, the execution task sequence can be expressed in various ways, such as, but not limited to, linear lists, directed acyclic graphs (DAGs), and tree structures with conditional branches.

[0048] Specifically, this embodiment can perform structured modeling of the task intent according to a pre-set structured task template to obtain a structured task description. Then, the structured task description is analyzed to determine one or more subtasks contained within it. If there is one subtask, it is directly used as the execution task sequence corresponding to the task information. If there are multiple subtasks, the dependencies between them need to be determined. Based on these dependencies, the subtasks are sequentially planned, and then sorted according to the planned order to form the execution task sequence corresponding to the task information. The analysis of the structured task description can include analyzing whether it is a complex or simple task. If it is a simple task, the task corresponding to it is directly used as a subtask, and the execution task sequence may contain only this one subtask. If it is a complex task, it needs to be broken down into several subtasks. Specifically, a subtask can be an email sending, a meeting room booking, a document generation, an instant messaging message push, etc.; it can also include a verification, completion, or query of existing information. For example, the task intent of "send the 'AIPC Functional Brainstorming Meeting' notification to Zhang San via email before next Monday" can be broken down into several sub-tasks, such as confirming meeting information, constructing email content, executing the sending, and handling exceptions, and organized into an ordered sequence according to their sequential dependencies.

[0049] This embodiment transforms task intent into an execution task sequence containing ordered subtasks, enabling the system to perform structured management and step-by-step execution of complex multi-step tasks, thus avoiding the limitation of only being able to handle a single task.

[0050] In one specific implementation, determining the execution task sequence corresponding to the task information based on the task intent includes the following steps: First, extract the task elements from the task intent to obtain the task elements corresponding to the task intent.

[0051] This embodiment first requires task element extraction from the task intent, that is, identifying and extracting the key elements constituting the task from the task intent. These task elements may include, but are not limited to, at least one of the following: time elements (e.g., absolute time, relative time, or period), subject elements (e.g., person, department, or role), object elements (e.g., meeting, document, or resource), behavioral elements (e.g., sending an email, booking a meeting room, or writing materials), constraint elements (e.g., duration, priority, or confidentiality level), and contextual elements (e.g., current conversation or historical tasks). By extracting multi-dimensional task elements from the task intent, the vague task intent expressed in natural language is transformed into a structured set of elements, providing clear input parameters for subsequent determination of subtasks and planning of the execution order based on these task elements.

[0052] Task element extraction from task intent can be performed using rule-based methods, where keyword matching rules are pre-defined and task elements are matched from the task intent based on these rules. Alternatively, machine learning-based methods can be used, where a task element extraction model is pre-trained and then used to extract task elements from the task intent. In this approach, sample intents labeled with task element tags are used to train the task element extraction model, and the training objective is to minimize the difference between the predicted task elements and the task element tags corresponding to the sample intents output by the task element extraction model.

[0053] Second, based on task elements, determine the sub-tasks corresponding to the task information.

[0054] This embodiment constructs a task based on task elements. For example, according to a pre-set task template, the task elements are matched with the task template and task slots are filled to construct a structured task description as the task. This embodiment needs to determine at least one subtask based on the constructed structured task description. For example, if the constructed task is a simple task, it is directly used as a subtask; if the constructed task is a complex task, it is broken down to obtain at least two subtasks.

[0055] Specifically, firstly, based on pre-set complex task judgment rules, the tasks constructed from task elements are judged as complex tasks, and the judgment results are obtained.

[0056] This embodiment pre-sets a complex task determination rule, which is a preset judgment condition used to distinguish between simple and complex tasks. For example, the complex task determination rule can be based on at least one of the following judgment dimensions: the number of sub-actions, whether cross-system calls are involved, whether multiple roles are involved, whether conditional branches exist, the strength of time constraints, the level of failure cost, and whether external dependencies exist. For example, if any of the following conditions are met: the number of sub-actions is greater than 1, cross-system calls are involved, multiple roles are involved, external resources are involved, conditional branches exist, time constraints are strong, failure cost is high, or dependence on others' behavior, it is determined to be a complex task; otherwise, it is determined to be a simple task (also known as an atomic task). For example, the task "Remind me" has 1 sub-action, no cross-system calls, single-person participation, no conditional branches, and loose time constraints, so it can be determined to be an atomic task; while the task "Organize an AIPC brainstorming meeting next Monday and notify Zhang San via email" involves multiple sub-actions, crosses multiple systems such as email and calendar, and involves multiple people and strong time constraints, so it can be determined to be a complex task.

[0057] This embodiment classifies tasks constructed from task elements into complex tasks according to the complex task determination rules, and obtains a determination result. This determination result includes whether the task constructed from task elements is a complex task or a simple task.

[0058] Secondly, if the determination result indicates that the task constructed from the task elements is a complex task, then the task constructed from the task elements is decomposed to obtain the sub-tasks corresponding to the task information.

[0059] If the determination result indicates that the task constructed from task elements is a complex task, the task constructed from task elements is decomposed into at least two finer-grained, independently executable subtasks. This embodiment only performs task decomposition for cases determined to be complex tasks. When processing simple tasks, they can be executed directly without additional decomposition. When processing complex tasks, decomposition transforms them into atomic tasks that can be independently scheduled and executed, achieving refined management of task processing strategies.

[0060] This embodiment of task decomposition can be achieved by performing structured task modeling, representing the task as a Directed Acyclic Graph (DAG), and then decomposing the task according to pre-defined decomposition dimensions. Subtasks are then generated, and dependencies between them are established. The pre-defined decomposition dimensions include at least one of the following: time dimension (preparation → execution → completion), role dimension (initiator → executor → approver), system dimension (email system, calendar, instant messaging system), and condition dimension (success path, failure path). This multi-dimensional task decomposition strategy ensures that complex tasks are fully decomposed from different dimensions such as time, role, system, and condition, guaranteeing that the decomposed subtasks cover all execution requirements of the original task and have clear dependencies.

[0061] In addition, the task decomposition in this embodiment can also construct task decomposition prompts, and input the task and task decomposition prompts into the large language model so that the large language model can decompose the task according to the task decomposition prompts.

[0062] After decomposing the task into several subtasks, this embodiment can further perform executability checks on the subtasks. The executability check items include at least one of the following: atomicity (i.e., whether it can be further decomposed), observability (i.e., whether there is a success / failure state), timeout, and compensability (i.e., whether there is a rollback or alternative solution). This embodiment ensures that each subtask meets the execution conditions of atomicity, observability, timeout, and compensability by checking the executability of the subtasks, avoiding execution failures caused by improper subtask granularity or lack of exception handling mechanisms.

