Intelligent agent wake-up opportunity identification method and device, equipment, medium and product

Abstract

The invention relates to the technical field of data processing technologies, large model technologies, large language model technologies and intelligent assistants, and discloses an intelligent agent wake-up opportunity identification method and device, equipment, a medium and a product. The method comprises the following steps: acquiring voice description information of a target object; analyzing the voice description information, and identifying whether wake-up information is carried in the voice description information; and if the voice description information carries wake-up information, waking up a dialogue agent based on the wake-up information, so that the dialogue agent generates matched dialogue information according to the wake-up information and the voice description information. According to the technical scheme, the response time delay of the dialogue agent is reduced, the awakening accuracy of the dialogue agent is improved, and false triggering of the dialogue agent is avoided.

Description

Technical Field

[0001] The present disclosure relates to the fields of data processing technology, large model technology, large language model technology, and intelligent assistant technology, and specifically to methods, devices, equipment, media, and products for identifying the timing of waking up an intelligent agent. Background Art

[0002] As a natural language processing model based on deep learning, the Large Language Model (LLM) has high language understanding and generation capabilities. It can generate natural and fluent text, answer various questions, and perform various natural language processing tasks. With the development of LLM, voice dialogue assistants are gradually upgraded. However, waking up the voice dialogue assistant requires accurate identification of the wake-up timing. How to enable the LLM model to accurately determine whether the current scene of the conversation contains the wake-up timing has become an urgent problem to be solved for actively waking up the voice dialogue assistant. Summary of the invention

[0003] In view of this, the present disclosure provides a method, device, equipment, medium and product for identifying the wake-up timing of an intelligent body to solve the problem of difficulty in accurately identifying the wake-up timing.

[0004] In a first aspect, the present disclosure provides a method for identifying the timing for waking up an intelligent agent, comprising: obtaining voice description information of a target object; parsing the voice description information, and identifying whether the voice description information carries wake-up information; if the voice description information carries wake-up information, waking up a dialogue intelligent agent based on the wake-up information, so that the dialogue intelligent agent generates matching dialogue information according to the wake-up information and the voice description information.

[0005] In a second aspect, the present disclosure provides a device for identifying the timing of awakening an intelligent agent, including: an acquisition module for acquiring voice description information of a target object; a recognition module for parsing the voice description information and identifying whether the voice description information carries wake-up information; and a wake-up module for waking up a dialogue intelligent agent based on the wake-up information if the voice description information carries the wake-up information, so that the dialogue intelligent agent generates matching dialogue information according to the wake-up information and the voice description information.

[0006] In a third aspect, the present disclosure provides a computer device, comprising: a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method for identifying the wake-up timing of an intelligent agent according to the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0007] In a fourth aspect, the present disclosure provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the method for identifying the timing of awakening an intelligent agent according to the first aspect or any corresponding embodiment thereof.

[0008] In a fifth aspect, the present disclosure provides a computer program product, including computer instructions, which are used to enable a computer to execute the method for identifying the wake-up timing of an intelligent agent according to the first aspect or any corresponding embodiment thereof.

[0009] The present invention provides a method, device, equipment, medium and product for identifying the timing of awakening an intelligent agent. The method analyzes the voice description information to determine whether it carries the awakening information, and determines whether to wake up the dialogue intelligent agent based on the result of the wake-up information. As long as the wake-up information is carried, the dialogue intelligent agent is automatically awakened, which reduces the response delay of the dialogue intelligent agent, improves the awakening accuracy of the dialogue intelligent agent, and avoids false triggering of the dialogue intelligent agent. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the specific embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the specific embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0011] Figure 1 is a schematic diagram of a keyword awakening conversation agent in an embodiment of the present disclosure;

[0012] Figure 2 is a schematic diagram of a system for wake-up timing identification according to an embodiment of the present disclosure;

[0013] Figure 3 is a schematic diagram of a two-stage calling flow of functional1 in an embodiment of the present disclosure;

[0014] Figure 4 is a schematic diagram of a classification head design for user identification according to an embodiment of the present disclosure;

[0015] Figure 5 is a flow chart of a method for identifying an intelligent agent wake-up timing according to an embodiment of the present disclosure;

[0016] Figure 6 is a flowchart of a method for training a recognition model according to an embodiment of the present disclosure;

[0017] Figure 7 is a flowchart of another method for training a recognition model according to an embodiment of the present disclosure;

[0018] Figure 8 is a structural schematic diagram of a recognition model according to an embodiment of the present disclosure;

[0019] Fig. 9 is a structural schematic diagram of a parameter fine-tuning branch according to an embodiment of the present disclosure;

[0020] Fig.10 is a flow chart of another method for identifying the wake-up timing of an intelligent agent according to an embodiment of the present disclosure;

[0021] Fig.11 is a specific schematic diagram of a method for identifying an intelligent agent wake-up timing according to an embodiment of the present disclosure;

[0022] Fig.12 is a curve diagram of a loss function according to an embodiment of the present disclosure;

[0023] Fig.13 is a structural block diagram of a device for identifying an intelligent agent wake-up timing according to an embodiment of the present disclosure;

[0024] Fig.14 It is a schematic diagram of the hardware structure of the computer device of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0026] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0027] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0028] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0029] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet the relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0030] It is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and relevant provisions.

[0031] With the development of LLM, voice dialogue assistants have been gradually upgraded, bringing great convenience to people's study, work and life. In most scenarios, voice dialogue assistants are awakened by keywords and then interacted. For example, Figure 1 As shown, however, there are cases where keywords are falsely triggered.

