Household appliance control method and household appliance

Abstract

The embodiment of the invention provides a household appliance control method and a household appliance, and relates to the technical field of large models, and the method comprises the steps: obtaining a control instruction related to the household appliance, carrying out the intention recognition of the control instruction through an intention distribution agent, and obtaining a control intention corresponding to the control instruction; calling a target equipment function extraction model corresponding to the control intention from preset equipment function extraction models through the equipment function extraction agent, and performing equipment function extraction on the control instruction through the target equipment function extraction model to obtain an extraction result; and generating target control information according to the extraction result by using the dialogue management agent, and controlling the household appliance according to the target control information. According to the invention, through multi-agent cooperation, the accuracy of control instruction identification of the household appliance can be improved, and the equipment control difficulty is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large models, in particular to a home appliance control method and a home appliance. BACKGROUND

[0002] With the popularization of intelligent home appliances, the use of home appliance control functions is also becoming more and more frequent. Existing home appliance control functions can usually only control home appliances according to a dialog module prepared in advance, lack precise understanding ability of domain-specific intent and semantics, and result in a high error rate of control instruction recognition. SUMMARY

[0003] Therefore, the present application aims to provide a home appliance control method and a home appliance to solve the problem of a high error rate of existing home appliance control instructions.

[0004] To achieve the above object, the technical scheme adopted by the embodiments of the present application is as follows: In a first aspect, the embodiments of the present application provide a home appliance control method, comprising: obtaining a control instruction related to a home appliance, performing intent recognition on the control instruction by an intent distribution agent to obtain a control intent corresponding to the control instruction; calling a target device function extraction model corresponding to the control intent by a device function extraction agent, and performing device function extraction on the control instruction by the target device function extraction model to obtain an extraction result; generating target control information according to the extraction result by a dialog management agent, and controlling the home appliance according to the target control information.

[0005] In an optional implementation, the step of performing intent recognition on the control instruction by the intent distribution agent to obtain a control intent corresponding to the control instruction comprises: obtaining device information and intent recognition similar cases of all home appliances from a database by the intent distribution agent; performing intent recognition on the control instruction according to the device information and the intent recognition similar cases to obtain the control intent, the control intent including a device type, a device function and a device model corresponding to the control instruction.

[0006] In an optional implementation, the step of obtaining device information and intent recognition similar cases of all home appliances from a database by the intent distribution agent comprises: determining all home appliances in the current environment and device functions corresponding to all home appliances according to the control instruction by the intent distribution agent; obtaining historical dialogue information and standard similar cases related to the home appliance from the database; generating background knowledge according to the home appliance, the device function, the historical dialogue information and the standard similar cases, and injecting the background knowledge into a preset intent recognition model as an output prompt; converting the control instruction into a vector code, and generating the intent recognition similar case corresponding to the vector code through the preset intent recognition model.

[0007] In an optional implementation, the step of generating the intent recognition similar case corresponding to the vector code through the preset intent recognition model comprises: obtaining a standard example with the highest similarity to the vector code from the standard similar cases; obtaining a historical example with the highest similarity to the vector code from the historical dialogue information; performing content splicing on the standard example and the historical example through the preset intent recognition model to obtain the intent recognition similar case.

[0008] In an optional implementation, the method further comprises a training step of the preset intent recognition model, comprising: obtaining device functions contained in all home appliances, and generating home appliance control dialogues according to the device functions and prompt words; generating user information corresponding to the home appliance control dialogues, the user information comprising user roles, personalities and dialogue target contents; performing dialogue expansion on the home appliance control dialogues according to the user information to obtain multi-round dialogue data; training the preset intent recognition model by using the multi-round dialogue data, so that the preset intent recognition model outputs control intents corresponding to the multi-round dialogue data.

