Multi-intention instruction execution method and device, equipment, medium and program product

Abstract

The invention discloses a multi-intention instruction execution method and device, equipment, a medium and a program product. Under the condition that the language instruction is a multi-intention instruction, generating a first type of prompt information based on a preset thinking prompt template; performing intention segmentation on the language instruction by utilizing a trained large language model according to a thinking mode in the first type of prompt information to obtain a plurality of sub-intention instructions; on the basis of the first type of prompt information, intention completion is conducted on the multiple sub-intention instructions, and multiple target sub-intention instructions are obtained; determining an execution relationship among the plurality of target sub-intention instructions by utilizing a large language model; and controlling the target device to execute the target sub-intention instruction according to the execution relation. According to the embodiment of the invention, the multi-intention instruction can be accurately understood in a complex semantic scene.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a method, apparatus, device, medium and program product for executing multi-intent instructions. Background Technology

[0002] In recent years, home smart assistant functions have not only been able to control home appliances but also provide users with various life services in different living scenarios.

[0003] As the number of devices integrated into a smart home increases and their functions become more complex, users often issue multiple commands in a single sentence when interacting with smart assistants, expecting the assistant to complete multiple tasks at once. However, such multi-intent commands are semantically complex, posing a significant challenge to the smart assistant's understanding and execution.

[0004] In simple interactive scenarios where sub-intents are clearly distinguishable, the relevant technologies can handle the situation. However, when encountering complex situations with semantic ambiguity caused by users' colloquial expressions, such as when sub-intents are nested, the boundaries between preceding and following intents are not clear, or the expressions are too colloquial, the accuracy of the relevant technologies will be severely limited, making it difficult to meet the actual needs of users.

[0005] Therefore, the accuracy of related technologies in understanding multi-intent instructions in complex semantic scenarios is significantly insufficient. Summary of the Invention

[0006] This application provides a method, apparatus, device, medium, and program product for executing multi-intent instructions, which can accurately understand multi-intent instructions in complex semantic scenarios.

[0007] In a first aspect, embodiments of this application provide a method for executing multi-intent instructions, comprising: When the language instruction is a multi-intent instruction, the first type of prompt information is generated based on the preset thought prompt template; Using a pre-trained large language model, the language instructions are segmented according to the thought process in the first type of prompt information to obtain multiple sub-intention instructions; based on the first type of prompt information, the multiple sub-intention instructions are completed to obtain multiple target sub-intention instructions. Using a large language model, the execution relationships between multiple target sub-intent instructions are determined; According to the execution relationship, control the target device to execute the target sub-intention instruction.

[0008] Based on the same inventive concept, in a second aspect, embodiments of this application also provide an execution device for multi-intent instructions, comprising: The generation module is used to generate the first type of prompt information based on a preset thought prompt template when the language instruction is a multi-intent instruction. The segmentation and completion module is used to segment language instructions according to the thought process in the first type of prompt information using a pre-trained large language model, resulting in multiple sub-intention instructions; and to complete the multiple sub-intention instructions based on the first type of prompt information, resulting in multiple target sub-intention instructions. The determination module is used to determine the execution relationship between multiple target sub-intent instructions using a large language model; The execution module is used to control the target device to execute the target sub-intent instructions according to the execution relationship.

[0009] Based on the same inventive concept, in a third aspect, embodiments of this application also provide a multi-intent instruction execution device, the device including a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the first aspect, or the multi-intent instruction execution method in any embodiment of the first aspect.

[0010] Based on the same inventive concept, in a fourth aspect, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the execution method of the first aspect, or the multi-intent instructions in any embodiment of the first aspect.

[0011] Based on the same inventive concept, in a fifth aspect, embodiments of this application also provide a computer program product, wherein when the instructions in the computer program product are executed by the processor of a device, the device is able to execute the execution method of the multi-intent instructions in the first aspect or any embodiment of the first aspect.

[0012] The multi-intent instruction execution method, apparatus, device, medium, and program product of this application, when the language instruction is a multi-intent instruction, can generate a first type of prompt information based on a pre-set thinking prompt template. This first type of prompt information serves as heuristic prompt information for a large language model, guiding the model on how to think when segmenting and completing multi-intent instructions. The large language model can then segment the language instruction according to the thinking method in the first type of prompt information, obtaining multiple sub-intent instructions. Next, based on the first type of prompt information, it performs intent completion on the multiple sub-intent instructions, obtaining multiple target sub-intent instructions. For example, the first type of prompt information can guide the large language model to segment and complete multi-intent instructions using a thinking chain of "language instruction → intent segmentation thinking process → intent segmentation answer → intent completion thinking process → intent completion answer → final answer," providing a heuristic reasoning approach for the large language model. Through step-by-step thinking, the potential of the large language model can be fully explored, improving the accuracy of multi-intent segmentation and completion. Subsequently, the large language model can also orchestrate multiple target sub-intent instructions to determine the execution relationship between them, and control the target device to execute the target sub-intent instructions according to this relationship. This effectively improves the accuracy and efficiency of understanding multiple intent instructions in complex semantic scenarios. This application's embodiment uses a large language model combined with thought chain technology. By formulating appropriate prompts (appropriate thinking methods) through the thought chain format, it inspires and guides the large language model to perform intent segmentation and completion. This overcomes the limitations of related technologies that are result-oriented, enabling the model to tap into its deeper potential in complex semantic scenarios, achieving a leap from "recognition result" to "understanding process," thereby significantly improving the accuracy and efficiency of understanding multiple intent instructions in complex semantic scenarios and meeting the actual needs of users. Attached Figure Description

[0013] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings, in which the same or similar reference numerals denote the same or similar features, and the drawings are not drawn to scale.

[0014] Figure 1 This is a flowchart illustrating a method for executing multi-intent instructions provided in an embodiment of this application; Figure 2 This is a flowchart illustrating intent understanding in the multi-intent instruction execution method provided in this application embodiment; Figure 3 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 4 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 5 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 6 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 7 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 8 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 9 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application; Figure 10 This is a schematic diagram of a multi-intent instruction execution system provided in an embodiment of this application; Figure 11 This is a schematic diagram of a multi-intent instruction execution device provided in an embodiment of this application; Figure 12 This is a schematic diagram of a multi-intent instruction execution device provided in an embodiment of this application. Detailed Implementation

[0015] The features and exemplary embodiments of various aspects of this application will now be described in detail. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only configured to explain this application and are not configured to limit this application. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples of this application.