[0032] In actual use, there are also some scenarios where the voice dialogue assistant needs to actively identify the wake-up opportunity and respond based on the user's dialogue, behavior, actions, etc., give the user some reminders and recommendations, and take the initiative to wake up to improve the usability of the intelligent dialogue system, such as Figure 2 For example, in the smart cockpit scenario of a car, when the user is driving tired or not paying attention, the driver can be reminded to rest in time. When the user mentions food during the conversation and shows a desire to eat, the user can be recommended a restaurant. The active wake-up timing recognition system can greatly improve the user experience of the dialogue system, but the accuracy of the LLM large language model for wake-up timing recognition is very high. How to enable the model to accurately determine whether the current scene and conversation contain wake-up timing is the biggest difficulty of the active wake-up dialogue system.

[0033] In related technologies, the most common method to improve the accuracy of a large model in identifying wake-up timing in a specific scenario is to fine-tune the large model using data from a specific scenario to improve the large model's sensitivity to the wake-up timing of the scenario, or to construct a classifier after the generative LLM and use the classifier to identify the wake-up timing.

[0034] The fine-tuning method that is most relevant to wake-up timing recognition is Functioncall fine-tuning, which uses conversation sample data with Functioncall for fine-tuning, such as Figure 3As shown. First, define some functional functions, and add an additional role named function to the fine-tuning sample. The model decides what function to call in what scenario during inference. Since this method needs to add new roles to the dialogue data, it has extremely high sample requirements. A large number of samples containing function input and output are required for training to achieve better results. When the number of samples is small and full parameter fine-tuning is not used, the effect is relatively poor. In addition, the LLM call with functioncall requires two complete LLM conversations. The first time, LLM decides whether to call the function and obtains the basic information of the function. The second time, the function call result and historical dialogue information are combined to obtain the final dialogue information. The two complete generation processes will make the latency of the entire link extremely high, which is very fatal to the dialogue system deployed on the edge device, such as the in-vehicle smart cockpit scenario.

[0035] Adding a classifier after the generative large language model (LLM) is a common practice in traditional deep learning. This method works well in most scenarios. However, for the autoregressive model, since only one tag is predicted each time it is generated, this means that a classification decision process needs to be performed each time forward reasoning is performed. Although the number of classifier parameters is not large, when the length of the generated sequence is very long, the number of calls increases, and the latency of the entire dialogue system will be very high. In addition, the design of the additional classification head is slightly abrupt compared to the LLM system. Figure 4 shown.

[0036] In addition to ensuring high accuracy, the dialogue system used to identify the user's wake-up intention also requires low latency. Especially for systems deployed at the edge, latency is an issue that needs to be considered during the design phase due to computing power limitations.

[0037] In view of this, the disclosed technical solution is different from the general function call fine-tuning and classifier design, which can greatly improve the recognition accuracy of wake-up information in edge scenarios at a lower latency. The core idea is to construct fine-tuning samples, fine-tune the model, and let the model contain specific flags in the normal output during a Chat inference process. <functioncall>, this flag is used to determine whether the current input contains wake-up information. In post-processing, the corresponding flag is used to determine the direction of the data flow, and the output of the model is used as the input of the next module. For example, in the smart cockpit scenario, the semantics of the flag indicates that if the current input contains wake-up information, the microphone will be woken up. After waking up, the output will be converted into speech through the TTS service and output through the microphone.

[0038] According to an embodiment of the present disclosure, an embodiment of a method for identifying the timing of waking up an intelligent agent is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] In this embodiment, a method for identifying the timing of waking up an intelligent agent is provided, which can be used in computer devices, such as computers, etc. Figure 5 is a flow chart of a method for identifying an agent wake-up timing according to an embodiment of the present disclosure, such as Figure 5 As shown, the process includes the following steps:

[0040] Step S101, obtaining voice description information of a target object.

[0041] The voice description information is the description information in the form of voice issued by the target object in the current scene, for example, "I am hungry", "Play a song", "Turn on the air conditioner in the car", etc.

[0042] Specifically, the voice description information can be obtained through a voice collection program deployed in a computer device, and the voice collection program is triggered and started when the target object is detected to be uttering voice; the voice description information can also be collected by an installed recording device and uploaded to the computer device. For example, a recording device is installed in the cabin of a car, and the recording device is communicatively connected to the computer device. The recording device can send the recorded voice description information to the computer device.

[0043] Of course, the voice description information may also be obtained in other ways, and the way of obtaining the voice description information is not limited here.

[0044] Step S102: parse the voice description information to identify whether the voice description information carries wake-up information.

[0045] The wake-up information is used to characterize the wake-up prompt information carried in the language description information sent by the target object. For example, when the target object sends a voice description information of "I'm hungry", it means that the target object may want to find a restaurant. At this time, it can be determined that the voice description information carries the wake-up information.

[0046] Specifically, speech recognition technology is used to convert speech description information into description information in text form, and natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic understanding on the text description information to extract feature information such as semantic relationship and context information from the text description information. By analyzing the correlation between the feature information and the wake-up information, it is determined whether the speech description information carries the wake-up information. If the speech description information carries the wake-up information, step S103 is executed, otherwise other operations are executed. The other operations may be not responding to the speech description information or continuing to collect the speech description information. There is no limitation here, as long as the dialogue agent is kept silent when the wake-up information is not detected.

[0047] Step S103: If the voice description information carries wake-up information, the dialogue agent is woken up based on the wake-up information, so that the dialogue agent generates matching dialogue information according to the wake-up information and the voice description information.