[0009] In an optional implementation, before the step of training the preset intent recognition model by using the multi-round dialogue data, the method further comprises a step of deduplicating the multi-round dialogue data; The step of deduplicating the multi-round dialogue data comprises: cleaning the multi-round dialogue data to obtain standardized texts; performing word segmentation on the standardized texts to obtain word sequences, and calculating weights and hash values of each word in the word sequences; converting the hash values into weighted vectors according to the weights, and obtaining a total feature vector of the word sequence according to the weighted vectors of all words; binarizing the total feature vector to obtain a similar hash fingerprint; If the Hamming distance of two similar hash fingerprints is less than the Hamming distance threshold, it is determined that the multi-round dialogue data corresponding to the two similar hash fingerprints is similar data, and the similar data is de-duplicated.

[0010] In an optional implementation, the step of calling, by the device function extraction agent, a target device function extraction model corresponding to the control intent, and performing device function extraction on the control instruction by the target device function extraction model to obtain an extraction result, includes: generating an intent tuple according to the device type, the device function, and the device model; Figure Three determining target device information corresponding to the intent tuple from device information of all the home appliances, and calling, by a device function extraction agent, a target device function extraction model corresponding to the intent tuple; Figure Three Figure Three performing device function extraction on the target device information by the target device function extraction model to obtain the extraction result.

[0011] In an optional implementation, the step of performing device function extraction on the target device information by the target device function extraction model to obtain the extraction result, includes: determining a corresponding function slot according to the target device function extraction model; obtaining a slot extraction similar case corresponding to the target device information from a database; performing device function extraction on the target device information according to the function slot and the slot extraction similar case to obtain the extraction result.

[0012] In an optional implementation, the step of performing intent recognition on the control instruction, further includes: if there is no device type, device function, or device model corresponding to the control instruction, generating a control instruction error prompt; and / or if the control instruction does not contain a control action keyword, generating the control instruction error prompt, and the format of the control instruction error prompt includes readable text.

[0013] In an optional implementation, the method further includes a training step of the preset intent recognition model, including: obtaining device functions contained by all home appliances, and generating a home appliance control dialogue according to the device functions and prompt words; generating user information corresponding to the home appliance control dialogue, the user information including user roles, personalities, and dialogue target content; ​​​dialogue expansion is performed on the household appliance control dialogue according to the user information, to obtain multi-round dialogue data; The preset intent recognition model is trained by using the multi-round dialogue data, so that the preset intent recognition model outputs a control intent corresponding to the multi-round dialogue data.

[0014] In an optional implementation, before the step of training the preset intent recognition model by using the multi-round dialogue data, the method further includes a step of deduplicating the multi-round dialogue data; The step of deduplicating the multi-round dialogue data includes: The multi-round dialogue data is cleaned to obtain normalized text; The normalized text is segmented to obtain a word sequence, and a weight and a hash value of each word in the word sequence are calculated; The hash value is converted into a weighted vector according to the weight, and a total feature vector of the word sequence is obtained according to the weighted vectors of all words; The total feature vector is binarized to obtain a similar hash fingerprint; If the Hamming distance of two similar hash fingerprints is less than a Hamming distance threshold, it is determined that the multi-round dialogue data corresponding to the two similar hash fingerprints is similar data, and the similar data is deduplicated.

[0015] In a second aspect, an embodiment of the present application provides a household appliance, including a processor and a memory, the memory stores machine executable instructions capable of being executed by the processor, and the processor can execute the machine executable instructions to implement the household appliance control method in the first aspect.

[0016] The household appliance control method and the household appliance provided by the embodiments of the present application can recognize the control intent of the control instruction of the household appliance related to the control instruction by the intent distribution agent, so as to determine the control intent, and can extract the device function of the control instruction by calling the target device function extraction model corresponding to the control intent by the device function extraction agent, and can control the household appliance according to the extraction result by the dialogue management agent. The embodiments of the present application can improve the response speed and the anti-interference ability of the system by the cooperation of the multiple agents, improve the recognition accuracy of the control instruction, and reduce the difficulty of the device control.

[0017] To make the object, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 A block diagram of a home appliance according to an embodiment of the present application is shown; Figure 2 A flow diagram of a home appliance control method according to an embodiment of the present application is shown; Figure 3 A principle diagram of a home appliance control method according to an embodiment of the present application is shown; Figure 4 A flow diagram of an intent recognition similar case obtaining method according to an embodiment of the present application is shown.

