Task decomposition method and computing device

Abstract

Embodiments of the present application provide a task decomposition method and a computing device, and relate to the technical field of artificial intelligence. The method comprises: obtaining intention data; the intention data is used to describe a task requirement to be implemented; obtaining global anchor data from the intention data; performing dimension division on the task based on the global anchor data to obtain a first decomposition result; the first decomposition result is used to indicate a plurality of independent subtask dimensions that need to be considered for implementing the task requirement; performing path planning on an execution process of the task requirement based on the global anchor data to obtain a second decomposition result; the second decomposition result is used to indicate an atomic action sequence that needs to be executed in sequence for implementing the task requirement; and generating a specific execution scheme for implementing the task and meeting the global anchor data according to the first decomposition result and the second decomposition result. The method can improve the reliability of the specific execution scheme obtained by task decomposition.

Description

Technical Field

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

[0002] With the widespread application of large language models in scenarios such as intelligent question answering, search assistants, operation and maintenance assistants, intelligent customer service, and professional decision support, and the gradual improvement of the automation functions of artificial intelligence systems, artificial intelligence systems are increasingly taking on the role of automating tasks on behalf of users. In this scenario, the large language model, as the "brain" of the artificial intelligence system, fully understands the user's intent and outputs specific execution plans that the artificial intelligence system can directly execute, thereby meeting the user's automated office needs.

[0003] However, current large language models suffer from drawbacks such as distracted attention, weak long-range reasoning ability, and imprecise logical steps when dealing with complex tasks. These shortcomings prevent large language models from outputting accurate and reliable specific execution solutions, and these problems have become key bottlenecks restricting the autonomous execution capability of large language models. Summary of the Invention

[0004] This application provides a task decomposition method and computing device, which can improve the reliability of the specific execution plan obtained from task decomposition.

[0005] In a first aspect, embodiments of this application provide a task decomposition method, the method comprising: acquiring intent data; the intent data describing the task requirements to be achieved; acquiring global anchoring data based on the intent data; the global anchoring data indicating the core objective of the task requirements; dividing the task into dimensions based on the global anchoring data to obtain a first decomposition result; the first decomposition result indicating multiple independent sub-task dimensions to be considered in achieving the task requirements; performing path planning on the execution process of the task based on the global anchoring data to obtain a second decomposition result; the second decomposition result indicating the sequence of atomic actions to be executed sequentially in achieving the task requirements; and generating a specific execution plan that achieves the task and satisfies the global anchoring data based on the first decomposition result and the second decomposition result.

[0006] In the task decomposition method provided in this application embodiment, the computing device acquires intent data describing task requirements, and obtains global anchoring data based on the intent data. This global anchoring data is used to indicate the core objective of the task requirements. Further, the task is dimensionally divided based on the global anchoring data to obtain a first decomposition result. The execution process of the task is then path-planned to obtain a second decomposition result. The first decomposition result ensures broad coverage during task decomposition, indicating what needs to be addressed during the task decomposition process. The second decomposition result ensures that the atomic action sequence obtained from the decomposition is logically executable, indicating how to specifically perform tasks during the decomposition process. Additionally, global anchoring data is introduced during the decomposition process to ensure that the decomposition result does not deviate from the core objective. Through the mutual constraints of global anchoring data, the first decomposition result, and the second decomposition result, the final execution plan output after task decomposition is realized, simultaneously possessing the effects of "broad coverage," "deep executableness," and "long-term logical conservation" (i.e., not deviating from the core objective), thereby effectively improving the accuracy and reliability of task decomposition.

[0007] In one possible implementation, the step of dividing the task into dimensions based on the global anchoring data to obtain a first decomposition result includes: decomposing the intent data using different decomposition strategies to obtain multiple candidate subtask sets; wherein each decomposition strategy corresponds to one candidate subtask set; performing integrity checks on the multiple candidate subtask sets based on the global anchoring data to obtain checked multiple candidate subtask sets; and taking the intersection of the checked multiple candidate subtask sets to obtain the first decomposition result. This method provides a specific implementation for obtaining the first decomposition result, ensuring the feasibility of this solution.

[0008] In another possible implementation, the completeness verification of the plurality of candidate subtask sets based on the global anchoring data includes: for each candidate subtask set in the plurality of candidate subtask sets, determining whether the subtask dimensions of the candidate subtask set cover all dimensions that should be considered by the global anchoring data; if the subtask dimensions of the candidate subtask set cannot cover all dimensions that should be considered by the global anchoring data, supplementing the candidate subtask set with subtasks of missing dimensions.

[0009] Another possible implementation method includes: for each candidate subtask set, if the functional overlap between the first and second subtasks exceeds a threshold, merging the first and second subtasks; the first and second subtasks are any two subtasks from the candidate subtask set. This approach ensures that the decomposed subtasks are mutually exclusive and exhaustive, guaranteeing the comprehensiveness of the subsequently generated execution plan.

[0010] In another possible implementation, the degree of functional overlap between the first subtask and the second subtask is determined based on the semantic similarity between the first subtask and the second subtask; the greater the semantic similarity, the greater the degree of functional overlap.