[0016] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0017] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0018] Various modifications and variations can be made to this application without departing from its spirit or scope, which will be apparent to those skilled in the art. Therefore, this application is intended to cover modifications and variations falling within the scope of the corresponding claims (the claimed technical solutions) and their equivalents. It should be noted that the implementation methods provided in the embodiments of this application can be combined with each other without contradiction.

[0019] Before describing the technical solutions provided in the embodiments of this application, in order to facilitate understanding of the embodiments of this application, this application first specifically explains the problems existing in the related technologies: In recent years, home smart assistant functions have not only been able to control home appliances but also provide users with various life services in different living scenarios.

[0020] As the number of devices integrated into a smart home increases and their functions become more complex, users often issue multiple commands in a single sentence when interacting with smart assistants, expecting the assistant to complete multiple tasks at once. However, such multi-intent commands are semantically complex, posing a significant challenge to the smart assistant's understanding and execution.

[0021] In simple interaction scenarios where sub-intents are clearly distinguishable, the relevant technologies can handle the situation. However, in actual interaction, users' colloquial expressions often lead to complex scenarios with semantic ambiguity, such as nested sub-intents, unclear boundaries between preceding and following intents, and overly colloquial expressions. In such complex semantic scenarios, the accuracy of the relevant technologies in understanding these scenarios is severely limited, making it difficult to meet the actual needs of users.

[0022] For example, in related technologies, intelligent assistants typically use a binary classification model to directly determine whether a message has multiple intents, followed by intent segmentation using a sequence labeling model. Sequence labeling models are a classic method in Natural Language Processing (NLP) for processing sequence data, suitable for scenarios with clear structure and easily distinguishable sub-intents. However, in complex multi-intent interactions, their lack of deep reasoning ability becomes amplified. In other words, sequence labeling models perform well for "simple" interactions with clearly distinguishable sub-intents, but multi-intent interactions often present semantically ambiguous and complex scenarios. Using sequence labeling models encounters accuracy bottlenecks, failing to accurately understand multi-intent commands in complex semantic contexts. This is because while sequence labeling models can identify the result, they do not reason out the thought process that led to that result from a deeper perspective, resulting in low accuracy in complex interaction scenarios.

[0023] Therefore, the accuracy of related technologies in understanding multi-intent instructions in complex semantic scenarios is significantly insufficient.

[0024] Based on this, embodiments of this application provide a method, apparatus, device, medium, and program product for executing multi-intent instructions. It employs a large language model combined with thought chain technology, using a thought chain format to formulate appropriate prompts (appropriate thinking methods) to inspire and guide the large language model in intent segmentation and completion. This overcomes the limitations of related technologies that are result-oriented, enabling the exploration of the deep potential of the large model in complex semantic scenarios. It achieves a leap from "recognizing results" to "understanding the process," thereby significantly improving the accuracy and efficiency of understanding multi-intent instructions in complex semantic scenarios and meeting the actual needs of users.

[0025] The method for executing multi-intent instructions provided in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0026] Figure 1 This is a schematic flowchart of a multi-intent instruction execution method provided in an embodiment of this application, such as... Figure 1 As shown, the method may include steps S110 to S140.

[0027] S110, when the language instruction is a multi-intent instruction, generate the first type of prompt information based on the preset thought prompt template.

[0028] Language commands are instructions issued by the user to instruct the device to perform specific tasks. Language commands can be single-intent commands or multi-intent commands.

[0029] A multi-intention instruction is a language instruction that contains multiple intentions, meaning that a single instruction expresses multiple desired goals.

[0030] For example, considering that multi-intent interactions account for a relatively small proportion in daily interactions, it is necessary to quickly reject most single-intent data through multi-intent classification. Therefore, a 0.5B parameter (containing 500 million parameters) large language model can be used. This model belongs to the medium-sized model category, between small models (such as those with a few million parameters) and ultra-large models (such as those with hundreds of billions of parameters). The 0.5B parameter large language model, through the balance of "moderate scale," can ensure a certain semantic understanding capability while avoiding the high computational cost of ultra-large models. It can perform fast classification and directly output results, with a single response time kept within 50ms.

[0031] The thought prompt template is a pre-set template used to generate prompt information. This template specifies the format for generating prompt information and provides thought guidance for the subsequent processing of multi-intent commands by the Large Language Model (LLM).

[0032] For example, a mind prompt template can include role requirements, target function requirements, device information (such as device name, device function, etc.), input format requirements, and output format requirements. For instance, a mind prompt template could be: "Please play the role of {role} in a family setting, and you can achieve functions such as {target} for the user. You have the ability to control devices such as {device name}. The specific capabilities of each device are as follows: {device function}. Both input and output are in pure JSON format, with input format being {...} and output format being {...}."

[0033] The first type of prompt is generated based on a preset thought prompt template when the language instruction is a multi-intent instruction. As a heuristic prompt for the large language model, it prompts the large language model on how to segment and complete multi-intent instructions.

[0034] For example, the first type of prompt could be: "Please play the role of {multi-intent segmentation and completion} in a home scenario, and can perform {intent segmentation and intent completion functions for language commands} for the user. You have the ability to control devices such as {Device 1, Device 2, ...}. The device information for each device is as follows: {...}. Please think step by step and provide the answer according to the thought process of 'language command -> intent segmentation thought process -> intent segmentation answer -> intent completion thought process -> intent completion answer -> final answer'. Both input and output are in pure JSON format, with the input format being {...} and the output format being {...}."

[0035] Specifically, step S110 can be understood as follows: after receiving the user's language instruction, first determine whether the language instruction is a multi-intent instruction. If the language instruction is a multi-intent instruction, then generate intention segmentation and completion prompt information based on the preset thought prompt template, which is the first type of prompt information.

[0036] S120: Using a pre-trained large language model, the language instruction is segmented according to the thought process in the first type of prompt information to obtain multiple sub-intention instructions; based on the first type of prompt information, the multiple sub-intention instructions are completed to obtain multiple target sub-intention instructions.

[0037] It should be noted that before executing step S120, sample data (such as historical multi-intent instructions and first-class historical prompt information generated based on thought prompt templates) can be collected as input to the initial large language model, and actual results (such as historical sub-intent instructions after segmentation of historical multi-intent instructions and historical target sub-intent instructions after completion) can be collected as ideal output. The initial large language model is then trained multiple times to obtain a trained large language model. The trained large language model includes pre-trained segmentation rules and completion rules.