[0048] The dialogue agent is a system that can communicate with the target object. When the target object sends a voice description message carrying a wake-up message, the dialogue agent can automatically wake up and generate corresponding dialogue information according to the wake-up message. The dialogue information can be played through a microphone or other playback device. Correspondingly, the target object can receive the dialogue information fed back by the dialogue agent and give its reply based on the dialogue information to realize the interaction between the target object and the dialogue agent.

[0049] In a specific example, when the target object sends a voice description message "It's so hot", the computer device determines that the target object may want to turn on the air conditioner by parsing "It's so hot". At this time, the dialogue agent can generate dialogue information based on the target object's wake-up information and play it through the microphone, such as "I detected that you expressed a potential need - the feeling of heat. As an intelligent voice assistant, I can turn on the air conditioner for you. Do you want to turn on the air conditioner in the car?". If the target object replies "Please help me turn on the air conditioner in the car", at this time, the dialogue agent will receive the target object's reply information and reply "OK, turn it on for you right away". Furthermore, the dialogue agent will send the request to turn on the air conditioner to the computer device, and the computer device will control the start of the air conditioner in the car.

[0050] The method for identifying the timing of waking up an intelligent agent provided in this embodiment analyzes the voice description information to determine whether it carries the wake-up information, and determines whether to wake up the dialogue intelligent agent based on the judgment result of carrying the wake-up information. As long as the wake-up information is carried, the dialogue intelligent agent is automatically woken up, which reduces the response delay of the dialogue intelligent agent, improves the wake-up accuracy of the dialogue intelligent agent, and avoids false triggering of the dialogue intelligent agent.

[0051] In this embodiment, a method for training a recognition model is provided, whereby a pre-trained recognition model is used to parse multimodal description information and identify whether the multimodal description information carries wake-up information. Specifically, the recognition model can be trained based on a machine learning or deep learning algorithm to learn and identify different types of wake-up information.

[0052] Figure 6 is a flowchart of a method for training a recognition model according to an embodiment of the present disclosure, such as Figure 6 As shown, the process includes the following steps:

[0053] Step S201, obtaining a multimodal description sample set, wherein the multimodal description sample set includes a plurality of description samples of the target object in different scenarios, and the description samples include wake-up intention description information.

[0054] The multimodal description sample set is a set of original description samples collected in advance for each edge scenario. The wake-up intention description information is the description information used to characterize the awakened intelligent agent. The multimodal description sample set includes description samples of multiple modalities such as original image description samples, original text description samples, and original voice description samples in different scenarios. Specifically, the multimodal description sample set can be obtained through recording or through online collection. The method of collecting the multimodal description sample set is not specifically limited here.

[0055] Multiple description samples in the multimodal description sample set are divided into positive samples and negative samples, where the positive samples are samples carrying the awakening intention description information, and the negative samples are samples not carrying the awakening intention description information.

[0056] In some optional embodiments, such as Figure 7 As shown, the above step S201 includes:

[0057] Step S2011, collect multimodal data of the target object, and annotate the multimodal data with the wake-up intention to obtain annotation information.

[0058] Multimodal data refers to data in multiple modal forms collected in multiple edge scenarios, including image data, text data (such as text input by the target object, text obtained by voice conversion), etc. Among them, edge scenarios are pre-set specific fields, applications or environments, such as smart homes, automation, cockpits, etc.

[0059] The collected multimodal data needs to be annotated to determine whether it contains the object's intent. The wake-up intent annotation for multimodal data can be done manually or automatically using annotation tools. It should be noted that after the automatic annotation of intent is completed using the annotation tool, manual assistance can be used to verify the accuracy of the wake-up intent annotation to ensure that the wake-up intent annotation is as accurate and consistent as possible, thereby ensuring the quality of the annotation information.

[0060] Step S2012, obtaining judgment prompt information, and adding the judgment prompt information to the labeled multimodal data to obtain prompt data.

[0061] Judgment prompt information is used to convert multimodal data into a question or prompt to guide the recognition model to classify or judge the multimodal data. For example, in a text classification task, the judgment prompt information can be "Is this a positive text or a negative text?".

[0062] The judgment prompt information is added to the end of the multimodal data to convert the multimodal data into a classification or judgment problem, so that the recognition model can process and analyze it. At the same time, the judgment prompt information can also help the recognition model better understand the meaning of the text input and the object intention carried in the text, thereby improving the performance and accuracy of the recognition model.

[0063] In a specific example, the mode for determining the prompt information is as follows:

[0064] prompt=If A print B else 1print C,and why?

[0065] If the multimodal data is the text "Singer A's new album is pretty good, there are several songs I like, why don't we listen to it in the car", the multimodal data is transformed according to the judgment prompt information to obtain prompt data containing the judgment prompt information, as follows:

[0066] Subject: Singer A’s new album is pretty good, and I like several of the songs. How about we listen to it in the car?

[0067] Judgment prompt information: Please ask if this text contains potential user needs. If so, please output <functioncall>, and respond to the demand, otherwise output <nofunctioncall>.

[0068] The text input by the subject and the judgment hint information are used as hint data.

[0069] Step S2013, perform reasoning according to the prompt data to obtain intention reasoning data corresponding to the prompt data.