[0011] In another possible implementation, based on the global anchoring data, path planning is performed on the execution process of the task requirement to obtain a second decomposition result. This includes: decomposing the task requirement using different decomposition strategies to obtain multiple candidate sequences; one decomposition strategy corresponds to one candidate sequence; the multiple candidate sequences are used to indicate different execution paths to achieve the task requirement; based on the global anchoring data, compliance verification is performed on the multiple candidate sequences to obtain verified candidate sequences; the intersection of the verified candidate sequences is taken to obtain the second decomposition result. The above method provides a specific implementation for obtaining a second decomposition result through decomposition, ensuring the feasibility of this solution.

[0012] In another possible implementation, the compliance verification of the multiple candidate sequences based on the global anchoring data includes: for each candidate sequence, determining whether the candidate sequence deviates from the core objective; if the candidate sequence deviates from the core objective, deleting the candidate sequence or adjusting the first atomic action in the candidate sequence; the first atomic action is the key node that caused the candidate sequence to deviate from the core objective. Through compliance verification, the system checks whether the task decomposition deviates from the core objective, deletes non-compliant sequences when deviations occur, or makes timely adjustments to ensure the stability of the long-range inference process.

[0013] In another possible implementation, the global anchoring data is also used to indicate the constraints that must be followed during the task requirements. Based on the global anchoring data, compliance checks are performed on multiple candidate sequences, including: for each candidate sequence, determining whether there is a second atomic action that violates the constraints; if there is a second atomic action that violates the constraints, deleting the candidate sequence or adjusting the second atomic action in the candidate sequence. Through compliance checks, it is possible to verify whether the task decomposition violates the constraints, and if so, deleting non-compliant sequences or making timely adjustments to ensure the stability of the long-range inference process.

[0014] In another possible implementation, obtaining intent data includes: obtaining explicit intent input from the user; obtaining latent intents associated with the explicit intent from a domain knowledge base; the domain knowledge base is used to store the knowledge system of the domain to which the task requirements belong; obtaining user preference data from the user's historical behavior database; and fusing explicit intent, latent intent, and user preference data to obtain intent data. It should be understood that the computing device fuses explicit intent, latent intent, and user preference data to form a high-dimensional information set, thus obtaining intent data. This intent data serves as the foundational data source for subsequent task decomposition processes, ensuring that the task decomposition process not only conforms to the current context but also compensates for any oversights by the user and aligns with the user's deep-seated habits.

[0015] In another possible implementation, the method further includes: obtaining user feedback data on the specific execution plan; updating user preference data based on the feedback data; and storing the updated user preference data in the historical behavior database.

[0016] Secondly, this application provides a task disassembly apparatus, which includes one or more functional modules for implementing the task disassembly method described in the first aspect above.

[0017] Thirdly, embodiments of this application provide a computing device including a processor and a memory; the processor is coupled to the memory; the memory is used to store computer instructions, which are loaded and executed by the processor to enable the computing device to implement the method described in the first aspect.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium comprising: computer software instructions; when the computer software instructions are executed in a computing device, they cause the computing device to implement the method described in the first aspect.

[0019] Fifthly, embodiments of this application provide a computer program product that, when run on a computing device, causes the computing device to execute the steps of the method related to the first aspect described above, so as to implement the method of the first aspect described above.

[0020] The beneficial effects of the second to fifth aspects mentioned above can be referred to the corresponding description of the first aspect, and will not be repeated here. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the composition of a computing device provided in an embodiment of this application; Figure 2 A flowchart illustrating a task decomposition method provided in an embodiment of this application; Figure 3 A flowchart illustrating another task decomposition method provided in an embodiment of this application; Figure 4 A flowchart illustrating yet another task decomposition method provided in this application embodiment; Figure 5 A flowchart illustrating yet another task decomposition method provided in this application embodiment; Figure 6 A schematic diagram illustrating a complete process provided for an embodiment of this application; Figure 7 This is a schematic diagram of the composition of a task disassembly device provided in an embodiment of this application. Detailed Implementation

[0022] 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.

[0023] It should be noted that in the embodiments of this application, the words "exemplarily" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplarily" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplarily" or "for example" is intended to present the relevant concepts in a specific manner.

[0024] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.

[0025] The following is a brief explanation of the technical terms used in the embodiments of this application: 1. Large Language Model (LLM): A type of artificial intelligence model that uses deep learning algorithms to learn language patterns and structures through training on a large amount of text data, thereby enabling it to understand and generate natural language text.

[0026] To address the aforementioned issues, this application provides a task decomposition method that enables the specific implementation schemes output by task decomposition to simultaneously possess the effects of "broad coverage," "deep executability," and "long-term logic conservation" (i.e., not deviating from the core objective), thereby effectively improving the accuracy and reliability of task decomposition.

