Request processing method and device, equipment, medium and program product

Abstract

The embodiment of the invention provides a request processing method and device, equipment, a medium and a program product, and relates to the technical field of intelligent driving. The method comprises the steps of obtaining a first request; identifying a user intention based on the first request, wherein the user intention comprises a query intention and / or an execution intention; and determining a target strategy based on the first request, the user intention and pre-acquired structured data. According to the method, the requirement of the user for scenarized service can be met, the user can be supported to inquire the historical behavior data of the user, the corresponding operation can be automatically executed according to the past behaviors of the user, and therefore the driving experience of the user is effectively improved.

Description

Technical Field

[0001] This application relates to intelligent driving technology, and more particularly to a request processing method, apparatus, device, medium, and program product. Background Technology

[0002] With the rapid development of new energy vehicles (such as pure electric vehicles, range-extended electric vehicles, and hybrid electric vehicles), the level of vehicle intelligence is constantly improving, and users' demand for scenario-based vehicle services is also growing.

[0003] In existing technologies, vehicle log systems primarily use a time-series approach to store operation records of various subsystems such as navigation, entertainment, and air conditioning, thereby forming log data. This log data generated by these subsystems is stored in a distributed manner within the vehicle's onboard storage, mainly for vehicle fault diagnosis and system maintenance.

[0004] However, the log system of the aforementioned vehicles only records the operation information of each subsystem, and does not perform structured integration and semantic processing on this scattered operation information. As a result, the system cannot support users in querying their own historical behavior data, nor can it automatically execute corresponding operations based on users' past behavior, thus failing to meet users' scenario-based service needs. Summary of the Invention

[0005] This application provides a request processing method, apparatus, device, medium, and program product to solve the problem that existing systems cannot support users in efficiently querying their own historical behavior data, nor can they automatically execute corresponding operations based on users' past behavior, thus significantly improving users' contextualized interactive experience.

[0006] In a first aspect, embodiments of this application provide a request processing method applied to an electronic device, the method comprising:

[0007] Get the first request;

[0008] Based on the first request, the user intent is identified, including query intent and / or execution intent;

[0009] Based on the first request, the user intent, and the pre-acquired structured data, a target strategy is determined;

[0010] The structured data includes the following correspondences: scene information, context information, and the historical execution action sequence of the electronic device; the historical execution action sequence is used to indicate the content and order of actions historically performed by the electronic device.

[0011] The target strategy includes: outputting information about the target execution action sequence, and / or executing the target execution action sequence; the target execution action sequence is a historical execution action sequence in the structured data that is adapted to the first request.

[0012] In one possible implementation, the scene information includes: first-level scene information and one or more second-level scene information corresponding to the first-level scene information, wherein the first-level scene information is used to describe the user's behavior on the electronic device, and the second-level scene information is a subdivision of the first-level scene information.

[0013] In one possible implementation, the structured data is obtained by processing data in the electronic device based on a target model; the target model is used to classify the data in the electronic device into scenarios, extract key information, and generate action sequences; the key information includes the context information.

[0014] In one possible implementation, the target model is trained using a dynamically weighted multi-task loss function, which includes a loss function that dynamically adjusts the weights of each task based on task relevance, dynamic weights, and confidence during model training.

[0015] In one possible implementation, the data in the electronic device includes: vehicle bus data, sensor data, application layer data, user interaction data, context data, and user profile data.

[0016] In one possible implementation, determining the target strategy based on the first request, the user intent, and pre-acquired structured data includes:

[0017] Based on the text information included in the first request, the target scenario is determined;

[0018] Obtain the target execution action sequence corresponding to the target scene from the structured data;

[0019] Verify the validity of each execution action parameter in the target execution action sequence;

[0020] If the user intent is the query intent and the parameters of each execution action in the target execution action sequence are valid, output the information of the target execution action sequence.

[0021] In one possible implementation, determining the target strategy based on the first request, the user intent, and pre-acquired structured data further includes:

[0022] If the user intent is the execution intent and each execution action parameter in the target execution action sequence is valid, output the information of the target execution action sequence;

[0023] Based on the user's confirmation of the target execution action sequence, the electronic device is controlled to sequentially execute the actions in the target execution action sequence.

[0024] In one possible implementation, the pre-acquired structured data is stored in the electronic device, and the method further includes:

[0025] The structured data is updated, wherein the update includes: retaining important data in the structured data, and / or deleting unimportant data after a preset time period has elapsed; the important data includes data in the structured data that has been used based on user requests, and the unimportant data includes data in the structured data that has not been used due to user requests.