[0038] In multi-intent scenarios, it is necessary to first identify whether the dialogue is multi-intent. If it is multi-intent, intent segmentation and completion are required. This process can be called multi-intent understanding.

[0039] In multi-intent understanding scenarios, semantically ambiguous and complex scenarios often arise. Although related algorithms can identify the results, they cannot provide the reasoning process for obtaining the results, and they will encounter accuracy bottlenecks in complex interaction scenarios.

[0040] For example, such as Figure 2 As shown, multi-intent understanding mainly consists of three functions: intent classification, intent segmentation, and intent completion. Multi-intent segmentation and completion can be performed within the same algorithm. The specific process of multi-intent recognition, intent segmentation, and completion can include steps S01 to S05.

[0041] S01 receives language commands from the user.

[0042] S02, determine whether the language instruction is a multi-intent language instruction. If yes, proceed to step S03; otherwise, proceed to step S05.

[0043] S03, when the language instruction is a multi-intent instruction, a first type of prompt information is generated based on a preset thought prompt template. Then, using a pre-trained large language model, the language instruction is segmented according to the thought process in the first type of prompt information to obtain multiple sub-intent instructions.

[0044] S04. Using the pre-trained large language model, based on the first type of prompt information, perform intent completion on multiple sub-intent instructions to obtain multiple target sub-intent instructions.

[0045] S05, return the result.

[0046] Specifically, step S120 can be understood as follows: using the pre-trained large language model, the language instruction can be segmented according to the intention segmentation thinking method in the first type of prompt information to obtain multiple sub-intention instructions. At this time, the information contained in each sub-intention instruction may be incomplete. Therefore, the pre-trained large language model can be used to complete the multiple sub-intention instructions according to the intention completion thinking method in the first type of prompt information to obtain multiple intention-completed sub-intention instructions, that is, the target sub-intention instruction.

[0047] Step S120 uses a pre-trained large language model to perform multi-intent understanding. Appropriate prompts (such as the first type of prompts) are formulated using the Chain of Thought format (thought prompt template). This allows the large language model to provide judgment and reasoning processes while performing intent understanding, thereby fully leveraging the powerful reasoning and inductive abilities of the large language model. It can simulate the human thought process, perform in-depth semantic analysis during thinking, judge and segment each complex sub-intent, and complete the semantics for still ambiguous expressions, thereby improving the accuracy of multi-intent understanding.

[0048] It should be noted that multi-intent segmentation and completion require complex semantic understanding. This application's embodiments utilize a chain of thought, enabling the large model to provide judgment criteria step-by-step through a heuristic reasoning process in a pipeline manner. Intent segmentation and completion prompts (first type of prompts) are generated using thought prompt templates, such as: language instruction + intent segmentation thought process + intent segmentation answer + intent completion thought process + intent completion answer + final answer. Inputting this thought process into the large language model allows for the full exploitation of the model's potential through step-by-step thinking, thereby improving the accuracy of multi-intent segmentation and completion.

[0049] S130 uses a large language model to determine the execution relationship between multiple target sub-intent instructions.

[0050] Specifically, large language models can be used for intent orchestration, that is, to determine the execution relationships between multiple target sub-intents. For example, multiple target sub-intents can be directly input into the large language model, allowing the model to analyze the execution relationships between them. Alternatively, multiple reference execution relationships can be pre-set, allowing the model to analyze which of the multiple reference execution relationships the execution relationships between the target sub-intents belong to.

[0051] S140, according to the execution relationship, control the target device to execute the target sub-intention instruction.

[0052] Specifically, the target device can be controlled to execute the target sub-intention instructions according to the execution relationship between multiple target sub-intention instructions, that is, the execution order of each target sub-intention instruction.

[0053] The multi-intent instruction execution method, apparatus, device, medium, and program product of this application, when the language instruction is a multi-intent instruction, can generate a first type of prompt information based on a pre-set thinking prompt template. This first type of prompt information serves as heuristic prompt information for a large language model, guiding the model on how to think when segmenting and completing multi-intent instructions. The large language model can then segment the language instruction according to the thinking method in the first type of prompt information, obtaining multiple sub-intent instructions. Next, based on the first type of prompt information, it performs intent completion on the multiple sub-intent instructions, obtaining multiple target sub-intent instructions. For example, the first type of prompt information can guide the large language model to segment and complete multi-intent instructions using a thinking chain of "language instruction → intent segmentation thinking process → intent segmentation answer → intent completion thinking process → intent completion answer → final answer," providing a heuristic reasoning approach for the large language model. Through step-by-step thinking, the potential of the large language model can be fully explored, improving the accuracy of multi-intent segmentation and completion. Subsequently, the large language model can also orchestrate multiple target sub-intent instructions to determine the execution relationship between them, and control the target device to execute the target sub-intent instructions according to this relationship. This effectively improves the accuracy and efficiency of understanding multiple intent instructions in complex semantic scenarios. This application's embodiment uses a large language model combined with thought chain technology. By formulating appropriate prompts (appropriate thinking methods) through the thought chain format, it inspires and guides the large language model to perform intent segmentation and completion. This overcomes the limitations of related technologies that are result-oriented, enabling the model to tap into its deeper potential in complex semantic scenarios, achieving a leap from "recognition result" to "understanding process," thereby significantly improving the accuracy and efficiency of understanding multiple intent instructions in complex semantic scenarios and meeting the actual needs of users.

[0054] The following describes the specific process of intent segmentation and intent completion in the multi-intent instruction execution method provided in the embodiments of this application.

[0055] Figure 3 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0056] In some embodiments, such as Figure 3As shown, the first type of prompt information includes prompt information for the first scene role, prompt information for the first function, prompt information for the first type of thought process, and device information for each device.

[0057] For example, the first type of prompt information could be: "Please play the role of {multi-intent segmentation and completion} in a family scenario, and be able to implement {intent segmentation and intent completion functions for language commands} for the user. You have the ability to control devices such as {Device 1, Device 2, ...}. The device information for each device is as follows: {...}. Please think and provide the answer step by step according to the thought process of 'language command -> intent segmentation thinking process -> intent segmentation answer -> intent completion thinking process -> intent completion answer -> final answer'. Both input and output are in pure JSON format, with the input format being {...} and the output format being {...}." Here, "Please play the role of multi-intent segmentation in a family scenario" is the first scenario role prompt information, "implement the intent segmentation function for language commands for the user" is the first function prompt information, and "language command -> intent segmentation thinking process -> intent segmentation answer -> intent completion thinking process -> intent completion answer -> final answer" is the first type of thought process prompt information. The device information for each device can include device name, device model, device function, etc.