[0063] Furthermore, during task decomposition, a multi-layered memory network can also participate collaboratively. The short-term memory layer can store structured task descriptions and the execution status of each subtask, while the long-term memory layer can provide decomposition experience from similar historical tasks (e.g., "how to decompose similar tasks last time," "a participant switched to instant messaging notifications after not receiving emails last time"), and can also provide task templates, decomposition rules, and default constraints (e.g., "remind the meeting 15 minutes in advance"). By comprehensively utilizing the information from the multi-layered memory network during task decomposition, the decomposition strategy is not only based on the static rules of the current task elements but also draws on historical experience to avoid known failure paths. Furthermore, domain knowledge is used to ensure that the decomposition results conform to business specifications, thereby improving the rationality and success rate of task decomposition.

[0064] Third, the execution order of the subtasks corresponding to the task information is planned to obtain the execution task sequence corresponding to the task information.

[0065] In this embodiment, after determining the subtasks corresponding to the task information, it is necessary to plan the execution order of the subtasks. The subtasks are then arranged according to the planned execution order, and the sorted subtasks are combined into an execution task sequence corresponding to the task information. In this embodiment, planning the execution order of subtasks refers to planning and arranging the execution order of subtasks based on factors such as dependencies, time constraints, and resource constraints. By planning the execution order of subtasks, it is ensured that subtasks with dependencies are executed in the correct logical order (e.g., confirming attendees before sending emails), avoiding task failures or resource conflicts due to improper execution order.

[0066] The dependencies between subtasks can be modeled as a Directed Acyclic Graph (DAG), where nodes represent subtasks and edges represent dependencies between them. Modeling these dependencies as a DAG allows the system to clearly represent and manage the pre- and post-task dependencies, parallel possibilities, and constraints between subtasks, providing a structured basis for the task scheduling engine's execution.

[0067] Execution sequence planning can be implemented using various algorithms. For example, a graph planning algorithm can be used, modeling the task as a directed graph, with nodes representing subtasks and edges representing dependencies. This embodiment can calculate the planning success rate. If the planning success rate is lower than a preset threshold, a backtracking algorithm is used to identify the nodes where task execution failed. Subtasks are adjusted based on the identified nodes; for example, if sending an email fails, it can be adjusted to sending a text message. The planning success rate is the ratio of the number of successfully executed tasks to the total number of tasks. Whether a task is executed successfully requires judging its syntactic correctness, semantic correctness, and runtime correctness. Syntactic correctness refers to whether the subtasks corresponding to the task information conform to the DAG constraints after graph planning. If they do not conform to the DAG constraints, it means the task execution does not meet the syntactic correctness constraints. Semantic correctness refers to whether the task execution has deviated from its target (e.g., sending an email but forgetting to book a meeting room), whether there is a time reversal (e.g., sending a notification and then confirming attendees), or whether there is a lack of pre-existing resources (e.g., not making an estimate before applying for a budget). If there is a target deviation, time reversal, or a lack of pre-existing resources, it means the task execution does not meet the syntactic correctness constraints. Runtime correctness refers to whether the task execution result is accurate. For example, if the subtasks corresponding to the task fail frequently, or a large number of compensation paths are triggered, or there is frequent manual intervention, it indicates that the task execution result is inaccurate and the task execution does not meet the constraints of runtime correctness.

[0068] Alternatively, execution sequence planning can also be implemented using reinforcement learning-based methods. For example, Q-learning can be used to learn the optimal sequence of subtasks. Reinforcement learning-based methods continuously learn the optimal sequence of subtasks through interaction with the environment, and can adaptively optimize the execution order based on historical execution experience. This approach is suitable for scenarios with complex or dynamically changing subtask dependencies.

[0069] S103. Execute the subtasks related to the task information according to the task sequence.

[0070] This embodiment, after determining the execution task sequence corresponding to the task information, executes each subtask step by step according to the arrangement order of the subtasks in the execution task sequence. This ensures that subtasks with dependencies are completed in the correct logical order, avoiding task failure due to disordered execution order. The execution of each subtask can directly utilize the corresponding API to call the relevant function.

[0071] In one specific implementation, the subtasks related to the task information are executed according to the task execution sequence, specifically including the following steps: First, construct a virtual environment that is isolated from the user's current work environment.

[0072] This embodiment can construct an isolated execution environment independent of the user's current operating system desktop or work interface. For example, the virtual environment can be built using containerization technology (such as Docker), ensuring the independence of task execution through resource isolation. By constructing an isolated virtual environment using containerization technology, subtasks execute in independent resource spaces, isolated from the user's current work environment in terms of computing resources, file systems, and networks, fundamentally avoiding interference from automated task execution on the user's normal operations.

[0073] Second, in accordance with the task sequence, subtasks related to the task information are executed in the virtual environment.

[0074] This embodiment executes subtasks related to task information within a virtual environment according to the task sequence. By constructing an isolated virtual environment and executing subtasks within it, the automated execution of tasks does not occupy or interfere with the user's current work interface and operating environment, achieving seamless task execution.

[0075] The isolation effect of a virtual environment can be measured using an interference index. For example, interference can be calculated using the following formula: in, Indicates the degree of interference. Indicates the user's operation time. This indicates the total time for task execution. This embodiment provides a measurable evaluation basis for the isolation effect of the virtual environment by quantifying the interference level index, enabling the system to monitor and ensure that the interference of task execution on user operations remains within an acceptable range.

[0076] This embodiment can dynamically adjust the resource allocation of the virtual environment based on execution feedback. Resource allocation may include, for example, at least one of the following: Central Processing Unit (CPU) quota, memory limits, Input / Output (I / O) weights, startup priority, concurrency, execution window, retry budget, and network bandwidth. By dynamically adjusting the multi-dimensional resource allocation parameters of the virtual environment based on execution feedback, the system can rationally utilize computing resources while ensuring low interference, promptly reducing task resource consumption when interference exceeds limits, and appropriately increasing resource availability to improve execution efficiency when interference is low.

[0077] When executing subtasks in a virtual environment, several execution methods can be employed. For example, virtualization can be used, simulating user actions (such as clicking and typing) within a virtual desktop to complete the subtask. Virtualization allows the system to execute subtasks requiring graphical interface interaction (such as operating an office system interface) without occupying the user's actual desktop environment, by simulating user actions in an isolated virtual desktop. Alternatively, an Application Programming Interface (API)-driven execution method can be used, directly calling functions through system APIs to complete subtasks (such as sending emails via an email API). API-driven execution completes subtasks by directly calling system interfaces, eliminating the need to simulate user interface operations, resulting in higher efficiency and stability. When an API call times out, a backup interface can be used, enhancing fault tolerance.

[0078] During the execution of subtasks in a virtual environment, a multi-layered memory network can collaboratively participate in resource management. The short-term memory layer stores the current disturbance level and resource status; the contextual memory layer within the long-term memory layer stores historical disturbance patterns for specific task types; and the semantic memory layer within the long-term memory layer stores resource allocation rules (e.g., "high-disturbance tasks should have reduced resource weight"). By collaboratively participating in resource management within the virtual environment through this multi-layered memory network, resource allocation strategies not only rely on current real-time feedback but also draw upon historical disturbance patterns and general allocation rules, achieving more intelligent and predictive resource scheduling.