[0026] Secondly, embodiments of this application provide a request processing apparatus applied to an electronic device, the apparatus comprising:

[0027] The acquisition module is used to acquire the first request;

[0028] The identification module is configured to identify user intent based on the first request, wherein the user intent includes query intent and / or execution intent;

[0029] The determination module is used to determine the target strategy based on the first request, the user intent, and pre-acquired structured data;

[0030] The structured data includes the following correspondences: scene information, context information, and the historical execution action sequence of the electronic device; the historical execution action sequence is used to indicate the content and order of actions historically performed by the electronic device.

[0031] The target strategy includes: outputting information about the target execution action sequence, and / or executing the target execution action sequence; the target execution action sequence is a historical execution action sequence in the structured data that is adapted to the first request.

[0032] In one possible implementation, the scene information includes: first-level scene information and one or more second-level scene information corresponding to the first-level scene information, wherein the first-level scene information is used to describe the user's behavior on the electronic device, and the second-level scene information is a subdivision of the first-level scene information.

[0033] In one possible implementation, the structured data is obtained by processing data in the electronic device based on a target model; the target model is used to classify the data in the electronic device into scenarios, extract key information, and generate action sequences; the key information includes the context information.

[0034] In one possible implementation, the target model is trained using a dynamically weighted multi-task loss function, which includes a loss function that dynamically adjusts the weights of each task based on task relevance, dynamic weights, and confidence during model training.

[0035] In one possible implementation, the data in the electronic device includes: vehicle bus data, sensor data, application layer data, user interaction data, context data, and user profile data.

[0036] In one possible implementation, the determining module is further configured to determine the target scene based on the text information included in the first request;

[0037] The acquisition module is specifically used to acquire the target execution action sequence corresponding to the target scene from the structured data;

[0038] The device further includes: a verification module;

[0039] The verification module is used to verify the validity of each execution action parameter in the target execution action sequence;

[0040] The device further includes: an output module;

[0041] The output module is specifically used to output information about the target execution action sequence when the user intent is the query intent and the parameters of each execution action in the target execution action sequence are valid.

[0042] In one possible implementation, the output module is specifically used to output information about the target execution action sequence when the user intent is the execution intent and the parameters of each execution action in the target execution action sequence are valid;

[0043] The device further includes: an execution module;

[0044] The execution module is used to control the electronic device to sequentially execute the actions in the target execution action sequence according to the user's confirmation action of the target execution action sequence.

[0045] In one possible implementation, the device further includes: an update module;

[0046] The update module is used to update the structured data, wherein the update includes: retaining important data in the structured data, and / or deleting unimportant data after a preset time period; the important data includes data in the structured data that has been used based on user requests, and the unimportant data includes data in the structured data that has not been used due to user requests.

[0047] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0048] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0049] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0050] The request processing method, apparatus, device, medium, and program product provided in this application first obtain a first request, then identify user intents containing query intents and / or execution intents based on the request, and finally determine a target strategy by combining the first request, user intents, and pre-acquired structured data. This method, by identifying user intents and combining the first request, user intents, and structured data, solves the problem that the system cannot support users in efficiently querying their own historical behavior data, nor can it automatically execute corresponding operations based on past user behavior, thus failing to meet users' scenario-based service needs. It can provide users with efficient, personalized, and scenario-specific services, thereby improving the user's scenario-based interaction experience. Attached Figure Description

[0051] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0052] Figure 1 A flowchart illustrating the request processing method provided in an embodiment of this application;

[0053] Figure 2 The flow of the request processing method provided in this embodiment Figure 2 ;

[0054] Figure 3A schematic diagram of the structure of the request processing apparatus provided in the embodiments of this application;

[0055] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0056] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0058] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0059] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0060] In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0061] In the embodiments of this application, the use of terms such as "first" and "second" is to distinguish between identical or similar items that have essentially the same function and effect. For example, "first electronic device" and "second electronic device" are merely used to distinguish different electronic devices and do not limit their order of execution. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different.

[0062] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0063] First, let me explain the terms used in this application:

[0064] Automatic Speech Recognition (ASR): ASR is a technology that converts human speech into computer-readable text in real time. ASR analyzes audio features (such as Mel-frequency spectrum and MFCC) using acoustic models and combines them with language models (such as N-gram and Transformer) to predict text sequences, ultimately outputting clean text content. For example, when a user says "Check tomorrow's weather" into their phone, the ASR system can recognize this speech as the corresponding text "Check tomorrow's weather".