[0020] Icon: 100 - home appliance; 110 - memory; 120 - processor; 130 - communication module. DETAILED DESCRIPTION

[0021] The technical solutions of the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0023] It should be noted that the terms "first", "second", and so on, and the like relational terms merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of additional identical elements in the process, method, article or device that includes the element.

[0024] Please refer to Figure 1 , Figure 1 is a block schematic diagram of a home appliance 100 provided by the present embodiment. The home appliance 100 comprises a memory 110, a processor 120 and a communication module 130. The memory 110, the processor 120 and the communication module 130 are electrically connected to each other directly or indirectly to realize data transmission or interaction. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines.

[0025] The memory 110 is configured to store programs or data. The memory 110 can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc.

[0026] The processor 120 is configured to read / write the data or programs stored in the memory 110 and perform corresponding functions.

[0027] The communication module 130 is configured to establish a communication connection between the home appliance 100 and other communication terminals through the network, and to receive / transmit data through the network.

[0028] It should be understood that Figure 1 The structure shown is only a structural schematic diagram of the home appliance 100, and the home appliance 100 can further comprise more elements than those shown in the figure. Figure 1more or less components than those shown in the figures, or configured differently from those shown in the figures. Figure 1 Figure 1 The components shown in the figures can be implemented in hardware, software, or a combination thereof.

[0029] Please refer to Figure 2 , Figure 2 A flowchart of a home appliance control method provided by the embodiment is shown in the figure. The method can be executed by the home appliance shown in the figure, and the method comprises: Figure 1 S201, obtaining a control instruction related to a home appliance, and performing intent recognition on the control instruction by an intent distribution agent to obtain a control intent corresponding to the control instruction.

[0030] Please refer to Figure 3 , Figure 3 A principle diagram of a home appliance control method provided by the embodiment.

[0031] The user issues a control instruction to the home appliance. Since there can be multiple home appliances, and each home appliance has multiple functions, it is necessary to determine the user's control intent according to the control instruction first.

[0032] Multiple agents can be built into each home appliance, or multiple agents can be set in the cloud. The home appliance sends the obtained control instruction to the cloud, and the cloud device calls the corresponding agent to process the control instruction. The agent refers to an autonomous computing system that can perceive, decide and execute actions in a specific environment. Through the cooperation between multiple agents, the control of the home appliance is realized.

[0033] As shown in Figure 3 The embodiment adopts three agents, including an intent distribution agent, a device function extraction agent, and a dialogue management agent. Each agent is responsible for executing the corresponding function.

[0034] After obtaining the control instruction, the intent distribution agent is responsible for intent recognition and distribution of the control instruction. The intent distribution agent can recognize the intent of the control instruction through the preset intent recognition model built in it to obtain the corresponding control intent. The control intent can be the device type, device model and device function that the user wants to control, etc. According to the control intent, it can be determined which function of which device the user wants to control through the control instruction.

[0035] S202, calling a target device function extraction model corresponding to the control intent through a device function extraction agent, and extracting the device function of the control instruction through the target device function extraction model to obtain an extraction result.​​

[0036] The control intent distribution agent can distribute the control intent to a device function extraction agent after obtaining the control intent. The device function extraction agent is built-in with multiple preset device function extraction models, and each preset device function extraction model can be used to extract the function of a control instruction corresponding to a control intent.

[0037] Since the types, functions, and models of different home appliances may be different, a corresponding device function extraction model can be trained for each type of home appliance. Then, after receiving the control intent, a target device function extraction model corresponding to or matching the control intent is selected from the multiple preset device function extraction models to extract the function of the control instruction to obtain an extraction result.

[0038] S203, generating target control information according to the extraction result by using a dialogue management agent, and controlling the home appliance according to the target control information.

[0039] After obtaining the extraction result, the extraction result needs to be arranged, spliced, or otherwise operated to correctly control the home appliance. For example, the extraction result may only contain part of the keywords, and the home appliance may not be able to accurately understand. At this time, a preset large language model built-in the dialogue management agent can be used to optimize the extraction result to obtain target control information that can be understood by the home appliance, and then the home appliance is controlled.