[0079] As described above, the task execution method proposed in this application determines the task intent corresponding to the task information based on the acquired task information and a pre-constructed multi-layer memory network. The multi-layer memory network includes a short-term memory layer and a long-term memory layer. Based on the task intent, an execution task sequence corresponding to the task information is determined, and the execution task sequence includes at least one subtask related to the task information. The subtasks related to the task information are executed according to the execution task sequence. Using the technical solution of this embodiment, automatic execution of subtasks related to task information can be achieved without employee intervention, improving employee work efficiency. Furthermore, the task intent recognition of the task information is based on the collaborative recognition of the short-term memory layer and the long-term memory layer, combined with historical data in the long-term memory layer, which improves the accuracy of task intent recognition and subtask determination, thereby improving task processing accuracy.

[0080] As an optional implementation, see [link to implementation details]. Figure 2 As shown, in another embodiment of this application, step S101, based on the acquired task information and the pre-constructed multi-layer memory network, determines the task intent corresponding to the task information, specifically including the following steps: S201. Extract the text content from the obtained task information to obtain the task text corresponding to the task information.

[0081] In this embodiment, after acquiring task information, text content extraction is performed to convert the content of the task information into text form, thereby obtaining the task text corresponding to the task information. Different text content extraction methods are used for task information of different modalities, enabling the system to uniformly convert the raw inputs of various modalities into text representations, providing a standardized input format for subsequent intent recognition. The task text obtained after text content extraction is the text data representing the content of the task information, serving as the input for subsequent intent recognition. The modal types include at least one of text modality, image modality, and speech modality.

[0082] When the modality of the task information is text, text content extraction may include extracting plain text from word-highlighted text and removing noise; when the modality of the task information is image, text content extraction may include recognizing text in an image using Optical Character Recognition (OCR) technology; when the modality of the task information is speech, text content extraction may include transcribing speech into text using Automatic Speech Recognition (ASR) technology.

[0083] In one specific implementation, the acquired task information is processed to extract the text content, resulting in the corresponding task text. This process includes the following steps: First, based on the modality type of the task information, the text content of the task information is extracted to obtain the task text corresponding to the task information, and the recognition accuracy of the task text is calculated.

[0084] A quality assurance mechanism can also be introduced during the text content extraction process to ensure that the extracted task text has sufficient accuracy and avoid subsequent intent recognition deviations due to text extraction errors. Specifically, this embodiment can extract the text content of the task information based on the modality type of the task information to obtain the task text, and calculate the recognition accuracy of the task text.

[0085] Recognition accuracy characterizes the accuracy of text content extraction and can be calculated as the ratio of correctly extracted content to the total content. When the modality of the task information is text, recognition accuracy... The calculation formula is as follows: in, This indicates the number of characters correctly extracted. This represents the total number of characters. By calculating the accuracy at the character level, we can accurately quantify the loss or errors of characters during the text extraction process, providing precise numerical data for determining whether to trigger the correction mechanism.

[0086] When the modality type of the task information is image modality, the recognition accuracy is... The calculation formula is as follows: in, This indicates the number of correctly recognized characters. This represents the total number of characters. By calculating the proportion of correctly recognized characters in the OCR recognition result out of the total number of characters, the accuracy of image text recognition can be quantified.

[0087] When the modal type of the task information is speech, the recognition accuracy is... The calculation formula is as follows: in, Indicates the number of characters correctly transcribed. This represents the total number of characters. By calculating the proportion of correctly transcribed content in the speech-to-text results, the transcription quality during the speech recognition process can be quantified.

[0088] Second, if the recognition accuracy does not reach the accuracy threshold corresponding to the modality type, the task text is corrected.

[0089] Different modality types may correspond to different accuracy thresholds. For example, the accuracy threshold corresponding to the text modality may be set to 100%, the accuracy threshold corresponding to the image modality may be set to 98%, and the accuracy threshold corresponding to the speech modality may be set to 95%. Setting differentiated accuracy thresholds for different modality types enables flexible control of correction triggering conditions based on the inherent characteristics and accuracy requirements of text extraction for each modality type, ensuring the quality of task texts while avoiding unnecessary correction processing overhead.

[0090] When the recognition accuracy does not reach the accuracy threshold corresponding to the modality type, correction processing is performed on the extracted task text to improve its accuracy. By correcting the task text in a timely manner when the accuracy does not meet the standard, it is ensured that the text input to the subsequent intent recognition link has sufficient quality, so as to avoid task execution deviation caused by the amplification of text extraction errors in the intent recognition stage.

[0091] For task information of different modality types, the corresponding correction methods for task texts are different. When the task information is in the text modality, the correction method for the task text may use a context association algorithm to compensate for missing characters in the task text. For example, this embodiment may use a neural network with a context association algorithm to implement compensation for the task text, for example, using sequence modeling (such as bidirectional LSTM / Transformer) combined with contextual semantic probability to infer and fill in missing or ambiguous characters. Sequence dependency modeling captures long-distance context dependencies through a bidirectional recurrent neural network (Bi-LSTM) or Transformer Encoder, and calculates the conditional probability distribution of the missing position in a specific context (for example, in "Beijing [] District", "[]" is most likely to be "Chao"). After compensating for the missing characters in the task text, the calculation formula for the recognition accuracy is improved as: in which, is the compensated recognition accuracy corresponding to the text modality, is the pre-compensation recognition accuracy corresponding to the text modality, represents a compensation factor (default 0.1), which can be adjusted through feedback to obtain its value. The context association algorithm uses the correlation between adjacent characters and semantics in the text to compensate for missing characters. The compensation factor can be dynamically adjusted through a feedback mechanism, so that the correction process can continuously optimize the compensation effect with changes in usage scenarios.