[0065] Text-to-Speech (TTS) is a technology that converts text information stored in a computer into natural and fluent speech output. For example, after a computer processes a user's request to check the weather and obtains weather information text such as "Tomorrow's weather will be sunny, with a temperature of 20-25℃," TTS can convert this weather information text into corresponding speech.

[0066] A Large Language Model (LLM) is a deep learning model trained on large amounts of text data, enabling it to generate natural language text or understand the meaning of language text. A LLM consists of multiple layers, including an input layer, hidden layers, and an output layer. The input layer receives and encodes the input text, converting it into a numerical representation. The hidden layers process the input information through multiple layers of neural networks, including self-attention mechanisms, to capture complex semantics and contextual relationships in the text. The output layer converts the processed information into understandable language output, such as generating sentences or answering questions.

[0067] Self-attention is a core computational module in the hidden layer of large language models. Essentially, it is an algorithm that allows the model to dynamically calculate the degree of association between each word in the input text sequence and all other words. By assigning differentiated weights to words with different degrees of association, the model can accurately capture the complex semantic logic and contextual dependencies within the text. For example, when processing the first request text "search for restaurants I visited last week", the self-attention mechanism will assign a higher weight to "search" to highlight the core need, while establishing a temporal association between "last week" and "restaurants", allowing the model to accurately understand the user's true intent.

[0068] Low-Rank Adaptation (LoRA) is a fine-tuning technique for large language models. Its core principle is to train only a subset of the model's parameters, rather than fine-tuning the entire model, thus significantly reducing computational cost and time while maintaining performance. With LoRA, multiple tasks can share the same foundational knowledge of the model and adapt to their respective needs, enabling the model to handle various tasks more efficiently, such as text generation, sentiment analysis, and question answering. This approach not only improves the model's flexibility and versatility in multi-task environments but also reduces the amount of data and resources required for training.

[0069] Vector retrieval is an information retrieval technology based on a vector space model. Its core is to convert unstructured data such as text, images, and audio into high-dimensional vector representations and achieve efficient matching by calculating the similarity between vectors (such as cosine similarity, Euclidean distance, etc.).

[0070] With the rapid development of new energy vehicles (such as pure electric vehicles, range-extended electric vehicles, and hybrid electric vehicles), the level of vehicle intelligence is constantly improving, and users' demand for scenario-based vehicle services is also increasing. For example, a user might tell the vehicle via voice to "repeat last night's navigation route," expecting the vehicle to automatically execute last night's navigation route, rather than simply returning relevant information or only executing the first operation in the relevant command sequence.

[0071] In the existing technology, vehicle log systems mainly use a time-series approach to record the independent operation information of each subsystem, and store the operation information of each system in the form of logs in the on-board memory, which is mainly used for vehicle fault diagnosis and system maintenance.

[0072] The aforementioned vehicle log system has the following drawbacks: First, the log only records the operation information of each subsystem, such as "navigation started," "music played," and "air conditioning set to 25°C." These operation records can only reflect changes in the vehicle's operating status and cannot reflect the user's true behavioral intentions and the scenario. For example, is navigation started for commuting to work or for a weekend trip? Is music played for the user to relax or for entertainment with passengers?

[0073] Secondly, the log formats of different subsystems are inconsistent, and data is scattered and stored in their respective log files. For example, navigation logs record route information, media logs record playback content, and air conditioning logs record temperature settings. There is no effective correlation between these dimensions of data, resulting in severe data fragmentation and an inability to form a structured user behavior data system. For instance, when a user queries "restaurants visited last week," the existing vehicle log system cannot provide an accurate answer because it cannot correlate navigation, occupant, and other multi-dimensional information.

[0074] Finally, the vehicle log system lacks natural language understanding capabilities and cannot comprehend users' natural language, forcing users to manually filter logs through the system interface (such as filtering by time or type). For example, users cannot directly ask "What was my navigation route last night?" or "When was my last lunch break?"

[0075] To address the aforementioned issues, this application provides a request processing method. This method involves acquiring a first request, identifying the user's intent based on the first request, and then determining a target strategy based on the user's first request, user intent, and pre-acquired structured data. This method, through the fusion analysis of user request intent identification and structured data, solves the problem that existing systems cannot support users in efficiently querying their own historical behavior data, nor can they automatically execute corresponding operations based on past user behavior, thus failing to meet users' scenario-based service needs and improving the user's driving experience.