[0027] In some implementations, the task decomposition method provided in this application involves a computing device acquiring intent data describing task requirements and obtaining global anchoring data based on the intent data. This global anchoring data indicates the core objective of the task requirements. Further, the task is dimensionally divided based on the global anchoring data to obtain a first decomposition result. The execution process of the task is then path-planned to obtain a second decomposition result. The first decomposition result ensures broad coverage during task decomposition, indicating what needs to be addressed during the task decomposition process. The second decomposition result ensures that the resulting atomic action sequence is logically executable, indicating how to specifically perform tasks during decomposition. Additionally, global anchoring data is introduced during the decomposition process to ensure that the decomposition result does not deviate from the core objective. Through the mutual constraints of global anchoring data, the first decomposition result, and the second decomposition result, a specific execution plan is finally output after task decomposition, achieving "broad coverage," "deep executableness," and "long-term logical conservation" (i.e., not deviating from the core objective), thereby effectively improving the accuracy and reliability of task decomposition.

[0028] The embodiments provided in this application will now be described in detail with reference to the accompanying drawings.

[0029] The technical solutions provided in this application can be applied to computing devices, which can be servers with displays or other computing devices. This application does not specifically limit the type of computing device.

[0030] For example, taking a computing device as a server, the server can be a physical or logical server, or it can be two or more physical or logical servers that share different responsibilities and work together to realize the various functions of the server.

[0031] In terms of form factor, the server can be a blade server, a high-density server, a rack server, or a full-rack server; in terms of function, the server can be a general-purpose server, a graphics processing unit (GPU) server, an artificial intelligence (AI) server, etc.

[0032] For example, taking a computing device as a terminal device, the specific form of the terminal device can be an ultra-mobile personal computer (UMPC), tablet computer, laptop computer, netbook, desktop computer, all-in-one computer, etc.

[0033] The task decomposition method provided in this application can be applied to scenarios such as planning and office automation. In planning scenarios, for example, a user can input basic travel needs, and the computing device can output a specific travel plan by executing the task decomposition method. Similarly, a user can input a basic vehicle problem, and the computing device can output a specific repair plan to solve the problem. In office automation scenarios, the more detailed the instructions given to the automation software, the better the automation effect. Based on the task decomposition method of this application, a user's brief office needs can be broken down into detailed steps, facilitating accurate execution by the automation software.

[0034] Figure 1 This is a schematic diagram illustrating the composition of a computing device provided in an embodiment of this application. Figure 1 As shown, the computing device may include a processor 101 and a memory 102; the memory 102 stores instructions executable by the processor 101; when the processor 101 is configured to execute instructions, the computing device executes the task decomposition method in the embodiments of this application.

[0035] In some implementations, the computing device is connected to the display device. The user can input intent data describing the task requirements through the display device. The computing device can execute the task decomposition method of this application embodiment based on the intent data to obtain a specific execution plan to achieve the task requirements, and then display the specific execution plan to the user through the display device.

[0036] Figure 2 This is a flowchart illustrating a task decomposition method provided in an embodiment of this application. For example, the task decomposition method provided in this embodiment can be applied to... Figure 1 In the computing device shown, in other words, it can be used... Figure 1 The computing device shown executes the method; specifically, the method can be executed by the processor of the computing device.

[0037] like Figure 2 As shown, the task decomposition method provided in this application embodiment may include the following steps: S201. Obtain intent data.

[0038] Intent data is used to describe the task requirements that need to be achieved.

[0039] In this embodiment of the application, the computing device can directly obtain the intent data input by the user through the display device, or the computing device can supplement the content input by the user to obtain the supplemented intent data.

[0040] In one possible implementation, for scenarios where the computing device supplements the user's input with intent, such as... Figure 3 As shown, the above S201 can be specifically implemented as follows: S2011. Obtain the explicit intent of the user input.

[0041] S2012. Obtain the potential intents associated with the explicit intents from the domain knowledge base.

[0042] The domain knowledge base is used to store the knowledge system of the domain to which the task requirements belong.

[0043] Computing devices can extract explicit intents—that is, the task requirements explicitly stated by the user—from the original dialogue records between the user and the large language model. Further, the computing device queries a domain knowledge base (such as a knowledge graph) to infer the user's unspoken but potentially implicit intents. For example, when a user inputs the explicit intent "how to repair a car," the computing device recognizes the "repair a car" action, determines that the explicit intent belongs to the vehicle repair domain, and retrieves unspoken potential intents such as nearby repair shop locations, "repair tutorial searches," and "parts purchase channels" from the corresponding domain knowledge base.

[0044] S2013. Obtain user preference data from the user's historical behavior database.

[0045] The computing device retrieves the user's historical behavior database based on the current user's identity. If the database detects frequent usage records of specific algorithms (such as "Pareto optimal solution" and "cost-first strategy"), the identifier of that specific algorithm is used as user preference data. This user preference data can be used in subsequent task decomposition processes.

[0046] S2014. Integrate explicit intent, latent intent, and user preference data to obtain intent data.

[0047] Finally, the computing device integrates explicit intent, latent intent, and user preference data to form a high-dimensional information set (which can be called an intent tensor), thus obtaining intent data. This intent data serves as the foundational data source for subsequent task decomposition processes, ensuring that the task decomposition process not only conforms to the current context but also compensates for any oversights by the user and aligns with the user's deep-seated habits.