[0070] Input the prompt data into the intent reasoning model, use the built-in algorithm in the intent reasoning model to perform data reasoning and calculation, and obtain the intent reasoning data corresponding to the prompt data. Specifically, the intent reasoning model is a model pre-trained based on deep learning or machine learning algorithms, and the intent reasoning model is deployed on a cloud server. After obtaining the prompt data, organize and preprocess the prompt data to ensure that the data format and content of the prompt data meet the requirements of the intent reasoning model, and use appropriate interfaces or tools to connect with the intent reasoning model deployed in the cloud to establish a data transmission channel between the computer device and the cloud server. Upload the prompt data to the cloud server through the data transmission channel, so that after receiving the prompt data, the cloud server can use the intent reasoning model to reason and calculate the prompt data, output the intent reasoning data corresponding to the prompt information, and feed the intent reasoning data back to the computer device.

[0071] Step S2014: add a wake-up intention flag to the intention inference data according to the annotation information to obtain target inference data.

[0072] The wake-up intention flag is used to indicate that the wake-up intention represented by the multimodal data is consistent with the intention inference data. Perform semantic analysis on the intention inference data to determine whether the wake-up information carried by the intention inference data is the wake-up intention represented by the multimodal data. If the wake-up information carried by the intention inference data is the wake-up intention represented by the multimodal data, add the corresponding wake-up intention flag to the intention inference data according to the annotation information of the multimodal data, and use the intention inference data with the wake-up intention flag added as the target inference data.

[0073] In some optional implementations, the multimodal data includes positive sample data and negative sample data, and the intention inference data includes first intention inference data corresponding to the positive sample data and second intention inference data corresponding to the negative sample data. The positive sample data is sample data containing wake-up information, and the negative sample data is sample data not containing wake-up information. The first intention inference data represents carrying wake-up information, and the second intention inference data represents not carrying wake-up information.

[0074] Accordingly, the above step S2014 includes:

[0075] Step a1: clean the intention inference data to obtain cleaned data.

[0076] Step a2: adding a first wake-up intention flag to the first intention inference data using the annotation information corresponding to the positive sample data, where the first wake-up intention flag is used to represent the wake-up information carried in the multimodal data.

[0077] Step a3: adding a second wake-up intention flag to the second intention inference data using the annotation information corresponding to the negative sample data, where the second wake-up intention flag is used to indicate that the multimodal data does not carry wake-up information.

[0078] The intention reasoning data is cleaned and preprocessed to remove data noise, duplicate data, etc. in the intention reasoning data to ensure the quality and availability of the intention reasoning data. The cleaned data obtained after cleaning is divided into positive sample data and negative sample data to obtain the cleaned first intention reasoning data and second intention reasoning data.

[0079] The first wake-up intention flag is added to the first intention inference data using the annotation information corresponding to the positive sample data to indicate that the first intention inference data carries the wake-up information. At the same time, the second wake-up intention flag is added to the second intention inference data using the annotation information corresponding to the negative sample data to indicate that the second intention inference data does not carry the wake-up information. Thus, a description sample set for model training can be constructed by combining the prompt information corresponding to the sample data, the first intention inference data with the first wake-up intention flag added, and the second intention inference data with the second wake-up intention flag added.

[0080] Step S2015, constructing a multimodal description sample set based on the prompt data and the target inference data.

[0081] Combine the prompt data and the target reasoning data to determine the scene where the conversation occurs. According to the scene where the conversation occurs, design the structure of the conversation (such as the beginning, middle and end of the conversation), construct the conversation data according to the conversation structure, so that the conversation data conforms to the logic and context of the conversation, and obtain a multimodal description sample set through the constructed conversation data.

[0082] A specific description sample in the multimodal description sample set is shown below;

[0083] Prompt data:

[0084] I'm a little hungry, shall we go eat something?

[0085] Does this text contain potential user needs? If so, please output <functioncall>, and respond to the demand, otherwise output <nofunctioncall>.

[0086] Target inference data:

[0087] <functioncall>

[0088] I detect that you are expressing an underlying need - hunger, and I hope someone can respond and help. As a voice assistant, I can provide you with information about nearby restaurants or recommend some delicious food. Do you want to know about nearby restaurants?

[0089] By adding judgment prompt information to the annotated multimodal data and constructing the expected input and output patterns, the recognition model can be facilitated to process and analyze, which is beneficial to improving the recognition accuracy of the recognition model.

[0090] Step S202: for any target description sample in the multimodal description sample set, extract object wake-up information from the target description sample.

[0091] The object awakening information is information in the target description sample used to characterize the awakening timing. For any target description sample in the multimodal description sample set, the text content in the target description sample is processed by semantic understanding, context processing, dialogue management, etc. to extract the object awakening information carried in the target description sample.

[0092] Step S203: Process the object wake-up information using the recognition model, and output predicted dialogue information corresponding to the target description sample.

[0093] The predicted dialogue information is intended to interact with the target object to meet the current needs of the target object or provide assistance to the target object. As described above, the recognition model can be a model based on deep learning or machine learning algorithms. The object awakening information is input into the recognition model for processing and analysis. The recognition model will try to understand the meaning of the object awakening information and predict the most likely dialogue information of the object awakening information.

[0094] Step S204: updating the model parameters of the recognition model based on the predicted dialogue information and the wake-up intention description information.

[0095] Analyze the predicted dialogue information and the wake-up intention description information, determine the matching degree between the predicted dialogue information and the wake-up intention description information, and generate a matching result between the dialogue and the intention. If the matching degree between the predicted dialogue information and the wake-up intention description information is lower than a preset value (a preset matching degree for determining that the predicted dialogue information matches the intention description information), it means that the predicted dialogue information does not match the wake-up intention description information, that is, there is a deviation between the predicted dialogue information and the wake-up intention description information.