[0058] Step S120 utilizes the pre-trained large language model to segment the language instructions according to the thought process in the first type of prompt information, obtaining multiple sub-intention instructions; based on the first type of prompt information, it performs intent completion on the multiple sub-intention instructions to obtain multiple target sub-intention instructions, which may include: S121. Using a pre-trained large language model, based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information, and the device information of each device, the language command is segmented to obtain multiple sub-intention commands.

[0059] S122: Using a pre-trained large language model, based on the first type of prompt information, perform intent completion on multiple sub-intent instructions to obtain multiple target sub-intent instructions.

[0060] Specifically, using a pre-trained large language model, intent segmentation can be performed on language instructions based on the first type of prompt information, including "first scenario role prompt information (e.g., please play a role in a family scene with multiple intents), first function prompt information (e.g., to implement the intent segmentation function of the language instruction for the user), first type of thought process prompt information (e.g., language instruction -> intent segmentation thought process -> intent segmentation answer -> intent completion thought process -> intent completion answer -> final answer), and device information of each device (e.g., device name, device model, device function, etc.)", resulting in multiple independent sub-intent instructions. Then, using the pre-trained large language model, based on the first type of prompt information, intent completion is performed on each of the multiple sub-intent instructions to obtain multiple target sub-intent instructions.

[0061] This application embodiment significantly improves the accuracy and interpretability of multi-intent instruction processing by constructing a structured first-class prompt information that includes scene roles, functional positioning, thought processes, and device information, and combining it with the step-by-step execution mechanism of a large language model. Specifically, scene role prompts (such as setting roles in a family scene) help the model quickly adapt to the task context, functional prompts (such as intent segmentation and completion capabilities) clarify the operational boundaries of the model, and detailed thought process prompts (such as phased thinking paths) guide the model to reason step by step according to the logical chain of "instruction parsing -> sub-intent decomposition -> semantic completion," effectively avoiding the black box problem of traditional end-to-end models. At the same time, device information such as device name, model, and function provides the model with an accurate profile of the operation object, making sub-intent segmentation and completion more in line with the actual needs of the scenario.

[0062] Figure 4 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0063] In some embodiments, such as Figure 4 As shown, the first type of prompt information also includes prompt information in the input format and prompt information in the output format. Step S121 uses the pre-trained large language model to perform intent segmentation on the language instructions based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information, and the device information of each device, to obtain multiple sub-intent instructions, which may include steps S1211 and S1212.

[0064] S1211 uses a pre-trained large language model to convert the format of language instructions into the input format to obtain the target language instructions.

[0065] S1212, based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information and the device information of each device, performs intent segmentation on the target language instruction and outputs multiple sub-intent instructions according to the output format.

[0066] Specifically, the prompts can include prompts for the input and output formats. That is, they can prompt the large language model to standardize the input information before each data processing and to standardize the output results after data processing, and then use them as the standardized input for the next data processing step.

[0067] In this embodiment, the input format prompts (such as JSON structure requirements) enable the large language model to automatically convert natural language instructions into a model-parsable standard format, giving the language instructions a standardized structure before segmentation. This makes it easier for the large language model to perform intent segmentation and completion, improving the input quality and accuracy of subsequent intent segmentation and completion. The output format prompts (such as JSON structure requirements) enable the large language model to automatically output multiple sub-intent instructions in a standard format, improving the input quality of subsequent intent completion and increasing the accuracy of intent completion.

[0068] Figure 5 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0069] In some embodiments, such as Figure 5 As shown, the first type of prompt information also includes second scene role prompt information and second function prompt information. Step S122, based on the first type of prompt information, performs intent completion on multiple sub-intent instructions to obtain multiple target sub-intent instructions, and may include step S1221.

[0070] S1221, using a pre-trained large language model, based on the second scene role prompt information, the second function prompt information, the first type of thought process prompt information, and the device information of each device, performs intent completion on multiple sub-intent instructions to obtain multiple target sub-intent instructions.

[0071] Specifically, using a pre-trained large language model, based on second-scene role prompts (e.g., please play a role in a family scene with multiple intent completions), second-function prompts (e.g., to provide users with the intent completion function for language commands), first-type thought process prompts (e.g., language command -> intent segmentation thought process -> intent segmentation answer -> intent completion thought process -> intent completion answer -> final answer), and device information of each device (e.g., device name, device model, device function, etc.), intent completion is performed on multiple sub-intent commands to obtain multiple target sub-intent commands.

[0072] This application's embodiments significantly improve the accuracy and interpretability of multi-intent command processing by constructing structured first-type prompt information with different scenario roles (such as intent segmentation roles and intent completion roles), different functional positioning (such as intent segmentation functions and intent completion functions), thought processes, and device information, and combining this with the step-by-step execution mechanism of a large language model. This chain-of-thinking approach to segmenting and completing multi-intent commands provides a heuristic reasoning approach for the large language model. Through step-by-step thinking, the potential of the large language model can be fully explored, improving the accuracy of multi-intent segmentation and completion.

[0073] It should be noted that in daily life at home, the number of devices a user binds can dynamically increase or decrease due to usage. Smart assistants in related technologies can only know the binding status of devices and cannot update device information based on changes in bound devices. Loading device information through a knowledge base would not only require significant manpower to maintain the knowledge base but also slow down the interaction response speed. The thought prompt template in this embodiment can be a dynamic template; for example, the device information in the template can be dynamically updated.

[0074] The inventors' research revealed that smart devices have now entered countless households, with an overwhelming number of devices and models. Furthermore, we are currently in the peak period of smart home technology replacing traditional devices, resulting in a very high frequency of device updates. This application's embodiments utilize a dynamic template design to dynamically update the user's bound smart device information. The large language model can promptly update user device knowledge, ensuring that changes in the user's bound devices do not affect the effectiveness of multi-intent understanding and orchestration.

[0075] Figure 6 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0076] In some embodiments, such as Figure 6 As shown, the thought prompt template includes device information for each device. Before generating the first type of prompt information based on the preset thought prompt template in step S110 when the language instruction is a multi-intent instruction, the execution method of the multi-intent instruction may further include steps S151 and / or S152.