[0092] When the task information is of image modality, the correction method for the task text may use a multi-model voting mechanism to correct the task text. Exemplarily, the multi-model voting mechanism may integrate four models including BERT, MacBERT, ELECTRA and ERNIE, which respectively assume different roles. The BERT model, as a basic semantic understanding model, is responsible for detecting semantic inconsistencies in context (for example, the improper use of "but" in the sentence "He went to the bank to deposit money, but withdrew cash"). The MacBERT model is improved based on the BERT model, and performs mask training by replacing some tokens to enhance the error correction capability for low-frequency words and domain-specific terms (for example, the spelling of professional terms in medical texts). The ELECTRA model adopts a generation-discrimination architecture, where the discriminator performs detection after the generator replaces tokens, so as to improve the sensitivity to spelling errors and local grammatical errors (for example, correcting "recieve" to "receive"). The ERNIE model is an enhanced semantic representation model integrated with a knowledge graph, and is good at handling entity confusion and logical contradictions (for example, verifying the consistency of "iPhone" with contextual entities in the sentence "Apple Inc. released a new iPhone"). Specifically, the task text is preprocessed first, that is, obvious errors (such as consecutive repeated characters) are filtered out through regular expressions and simple rules, so as to reduce the subsequent calculation load of the models. Then multi-model parallel detection is performed, that is, the multiple models simultaneously perform error correction on the input preprocessed task text to generate candidate error correction results and confidence scores. Then the outputs of the models are integrated based on a weighted voting mechanism. For example, for error correction suggestions at the same position, if the confidence levels of both the BERT model and the ERNIE model exceed a threshold, the suggestion is preferentially adopted; if the ELECTRA detects a spelling error and the confidence level is significantly higher than that of other models, the suggestion is directly adopted. Finally, syntax and semantic consistency verification is performed on the fused result to avoid "over-correction" (for example, mistakenly correcting "Chongqing hot pot" to "Chengdu hot pot").

[0093] When the task information is of speech modality, the correction method for the task text may enable an acoustic model, and introduce a speech enhancement module at the front end of Automatic Speech Recognition (ASR) to compensate for environmental degradation. The acoustic model enhancement corrects the speech transcription result by introducing stronger acoustic feature modeling capability.

[0094] S202: Performing intent recognition on the task text based on a pre-constructed multi-layer memory network to obtain the task intent corresponding to the task information.

[0095] By leveraging short-term contextual information and long-term experiential knowledge stored in a multi-layered memory network, a comprehensive analysis of the task text is performed to determine the task intent reflected in the text. Introducing a multi-layered memory network into the intent recognition stage allows intent recognition to not only rely on the literal meaning of the current task text but also to combine the dialogue context in the short-term memory layer and historical experience and domain knowledge in the long-term memory layer for deep reasoning, thereby significantly improving the accuracy and personalization of intent recognition.

[0096] In one specific implementation, based on a pre-constructed multi-layer memory network, intent recognition is performed on the task text to obtain the task intent corresponding to the task information, specifically including the following steps: First, based on the short-term memory layer, the task text is classified into intents to obtain the initial task intents corresponding to the task information.

[0097] In this embodiment, the short-term memory layer stores the context information of the task information. Based on the task information and the context information of the task information, the task text is initially parsed and classified to obtain the initial task intent at the surface level. This can quickly extract the surface intent of the task text and provide basic input for the subsequent deep intent understanding in the long-term memory layer.

[0098] Intent classification of task text can be achieved in various ways. For example, it can be based on rules, using a predefined rule base (such as keyword matching) for intent recognition. It can also be based on machine learning, using open-source large models or pre-trained intent classification models for intent classification. Alternatively, it can be a hybrid approach that combines rules and machine learning, setting corresponding weights for the rule-based approach and the machine learning approach respectively, and combining the intent recognition results of the two approaches to determine the initial task intent.

[0099] Second, based on the long-term memory layer and the initial task intent, determine the supplementary intent data associated with the initial task intent.

[0100] After obtaining the initial task intent, supplementary intent data associated with it is determined based on the long-term memory layer and the initial task intent. Specifically, historical experience and domain knowledge stored in the long-term memory layer are used to deepen and supplement the initial task intent, obtaining supplementary information that helps to accurately understand the user's true intent. By retrieving supplementary data associated with the initial task intent from the long-term memory layer, the surface-level initial intent can be combined with deep historical experience and domain knowledge, thereby eliminating intent ambiguity and deepening intent understanding.

[0101] Third, the initial task intent and supplementary intent data are fused together, and based on the fused intent information, the task intent corresponding to the task information is generated.

[0102] The initial task intent and supplementary intent data are fused together to generate a final, refined task intent based on the fused, comprehensive intent information. By integrating multi-dimensional information from short-term and long-term memory layers, the final generated task intent not only reflects the literal meaning of the current task text but also incorporates deeper knowledge such as historical preferences and domain rules, thus generating a more accurate and personalized task intent. Methods for fusing the initial task intent and supplementary intent data include: weighted summation, attention mechanism fusion, or cue concatenation based on a large language model.

[0103] For example, when the task text corresponding to the task information entered by the user is "Remind me this afternoon", the short-term memory layer can first extract the action ("remind"), time ("this afternoon"), and object ("me") to obtain the initial task intent as "set a reminder for this afternoon". Then, using the initial task intent and context such as the current time and dialogue history as query vectors, a retrieval is initiated to the long-term memory layer, which can obtain supplementary intent data such as "the user said yesterday: 'I have a project review meeting with the client at 3 pm tomorrow'" and "the user always sets reminders 15 minutes in advance when setting reminders last week". Combined with common sense such as "reminders are usually associated with events" and "this user is a project manager, and meeting-related reminders have high priority" and user profile information, a refined intent is generated after fusion, such as "set a reminder for the user about the 'project review meeting' at 2:45 pm this afternoon".

[0104] As an optional implementation, see [link to implementation details]. Figure 3 As shown, another embodiment of this application discloses determining supplementary intent data associated with the initial task intent based on the long-term memory layer and the initial task intent, specifically including the following steps: S301. Based on the context information corresponding to the initial task intent and task information, retrieve the historical association information corresponding to the initial task intent from the context memory layer.

[0105] In this embodiment, the long-term memory layer in the multi-layer memory network includes a contextual memory layer, which stores historical task information and the corresponding task intentions and execution sequences. For example, the information stored in the contextual memory layer may include timestamp-stored records of past interactions (e.g., the last 100 interactions), specific action sequences performed by the user and their results (success or failure), explicit preferences expressed by the user (e.g., "I usually have meetings on Tuesdays"), and important interaction events (e.g., "The user complained about network problems yesterday"). By storing specific interaction history and experiences with spatiotemporal tags, the contextual memory layer can utilize historical similarities for disambiguation and personalized inference during intent recognition. For example, when a user expresses a vague intent (e.g., "like last time"), specific historical execution details can be retrieved from the contextual memory layer to clarify the user's true intent.

[0106] In terms of storage implementation, the context memory layer can use a vector database for storage. During retrieval, a query vector can be generated based on the current context (e.g., encoding the context information corresponding to the initial task intent and task information as the current context to generate a query vector), and relevant historical contexts (e.g., historical association information corresponding to the initial task intent) can be retrieved through similarity search. Using a vector database to store context memory and performing retrieval through vector similarity search can efficiently find historical contexts that are semantically closest to the current context from a large number of historical interaction records, improving retrieval efficiency and relevance.

[0107] S302. Based on the initial task intent and historical association information, retrieve the associated knowledge corresponding to the initial task intent from the semantic memory layer.