[0096] The error information between the predicted dialogue information and the wake-up intention description information is determined through the matching results (the error information is represented by the value of the loss function), and the gradient of the model parameters is calculated by back propagation (the gradient represents the degree of influence of the model parameters on the loss function). The recognition model is iteratively trained based on the error information and the gradient of the model parameters to achieve parameter updates for the recognition model, making the output results of the recognition model more accurate.

[0097] In some optional embodiments, such as Figure 7 As shown, the above step S204 includes:

[0098] Step S2041, obtaining pre-trained weight parameters of the recognition model.

[0099] Pre-trained weight parameters are model parameters of the recognition model. Pre-trained weight parameters are representations of the features and patterns of the data learned by the model, and can be used to initialize new model training or as a starting point in the fine-tuning process to speed up the convergence of the model and improve performance.

[0100] Specifically, if the recognition model is a pre-trained model, its pre-trained weight parameters can be directly downloaded through the relevant model library, for example, Transformer series models (such as GPT, BERT, etc.), CNN series models (such as ResNet, VGG, etc.), etc.

[0101] If the recognition model is a model trained based on a public data set, the corresponding model weight parameters will be generated and saved during the training process. At this time, the data space of the training model can be read to read the pre-trained weight parameters of the model from the data space.

[0102] The method for obtaining the pre-training weight parameters is determined based on the recognition model, and the method for obtaining the pre-training weight parameters is not specifically limited here.

[0103] Step S2042: construct a parameter fine-tuning branch according to the pre-trained weight parameters.

[0104] Load the obtained pre-trained weight parameters into the pre-trained recognition model, analyze the structure of the recognition model, and determine the location where the parameter fine-tuning branch (LORA branch) needs to be added. Figure 8 As shown, taking the generative large language model LLM as an example, parameter fine-tuning branches are added to each feature conversion layer in the linear layer and the feature conversion module (Transformer Block).

[0105] A parameter fine-tuning branch is constructed according to the pre-trained weight parameters of the recognition model, and the parameter fine-tuning branch is connected with the corresponding part of the original recognition model to form a complete model. The LORA branch is trained using data in the multimodal description sample set to adjust its parameters to adapt it to the intent recognition task.

[0106] The LORA branch is composed of trainable parameters and can be implemented by matrix decomposition. In some optional implementations, the above step S2042 includes:

[0107] Step b1, constructing a first low-rank matrix and a second low-rank matrix for pre-training weight parameters, wherein the first low-rank matrix is ​​a downsampling matrix and the second low-rank matrix is ​​an upsampling matrix.

[0108] Step b2: determine the first low-rank matrix and the second low-rank matrix as parameter fine-tuning branches.

[0109] The low-rank adaptive LORA branch is constructed according to the pre-trained weight parameters. The LORA branch includes two low-rank matrices, namely the first low-rank matrix and the second low-rank matrix, such as Fig. 9 As shown. The first low-rank matrix A is initialized using the normal distribution N(0,0.1), that is, the elements in the first low-rank matrix are randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.1. This initialization method can introduce a certain degree of randomness and diversity, which is helpful for the learning and generalization ability of the model. Moreover, the downsampling coefficient of the first low-rank matrix depends on the specifications of the pre-trained recognition model. Usually, the scaling factor scale=0.25 can be set, that is, the size of the first low-rank matrix A will be 0.25 times that of the original pre-trained weight matrix, thereby reducing the number of parameters of the recognition model while maintaining a certain degree of expression.

[0110] The second low-rank matrix B is initialized with 0. That is, all elements in the second low-rank matrix are initialized to 0, so that during the training process, these elements can be adjusted according to the learning situation of the recognition model.

[0111] By constructing a LORA branch consisting of a first low-rank matrix and a second low-rank matrix, the recognition model can be fine-tuned to adapt to a specific intent recognition task without significantly increasing the number of recognition model parameters.

[0112] Step S2043, when training the recognition model using the predicted dialogue information and the wake-up intention description information, freeze the pre-trained weight parameters and update the parameter weights of the parameter fine-tuning branch.

[0113] When the recognition model is iteratively trained using the predicted dialogue information and the wake-up intention description information, the pre-trained weight parameters are set to be non-trainable using the method provided by the model library of the recognition model, that is, the pre-trained weight parameters are frozen. Thus, during the training process, the pre-trained weight parameters will not be updated. At the same time, according to the difference between the predicted dialogue information and the wake-up intention description information output by the recognition model, the parameter weight update of the parameter fine-tuning branch (i.e., the LORA branch) is guided.

[0114] During the training process, only the parameter weights of the LORA branch will be updated, while the pre-trained weight parameters remain unchanged, so that the knowledge and features of the pre-trained recognition model can be used while fine-tuning specific tasks. This reduces the training time and computing resource requirements, and also helps avoid overfitting of the recognition model during training.

[0115] In some optional embodiments, such as Figure 8 As shown, the recognition model includes an output linear layer and multiple feature conversion layers, and the parameter weights of the above-mentioned update parameter fine-tuning branch include:

[0116] Step c1, during the process of training the recognition model, the first parameter weight of the output linear layer is always updated.

[0117] Step c2: randomly selecting a feature conversion layer from the multiple feature conversion layers based on a preset interval, and updating the second parameter weight of the randomly selected feature conversion layer.

[0118] The parameter weight of the parameter fine-tuning branch includes a first parameter weight and a second parameter weight.

[0119] In each training step, the parameter weights of the LORA branch of the output linear layer are updated, which ensures that the recognition model can be adjusted according to the latest input data and improves the output accuracy of the recognition model.