[0077] S151: Obtain device information for each device at preset time intervals. Update the device information for each device to the thought prompt template.

[0078] The preset time interval is the update interval for device information in the thought prompt template. This preset time interval can be set according to actual needs, and this embodiment does not limit it.

[0079] Specifically, the device information in the mind prompt template can be updated at regular intervals, such as once an hour, thus realizing the dynamic loading and updating of device information.

[0080] And / or, in response to the device information update request, obtain the updated device information and update the mind prompt template with the updated device information.

[0081] Specifically, when a device information update request (such as adding, modifying, or deleting device information) is received, the updated device information can be obtained and synchronized to the mind prompt template to achieve real-time device information in the mind prompt template.

[0082] The device information-dynamically updated thought prompt template provided in this application embodiment can provide the latest device information to the large language model, enabling the large language model to effectively adapt to device changes and provide the most real-time multi-intent understanding and multi-intent orchestration capabilities, thereby improving the accuracy of the large language model in multi-intent segmentation and completion of language instructions.

[0083] This application embodiment uses a dynamic template to dynamically update the user's bound smart device information into the template to generate the latest prompt information. The large language model can perform multi-intent understanding (intent segmentation and completion) and multi-intent orchestration (determining the execution relationship between multiple sub-intent instructions) based on the timely updated user device information in the prompt information, ensuring the accuracy of multi-intent understanding and multi-intent orchestration.

[0084] The following describes the specific process of intent orchestration in the multi-intent instruction execution method provided in the embodiments of this application. In the embodiments of this application, not only multi-intent segmentation and completion use thought templates, but orchestration (determining execution relationships) also uses thought templates.

[0085] The inventors discovered that to handle multi-intent interaction issues, multiple intents need to be divided into sub-intents. While related technologies can combine the sub-intents to fully represent user needs, problems such as disordered execution order, incomplete execution conditions, and conflicting execution content can occur under different semantic scenarios during execution. Related technologies lack systematic analysis of this situation. The embodiments of this application utilize the powerful orchestration capabilities of a large language model to solve the problem of sub-intent execution in multi-intent states for intelligent assistants.

[0086] Figure 7 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0087] In some embodiments, such as Figure 7 As shown, step S130 uses a large language model to determine the execution relationship between multiple target sub-intent instructions, and may include steps S131 and S132.

[0088] S131, Based on the preset thought prompt template, generate the second type of prompt information.

[0089] Specifically, the scene roles and target functions in the thought prompt template can be limited. After the thought process is limited, the device information of each device in the template is combined to generate prompt information for sub-intent arrangement, which is the second type of prompt information.

[0090] For example, the second type of prompt message is: "Please play the role of {multi-intent orchestration} in a home scenario, capable of {determining the execution relationship between multiple sub-intent instructions}. You have the ability to control devices such as {Device 1, Device 2, ...}. The device information for each device is as follows: {...}. Please think step by step and provide the answer according to the thought process of 'multiple sub-intents -> orchestration thought process -> orchestration answer'. Both input and output are in pure JSON format, with the input format being {...} and the output format being {...}." S132, using the large language model, determine the execution relationship between multiple target sub-intention instructions according to the thinking patterns in the second type of prompt information.

[0091] It should be noted that before executing step S132, sample data (such as multiple historical sub-intention instructions and second-type historical prompt information generated based on the thought prompt template) can be collected as input to the large language model, and actual results (such as the execution relationship between multiple historical sub-intention instructions) can be collected as the ideal output. The large language model is trained in multiple rounds to enable it to have multi-intention arrangement capability, that is, to train the arrangement rules of sub-intention arrangement. The trained large language model can be obtained, and the trained large language model includes the trained arrangement rules.

[0092] Specifically, a pre-trained large language model (including pre-trained arrangement rules) can be used to arrange multiple target sub-intention instructions according to the thinking patterns in the prompts arranged according to sub-intentions, thereby obtaining the execution relationship between multiple target sub-intention instructions.

[0093] This application's embodiments leverage the powerful orchestration capabilities of large language models to solve the problem of sub-intent execution in multi-intent states for intelligent assistants.

[0094] Figure 8 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0095] In some embodiments, such as Figure 8 As shown, the second type of prompt information includes third scene role prompt information, third function prompt information, second type of thought process prompt information, and device information for each device.

[0096] Step S132 utilizes a large language model to determine the execution relationship between multiple target sub-intent instructions based on the second type of prompt information, and may include step S1321.

[0097] S1321, using a large language model, based on third-scene role prompts (e.g., please play a role in a multi-intent orchestration in a family scene), third-function prompts (e.g., please implement the function of determining the execution relationship between multiple sub-intent instructions), second-type thought process prompts (e.g., please think and give the answer step by step according to the thought process of 'multiple sub-intents -> orchestration of thought process -> orchestration of answer'), and device information of each device (e.g., you have the ability to operate devices {device 1, device 2...}, etc., and the device information of each device is as follows: {...}), determine the execution relationship between multiple target sub-intent instructions.

[0098] For example, the second type of prompt information may also include prompt information in input format and prompt information in output format, ensuring the structuring and standardization of input information (such as target sub-intention instructions) and output information (execution relationships between multiple target sub-intentions).

[0099] This application embodiment constructs an efficient and accurate multi-intent execution relationship determination mechanism by introducing a second type of prompt information that includes third scene role prompt information, third function prompt information, second type of thought process prompt information, and dynamic device information.

[0100] Figure 9 This is another flowchart illustrating the method for executing multi-intent instructions provided in the embodiments of this application.

[0101] In some embodiments, such as Figure 9 As shown, step S132 utilizes the large language model and, based on the second type of prompt information, determines the execution relationship between multiple target sub-intent instructions, which may include steps S1322 to S1324.

[0102] S1322, using a large language model, based on the second type of prompt information and information on multiple preset reference execution relationships, determines the target reference execution relationship to which multiple target sub-intent instructions belong.

[0103] The preset reference execution relationship refers to the possible execution relationships between pre-defined sub-intention instructions, which may include at least one of simultaneous execution, selective execution, merged execution, compromise execution, sequential execution, and conditional execution.

[0104] Specifically, a large language model can be used to classify the execution relationships between multiple target sub-intention instructions. For example, the reference execution relationship with the highest similarity to the "execution relationship between multiple target sub-intention instructions" among multiple reference execution relationships can be taken as the target reference execution relationship.