[0108] In this embodiment, the long-term memory layer in the multi-layer memory network also includes a semantic memory layer. This semantic memory layer stores domain knowledge and rules corresponding to task information. The information stored in the semantic memory layer can include, for example, structured domain knowledge (e.g., product manuals, business rules), deep user profile features abstracted from a large amount of history (e.g., long-term goals, core interests, personality traits), and behavioral strategies and rules for dealing with specific situations. By storing compressed, decontextualized domain knowledge and rules, the semantic memory layer provides common-sense support and logical reasoning for intent recognition. For example, when a user says "I'm hungry," the common-sense knowledge in the semantic memory layer can be used to infer that the user's intent is "to find food or restaurant recommendations." In terms of storage format, the information in the semantic memory layer can be a structured knowledge graph, parameterized neural network weights (e.g., the knowledge of the large model itself), or refined text archives. The semantic memory layer supports multiple storage formats and can select the most suitable representation method according to the different characteristics of domain knowledge, thus balancing the accuracy of structured queries and the flexibility of semantic matching during retrieval.

[0109] After obtaining historical association information, a retrieval process is performed from the semantic memory layer based on the initial task intent and historical association information to obtain the associated knowledge corresponding to the initial task intent. This associated knowledge may include at least one of the following: domain knowledge, user profiles, and rules related to the initial task intent. Using the retrieval results from the context memory layer as one of the input conditions for further retrieval of the semantic memory layer creates a cascading relationship between the two memory layers. This allows for targeted acquisition of the most relevant domain knowledge and rules for the current task based on historical experience, avoiding information redundancy caused by broad retrieval based solely on the initial task intent.

[0110] S303. Use the historical association information and association knowledge corresponding to the initial task intent as supplementary intent data associated with the initial task intent.

[0111] Historical association information retrieved from the contextual memory layer and association knowledge retrieved from the semantic memory layer are used together as supplementary intent data to provide multi-dimensional information support for subsequent fusion with the initial task intent. Personalized historical experience from the contextual memory layer and general domain knowledge from the semantic memory layer are aggregated as supplementary intent data, enabling the subsequent intent fusion process to receive dual support from both the experience and knowledge levels, thereby generating a more comprehensive and accurate task intent.

[0112] As an optional implementation, see [link to implementation details]. Figure 4 As shown, in another embodiment of this application, step S202, based on a pre-constructed multi-layer memory network, performs intent recognition on the task text to obtain the task intent corresponding to the task information, specifically including the following steps: S401. Based on the short-term memory layer, perform intent classification on the task text to obtain the initial task intent corresponding to the task information.

[0113] S402. Based on the long-term memory layer and the initial task intent, determine the supplementary intent data associated with the initial task intent.

[0114] S403. Based on the initial task intent and supplementary intent data, use the pre-built network of other models corresponding to the relevant personnel of the task information to determine the intent data of the relevant personnel associated with the task information.

[0115] In this embodiment, an "other model network" corresponding to relevant personnel in the task information is pre-constructed. This network is a model built based on the behavioral data of relevant personnel, capable of predicting their intentions. In this embodiment, the relevant personnel in the task information can be collaborators, approvers, or other individuals involved in the task information. For example, the "other model network" is a dynamic mental model maintained by other relevant personnel within the group. Its content may include, for example, at least one of the following: a predicted intention model (predicting possible current goals and intentions based on the relevant personnel's public behavior and history), a predicted knowledge or belief model (predicting what the relevant personnel know and don't know), a predicted ability or resource model (predicting what the relevant personnel are good at and what resources they possess), and social relationships and trust levels (the level of trust and the closeness of the relationship with relevant personnel). The "other model network" models relevant personnel from multiple dimensions such as intention, knowledge, ability, and social relationships, enabling a comprehensive understanding of their internal states and thus allowing for more reasonable social reasoning and collaborative decision-making. Specific implementations of the "other model network" may include, but are not limited to, recurrent neural networks trained based on historical behavioral sequences, social relationship modeling based on graph neural networks, or behavioral prediction prompts based on large language models.

[0116] By using other-model networks to predict the intention data of relevant personnel, the potential intentions of relevant personnel can be taken into account when dealing with tasks involving multi-party collaboration. This allows for the prediction of collaboration needs and potential conflicts before task execution, thereby improving the success rate of task execution in multi-person collaboration scenarios.

[0117] Other-based model networks can be updated in ways that include at least one of the following: observation-driven (updating the model by observing the actions and outcomes of relevant individuals), dialogue-driven (updating the model through direct communication with relevant individuals), and reflection-driven (correcting the model by analyzing the discrepancies between predictions and actual behavior). These multiple update-driven approaches ensure that the other-based model network can continuously track changes in the behavior of relevant individuals, promptly correcting prediction biases and thus maintaining the model's accuracy and timeliness.

[0118] This embodiment, based on the initial task intent and supplementary intent data, utilizes a pre-constructed other-person model network corresponding to relevant personnel of the task information to determine the intent data of relevant personnel associated with the task information. In this embodiment, the pre-constructed other-person model network corresponding to relevant personnel can be an intelligent agent. The initial task intent and supplementary intent data are input into the other-person model network, which can then predict the current intent, knowledge state, and trust level of relevant personnel towards the task executor, etc., and use the predicted information as the intent data of relevant personnel associated with the task information.

[0119] S404. Based on the initial task intent and intent supplement data, query public information associated with the task information from the pre-built shared collective memory.

[0120] In this embodiment, a shared group memory is pre-constructed to store group public information, which is a public memory accessible to all agents. The group public information may include at least one of the following: (1) Public facts and knowledge base: facts, maps, task breakdowns, and protocols that are commonly recognized by the group (e.g., "The base is at point A", "Report danger signals"). (2) Group event history: Records of important shared experiences of the group (e.g., "We completed task X together yesterday", "Member j performed well in task Y"). This provides the group with a shared narrative and context. (3) Social relationship graph: Public or semi-public relationships, roles, and authority structures among members (e.g., "A is the captain", "B and C often cooperate"). (4) Group norms, rules, and protocols: Conventional or explicitly established behavioral guidelines (e.g., "Take turns speaking", "Distribute spoils equally"). (5) Task status dashboard: The progress, division of labor, and dependencies of the current group tasks. Shared collective memory provides information support for group collaboration from multiple levels, including public knowledge, group history, social relations, behavioral norms, and task status, enabling the system to make rational decisions based on comprehensive group information in tasks involving multiple people. Specific implementation methods for shared collective memory may include, but are not limited to, centralized knowledge bases, distributed databases, or blockchain-based public ledgers.

[0121] Based on the initial task intent and supplementary intent data, public information associated with the task information is retrieved from a pre-built shared collective memory. By acquiring public information from the shared collective memory, decisions can be made based on commonly accepted facts and rules when handling collaborative tasks, ensuring that task execution conforms to group consensus and established norms. Retrieving public information associated with task information from the pre-built shared collective memory involves encoding the initial task intent and supplementary intent data into query vectors. These query vectors are then matched against all information in the shared collective memory, and information matching the query vectors is considered public information associated with the task information. Alternatively, public information query prompts can be pre-built. These prompts, along with the initial task intent and supplementary intent data, are input into a large language model, enabling the model to retrieve public information associated with the initial task intent and supplementary intent data from the shared collective memory according to the public information query prompts.