[0048] For example, a computing device may be equipped with an intent supplementation module (software module) that can supplement the explicit intent input by the user with potential intent and user preference data, ultimately resulting in richer intent data.

[0049] As an example, intent data supplemented with intent can be represented in JSON format: { Explicit intention: How to fix the car; Implicit intent: The attachment contains information on repair shops, repair manuals, and parts purchasing channels; User preference data: Cost-first strategy; } S202. Obtain global anchoring data from the intent data.

[0050] Among them, global anchoring data is used to indicate the core objectives of task requirements.

[0051] In this embodiment, the computing device analyzes the intent data to identify the core objective indicating the task requirements. This core objective serves as the benchmark for subsequent task decomposition, ensuring that the task decomposition results do not deviate from this core objective and guaranteeing the reliability of the task decomposition results.

[0052] For example, a user inputs a task requirement as: "Develop a travel plan to take the train to city A on Sunday, ensuring comfort and affordability." Here, the core objective of the task is: "Go to city A."

[0053] For example, the core objective can be transformed into a semantic vector for representation and storage, facilitating subsequent comparative analysis based on vector similarity.

[0054] In some implementations, where intent data is obtained through intent supplementation, the computing device can acquire global anchoring data based on the explicit intents within the intent data. It should be understood that explicit intents reflect the user's core needs; therefore, determining global anchoring data based on explicit intents ensures that subsequent task breakdowns do not deviate from the user's core needs.

[0055] In some implementations, global anchoring data is also used to indicate the constraints (or preset principles) that should be followed in achieving the aforementioned task requirements. Specifically, global anchoring data includes semantic vectors corresponding to the core objectives and a set of constraints. The computing device can extract these constraints from the user's short-term memory (such as historical conversation records) (e.g., the user explicitly mentions: repair costs should not exceed 1000 yuan), or from inherent long-term memory (e.g., laws that must be followed during repairs, prohibiting excessive vehicle modifications). This long-term memory can originate from user-uploaded rule documents or frequently referenced documents from the user's historical behavior.

[0056] The constraints obtained here can ensure the compliance of task decomposition in the subsequent task decomposition process, thereby improving the feasibility and reliability of the task decomposition results.

[0057] S203. Based on the global anchored data, the task is divided into dimensions to obtain the first decomposition result.

[0058] The first decomposition result is used to indicate the multiple independent sub-task dimensions that need to be considered to achieve the task requirements.

[0059] In this embodiment, the computing device can decompose the intent data based on the mutually exclusive and collectively exhaustive (MECE) principle to obtain a first decomposition result. During the decomposition process, globally anchored data is used to ensure that the task decomposition does not deviate from its intended path.

[0060] The MECE principle means mutually exclusive and exhaustive. This means that for a baseline problem, it should be decomposed into multiple simpler subtasks without omission or overlap, in order to find a solution to the baseline problem. The MECE principle ensures that all possible solutions to the baseline problem are considered, avoiding omissions and duplications, and guaranteeing the accuracy of problem solving.

[0061] For example, a MECE breadth engine (software module) is deployed in the computing device. This MECE breadth engine can decompose the intent data using a built-in breadth decomposition model to obtain a first decomposition result. The breadth decomposition model can be a pre-trained lightweight model, and the specific structure of the model is not limited in the embodiments of this application.

[0062] Specifically, such as Figure 4 As shown, the above S203 can be specifically implemented as follows: S2031. Different decomposition strategies are used to decompose the intent data to obtain a set of multiple candidate subtasks.

[0063] In this approach, each decomposition strategy corresponds to a set of candidate subtasks. Different decomposition strategies can be preference rules indicated by user preference data, such as efficiency-first strategies, cost-first strategies, etc.

[0064] Due to the different decomposition strategies employed, computing devices will decompose the same intent data into different sets of candidate subtasks. Each set of candidate subtasks includes multiple subtasks, and each subtask corresponds to a different dimension.

[0065] For example, a computing device can pre-configure a range of random seed values, with each range corresponding to a different splitting strategy. Different splitting strategies are randomly selected using the random seed to split the intent data.

[0066] For example, for the intent data of "specifying a travel plan", using a "cost-first strategy", the resulting set of candidate subtasks is as follows: 1. Major transportation (round-trip airfare) 2. Local transportation (bus + subway) 3. Accommodation (budget hotel) 4. Sightseeing (tickets + route) 5. Food and Beverages (Snacks + Light Meals) 6. Overall Itinerary Control (Timetable) 7. Budget Management For example, taking the same intent data of "specifying a travel plan" as an example, if an "experience-first strategy" is adopted, the resulting set of candidate subtasks includes the following subtasks: 1. Transportation (round-trip airfare + airport transfer) 2. Local transportation (chartered vehicle for the entire trip) 3. Accommodation (Seaview Family Hotel) 4. Scenic Spot Tour (In-depth Experience + Travel Photography) 5. Dining (Specialty Restaurants + Afternoon Tea) 6. Overall Itinerary Control (Daily Route) 7. Value-added services (tour guide + insurance) S2032. Based on the global anchoring data, perform integrity verification on multiple candidate subtask sets to obtain multiple verified candidate subtask sets.