[0105] Step S140, according to the execution relationship, controls the target device to execute the target sub-intent instruction, which may include step S141.

[0106] S141, according to the target reference execution relationship (such as simultaneous execution relationship, selective execution relationship, merged execution relationship, compromise execution relationship, sequential execution relationship and conditional execution relationship, etc.), control the target device to execute the target sub-intent instruction.

[0107] This application predefines various typical execution relationships (reference execution relationships) covering simultaneous, selection, merging, compromise, sequence, and condition, enabling the large language model to quickly match the most suitable target reference execution relationship from multiple reference execution relationships during inference, avoiding the complexity of reasoning from scratch. Furthermore, combining the scene roles, functional positioning, and device information in the second type of prompt information improves the accuracy of the large language model in determining execution relationships.

[0108] For example, when training a large language model, typical samples (sub-intention instructions) of various execution methods can be used as input, and the actual execution method can be used as the ideal output to train the large language model. By adopting the few-shot learning technique, the large language model can learn how to handle more complex instructions by observing these few samples.

[0109] In some embodiments, the plurality of preset reference execution relationships include at least one of single execution relationship and multiple execution relationship. The single execution relationship includes at least one of simultaneous execution, selective execution, merged execution, and compromise execution. The multiple execution relationship includes at least one of sequential execution and conditional execution.

[0110] Among them, a single execution relationship means that only one control command needs to be issued to the device. A single execution relationship can include at least one of the following: simultaneous execution, selective execution, combined execution, and compromise execution. In other words, simultaneous execution, selective execution, combined execution, and compromise execution all require issuing only one control command to the device and belong to the single execution relationship type.

[0111] Simultaneous execution: Multiple sub-intents can be executed concurrently, without considering the order in which they are executed. Execution failure does not affect the execution of other sub-intents. For example, the language command "Turn on both the bedroom light and the humidifier" can be issued simultaneously with other sub-intent commands, and their execution results will not affect each other.

[0112] Selective Execution: If multiple sub-intents cannot be executed simultaneously, or if two intents have identical meanings, only one needs to be executed to fulfill the user's intent. For intents that cannot be executed simultaneously, the system will only execute the last intent. For example, if the voice command is "Set the refrigerator to smart mode and energy-saving mode," the refrigerator cannot simultaneously activate smart mode and energy-saving mode; only one mode can be selected for adjustment. Another example is the voice command "Turn the TV volume down," where "TV volume is too high" and "Turn the TV volume down" are two commands performing the same operation; only one command needs to be executed.

[0113] Combined execution: When multiple sub-intents require one to complement another, they can be combined for execution. For example, the verbal instruction "Remind me to drink water at 7 PM every day" can be broken down into "Remind me to drink water at 7 PM" and "Remind me to drink water every day," and these two intents can be combined into "Remind me to drink water at 7 PM every day." Similarly, the verbal instruction "Turn on some relaxing music" can be combined into "Turn on some relaxing music."

[0114] Compromise Execution: When multiple sub-intents conflict, a compromise solution is needed. For example, the voice command "Turn up the TV volume, but not too loud" can be broken down into "Turn up the TV volume" and "Don't turn the TV volume too loud." These two intents conflict to some extent, but ultimately only "Turn the TV volume to medium" needs to be executed. Similarly, the voice command "Turn up the bedroom light, but not too bright" can be replaced with "Turn the bedroom light brightness to medium."

[0115] A multiple execution relationship indicates that after issuing a control command to the device, feedback information from the device is needed as a condition for issuing the next control command; that is, multiple interactions with the device are required. A multiple execution relationship can include at least one of sequential execution and conditional execution.

[0116] Sequential execution: Multiple sub-intents need to be executed in a specific order, with later sub-intents depending on the results of earlier intents. For example, "Will it be cold tomorrow? How much should I wear?" or "Check recipes for dinner recommendations." Both examples require the previous intent to return a result; for instance, the weather needs to be checked before determining what to wear, and recipes need to be found before dinner recommendations can be made.

[0117] Conditional execution: Sub-intents within a sentence need to be executed only under certain conditions. For example: "Open the window and wait for the wind to pick up before closing it." These two intents include opening and closing the window, but the action of closing the window can only be executed after "the wind picks up."

[0118] These six execution methods can cover most scenarios of multi-intent interaction. When orchestrating intents, the large language model can refer to these six execution relationships. Through the powerful orchestration capabilities of the large language model, the execution methods of each sub-intent can be accurately planned.

[0119] The above execution relationship is merely illustrative and is not intended to limit this application.

[0120] This application provides a systematic analysis of multi-intent execution relationships, categorizing multi-intent execution methods into six main types: simultaneous execution, sequential execution, selective execution, merged execution, compromise execution, and conditional execution. Leveraging the powerful orchestration capabilities of the large language model, the execution methods (relationships) between various sub-intents can be accurately planned. This refines the capabilities of the large language model from intent understanding (multi-intent segmentation and completion) to the level of execution method understanding, truly achieving user understanding.

[0121] In one embodiment, such as Figure 10 As shown, a multi-intent instruction execution system may include a multi-intent understanding module, a multi-intent orchestration module, a multi-intent execution module, and a thought prompt template.

[0122] The thought prompt template can include multiple prompts, including prompts for scene roles, target functions, thought processes, and device information. Device information can be dynamically updated to the thought prompt template in a structured format. For example, it can query the home smart terminal hourly to obtain device binding information such as device name, model, and functions, and then update the thought prompt template accordingly. The language instructions are processed into two main parts: the multi-intent understanding module and the multi-intent orchestration module. The multi-intent understanding module includes algorithms for multi-intent classification, multi-intent segmentation, and multi-intent completion. It can accurately determine whether a language instruction is a multi-intent instruction, segment the multi-intent into individual sub-intents, and complete the intent to obtain multiple target sub-intent instructions.

[0123] Then, the multi-intent orchestration module orchestrates these target sub-intent instructions at the execution level. For example, it categorizes multi-intent execution relationships (execution order) into six main types: simultaneous execution, sequential execution, selective execution, merged execution, compromise execution, and conditional execution, planning the most suitable execution relationship for the user. During the algorithm reasoning process, the multi-intent understanding module and the multi-intent orchestration module call upon dynamically updated thought prompt templates. The prompt words in the templates (such as device information) can be updated in real time through dynamic prompt word technology, enabling the large language model to respond promptly to changes in device binding, making the intelligent assistant truly a user's exclusive helper.