[0076] This application can be applied to automobiles and mobile terminals, and is suitable for solving the problem that existing systems cannot support users in querying their own historical behavior data, nor can they automatically execute corresponding operations based on users' past behavior, thus failing to meet users' scenario-based service needs.

[0077] The entity implementing this application can be the log system of the aforementioned vehicle, or it can be the log system of a mobile terminal. This application does not impose any special restrictions in this regard.

[0078] The technical solutions of this application will be described in detail below with reference to specific embodiments. The specific embodiments described below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0079] Figure 1 This is a flowchart illustrating the request processing method provided in an embodiment of this application. Figure 1 As shown, the method includes:

[0080] S101, Obtain the first request.

[0081] Understandably, the first request refers to the instruction initiated by the user to the electronic device to trigger a contextualized service. For example, the first request might be "repeat last night's navigation route".

[0082] The first request can be obtained, for example, through natural language input, where the user simply speaks the corresponding voice command. For instance, when a user says, "Find the restaurants I visited last week," the electronic device's built-in ASR module first receives the user's voice signal, performs preprocessing such as noise reduction and frame segmentation, and then uses a speech recognition algorithm to convert the voice signal into a corresponding text format request, thus obtaining the first request.

[0083] It can also be obtained through the control panel. Various function buttons are pre-set on the control panel of electronic devices, such as "historical trip query" and "navigation settings". When the user has an operation need, he / she can directly click the corresponding button. The sensor built into the control panel will sense the pressing action and convert the physical signal into an electrical signal, thereby obtaining the first request.

[0084] Since the first request is the trigger for the scenario-based service and can directly trigger the system to start the scenario-based service process, obtaining the first request can provide input basis for subsequent matching of scenario-based service strategies.

[0085] S102. Identify user intent based on the first request, whereby user intent includes query intent and / or execution intent.

[0086] Understandably, query intent refers to a user's desire to obtain specific information. For example, assuming the electronic device is a car, when the first request is "find all the restaurants I visited last week," it indicates that the user wants the car to display the specific names and locations of the restaurants visited last week via a display module or voice prompts.

[0087] An execution intent refers to a user's expectation that an electronic device will perform a specific action. For example, suppose the electronic device is a car. If the first request is "Go to the restaurant I went to yesterday again," it means that the user expects the car to automatically plan a navigation route, guide the user to the restaurant they visited yesterday, and display a pop-up window asking the user if they want to make a reservation.

[0088] By identifying the first request and obtaining the user's intent, the system can not only transform the raw, unstructured data input by the user, such as spoken voice commands, into requirements that the system can recognize, thus solving the problem of existing systems lacking natural language understanding capabilities and being unable to understand the user's natural language, but also provide a reference for subsequent determination of target strategies by combining structured data, ensuring that the system can match the user's actual needs and avoiding service deviations.

[0089] This step identifies the first request, and the user intent can be obtained, for example, through a large language model. First, the first request is preprocessed, such as through text cleaning and text conversion, to obtain a text-formatted first request. Then, the converted first request is input into the input layer of the large language model. This layer receives the text data and encodes it, converting the discrete text content into a numerical representation that the large language model can compute. Next, the numerical information is passed to a hidden layer composed of multiple neural networks. This layer uses a self-attention mechanism to capture the core semantics and contextual relationships in the text content. After that, the hidden layer passes the processed information to the output layer. Finally, the output layer converts the information processed by the hidden layer into a clear and understandable result, outputting an intent determination result containing the query intent or execution intent.

[0090] S103. Determine the target strategy based on the first request, user intent, and pre-acquired structured data;

[0091] The structured data includes the following correspondences: scene information, context information, and the historical execution sequence of the electronic device; the historical execution sequence is used to indicate the content and order of actions performed by the electronic device in the past.

[0092] The target strategy includes: outputting information about the target execution action sequence, and / or executing the target execution action sequence; the target execution action sequence is a historical execution action sequence in structured data that is adapted to the first request.

[0093] As is understandable, scenario information refers to the hierarchical scenario attribution determination results related to user requests, including first-level scenario information and one or more second-level scenario information corresponding to the first-level scenario information. For example, the "dining-restaurant" scenario in the context of driving.

[0094] Contextual information refers to background auxiliary information directly associated with the user's request event, including but not limited to: the time of the event, vehicle location, real-time weather, and destination.

[0095] Historical execution action sequences refer to the set of instructions executed sequentially in response to past user requests, which can completely record the content and execution order of actions during the historical service process. For example, after a user makes a request to "find nearby gas stations", a historical execution action sequence is generated as follows: start navigation → search for nearby gas stations → filter for the nearest station → push navigation route.