[0067] Specifically, the integrity verification process is as follows: For each candidate subtask set in the multiple candidate subtask sets, determine whether the subtask dimensions of the candidate subtask set cover all dimensions that should be considered in the global anchor data; if the subtask dimensions of the candidate subtask set cannot cover all dimensions that should be considered in the global anchor data, supplement the candidate subtask set with subtasks of missing dimensions.

[0068] Among these, all dimensions that should be considered for globally anchored data can be pre-defined according to the task type. For example, for travel-related tasks, transportation methods, accommodation, and dining all need to be considered. For equipment testing tasks, equipment operational stability and the correctness of the results all need to be considered.

[0069] Computing devices can use keyword coverage algorithms to determine whether the subtask dimensions of a candidate subtask set cover all dimensions that should be considered in the global anchoring data. For example, if the current candidate subtask set includes keywords in dimensions A, B, and C, while the global anchoring data should consider keywords in dimensions A, B, C, and D, and it is determined that the current candidate subtask set is insufficient in considering the user's original intent, then subtasks in dimension D will be added to the candidate subtask set.

[0070] S2033. Take the intersection of the multiple candidate subtask sets after verification to obtain the first decomposition result.

[0071] It should be understood that if different decomposition strategies can be used to decompose each subtask in each dimension, then the subtask is the most reliable and essential for fulfilling the task requirements. Therefore, the solution in this application takes the intersection of multiple candidate subtask sets to obtain a more stable and reliable first decomposition result.

[0072] In some implementations, during the process of obtaining the first decomposition result, the computing device also performs the following: for each candidate subtask set, if the functional overlap between the first subtask and the second subtask is greater than a threshold, the first subtask and the second subtask are merged; the first subtask and the second subtask are any two subtasks in the candidate subtask set.

[0073] In some implementations, the degree of functional overlap can be determined based on the semantic similarity between the first and second subtasks. Higher semantic similarity indicates a greater degree of functional overlap.

[0074] In other words, for each set of candidate subtasks, the computing device determines the semantic similarity between the subtasks (e.g., using a cosine similarity algorithm). If the semantic similarity is greater than a threshold, the two subtasks are considered to have overlapping functions, triggering subtask merging. Alternatively, the computing device re-executes the decomposition process. This approach ensures that the decomposed subtasks are mutually exclusive and exhaustive, guaranteeing the comprehensiveness of subsequent execution plans.

[0075] For example, in a device testing scenario, the candidate subtask set includes two subtasks: "hardware testing" and "reliability verification." However, these two subtasks essentially convey the same meaning, so they are merged.

[0076] S204. Based on global anchoring data, perform path planning on the task execution process to obtain the second decomposition result.

[0077] The second decomposition result indicates the sequence of atomic actions that need to be executed sequentially to fulfill the task requirements. An atomic action refers to an executable method at the smallest granularity.

[0078] In this embodiment, the computing device can decompose the intent data based on a logical depth-first principle to obtain a second decomposition result. During the decomposition process, globally anchored data is used to ensure that the task decomposition does not deviate from its intended path.

[0079] The computing device can employ a planning algorithm from a hierarchical task network (HTN). By relying on a predefined method library (which defines the possible decomposition methods for an abstract task to obtain a sequence of atomic actions), the computing device understands how to perform path planning on an abstract task to obtain a specific sequence of atomic actions. For example, "travel" can be decomposed into "booking tickets" - "booking a hotel" - "packing a tour". Through the above decomposition method, it can be ensured that the second decomposition result is logically executable and conforms to the methodological specifications of this field.

[0080] It should be noted that during the decomposition process, if a certain atomic action obtained is still an abstract task, then the decomposition continues recursively until the smallest granularity atomic action is obtained. Moreover, because the decomposition is reduced to the smallest granularity, the computational complexity of subsequent processing can be reduced, thus lowering inference latency and computing power costs.

[0081] For example, a computing device may deploy an HTN deep engine (software module). This HTN deep engine can decompose intent data into a second decomposition result through a built-in hierarchical task network. The hierarchical task network may be a pre-trained network model, and the specific structure of the model is not limited in this embodiment.

[0082] In one possible implementation, such as Figure 5 As shown, the above S204 can be specifically implemented as follows: S2041. Different decomposition strategies are used to decompose the task requirements to obtain multiple candidate sequences.

[0083] In this approach, one decomposition strategy corresponds to one candidate sequence, and multiple candidate sequences are used to indicate different execution paths to achieve the task requirements.

[0084] Due to different decomposition strategies employed, computing devices will decompose the same intent data into different candidate sequences, each corresponding to a different execution path. For example, for the intent data "go to a certain city," if the decomposition strategy prioritizes low cost, the resulting candidate sequence could be travel by train. If the decomposition strategy prioritizes high efficiency, the resulting candidate sequence could be travel by plane.

[0085] S2042. Based on global anchoring data, perform compliance verification on multiple candidate sequences to obtain multiple verified candidate sequences.