[0040] The embodiment determines the control intent by using the intent distribution agent to perform intent recognition on the control instruction related to the home appliance, extracts the function of the control instruction by using the device function extraction agent to call a target device function extraction model corresponding to the control intent, and controls the home appliance according to the extraction result by using the dialogue management agent. The application significantly improves the system response speed and anti-interference ability while ensuring high-precision intent recognition and parameter extraction of the control instruction through the cooperation of multiple agents.

[0041] In one embodiment, the step of performing intent recognition on the control instruction by using the intent distribution agent to obtain a control intent corresponding to the control instruction includes: The intent distribution agent obtains device information and intent recognition similar cases of all the home appliances from a database; The control intent is obtained by performing intent recognition on the control instruction according to the device information and the intent recognition similar cases, and the control intent includes a device type, a device function, and a device model corresponding to the control instruction.

[0042] After receiving the control instruction, the controllable household appliances connected in the user's home can be obtained from the user's household appliance database or some APP, and then the device functions of the controllable household appliances are obtained. The controllable household appliances can be household appliances that can recognize the control instruction and make corresponding responses. For example, the control instruction is a voice control instruction, and only household appliances with voice control function can be regarded as controllable household appliances.

[0043] After determining the controllable household appliances, the device information of all controllable household appliances is obtained from the database. Since the embodiment mainly controls the household appliances, the device information can be information for indicating the identity of the household appliances and function information.

[0044] The intent recognition similar example can be an example obtained by performing intent recognition according to historical control intent. For example, the user controls the air conditioner through the control instruction, and at this time, the record of the user's past control of the air conditioner through the control instruction can be obtained from the database, and the record is taken as the intent recognition similar example. In order to ensure the accuracy and success of intent recognition, the intent similar example is a record of past successful control of the household appliances in the database, and the record of the failed control of the household appliances can be used to train the preset intent recognition model to improve the stability of the intent recognition model. The device function is the function possessed by each controllable household appliance. For example, when the controllable household appliances include an oven and an air conditioner, the functions of the oven include the baking mode, the temperature setting function, etc., and the air conditioner includes the refrigeration function, the heating function, the temperature setting function, etc. When the device type is not included in the control instruction, the device type can also be determined according to the function of the control instruction.

[0045] For example, in Figure 3 When the device is not an oven in the control instruction, but the control instruction contains the baking mode, and only the oven has the baking mode, and the air conditioner does not have the baking mode, it can also be determined that the control target of the control instruction is the oven.

[0046] The embodiment performs intent recognition on the control instruction according to the device information of the household appliances and the intent recognition similar example, so as to determine the device type, the device model and the device function of the control instruction, and realizes accurate interpretation of the control instruction.

[0047] Please refer to Figure 4 In an embodiment, the step of obtaining, by the intent distribution agent, the device information and the intent recognition similar example of all the household appliances from the database comprises: S401, determining, by the intent distribution agent, all the household appliances in the current environment according to the control instruction, and the device functions corresponding to all the household appliances; S402, obtain historical dialogue information and standard similar cases related to the home appliance from the database; S403, generate background knowledge according to the home appliance, the device function, the historical dialogue information and the standard similar cases, and inject the background knowledge into the preset intent recognition model as an output prompt; S404, convert the control instruction into vector encoding, and generate the intent recognition similar case corresponding to the vector encoding through the preset intent recognition model.

[0048] The historical dialogue information can be dialogue information generated by the home appliance according to the received historical control instruction. The key words of the historical dialogue information are extracted, and the information such as the function commonly used by the home appliance can be determined.

[0049] The standard similar case can be an intent recognition example automatically generated by the system. For example, the control instruction of the user is simulated through the model, and the intent recognition of the control instruction is recorded as a standard similar case.

[0050] In some control processes, after the user issues a control instruction, the home appliance may reply to the control instruction accordingly, and after several rounds of dialogue, the home appliance can be controlled. Therefore, the historical dialogue information can also include the reply information of the home appliance, and the reply information of the home appliance can cause a lot of noise to the retrieval of the control instruction or the dialogue use case. Therefore, when obtaining the historical dialogue information and the standard similar case, only the user control instruction can be obtained, and the reply information of the home appliance can be removed.