[0120] The preset interval is the number of preset interval training steps, for example, 5 training steps, 10 training steps, etc. According to the preset interval, a feature conversion layer is randomly selected from multiple feature conversion layers of the feature conversion module (Transformer Block) to update the parameter weight of the LORA branch. In this way, it can be ensured that different parts of the recognition model can be properly adjusted, thereby improving the generalization ability of the recognition model.

[0121] Through the above weight updating strategy, the recognition model can converge faster during the fine-tuning process and improve the recognition performance of the recognition model.

[0122] Step S2044: Update the model parameters of the recognition model using the updated parameter weights of the parameter fine-tuning branch.

[0123] The LORA branch is connected to the corresponding part of the original recognition model to form a complete recognition model. While training the LORA branch and updating the parameter weights of the LORA branch, the model parameters of other parts of the original recognition model can be fine-tuned to further improve the model's recognition performance.

[0124] The training method of the recognition model provided in this embodiment continuously updates the model parameters of the recognition model according to the predicted dialogue information and the intention description information, so that the recognition model can learn the wake-up timing in different scenarios, thereby improving the recognition performance and recognition accuracy of the recognition model.

[0125] In this embodiment, a method for identifying the timing of waking up an intelligent agent is provided, which can be used in computer devices, such as computers, etc. Fig.10 is a flow chart of a method for identifying an agent wake-up timing according to an embodiment of the present disclosure, such as Fig.10 As shown, the process includes the following steps:

[0126] Step S301, obtaining multimodal description information of the target object. Please refer to the relevant description of the corresponding steps in the above embodiment for details, which will not be repeated here.

[0127] Step S302, parse the multimodal description information and identify whether the multimodal description information carries wake-up information. If the multimodal description information carries wake-up information, execute step S303, for details, please refer to the relevant description of the corresponding steps in the above embodiment, which will not be repeated here; otherwise, execute step S304.

[0128] Step S303: If the multimodal description information carries wake-up information, the dialog agent is awakened based on the wake-up information, so that the dialog agent generates matching dialog information according to the wake-up information and the multimodal description information. For details, please refer to the relevant description of the corresponding steps in the above embodiment, which will not be repeated here.

[0129] Step S304: If the voice description information does not carry the wake-up information, the dialog agent is controlled to remain in a non-awakened state.

[0130] When it is detected that the voice description information sent by the target object does not carry the wake-up information, the dialogue agent will not be awakened, that is, it will continue to remain in the non-awakened state and will not respond to the voice description information sent by the target object. Fig.11 As shown in the figure, the output of the recognition model is split into two parts: the wake-up intention recognition flag tag and the dialogue information response. The tag is used to guide the flow of subsequent data, and the dialogue information response is used as the downstream input. In the smart cockpit scenario, after obtaining the flag, you can consider whether the current car system should remain silent or wake up the car system. After waking up the car system, the dialogue information response is used as input and fed back to the user in the form of dialogue.

[0131] The method for identifying the timing of awakening an intelligent agent provided in this embodiment keeps the dialogue intelligent agent in an unawakened state when it is detected that the voice description information emitted by the target object does not carry the awakening information. This can prevent the dialogue intelligent agent from being awakened due to erroneous triggering of keywords, thereby improving the awakening intelligence of the dialogue intelligent agent.

[0132] As a specific application example of the present disclosure, the open source Qwen-VL is used for fine-tuning experiments. Fig.12 The curve of the fine-tuning loss function is shown. The data set uses hotel check-in evaluation data, uses the method in the disclosure to construct data, and identifies users' positive reviews. 6,000 plain text data are used for training, and an additional 2,000 data are used for testing. The comparison of the wake-up timing recognition result data before and after fine-tuning using the disclosed method is shown in Table 1.

[0133] Table 1 Data comparison

[0134]

[0135] Among them, P represents the total number of samples with positive wake-up timing in the test set; N represents the total number of samples with negative wake-up timing in the test set; TP represents the number of samples with positive wake-up timing, and the generated results contain <functioncall>TN means that the generated results contain <nofunctioncall>The number of; FN means that the awakening timing is a positive evaluation sample, the generated results include <nofunctioncall>FP means that the generated results contain <functioncall>The number of samples with negative evaluation when they are woken up is N-Recall, N-Recall = TN / (TN+FP); N-Precision, N-Precision = TN / (TN+FN); P-Recall, P-Recall = TP / (TP+FN); P-Precision, P-Precision = TP / (TP+FP); Accuracy, Accuracy = (TP+TN) / (TP+TN+FP+FN).

[0136] It can be seen from the above data table that by using the method provided by the present invention to construct data and perform fine-tuning, on the hotel check-in evaluation data set, the wake-up timing recognition accuracy is improved from 9.97% to 88.5%, the positive sample precision is improved from 21.2% to 93.6%, the recall rate is improved from 11.5% to 89.4%, the negative sample precision is improved from 3.29% to 79%, and the recall rate is improved from 6.54% to 86.7%. The improvement of each indicator is huge.

[0137] Judging from the experimental results, the use of this public method can effectively enable the large language model to be familiar with the selection and judgment mode of wake-up timing recognition in edge scenarios, thereby greatly improving the recognition accuracy of wake-up timing in combination with the context.

[0138] The disclosed method has a certain degree of universality. In the use scenario of the in-vehicle smart cockpit, unlike the traditional keyword wake-up, the use of the publicly provided method to actively identify user needs and wake up the vehicle system can greatly improve the convenience of user interaction, and the usability and intelligence of the intelligent dialogue agent are also further improved.