[0122] S405. Integrate the initial task intent, supplementary intent data, relevant personnel intent data, and public information, and generate the task intent corresponding to the task information based on the integrated intent information.

[0123] After obtaining relevant personnel intent data and public information, the system integrates the initial task intent, supplementary intent data, relevant personnel intent data, and public knowledge, generating the final task intent based on the integrated intent information. By comprehensively integrating multi-dimensional information from short-term memory, long-term memory, other model networks, and shared group memory, the final generated task intent not only possesses accuracy and personalization at the individual level but also sociality and collaboration at the group level, making it suitable for complex task scenarios requiring multi-party collaboration.

[0124] For example, in a multi-agent collaborative scenario, when agent i receives the instruction "Let's capture the resource point in the north," it infers, through the other-model network, that teammate j's current intention might be to report enemy activity or regroup, that their knowledge state might already include understanding of the enemy situation in the north, and assesses their trust level. Simultaneously, it can query the shared collective memory for map details and strategic value of the northern resource point, the standard procedure for capturing the resource point (e.g., reconnaissance, assembly, and attack), and the relevant tasks already assigned on the current task board. By integrating the aforementioned relevant personnel intention data and public information with the initial task intention and supplementary intention data, a socialized intention that considers individual experience, other-state, and group norms can be generated, such as "cooperatively capture the northern resource point," and derived sub-intentions such as "communicate with j to confirm enemy activity," "declare action objectives on the shared task board," "propose forming a squad depending on the severity of the enemy situation," and "execute the capture operation." By introducing other-model networks and shared collective memory, the task intent not only answers "what to do" but also includes "how to do it in collaboration with others," thereby further improving task execution efficiency and collective intelligence in collaborative work or multi-agent collaboration scenarios.

[0125] The specific execution method of steps S401-S402 in this embodiment has been specifically described in the above embodiments, and will not be repeated in this embodiment.

[0126] As an optional implementation, see [link to implementation details]. Figure 5 As shown, another embodiment of this application discloses a task execution method, which further includes the following steps: S501. Obtain the task execution result and verify the task execution result to obtain the verification result.

[0127] After the subtask is completed, the execution result is obtained and validated to obtain the validation result. Validation can include multiple layers: The first layer is execution-level validation, used to verify whether the task action was truly completed (e.g., checking return codes, status codes, exception handling, and idempotency). The second layer is result-level validation, used to verify whether the result after execution is correct (e.g., querying email delivery logs, checking if there are created events in the calendar, verifying the completeness of document content). The third layer is semantic-level validation, used to verify whether the execution result meets the user's actual expectations (e.g., validating whether key fields are included through rules, scoring the result through LLM, confirming through user feedback, or aligning and comparing with the original task description). This three-layer validation structure comprehensively evaluates the task execution result from three progressive levels: execution completion, result correctness, and semantic matching. This ensures that the validation process not only detects technical faults at the execution level but also data errors at the result level and deviations from the user's actual expectations at the semantic level.

[0128] If the execution state check indicates that the task action has not been truly completed, or the result state check indicates that the result after completion is incorrect, or the semantic state check indicates that the execution result does not meet the user's actual expectations, then the check result indicates that the task execution has failed. If the execution state check indicates that the task action has been truly completed, the result state check indicates that the result after completion is correct, and the semantic state check indicates that the execution result meets the user's actual expectations, then the check result indicates that the task execution has succeeded.

[0129] S502. If the verification result indicates that the task execution has failed, then the task will be executed again and / or a prompt message indicating the task execution failure will be output.

[0130] When the verification result indicates that the task execution has failed, the following actions can be taken as compensation: retrying the task, outputting a prompt message indicating the task execution failure (manual intervention), or taking at least one of the two measures simultaneously. Two flexible compensation strategies, retry and manual intervention, are provided to address different types and severity of verification failures. While possessing automatic retry and self-healing capabilities, the system also promptly notifies the user to intervene when automatic repair fails, preventing the error from escalating indefinitely.

[0131] Compensation actions can be differentiated based on the type of verification failure. For example, a limited number of retries (e.g., a maximum of 3) can be performed when execution fails; rewriting or rescheduling can be triggered when the result is incorrect; user confirmation or manual intervention can be requested when there is a semantic deviation; and degraded execution can be performed when a high-risk task fails. Adopting differentiated compensation strategies based on different types of verification failures makes the compensation mechanism targeted, avoiding the inefficiency or loss of risk that can result from using a uniform approach for all failure types.

[0132] S503. If the verification result indicates that the task was executed successfully, the task information, the task intent corresponding to the task information, and the execution task sequence corresponding to the task information are written into the short-term memory layer and the long-term memory layer, respectively.

[0133] When the verification result indicates that the task was successfully executed, the complete information of this task (including the original task information, the identified task intent, and the execution sequence) is stored in the multi-layer memory network for reference in subsequent task processing. The complete record of successfully executed tasks is written into the short-term memory layer and the long-term memory layer, so that the multi-layer memory network can continuously accumulate successful task processing experience, providing richer historical references for intent recognition and task planning of similar tasks in the future, and achieving continuous improvement in system performance.

[0134] Information written to the short-term memory layer is used for immediate reference during the current session, such as for contextual understanding of subsequent tasks. Information written to the long-term memory layer is used for persistent storage across sessions, such as for writing to the contextual memory layer as historical experience for similar future tasks, and for consolidating information into the semantic memory layer after abstraction and refinement to update domain knowledge and user preference models. By writing successful experiences into different sub-layers of the short-term and long-term memory layers respectively, the system can both utilize the latest task records immediately in the current session and continuously accumulate and refine experience in the long term, achieving a gradual consolidation from concrete contextual memory to abstract semantic knowledge, and completing a closed-loop optimization of intent recognition and task execution.

[0135] Exemplary device Accordingly, embodiments of this application also provide a task execution device, see below. Figure 6 As shown, the device includes: The intent determination module 100 is used to determine the task intent corresponding to the task information based on the acquired task information and the pre-constructed multi-layer memory network; the multi-layer memory network includes a short-term memory layer and a long-term memory layer. The task determination module 110 is used to determine the execution task sequence corresponding to the task information according to the task intent, wherein the execution task sequence includes at least one subtask related to the task information; The task execution module 120 is used to execute subtasks related to task information according to the task execution sequence.