[0086] In one possible implementation, the compliance verification process is as follows: for each candidate sequence among multiple candidate sequences, determine whether the candidate sequence deviates from the core objective; if the candidate sequence deviates from the core objective, delete the candidate sequence or adjust the first atomic action in the candidate sequence; the first atomic action is the key node that caused the candidate sequence to deviate from the core objective.

[0087] In other words, if a candidate sequence obtained from the decomposition deviates from the core objective, the candidate sequence can be removed, or the key node causing the deviation can be fed back to the HTN depth engine so that the HTN depth engine can adjust the candidate sequence to regenerate a more reasonable candidate sequence.

[0088] For example, a user's original intention is to travel to the more distant city B, but the task is broken down too much to keep costs down, and the output candidate sequence is the steps to travel to the closer city A. Although this partially meets the needs in terms of cost, it has deviated from the user's core goal in the whole.

[0089] In another possible implementation, the globally anchored data is also used to indicate the constraints that should be followed in fulfilling the task requirements. The compliance verification process also includes the following: for each of the multiple candidate sequences, determine whether there is a second atomic action in the candidate sequence that violates the constraints; if there is a second atomic action in the candidate sequence that violates the constraints, delete the candidate sequence or adjust the second atomic action in the candidate sequence.

[0090] In other words, if a candidate sequence obtained from the decomposition contains a second atomic action that violates the constraints (for example, the constraint is safety first, but the candidate sequence involves playing dangerous activities), then the candidate sequence can be removed. Alternatively, the second atomic action that violates the constraints can be reported back to the HTN depth engine so that the HTN depth engine can adjust the candidate sequence and regenerate a more reasonable candidate sequence.

[0091] In summary, compliance verification can, on the one hand, check whether the task breakdown deviates from the core objective, and on the other hand, check whether the atomic actions derived from the breakdown meet the constraints. This addresses the issue of large language models paying less attention to the content in the middle of the text, ensuring that the task breakdown can output reliable content that meets user needs in sequence, without requiring users to repeatedly interact and modify it, thus improving the user experience.

[0092] S2043. Take the intersection of the multiple candidate sequences after verification to obtain the second decomposition result.

[0093] It should be understood that if a certain part of the atomic action sequence can be decomposed using different decomposition strategies, then this part of the atomic action sequence is the most reliable and is the necessary path to achieve the task requirements. Therefore, the solution in this application takes the intersection of multiple candidate sequences to obtain structurally similar atomic action sequences, thereby obtaining a more stable and reliable second decomposition result.

[0094] It should be understood that breaking down a task into a sequence of atomic actions, ensuring that it is executable at the logical depth level, is a prerequisite for ensuring that the specific execution plan can be executed effectively.

[0095] For example, a game theory engine (software module) is deployed in the computing device. This engine acts as a referee for both the MECE breadth engine and the HTN depth engine, using globally anchored data to guide the decomposition process of both engines. This allows the engine to automatically identify and eliminate results that, while locally reasonable, deviate from the overall objective. Alternatively, it can automatically identify and eliminate results that, while not deviating from the overall objective, violate local constraints. In this way, the task decomposition method of this solution, like a human expert with "short-term memory + long-term memory," always stays focused on the core objective, eliminates logical illusions, and makes the output more accurate and reliable.

[0096] S205. Based on the first and second decomposition results, generate a specific execution plan to implement the task and satisfy the global anchoring data.

[0097] In this embodiment, a first decomposition result of "breadth coverage" and a second decomposition result of "depth executability" are obtained. Both the first and second decomposition results have been validated using globally anchored data and maintain a long-chain thinking effect. Therefore, by fusing the first and second decomposition results, the computing device can obtain a logically executable and comprehensively considered specific execution plan.

[0098] Optionally, the specific execution plan output can be displayed to the user for reference via a display device, or the computing device can input the specific execution plan into the automated execution device, which can then automatically execute the plan to achieve the task requirements.

[0099] In some implementations, the game engine deployed in the computing device is also used to supplement the content of the second decomposition result based on the first decomposition result. For example, continuing with the aforementioned tourism scenario, the first decomposition result indicates the subtask of "considering accommodation when traveling," while the atomic action sequence of the second decomposition result does not contain execution logic related to "accommodation." Therefore, the game engine feeds back to the HTN deep engine, instructing it to supplement the logical steps related to "booking accommodation."

[0100] In some implementations, the method further includes obtaining user feedback: obtaining user feedback data on a specific implementation plan; updating user preference data based on the feedback data; and storing the updated user preference data in the historical behavior database.

[0101] The computing device can capture user feedback data such as "accept," "modify," or "reject" based on the user's final selection. It then updates user preference data, such as the user's final task breakdown strategy and preference algorithm, based on this feedback. The computing device maps the updated user preference data into feature codes and stores them in a historical behavior database. During the next task breakdown, this updated user preference data can serve as the user's historical preferences, automatically adjusting the breakdown strategy (e.g., automatically switching from "time-first" to "cost-first") to achieve personalized adaptive service.