[0051] Then the background knowledge is generated according to the home appliance, the device function, the historical dialogue information and the standard similar case, and the background knowledge is used as the output prompt of the preset intent recognition model built in the intent distribution agent.

[0052] The background knowledge generated according to the home appliance, the device function, the historical dialogue information and the standard similar case can be injected into the preset intent recognition model through the Prompt mode.

[0053] The Prompt technology is a method for guiding a large language model or a generative AI to output a desired result through natural language, structured format or example, which is used to determine the task and output format of the model. Similarly, the large language model used by the intent distribution agent, the device function extraction agent and the dialogue management agent can be adjusted through the Prompt technology.

[0054] For example, the Prompt technology used in the intent distribution agent needs to include the following points: The supported device type and function definition and brief, non-control rejection identification prompt, control device corresponding model inferred from user access device circle, structured prompt (JSON output format problem) and Few-shot reference example.

[0055] The prompt technology used in the device function extraction agent needs to include the following points: Non-current device or non-current function rejection identification prompt, user input involves multiple devices and multiple functions, model needs to have filtering ability (focus on current function extraction), emphasize that automatic parameter filling is not allowed (reduce mis-extraction rate), enumerate class parameters emphasize mapping to enumeration values, numerical value class parameters emphasize only extracting explicit numerical values, structured prompt (JSON output format problem) and Few-shot reference example.

[0056] The prompt technology used in the dialog management agent needs to include the following points: Rule module output JSON data interpretation ability and high efficiency of generating guiding language, rationality.

[0057] Among them, Few-shot reference example refers to inserting a small amount of “task input→target output” paired examples in Prompt, so that the model can quickly understand the rules of complex tasks by imitating the mapping relationship of the examples, without fine-tuning to generate the expected results.

[0058] Background knowledge can ensure that the output of the preset intent recognition model does not exceed the specified range. For example, there is no washing machine included in the home appliance device, so when the control instruction is intent-recognized, it will not be considered that the user wants to control the home appliance device to include the washing machine.

[0059] Then the vector code corresponding to the control instruction is input into the preset intent recognition model, so that the preset intent recognition model outputs an intent recognition similar example corresponding to the vector code. The intent recognition similar example can be the process and result of intent recognition of the past user's control instruction.

[0060] The embodiment realizes automatic filtering of intent recognition similar examples and control intent by obtaining historical dialogue information and standard similar examples and using the historical dialogue information and the standard similar examples as output prompts of the preset intent recognition model, thereby improving the accuracy of intent recognition.

[0061] In one implementation, the step of generating, by the preset intent recognition model, the intent recognition similar example corresponding to the vector code includes: Obtain a standard example with the highest similarity to the vector code from the standard similar examples; obtain a historical example with the highest vector coding similarity from the historical dialogue information; perform content splicing on the standard example and the historical example through the preset intent recognition model to obtain the intent recognition similar example.

[0062] Each standard similar example and historical dialogue information can be converted into a corresponding vector. A standard example with the highest vector coding similarity is obtained from the standard similar example and a historical example with the highest vector coding similarity is obtained from the historical dialogue information through vector calculation. The number of standard examples and historical examples can include multiple, and then the historical example and the standard example are spliced in content, that is, the intent recognition similar example is obtained. When splicing in content, the dialogue information of the home appliance device can be filtered first, and only the content related to the control instruction of the user is spliced.

[0063] In an implementation manner, the step of calling, by the device function extraction intelligent agent, a target device function extraction model corresponding to the control intent through the device function of the target home appliance, and performing device function extraction on the target device information through the target device function extraction model to obtain the extraction result, includes: generating an intent Figure Three triple according to the device type, the device function and the device model; determining target device information corresponding to the intent Figure Three triple from all device information of the home appliance device, and calling, by the device function extraction intelligent agent, the target device function extraction model corresponding to the intent Figure Three triple; performing device function extraction on the target device information through the target device function extraction model to obtain the extraction result.

[0064] After the preset intent recognition model obtains the control intent, a triple: [device type, device function, device model] can be generated according to the control intent. Then, the target device information is determined according to the triple. The target device information can include the current state of the home appliance device, the device function and slot description supported in the device protocol, slot extraction similar example and the like.