[0136] As can be seen from the above description, the task execution device proposed in this application embodiment can automatically execute sub-tasks related to task information without the need for employee processing, thereby improving employee work efficiency. Furthermore, the task intent recognition of task information is based on the collaborative recognition of short-term memory and long-term memory layers. Combining historical data in the long-term memory layer for task intent recognition can improve the accuracy of task intent recognition and sub-task determination, thereby improving the accuracy of task processing.

[0137] As an optional implementation, another embodiment of this application discloses an intent determination module 100, including a text extraction unit and an intent recognition unit.

[0138] The text extraction unit is used to extract the text content from the acquired task information to obtain the task text corresponding to the task information. The intent recognition unit is used to recognize the intent of the task text based on a pre-built multi-layer memory network, so as to obtain the task intent corresponding to the task information.

[0139] As an optional implementation, another embodiment of this application discloses a text extraction unit, specifically used for: Based on the modality type of the task information, the text content of the task information is extracted to obtain the task text corresponding to the task information, and the recognition accuracy of the task text is calculated. If the recognition accuracy does not reach the accuracy threshold corresponding to the modality type, the task text is corrected.

[0140] As an optional implementation, another embodiment of this application discloses that the intent recognition unit includes: an intent classification unit, an intent supplementation unit, and an intent fusion unit.

[0141] The intent classification unit is used to classify the intent of the task text based on the short-term memory layer to obtain the initial task intent corresponding to the task information. The intent supplementation unit is used to determine intent supplementation data associated with the initial task intent based on the long-term memory layer and the initial task intent; The intent fusion unit is used to fuse the initial task intent and supplementary intent data, and generate the task intent corresponding to the task information based on the fused intent information.

[0142] As an optional implementation, another embodiment of this application discloses that the long-term memory layer includes: a contextual memory layer and a semantic memory layer, wherein the contextual memory layer is used to store historical task information and the corresponding task intents and execution task sequences, and the semantic memory layer is used to store the domain knowledge and rules corresponding to the task information; the intent supplementation unit is specifically used for: Based on the contextual information corresponding to the initial task intent and task information, the historical association information corresponding to the initial task intent is retrieved from the contextual memory layer. Based on the initial task intent and historical association information, the associated knowledge corresponding to the initial task intent is obtained by retrieving from the semantic memory layer. The historical association information and the associated knowledge corresponding to the initial task intent are used as supplementary intent data associated with the initial task intent.

[0143] As an optional implementation, another embodiment of this application discloses that the intent identification unit further includes: a relevant person intent determination unit and a public information query unit.

[0144] The relevant personnel intent determination unit is used to determine the relevant personnel intent data associated with the task information based on the initial task intent and intent supplementary data, using a pre-built other-person model network corresponding to the relevant personnel of the task information; wherein, the other-person model network corresponding to the relevant personnel is a model built based on the behavioral data of the relevant personnel that can predict the intent data of the relevant personnel. The public information query unit is used to query public information associated with task information from a pre-built shared group memory based on the initial task intent and intent supplementary data; wherein, the shared group memory is used to store group public information; The intent fusion unit is also used to fuse the initial task intent, supplementary intent data, relevant personnel intent data, and public information, and generate the task intent corresponding to the task information based on the fused intent information.

[0145] As an optional implementation, another embodiment of this application discloses that the task determination module 110 includes: an element extraction unit, a subtask determination unit, and a sequence planning unit.

[0146] The element extraction unit is used to extract task elements from the task intent and obtain the task elements corresponding to the task intent. The subtask determination unit is used to determine the subtasks corresponding to the task information based on task elements. The sequence planning unit is used to plan the execution order of the subtasks corresponding to the task information, so as to obtain the execution task sequence corresponding to the task information; wherein, the execution task sequence contains subtasks arranged according to the planned execution order.

[0147] As an optional implementation, another embodiment of this application discloses a subtask determination unit, specifically used for: Based on pre-set complex task judgment rules, the task constructed from task elements is judged as a complex task, and the judgment result is obtained. If the determination result indicates that the task constructed from the task elements is a complex task, then the task constructed from the task elements is decomposed to obtain the sub-tasks corresponding to the task information.

[0148] As an optional implementation, another embodiment of this application discloses a task execution module 120, specifically used for: Construct a virtual environment isolated from the user's current work environment; In accordance with the task sequence, subtasks related to the task information are executed in the virtual environment.

[0149] As an optional implementation, another embodiment of this application discloses that the task execution device further includes: a verification module, a compensation module, and a writing module.

[0150] The verification module is used to obtain the task execution results and verify them to obtain the verification results. The compensation module is used to re-execute the task and / or output a prompt message indicating that the task execution failed if the verification result indicates that the task execution failed. The writing module is used to write the task information, the task intent corresponding to the task information, and the execution task sequence corresponding to the task information into the short-term memory layer and the long-term memory layer, respectively, if the verification result indicates that the task was executed successfully.

[0151] The task execution apparatus provided in this embodiment belongs to the same concept as the task execution method provided in the above embodiments of this application. It can execute the task execution method provided in any of the above embodiments of this application and has the corresponding functional modules and beneficial effects for executing the task execution method. Technical details not described in detail in this embodiment can be found in the specific processing content of the task execution method provided in the above embodiments of this application, and will not be repeated here.

[0152] Exemplary electronic devices Another embodiment of this application also provides an electronic device, see [link to relevant documentation] Figure 7 As shown, the device includes: Memory 200 and processor 210; The memory 200 is connected to the processor 210 and is used to store programs; The processor 210 is configured to implement the task execution method disclosed in any of the above embodiments by running the program stored in the memory 200.

[0153] Specifically, the aforementioned electronic device may also include: a bus, a communication interface 220, an input device 230, and an output device 240.

[0154] The processor 210, memory 200, communication interface 220, input device 230, and output device 240 are interconnected via a bus. Among them: A bus can include a pathway for transmitting information between various components of a computer system.

[0155] Processor 210 can be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. It can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0156] Processor 210 may include a main processor, as well as a baseband chip, modem, etc.

[0157] The memory 200 stores a program that executes the technical solution of this invention, and may also store an operating system and other key business functions. Specifically, the program may include program code, which includes computer operation instructions. More specifically, the memory 200 may include read-only memory (ROM), other types of static storage devices capable of storing static information and instructions, random access memory (RAM), other types of dynamic storage devices capable of storing information and instructions, disk storage, flash memory, etc.

[0158] Input device 230 may include a device for receiving user input data and information, such as a keyboard, mouse, camera, scanner, light pen, voice input device, touch screen, pedometer, or gravity sensor.

[0159] Output device 240 may include devices that allow information to be output to a user, such as a display screen, printer, speaker, etc.

[0160] The communication interface 220 may include a device that uses any transceiver to communicate with other devices or communication networks, such as Ethernet, Radio Access Network (RAN), Wireless Local Area Network (WLAN), etc.

[0161] The processor 210 executes the program stored in the memory 200 and calls other devices, which can be used to implement the various steps of any of the task execution methods provided in the above embodiments of this application.