[0124] The multi-intent execution module is based on the orchestration results of a large language model (the execution relationship between multiple target sub-intent instructions), and needs to align with device-side instructions. A single-execution intent only requires issuing the instruction normally. For intents that are executed multiple times, the results of previous executions will affect the execution strategy of the current intent. For example, the instruction "Will it be cold tomorrow? How much should I wear?" requires first calling the weather forecast interface to return tomorrow's weather, and then executing a clothing plan based on tomorrow's temperature. This provides users with a more realistic response, truly becoming a considerate assistant to the user.

[0125] For example, sub-intent instructions 1 and 2 are in a selective execution relationship, requiring only a single interaction with the device, and instructions can be sent to the device in one go. Sub-intent instructions 3 and 4 are in a conditional execution relationship, requiring multiple interactions with the device. The control instruction for sub-intent instruction 3 needs to be sent first, then the result needs to be waited for, and then the control instruction for sub-intent instruction 4 needs to be sent based on the result.

[0126] The multi-intent orchestration method based on a large language model proposed in this application adds the large language model's own thinking before the result is identified at a higher semantic level, enabling accurate identification, segmentation, completion, and orchestration of each sub-intent. It truly achieves a comprehensive understanding of user intentions from comprehension to execution. Simultaneously, through dynamic prompts (such as dynamically updated device information in templates), the large language model can update device information in real time to produce the most suitable multi-intent understanding and orchestration effect for the user's intentions. It is capable of recognizing ambiguous dialogues, accurately segmenting multi-intent boundaries, and orchestrating execution flows considering actual scenarios when handling multi-intent issues.

[0127] Based on the same inventive concept, such as Figure 11 As shown, this application embodiment also provides a multi-intent instruction execution device, which 1100 may include a generation module 1110, a segmentation and completion module 1120, a determination module 1130, and an execution module 1140: The generation module 1110 is used to generate a first type of prompt information based on a preset thought prompt template when the language instruction is a multi-intent instruction. The segmentation and completion module 1120 is used to segment language instructions according to the thought process in the first type of prompt information using a pre-trained large language model to obtain multiple sub-intention instructions; and to complete the multiple sub-intention instructions based on the first type of prompt information to obtain multiple target sub-intention instructions. The determination module 1130 is used to determine the execution relationship between multiple target sub-intent instructions using a large language model; The execution module 1140 is used to control the target device to execute the target sub-intent instruction according to the execution relationship.

[0128] In some embodiments, the first type of prompting information includes first scene role prompting information, first function prompting information, first type of thought process prompting information, and device information of each device; The segmentation and completion module utilizes a pre-trained large language model to segment language instructions according to the thought process described in the first type of prompt, resulting in multiple sub-intention instructions. Specifically, it can be used for: Using a pre-trained large language model, based on the first scene role prompts, the first function prompts, the first type of thought process prompts, and the device information of each device, the language commands are segmented to obtain multiple sub-intention commands.

[0129] In some embodiments, the first type of prompt information further includes prompt information in the input format and prompt information in the output format; the segmentation and completion module is used to utilize a pre-trained large language model, based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information, and the device information of each device, to perform intent segmentation on the language instructions, obtaining multiple sub-intent instructions, which can be specifically used for: Using a pre-trained large language model, the format of the language instructions is converted into the input format to obtain the target language instructions; Based on the first scene role prompts, the first function prompts, the first type of thought process prompts, and the device information of each device, the target language command is segmented into intents, and then multiple sub-intent commands are output according to the output format.

[0130] In some embodiments, the first type of prompt information further includes second scene role prompt information and second function prompt information; the segmentation and completion module is used to perform intent completion on multiple sub-intent instructions based on the first type of prompt information to obtain multiple target sub-intent instructions, specifically for: Using a pre-trained large language model, based on second-scene role prompts, second-function prompts, first-type thought process prompts, and device information of each device, intent completion is performed on multiple sub-intent commands to obtain multiple target sub-intent commands.

[0131] In some embodiments, the thought prompt template includes device information for each device; before the generation module generates a first type of prompt information based on the preset thought prompt template when the language instruction is a multi-intent instruction, the device further includes an update module: The update module is used to obtain the device information of each device at preset time intervals and update the device information of each device to the mind prompt template. And / or, the update module is also used to respond to a device information update request, obtain the updated device information, and update the mind prompt template with the updated device information.

[0132] In some embodiments, the determining module is used to determine the execution relationship between multiple target sub-intent instructions using a large language model, specifically for: Based on the preset thought prompt template, a second type of prompt information is generated; Using a large language model, and following the thought process in the second type of prompt information, the execution relationship between multiple target sub-intention instructions is determined.

[0133] In some embodiments, the second type of prompting information includes third scene role prompting information, third function prompting information, second type of thought process prompting information, and device information of each device; The determination module utilizes a large language model and, based on the second type of prompt information, determines the execution relationship between multiple target sub-intent instructions. Specifically, it can be used for: By utilizing a large language model, and based on third-scene role prompts, third-function prompts, second-type thought process prompts, and device information of each device, the execution relationships between multiple target sub-intention instructions are determined.

[0134] In some embodiments, the determining module is used to utilize a large language model and, based on the second type of prompting information, determine the execution relationship between multiple target sub-intent instructions, specifically for: Using a large language model, based on the second type of prompt information and information from multiple preset reference execution relationships, the target reference execution relationship to which multiple target sub-intent instructions belong is determined; The execution module is used to control the target device to execute target sub-intent instructions according to the execution relationship. Specifically, it can be used for: According to the target reference execution relationship, control the target device to execute the target sub-intent instructions.

[0135] In some embodiments, the plurality of preset reference execution relationships include at least one of single execution relationship and multiple execution relationship. The single execution relationship includes at least one of simultaneous execution, selective execution, merged execution, and compromise execution. The multiple execution relationship includes at least one of sequential execution and conditional execution.

[0136] The various modules in the multi-intent instruction execution device provided in this application embodiment can achieve... Figures 1 to 9 The functions of each step in the execution method of the provided multi-intent instructions, and the corresponding technical effects they achieve, will not be elaborated here for the sake of brevity.

[0137] Figure 12 A schematic diagram of the hardware structure of the multi-intent instruction execution device provided in an embodiment of this application is shown.