[0065] Then, the device function extraction intelligent agent calls the target device function extraction model corresponding to the intent Figure Three triple from the multiple device function extraction models built-in in the device function extraction intelligent agent, and performs device function slot extraction on the triple and the target device information through the target device function extraction model, so as to obtain the corresponding extraction result.

[0066] Slot Filling is one of the core tasks of information extraction in natural language processing (NLP), mainly used to extract specific types of key information (i.e., "slots") from text, usually combined with intent recognition to support dialog systems, information retrieval, and other scenarios.

[0067] As shown in Figure 3 The device function extraction agent group includes a plurality of device function extraction agents, each of which can correspond to a type or a model of household appliance, and the extraction result can be each function contained by the device and the relationship between each function.

[0068] If the device control triple includes multiple, the multiple extraction results can be summarized and then sent to the dialog management agent.

[0069] The embodiment extracts target device information through a target device function extraction model to obtain key information related to device functions, providing a data foundation for subsequent dialog script generation.

[0070] In an implementation manner, the step of extracting device functions of the target device information through the target device function extraction model to obtain the extraction result comprises: determining a corresponding function slot according to the target device function extraction model; obtaining a slot extraction similar case corresponding to the target device information from a database; extracting device functions of the target device information according to the function slot and the slot extraction similar case to obtain the extraction result.

[0071] The functions of different household appliances can be different, resulting in different function slots that need to be set when extracting functions of different target device information. The function slot can be understood as a function-related keyword. Therefore, after determining the target device function extraction model, the function slot corresponding to the device function extraction model can be determined, and then the target device information is extracted through the function slot. The slot extraction similar case can be used to verify and adjust the extraction result after extraction, or can be adjusted in real time during the extraction process, so as to obtain the final extraction result.

[0072] The embodiment extracts device functions of target device information through a target device function extraction model, and adjusts in real time by using a slot extraction similar case, thereby improving the accuracy of the extraction result.

[0073] In an implementation manner, the step of identifying the intent of the control instruction further comprises: generate a control instruction error prompt if the control instruction does not contain a control action keyword, and the format of the control instruction error prompt includes readable text. generate a control instruction error prompt if the control instruction does not contain a control action keyword, and the format of the control instruction error prompt includes readable text.

[0074] The intent distribution agent, the device function extraction agent, and the dialogue management agent each include a built-in large model, and each agent can verify the information received through a corresponding prompt engineering after receiving the information.

[0075] The intent distribution agent is mainly used to receive control instructions, and if there is no device type, device function, or device model corresponding to the control instruction after receiving the control instruction, a control instruction error prompt is generated; and / or, if the control instruction does not contain a control action keyword, the control instruction error prompt is generated.

[0076] For example, the control instruction is to turn on the third gear of the air conditioner, but the air conditioner does not have a third gear option, and a corresponding prompt can be issued.

[0077] Similarly, the device function extraction agent and the dialogue management agent can also be configured with the same or similar prompt engineering to verify and feedback the received information.

[0078] The embodiments configure corresponding prompt engineering for each intelligent agent in the control process of kitchen electrical equipment, which can verify the control instruction and other information, and remind the user when the instruction is incorrect, thereby improving the success rate of device control.

[0079] In an implementation, the ReAct reflection process can be cancelled when training the device function extraction model built in the device function extraction agent.

[0080] ReAct reflection is a method that allows large language models or intelligent agents to dynamically adjust strategies to achieve goals through a closed loop of “step-by-step reasoning → executing actions → reflecting results → iterative optimization” when completing complex tasks. Although cancelling ReAct reflection can lose the intelligence of the extraction results, it can effectively improve the response speed, and the device function extraction agent needs to be polished and adjusted by the dialogue management agent after obtaining the extraction results, and the lost intelligence can be reproduced by the dialogue management agent.