[0102] Figure 6 This is a schematic diagram illustrating a complete process provided for an embodiment of this application. For example... Figure 6 As shown, the process begins with: 1. The user inputs the original task (explicit intent). 2. The computing device completes the intent based on the knowledge graph and historical behavior database, obtaining intent data. 3. The computing device decomposes the intent data into tasks. 3.1. The MECE breadth engine is used to decompose the intent data, obtaining the first decomposition result. 3.2. The HTN depth engine is used to decompose the intent data, obtaining the second decomposition result. 4. The game engine, based on the results output by the two engines, performs long-chain thinking based on global anchoring data to ensure the rationality of the decomposition process. Global anchoring data can come from short-term memory (mainly providing core objectives) and persistent memory (mainly providing constraints). 5. For the returned specific execution plan, the system interacts with the user to obtain user feedback data. 6. New user preference data is generated based on the feedback data and stored in the historical behavior database for future use in decomposition. 7. It is determined whether the specific execution plan meets the user's expectations. If so, the process ends. If not, the process returns to step 1 to readjust the original task.

[0103] In the task decomposition method provided in this application embodiment, the computing device acquires intent data describing task requirements, and obtains global anchoring data based on the intent data. This global anchoring data is used to indicate the core objective of the task requirements. Further, the task is dimensionally divided based on the global anchoring data to obtain a first decomposition result. The execution process of the task is then path-planned to obtain a second decomposition result. The first decomposition result ensures broad coverage during task decomposition, indicating what needs to be addressed during the task decomposition process. The second decomposition result ensures that the atomic action sequence obtained from the decomposition is logically executable, indicating how to specifically perform tasks during the decomposition process. Additionally, global anchoring data is introduced during the decomposition process to ensure that the decomposition result does not deviate from the core objective. Through the mutual constraints of global anchoring data, the first decomposition result, and the second decomposition result, the final execution plan output after task decomposition is realized, simultaneously possessing the effects of "broad coverage," "deep executableness," and "long-term logical conservation" (i.e., not deviating from the core objective), thereby effectively improving the accuracy and reliability of task decomposition.

[0104] This application's embodiments ensure broad coverage (no omissions, no overlaps) through the MECE breadth engine, ensure deep execution (conforming to domain specifications, atomicity) through the HTN depth engine, and ensure long-term logic conservation (anti-drift, high robustness) through the game engine, enabling the output of more comprehensive and reliable specific execution schemes, supporting the further development of large language models in autonomous execution capabilities.

[0105] As can be seen, the above mainly describes the solutions provided by the embodiments of this application from a methodological perspective. To achieve the above functions, the embodiments of this application provide corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the modules and algorithm steps of the various examples described in the embodiments disclosed herein, the embodiments of this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware 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.

[0106] In an exemplary embodiment, this application also provides a task breakdown apparatus. The task breakdown apparatus may be the aforementioned computing device or a processor within the computing device. The task breakdown apparatus may include one or more functional modules for implementing the task breakdown method of the above method embodiments.

[0107] For example, Figure 7 This is a schematic diagram illustrating the composition of a task disassembly device provided in an embodiment of this application. Figure 7As shown, the task disassembly device 700 includes: an acquisition module 701, a disassembly module 702, and a generation module 703. The acquisition module 701, the disassembly module 702, and the generation module 703 are interconnected.

[0108] The acquisition module 701 is used to acquire intent data; the intent data is used to describe the task requirements that need to be achieved. The acquisition module 701 is also used to acquire global anchoring data based on the intent data; the global anchoring data is used to indicate the core objectives of the task requirements. The decomposition module 702 is used to divide the task into dimensions based on global anchor data to obtain the first decomposition result; the first decomposition result is used to indicate the multiple independent sub-task dimensions that need to be considered to achieve the task requirements. The decomposition module 702 is also used to perform path planning on the execution process of the task based on global anchoring data to obtain a second decomposition result; the second decomposition result is used to indicate the sequence of atomic actions that need to be executed sequentially to achieve the task requirements; The generation module 703 is used to generate a specific execution plan for implementing tasks and satisfying the global anchoring data based on the first and second decomposition results.

[0109] Other optional implementation methods are described in the foregoing method embodiments.

[0110] The method provided in this application embodiment can be applied to the following scenarios: Scenario 1: Auditing the reasonableness of suggestions in an intelligent customer service system In intelligent customer service scenarios, customer service models often need to output processing suggestions based on user descriptions. Without this solution, the customer service model struggles to provide comprehensive and accurate processing suggestions when answering complex user requests. By introducing this solution, the customer service model can rationally break down user needs, improving the reliability of the customer service suggestions output.

[0111] Scenario 2: Automated office work using automated systems.

[0112] In the field of automation, users may only input brief requirements, such as "Please create an order form." Due to the lack of detail, the output of the automation system is likely to not meet the user's expectations and requires multiple adjustments. By introducing this application, the automation system can reasonably break down the user's requirements, obtain accurate and specific execution plans, and automatically execute tasks based on these plans to fulfill the user's needs.