[0138] The execution device for multi-intent instructions may include a processor 1201 and a memory 1202 storing computer program instructions.

[0139] Specifically, the processor 1201 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0140] Memory 1202 may include mass storage for data or instructions. For example, and not limitingly, memory 1202 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where suitable, memory 1202 may include removable or non-removable (or fixed) media. Where suitable, memory 1202 may be internal or external to a multi-intent instruction execution device. In a particular embodiment, memory 1202 is a non-volatile solid-state memory.

[0141] Memory may include read-only memory (ROM), random access memory (RAM), disk storage media, optical storage media, flash memory, and electrical, optical, or other physical / tangible storage devices. The processor 1201 implements any of the multi-intent instruction execution methods in the above embodiments by reading and executing computer program instructions stored in the memory 1202.

[0142] In one example, the device for executing multi-intent instructions may further include a communication interface 1203 and a bus 1204. Wherein, as Figure 12 As shown, the processor 1201, memory 1202, and communication interface 1203 are connected through bus 1204 and complete communication with each other.

[0143] The communication interface 1203 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0144] Bus 1204, including hardware, software, or both, couples together components of a multi-intent instruction execution device. This device can execute the multi-intent instruction execution method of this application based on the individual units / components in the multi-intent instruction execution device, thereby achieving a combination... Figures 1 to 9 The method for executing multi-intent instructions is described.

[0145] Furthermore, in conjunction with the multi-intent instruction execution method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the multi-intent instruction execution methods in the above embodiments.

[0146] This application also provides a computer program product in which the instructions, when executed by a processor of an electronic device, cause the electronic device to perform various processes implementing any of the above-described multi-intent instruction execution method embodiments.

[0147] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0148] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0149] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0150] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for executing multi-intent instructions, characterized in that, include: When the language instruction is a multi-intent instruction, the first type of prompt information is generated based on the preset thought prompt template; Using a pre-trained large language model, the language instruction is segmented according to the thought process in the first type of prompt information to obtain multiple sub-intention instructions; based on the first type of prompt information, the multiple sub-intention instructions are completed to obtain multiple target sub-intention instructions; Using the large language model, the execution relationships between the multiple target sub-intent instructions are determined; According to the aforementioned execution relationship, the target device is controlled to execute the target sub-intent instruction.

2. The method according to claim 1, characterized in that, The first type of prompt information includes first scene role prompt information, first function prompt information, first type of thought process prompt information, and device information of each device; The process involves using a pre-trained large language model to segment the language instructions according to the thought patterns in the first type of prompt information, resulting in multiple sub-intention instructions, including: Using a pre-trained large language model, based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information, and the device information of each device, the language instruction is segmented to obtain multiple sub-intention instructions.

3. The method according to claim 2, characterized in that, The first type of prompt information also includes prompt information in the input format and prompt information in the output format; the step of using a pre-trained large language model, based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information, and the device information of each of the devices, to perform intent segmentation on the language instructions, obtaining multiple sub-intent instructions, including: The pre-trained large language model is used to convert the format of the language instructions into the input format to obtain the target language instructions; Based on the first scene role prompt information, the first function prompt information, the first type of thought process prompt information, and the device information of each device, the target language instruction is segmented into intents, and multiple sub-intent instructions are output according to the output format.

4. The method according to claim 2, characterized in that, The first type of prompt information also includes second scene role prompt information and second function prompt information; the intention completion of multiple sub-intention instructions based on the first type of prompt information to obtain multiple target sub-intention instructions includes: Using a pre-trained large language model, based on the second scene role prompt information, the second function prompt information, the first type of thought process prompt information, and the device information of each device, intent completion is performed on multiple sub-intent instructions to obtain multiple target sub-intent instructions.

5. The method according to claim 2, characterized in that, The thought prompt template includes device information for each of the devices; before generating the first type of prompt information based on the preset thought prompt template when the language command is a multi-intent command, the method further includes: According to a preset time interval, acquire the device information of each device; update the device information of each device to the thought prompt template; And / or, in response to a device information update request, obtain the updated device information and update the updated device information to the thought prompt template.

6. The method according to claim 1, characterized in that, The step of using the large language model to determine the execution relationship between the multiple target sub-intent instructions includes: Based on the preset thought prompt template, a second type of prompt information is generated; Using the large language model, the execution relationship between the multiple target sub-intention instructions is determined according to the thinking patterns in the second type of prompt information.

7. The method according to claim 6, characterized in that, The second type of prompt information includes third-scene role prompt information, third-function prompt information, second-type thought process prompt information, and device information for each device; The step of using the large language model to determine the execution relationship between the multiple target sub-intent instructions based on the second type of prompt information includes: Using the large language model, based on the third scene role prompt information, the third function prompt information, the second type of thought process prompt information, and the device information of each device, the execution relationship between the multiple target sub-intention instructions is determined.

8. The method according to claim 6 or 7, characterized in that, The step of using the large language model to determine the execution relationship between the multiple target sub-intent instructions based on the second type of prompt information includes: Using the large language model, based on the second type of prompt information and information from multiple preset reference execution relationships, the target reference execution relationship to which the multiple target sub-intent instructions belong is determined; The step of controlling the target device to execute the target sub-intent instruction according to the execution relationship includes: According to the target reference execution relationship, control the target device to execute the target sub-intent instruction.

9. The method according to claim 8, characterized in that, The plurality of preset reference execution relationships include at least one of single execution relationship and multiple execution relationship. The single execution relationship includes at least one of simultaneous execution, selective execution, merged execution, and compromise execution. The multiple execution relationship includes at least one of sequential execution and conditional execution.

10. An execution device for multi-intent instructions, characterized in that, include: The generation module is used to generate the first type of prompt information based on a preset thought prompt template when the language instruction is a multi-intent instruction. The segmentation and completion module is used to use a pre-trained large language model to segment the language instruction according to the thought process in the first type of prompt information to obtain multiple sub-intention instructions; and to complete the multiple sub-intention instructions based on the first type of prompt information to obtain multiple target sub-intention instructions. The determination module is used to determine the execution relationship between the multiple target sub-intent instructions using the large language model; An execution module is used to control the target device to execute the target sub-intent instruction according to the execution relationship.

11. A device for executing multi-intent instructions, characterized in that, The device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the multi-intent instruction execution method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the multi-intent instruction execution method as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the device, the device is able to perform the method of executing multi-intent instructions as described in any one of claims 1 to 9.

Images