[0139] In this embodiment, a device for identifying the time for waking up an intelligent agent is also provided, and the device is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware for a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0140] This embodiment provides a device for identifying the time when an intelligent agent wakes up. Fig.13 As shown, including:

[0141] The acquisition module 401 is used to acquire the speech description information of the target object.

[0142] The recognition module 402 is used to parse the voice description information and recognize whether the voice description information carries the wake-up information.

[0143] The wake-up module 403 is used to wake up the dialogue agent based on the wake-up information if the voice description information carries the wake-up information, so that the dialogue agent generates matching dialogue information according to the wake-up information and the voice description information.

[0144] In some optional implementations, the above-mentioned agent wake-up timing identification device further includes:

[0145] The model training module is used to train the recognition model so as to parse the multimodal description information through the pre-trained recognition model and identify whether the multimodal description information carries the wake-up information.

[0146] In some optional implementations, the model training module includes:

[0147] The sample acquisition unit is used to acquire a multimodal description sample set, wherein the multimodal description sample set includes multiple description samples of the target object in different scenarios, and the description samples include wake-up intention description information.

[0148] The wake-up information extraction unit is used to extract object wake-up information from any target description sample in the multimodal description sample set.

[0149] The prediction unit is used to process the object wake-up information using the recognition model and output the predicted dialogue information corresponding to the target description sample.

[0150] The first parameter updating unit is used to update the model parameters of the recognition model based on the predicted dialogue information and the wake-up intention description information.

[0151] The second parameter updating unit is used to update the model parameters of the recognition model using the parameter weights of the updated parameter fine-tuning branch.

[0152] In some optional implementations, the sample acquisition unit includes:

[0153] The data collection subunit is used to collect multimodal data of the target object and annotate the multimodal data with the wake-up intention to obtain annotation information.

[0154] The prompt information acquisition subunit is used to acquire judgment prompt information and add the judgment prompt information to the annotated multimodal data to obtain prompt data.

[0155] The reasoning subunit is used to perform reasoning according to the prompt data to obtain the intention reasoning data corresponding to the prompt data.

[0156] The flag adding subunit is used to add the wake-up intention flag to the intention reasoning data according to the annotation information to obtain the target reasoning data.

[0157] The construction subunit is used to construct a multimodal description sample set based on the prompt data and the target reasoning data.

[0158] In some optional implementations, the multimodal data includes positive sample data and negative sample data, and the intention inference data includes first intention inference data corresponding to the positive sample data and second intention inference data corresponding to the negative sample data. The above-mentioned flag adding subunit is specifically used to: clean the intention inference data to obtain cleaned data; add a first wake-up intention flag to the first intention inference data using the annotation information corresponding to the positive sample data, and the first wake-up intention flag is used to represent that the multimodal data carries wake-up information; add a second wake-up intention flag to the second intention inference data using the annotation information corresponding to the negative sample data, and the second wake-up intention flag is used to represent that the multimodal data does not carry wake-up information.

[0159] In some optional implementations, the parameter updating unit includes:

[0160] The weight parameter acquisition subunit is used to obtain the pre-trained weight parameters of the recognition model.

[0161] The branch construction subunit is used to construct parameter fine-tuning branches according to the pre-trained weight parameters.

[0162] The update subunit is used to freeze the pre-trained weight parameters and update the parameter weights of the parameter fine-tuning branch when training the recognition model using the predicted dialogue information and the wake-up intent description information.

[0163] In some optional embodiments, the branch construction subunit is specifically used to: construct a first low-rank matrix and a second low-rank matrix for pre-trained weight parameters, wherein the first low-rank matrix is ​​a downsampling matrix and the second low-rank matrix is ​​an upsampling matrix; determine the first low-rank matrix and the second low-rank matrix as parameter fine-tuning branches.

[0164] In some optional embodiments, the recognition model includes an output linear layer and multiple feature conversion layers, and the updating subunit is specifically used to: during the training of the recognition model, always update the first parameter weight of the output linear layer; randomly select a feature conversion layer from the multiple feature conversion layers based on a preset interval, and update the second parameter weight of the randomly selected feature conversion layer; wherein the parameter weights of the parameter fine-tuning branch include the first parameter weight and the second parameter weight.

[0165] In some optional implementations, the above-mentioned agent wake-up timing identification device further includes:

[0166] The state maintenance module controls the dialogue agent to remain in an unawakened state if the voice description information does not carry wake-up information.

[0167] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0168] The device for identifying the wake-up timing of the intelligent agent in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0169] The device for identifying the timing of awakening an intelligent agent provided by the embodiment of the present disclosure analyzes the voice description information to determine whether it carries the awakening information, and determines whether to wake up the dialogue intelligent agent based on the judgment result of carrying the awakening information. As long as the awakening information is carried, the dialogue intelligent agent is automatically awakened, which reduces the response delay of the dialogue intelligent agent, improves the awakening accuracy of the dialogue intelligent agent, and avoids false triggering of the dialogue intelligent agent.

[0170] The present disclosure also provides a computer device having Fig.13 The device for identifying the time to wake up an intelligent agent is shown.

[0171] See also Fig.14 , Fig.14 is a schematic diagram of a computer device provided by an optional embodiment of the present disclosure, such as Fig.14 As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Fig.14 A processor 10 is taken as an example.

[0172] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0173] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0174] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0175] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0176] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Fig.14 The example of connecting through bus is taken in the following.