[0113] This application also provides a computer-readable storage medium. All or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware; for example, the related hardware can be a processor of a computing device. The program instructions can be stored in the above-described computer-readable storage medium, and when executed, they can implement the processes of the above method embodiments. The computer-readable storage medium can be memory. The above-described computer-readable storage medium can also be an external storage device, such as a hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Further, the above-described computer-readable storage medium can include both memory and external storage devices. The above-described computer-readable storage medium is used to store the above-described computer program instructions and other programs and data required for the above-described software package translation.

[0114] This application also provides a computer program product, which includes a computer program that, when run on a computing device, causes the computing device to execute any of the task breakdown methods provided in the above embodiments.

[0115] The modeling method based on structural connectors in this application can be widely applied to the following industry scenarios: It is applicable to multiple fields such as mechanical manufacturing, aerospace, automotive industry, and construction engineering, and shows significant advantages, especially in complex assemblies with a large number of structural connectors (such as bolts and rivets).

[0116] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0117] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

[0118] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A task decomposition method, characterized in that, The method includes: Acquire intent data; the intent data is used to describe the task requirements that need to be achieved. Global anchoring data is obtained based on the intent data; the global anchoring data is used to indicate the core objectives of the task requirements; Based on the global anchoring data, the task is divided into dimensions to obtain a first decomposition result; the first decomposition result is used to indicate the multiple independent sub-task dimensions that need to be considered to achieve the task requirements. Based on the global anchoring data, the execution process of the task is path-planned to obtain a second decomposition result; the second decomposition result is used to indicate the sequence of atomic actions that need to be executed sequentially to achieve the task requirements; Based on the first and second decomposition results, a specific execution plan is generated to implement the task and satisfy the global anchoring data.

2. The method according to claim 1, characterized in that, Based on the global anchoring data, the task is divided into dimensions to obtain a first decomposition result, including: The intent data is decomposed using different decomposition strategies to obtain multiple candidate subtask sets; wherein, one decomposition strategy corresponds to one candidate subtask set. Based on the global anchoring data, the integrity of the multiple candidate subtask sets is verified to obtain the verified multiple candidate subtask sets. The first decomposition result is obtained by taking the intersection of the multiple candidate subtask sets after verification.

3. The method according to claim 2, characterized in that, The integrity verification of the multiple candidate subtask sets based on the global anchoring data includes: For each candidate subtask set in the plurality of candidate subtask sets, determine whether the subtask dimensions of the candidate subtask set cover all dimensions that should be considered in the global anchoring data; If the subtask dimensions of the candidate subtask set cannot cover all dimensions that should be considered for the global anchoring data, the candidate subtask set is supplemented with subtasks that have missing dimensions.

4. The method according to claim 2 or 3, characterized in that, The method further includes: For each set of candidate subtasks, if the functional overlap between the first subtask and the second subtask is greater than a threshold, the first subtask and the second subtask are merged; the first subtask and the second subtask are any two subtasks in the set of candidate subtasks.

5. The method according to any one of claims 1-4, characterized in that, Based on the global anchoring data, the execution process of the task requirements is path-planned to obtain a second decomposition result, including: The task requirements are decomposed using different decomposition strategies to obtain multiple candidate sequences; one decomposition strategy corresponds to one candidate sequence; the multiple candidate sequences are used to indicate different execution paths to achieve the task requirements; Based on the global anchoring data, compliance verification is performed on the multiple candidate sequences to obtain multiple verified candidate sequences; The intersection of the multiple candidate sequences after verification is taken to obtain the second decomposition result.

6. The method according to claim 5, characterized in that, The compliance verification of the multiple candidate sequences based on the global anchoring data includes: For each of the plurality of candidate sequences, determine whether the candidate sequence deviates from the core objective; If the candidate sequence deviates from the core objective, the candidate sequence is deleted, or the first atomic action in the candidate sequence is adjusted; the first atomic action is the key node that causes the candidate sequence to deviate from the core objective.

7. The method according to claim 6, characterized in that, The global anchoring data is also used to indicate the constraints that should be followed in the process of achieving the task requirements; The compliance verification of the multiple candidate sequences based on the global anchoring data includes: For each of the plurality of candidate sequences, determine whether there is a second atomic action in the candidate sequence that violates the constraint condition; If a second atomic action in the candidate sequence violates the constraint, delete the candidate sequence or adjust the second atomic action in the candidate sequence.

8. The method according to any one of claims 1-7, characterized in that, The acquisition intent data includes: Obtain the explicit intent of the user's input; Retrieve potential intentions associated with the explicit intentions from a domain knowledge base; the domain knowledge base is used to store the knowledge system of the domain to which the task requirement belongs; Obtain user preference data from the user's historical behavior database; The explicit intent, the latent intent, and the user preference data are combined to obtain the intent data.

9. The method according to claim 8, characterized in that, The method further includes: Obtain user feedback data on the specific execution plan; The user preference data is updated based on the feedback data, and the updated user preference data is stored in the historical behavior database.

10. A computing device, characterized in that, The computing device includes a processor and a memory; the processor is coupled to the memory. The memory is used to store computer instructions; The computer instructions are loaded and executed by the processor to enable the computing device to implement the task breakdown method as described in any one of claims 1-9.

Images