[0081] In addition, the required parameter of the device function extraction agent can be cancelled to enable the agent to focus on semantic extraction. The required parameter refers to the core input information that must be obtained and cannot be missing when the agent performs a specific task (such as calling a tool, completing a function call, or triggering a process). Cancelling the required parameter can avoid unnecessary parameter definition affecting slot parameter extraction, enabling the device function extraction agent to extract the user's true intent, and the necessity check can be performed by the subsequent dialogue management agent.

[0082] In agent technology, enumeration (Enumeration, abbreviated as Enum) is a data type / parameter constraint rule that defines "finite, fixed, and enumerable values". It explicitly limits all possible legal values for a parameter / variable, and values outside this range are considered invalid. Therefore, by adding an Enum parameter to the device function extraction agent, the extraction range of the slot can be limited, which can improve the recognition accuracy when the user cannot completely remember the specific slot value or faces minor ASR errors.

[0083] In addition, when extracting slots through the device function extraction agent, the final extraction result can be integrated by combining the extraction results of the same device and the same function in the history, improving the slot integration capability. In one embodiment, the method further includes a training step of the preset intent recognition model, comprising: Obtaining all device functions contained in the household appliance, generating a household appliance control dialogue according to the device functions and prompt words; Generating user information corresponding to the household appliance control dialogue, the user information including user roles, personalities, and dialogue target content; Expanding the household appliance control dialogue according to the user information to obtain multi-round dialogue data; Training the preset intent recognition model using the multi-round dialogue data to enable the preset intent recognition model to output a control intent corresponding to the multi-round dialogue data.

[0084] First, sort out all the function lists of household appliances, and use a large model to generate 2-10 standard single-round dialogue data according to the list to ensure coverage under the current function. Only single-round data is required in this stage to ensure that the model has the ability to follow the rules in the prompt engineering without excessive interference.

[0085] Then, expand each single-round dialogue data. Different user information, such as different user roles, personalities, and dialogue target content, can be generated by a closed-source model to expand the single-round dialogue data and obtain multi-round dialogue data.

[0086] Reuse single-turn dialogue data and multi-turn dialogue data to train the preset intent recognition model, so that the preset intent recognition model can recognize both simple control dialogue and complex control dialogue between the user and the home appliance.

[0087] During training, LoRA (Low-Rank Adaptation) can be used for training, all parameters of the original model are frozen, only a small number of “low-rank matrices” are trained to simulate the update amount of model parameters, and finally the output of the low-rank matrix is fused with the output of the original model to realize the adaptation of the model in a specific task / domain.

[0088] In an embodiment, before the step of training the preset intent recognition model using the multi-turn dialogue data, the step of de-duplicating the multi-turn dialogue data is further included. The step of de-duplicating the multi-turn dialogue data includes: The multi-turn dialogue data is cleaned to obtain normalized text; The normalized text is segmented to obtain a word sequence, and the weight and hash value of each word in the word sequence are calculated; The hash value is converted into a weighted vector according to the weight, and the overall feature vector of the word sequence is obtained according to the weighted vectors of all words; The overall feature vector is binarized to obtain a similar hash fingerprint; If the Hamming distance of two similar hash fingerprints is less than a Hamming distance threshold, it is determined that the multi-turn dialogue data corresponding to the two similar hash fingerprints is similar data, and the similar data is de-duplicated.

[0089] Among them, the data of the preset intent recognition model, the preset device function extraction model and the preset large language model can be processed in the above manner, and then the processed model is used for training, so that the accuracy of the trained model for information extraction or recognition is higher.

[0090] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can also be implemented by other manners. The apparatus embodiments described above are merely illustrative, for example, the flowcharts and block diagrams in the drawings show the possible implementation architecture, function and operation of the apparatus, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowchart or block diagram can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementation manners, the functions noted in the blocks can also occur in different order from that noted in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can also be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and the combination of blocks in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system for executing the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0091] In addition, each functional module in the embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0092] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or the part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0093] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for controlling household appliances, characterized in that, include: The system acquires control commands related to home appliances, and uses an intent distribution agent to perform intent recognition on the control commands to obtain the control intent corresponding to the control commands. The device function extraction agent calls the target device function extraction model corresponding to the control intention, and the device function extraction result is obtained by extracting the control instruction through the target device function extraction model. The dialogue management agent generates target control information based on the extraction results, and controls the home appliances based on the target control information.