[0162] Exemplary computer program products and storage media In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the task execution methods according to various embodiments of this application as described in the "Exemplary Methods" section of this specification.

[0163] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0164] Furthermore, embodiments of this application may also be storage media storing a computer program, which is executed by a processor using the steps of the task execution methods according to various embodiments of this application described in the "Exemplary Methods" section above.

[0165] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0166] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0167] The steps in the methods of the various embodiments of this application can be adjusted, merged, or deleted in order according to actual needs, and the technical features described in each embodiment can be replaced or combined.

[0168] The modules and sub-modules in the various embodiments of the present application's devices and terminals can be merged, divided, and deleted according to actual needs.

[0169] It should be understood that the disclosed terminals, devices, and methods can be implemented in other ways, given the several embodiments provided in this application. For example, the terminal embodiments described above are merely illustrative. For instance, the division of modules or sub-modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple sub-modules or modules may be combined or integrated into another module, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0170] The modules or submodules described as separate components may or may not be physically separate. The components that constitute a module or submodule may or may not be physical modules or submodules; that is, they may be located in one place or distributed across multiple network modules or submodules. Some or all of the modules or submodules can be selected to achieve the purpose of this embodiment's solution, depending on actual needs.

[0171] Furthermore, the functional modules or sub-modules in the various embodiments of this application can be integrated into one processing module, or each module or sub-module can exist physically separately, or two or more modules or sub-modules can be integrated into one module. The integrated modules or sub-modules described above can be implemented in hardware or in the form of software functional modules or sub-modules.

[0172] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0173] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software unit executed by a processor, or a combination of both. The software unit can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0174] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0175] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A task execution method, characterized in that, include: Based on the acquired task information and the pre-constructed multi-layer memory network, the task intent corresponding to the task information is determined. The multi-layered memory network includes a short-term memory layer and a long-term memory layer; Based on the task intent, an execution task sequence corresponding to the task information is determined, wherein the execution task sequence includes at least one subtask related to the task information; According to the execution task sequence, subtasks related to the task information are executed.

2. The task execution method according to claim 1, characterized in that, Based on the acquired task information and a pre-constructed multi-layer memory network, the task intent corresponding to the task information is determined, including: The obtained task information is processed by extracting its text content to obtain the task text corresponding to the task information. Based on a pre-constructed multi-layer memory network, the task text is subjected to intent recognition to obtain the task intent corresponding to the task information.

3. The task execution method according to claim 2, characterized in that, The obtained task information is processed by extracting its text content to obtain the task text corresponding to the task information, including: Based on the modality type of the task information, the text content of the task information is extracted to obtain the task text corresponding to the task information, and the recognition accuracy of the task text is calculated. If the recognition accuracy does not reach the accuracy threshold corresponding to the modality type, the task text is corrected.

4. The task execution method according to claim 2, characterized in that, Based on a pre-constructed multi-layer memory network, intent recognition is performed on the task text to obtain the task intent corresponding to the task information, including: Based on the short-term memory layer, the task text is classified into intents to obtain the initial task intent corresponding to the task information. Based on the long-term memory layer and the initial task intent, determine the intent supplementary data associated with the initial task intent; The initial task intent and the supplementary intent data are fused together, and the task intent corresponding to the task information is generated based on the fused intent information.

5. The task execution method according to claim 4, characterized in that, The long-term memory layer includes a contextual memory layer and a semantic memory layer, wherein the contextual memory layer is used to store historical task information and the task intent and execution sequence corresponding to the historical task information, and the semantic memory layer is used to store the domain knowledge and rules corresponding to the task information; Based on the long-term memory layer and the initial task intent, supplementary intent data associated with the initial task intent is determined, including: Based on the context information corresponding to the initial task intent and the task information, historical association information corresponding to the initial task intent is obtained by retrieving from the context memory layer. Based on the initial task intent and the historical association information, the associated knowledge corresponding to the initial task intent is obtained by retrieving from the semantic memory layer. The historical association information and the associated knowledge corresponding to the initial task intent are used as supplementary intent data associated with the initial task intent.

6. The task execution method according to claim 4, characterized in that, After determining the supplementary intent data associated with the initial task intent based on the long-term memory layer and the initial task intent, the method further includes: Based on the initial task intent and the supplementary intent data, the intent data of relevant personnel associated with the task information is determined using a pre-constructed other-person model network corresponding to the relevant personnel of the task information; wherein, the other-person model network corresponding to the relevant personnel is a model constructed based on the behavioral data of the relevant personnel that can predict the intent data of the relevant personnel. Based on the initial task intent and the intent supplementary data, public information associated with the task information is queried from a pre-built shared group memory; wherein, the shared group memory is used to store group public information; Correspondingly, the initial task intent and the supplementary intent data are fused, and based on the fused intent information, a task intent corresponding to the task information is generated, including: The initial task intent, the supplementary intent data, the relevant personnel intent data, and the public information are merged, and the task intent corresponding to the task information is generated based on the merged intent information.

7. The task execution method according to claim 1, characterized in that, Based on the task intent, determine the execution task sequence corresponding to the task information, including: Extract task elements from the task intent to obtain the task elements corresponding to the task intent; Based on the task elements, determine the sub-tasks corresponding to the task information; The execution order of the subtasks corresponding to the task information is planned to obtain the execution task sequence corresponding to the task information; wherein, the execution task sequence contains subtasks arranged in the planned execution order.

8. The task execution method according to claim 7, characterized in that, Based on the task elements, the sub-tasks corresponding to the task information are determined, including: Based on pre-set complex task determination rules, the task constructed from the task elements is determined to be a complex task, and the determination result is obtained. If the determination result indicates that the task constructed from the task elements is a complex task, then the task constructed from the task elements is decomposed to obtain the sub-tasks corresponding to the task information.

9. The task execution method according to claim 1, characterized in that, According to the task execution sequence, subtasks related to the task information are executed, including: Construct a virtual environment isolated from the user's current work environment; In accordance with the execution task sequence, subtasks related to the task information are executed in the virtual environment.

10. The task execution method according to claim 1, characterized in that, Also includes: Obtain the task execution result and verify the task execution result to obtain the verification result; If the verification result indicates that the task execution has failed, then the task will be executed again and / or a prompt message indicating the task execution failure will be output. If the verification result indicates that the task was executed successfully, the task information, the task intent corresponding to the task information, and the execution task sequence corresponding to the task information are written into the short-term memory layer and the long-term memory layer, respectively.

11. An electronic device, characterized in that, include: Memory and processor; The memory is connected to the processor and is used to store programs; The processor is configured to implement the task execution method as described in any one of claims 1 to 10 by running a program in the memory.

12. A computer program product, characterized in that, It includes computer program instructions that, when executed by a processor, cause the processor to implement the task execution method as described in any one of claims 1 to 10.

Images