[0177] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0178] The computer device also includes a communication interface, which is used for the computer device to communicate with other devices or a communication network.

[0179] The embodiments of the present disclosure also provide a computer-readable storage medium. The above-mentioned method according to the embodiments of the present disclosure can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium and downloaded through a network, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state drive, etc.; further, the storage medium can also include a combination of the above-mentioned types of memory. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0180] A part of the present disclosure may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present disclosure through the operation of the computer. Those skilled in the art should understand that the existence of computer program instructions in computer-readable media includes, but is not limited to, source files, executable files, installation package files, etc., and accordingly, the way in which computer program instructions are executed by a computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to the computer.

[0181] Although the embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations are all within the scope defined by the appended claims.< / functioncall> < / nofunctioncall> < / nofunctioncall> < / functioncall> < / functioncall> < / nofunctioncall> < / functioncall> < / nofunctioncall> < / functioncall> < / functioncall>

Claims

1. A method for identifying the timing of waking up an intelligent agent, characterized in that: The method comprises: Obtaining voice description information of the target object; Parsing the voice description information, and identifying whether the voice description information carries wake-up information; If the voice description information carries the wake-up information, the dialogue agent is woken up based on the wake-up information, so that the dialogue agent generates matching dialogue information according to the wake-up information and the voice description information.

2. The method according to claim 1, characterized in that The parsing of the voice description information and identifying whether the voice description information carries the wake-up information is implemented by a pre-trained recognition model; wherein the recognition model is trained in the following manner: Acquire a multimodal description sample set, wherein the multimodal description sample set includes a plurality of description samples of the target object in different scenarios, and the description samples include intention description information; For any target description sample in the multimodal description sample set, extracting object wake-up information from the target description sample; Processing the object awakening information using the recognition model, and outputting predicted dialogue information corresponding to the target description sample; Based on the predicted dialogue information and the intention description information, model parameters of the recognition model are updated.

3. The method according to claim 2, characterized in that The obtaining of a multimodal description sample set includes: Collecting multimodal data of the target object, and annotating the multimodal data with the wake-up intention to obtain annotation information; Acquire judgment prompt information, and add the judgment prompt information to the annotated multimodal data to obtain prompt data; Perform reasoning according to the prompt data to obtain intention reasoning data corresponding to the prompt data; Adding a wake-up intention flag to the intention inference data according to the annotation information to obtain target inference data; The multimodal description sample set is constructed based on the prompt data and the target inference data.

4. The method according to claim 3, characterized in that The multimodal data includes positive sample data and negative sample data, and the intention inference data includes first intention inference data corresponding to the positive sample data and second intention inference data corresponding to the negative sample data; The step of adding a wake-up intention flag bit to the intention inference data according to the annotation information to obtain target inference data includes: Cleaning the intention inference data to obtain cleaned first intention inference data and second intention inference data; Adding a first wake-up intention flag to the first intention inference data using the annotation information corresponding to the positive sample data, where the first wake-up intention flag is used to indicate that the multimodal data carries wake-up information; A second wake-up intention flag is added to the second intention inference data using the annotation information corresponding to the negative sample data, where the second wake-up intention flag is used to indicate that the multimodal data does not carry wake-up information.

5. The method according to claim 2, characterized in that: The updating of the model parameters of the recognition model based on the predicted dialogue information and the intention description information includes: Obtaining pre-trained weight parameters of the recognition model; Construct a parameter fine-tuning branch according to the pre-trained weight parameters; When training the recognition model using the predicted dialogue information and the intent description information, freezing the pre-trained weight parameters and updating the parameter weights of the parameter fine-tuning branch; The model parameters of the recognition model are updated using the updated parameter weights of the parameter fine-tuning branch.

6. The method according to claim 5, characterized in that The step of constructing a parameter fine-tuning branch according to the pre-trained weight parameters includes: Constructing a first low-rank matrix and a second low-rank matrix for the pre-trained weight parameters, wherein the first low-rank matrix is ​​a downsampling matrix and the second low-rank matrix is ​​an upsampling matrix; The first low-rank matrix and the second low-rank matrix are determined as the parameter fine-tuning branches.

7. The method according to claim 5, characterized in that The recognition model includes an output linear layer and a plurality of feature conversion layers; and updating the parameter weights of the parameter fine-tuning branch includes: During the training of the recognition model, the first parameter weight of the output linear layer is always updated; Randomly selecting a feature conversion layer from the plurality of feature conversion layers based on a preset interval, and updating a second parameter weight of the randomly selected feature conversion layer; The parameter weight of the parameter fine-tuning branch includes the first parameter weight and the second parameter weight.

8. The method according to claim 1, characterized in that Also includes: If the voice description information does not carry the wake-up information, the dialogue agent is controlled to remain in an unawakened state.

9. A device for identifying the timing of waking up an intelligent agent, characterized in that: The device comprises: An acquisition module, used to acquire the voice description information of the target object; A recognition module, used to parse the voice description information and identify whether the voice description information carries wake-up information; A wake-up module is used to wake up the dialogue agent based on the wake-up information if the voice description information carries the wake-up information, so that the dialogue agent generates matching dialogue information according to the wake-up information and the voice description information.

10. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method for identifying the wake-up timing of an intelligent body as described in any one of claims 1 to 8 by executing the computer instructions.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method for identifying the wake-up timing of an intelligent agent according to any one of claims 1 to 8.

12. A computer program product, characterized in that The method comprises computer instructions for causing a computer to execute the method for identifying the wake-up timing of an intelligent agent according to any one of claims 1 to 8.

Images