2. The home appliance control method according to claim 1, characterized in that, The step of performing intent recognition on the control command through an intent distribution agent to obtain the control intent corresponding to the control command includes: The intent distribution agent retrieves device information and intent recognition similarities for all the home appliances from the database. The control intent is obtained by performing intent recognition on the control command based on the device information and the intent recognition similarity example. The control intent includes the device type, device function and device model corresponding to the control command.

3. The home appliance control method according to claim 2, characterized in that, The step of obtaining device information and intent recognition similarities of all home appliances from the database through the intent distribution agent includes: The intent distribution agent determines all home appliances in the current environment and their corresponding device functions based on the control instructions. Retrieve historical dialogue information and standard similarity examples related to the home appliance from the database; Background knowledge is generated based on the home appliances, the functions of the appliances, the historical dialogue information, and the standard similar examples. This background knowledge is then injected into the preset intent recognition model as an output prompt. The control command is converted into a vector code, and the intent recognition similarity example corresponding to the vector code is generated through the preset intent recognition model.

4. The home appliance control method according to claim 3, characterized in that, The step of generating intent recognition similarity examples corresponding to the vector encoding through the preset intent recognition model includes: Obtain the standard example with the highest similarity to the vector encoding from the standard similar examples; Obtain the historical example with the highest similarity to the vector encoding from the historical dialogue information; The standard example and the historical example are spliced ​​together using the preset intent recognition model to obtain the intent recognition similar example.

5. The home appliance control method according to claim 4, characterized in that, The method further includes a training step for the preset intent recognition model, comprising: Obtain the device functions of all home appliances, and generate home appliance control dialogue based on the device functions and prompts; Generate user information corresponding to the home appliance control dialogue, the user information including user role, personality and dialogue target content; Based on the user information, the dialogue for controlling the home appliances is expanded to obtain multi-round dialogue data. The preset intent recognition model is trained using the multi-turn dialogue data so that the preset intent recognition model outputs the control intent corresponding to the multi-turn dialogue data.

6. The home appliance control method according to claim 5, characterized in that, Before the step of training the preset intent recognition model using the multi-turn dialogue data, the method further includes a step of deduplicating the multi-turn dialogue data. The step of deduplicating the multi-turn dialogue data includes: The multi-turn dialogue data is cleaned to obtain standardized text; The normalized text is segmented to obtain a word sequence, and the weight and hash value of each word in the word sequence are calculated. The hash value is converted into a weighted vector according to the weight, and the overall feature vector of the word sequence is obtained according to the weighted vector of all words. Binarizing the overall feature vector yields a similar hash fingerprint; If the Hamming distance between two similar hash fingerprints is less than the Hamming distance threshold, then the multi-turn dialogue data corresponding to the two similar hash fingerprints are determined to be similar data, and the similar data is deduplicated.

7. The home appliance control method according to claim 2, characterized in that, The step of invoking the target device function extraction model corresponding to the control intention through the device function extraction agent, and extracting the device function of the control command through the target device function extraction model to obtain the extraction result includes: Generate an intent triple based on the device type, the device function, and the device model; Determine the target device information corresponding to the intent triple from the device information of all the home appliances, and call the target device function extraction model corresponding to the intent triple through the device function extraction agent; The extraction result is obtained by extracting the device function of the target device information using the target device function extraction model.

8. The home appliance control method according to claim 7, characterized in that, The step of extracting device functions from the target device information using the target device function extraction model to obtain the extraction result includes: The corresponding functional slots are determined based on the target device function extraction model; Retrieve similar examples of slot extraction from the database that correspond to the target device information; Based on the functional slots and similar examples extracted from the slots, the target device information is used to extract device functions to obtain the extraction result.

9. The home appliance control method according to claim 1, characterized in that, The step of performing intent recognition on the control command further includes: If no device type, device function, or device model corresponds to the control command, a control command error message is generated; and / or If the control instruction does not contain a control action keyword, an error message for the control instruction is generated. The format of the error message includes readable text.

10. A household appliance, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the home appliance control method according to any one of claims 1-9.

Images