[0096] A target strategy refers to the final response plan determined based on user requests, intentions, and structured data analysis. Specifically, it can be divided into two forms: one is to output information about the target action sequence, and the other is to directly execute the action sequence, or both.

[0097] The target execution action sequence refers to the historical execution action sequence retrieved from structured data that is highly adapted to the current user's first request, and can be directly used to respond to or execute user needs.

[0098] The pre-acquired structured data can be obtained, for example, from the vector database built into the electronic device. This vector database contains the correspondence data between scene information, context information, and the historical execution sequence of the electronic device, which can provide efficient retrieval support for subsequent intent matching and strategy decision-making.

[0099] By matching the user's initial request, the parsed user intent, and pre-processed structured data with the historical execution sequence that matches the current needs, and then determining the target strategy that can effectively respond to user needs, it can support users to efficiently query their own historical behavior data, and enable electronic devices to automatically execute corresponding operations based on the user's past interaction habits, thereby realizing the intelligent and efficient processing of user requests by electronic devices.

[0100] Optionally, this application provides a method for outputting information about a target's execution action sequence, including: verbally broadcasting the specific content of the target's execution action sequence to the user via a TTS speech synthesis module built into an electronic device, or displaying it via a display module of the electronic device.

[0101] In one possible implementation, the scenario information provided in this application includes: first-level scenario information and one or more second-level scenario information corresponding to the first-level scenario information. The first-level scenario information is used to describe the user's behavior on the electronic device, and the second-level scenario information is a subdivision of the first-level scenario information.

[0102] For example, this application provides a classification table of first-level scene information, as shown in Table 1.

[0103] Table 1

[0104]

[0105] As shown in Table 1, the first-level scenario information includes the scenario name and related descriptions, which can intuitively reflect the user's specific behavior. For example, when the scenario name is "driving scenario," the corresponding description can reflect the user's behavior of driving to their destination.

[0106] For example, if the name of the first-level scenario is a driving scenario, this application provides a classification table of second-level scenario information under the driving scenario, as shown in Table 2.

[0107] Table 2

[0108]

[0109] As shown in Table 2, the second-level scenario information is a further refinement of the first-level scenario information. For example, when the first-level scenario information is "behavioral scenario", the specific scenario can be accurately determined through the second-level scenario information, such as "user commuting - going to the company".

[0110] For example, if the name of the first-level scene is "rest scene", this application provides a classification table of second-level scene information under the rest scene, as shown in Table 3.

[0111] Table 3

[0112]

[0113] In one possible implementation, the present application can determine structured data in the following way: the structured data is obtained by processing data in an electronic device based on a target model; the target model is used to classify the data in the electronic device into scenarios, extract key information, and generate action sequences; the key information includes context information.

[0114] Understandably, the target model could be, for example, a large language model fine-tuned using LoRA techniques.

[0115] Scene classification is a hierarchical scene attribution determination performed by the target model on data from electronic devices, used to determine the corresponding behavioral scene when a user uses the electronic device. For example, assuming the data from the electronic device contains the user's voice command "Find me a nearby restaurant with high ratings," the target model can determine the first-level scene information as a driving scene, and the second-level scene information as the sub-scene "dining—restaurant" within the "driving scene."

[0116] Key information is a set of elements related to the current scene extracted by the target model from the input data of the electronic device, which includes contextual information. For example, the target model extracts the following contextual information from the relevant data of the above driving scene: current time (e.g., 18:00), vehicle location (e.g., intersection of XX Road and XX Road), real-time weather (e.g., real-time weather is light rain), and destination (e.g., restaurant).

[0117] An execution action sequence is a set of instructions generated by the target model based on scene classification results and key information, which can be executed sequentially by electronic devices. For example, for the "dining-restaurant" sub-scenarios in the above driving scenario, and combined with the key information in this scenario, the target model generates the following execution action sequence: Activate the in-vehicle navigation function → Search for restaurants within 1 kilometer of the vehicle's current location → Filter for shops that are open and have parking spaces → Push the optimal route to the central control display module → Voice broadcast the navigation start and end points.

[0118] The process by which the target model processes data from electronic devices includes the following steps: First, the model input layer receives data from the electronic device and directly converts the discrete multi-source data into a numerical vector that the model can compute, without requiring manual feature selection. Then, this numerical vector is passed to a hidden layer composed of multiple neural networks. This layer uses LoRA fine-tuning technology to embed pre-defined multi-scene classification knowledge into the model parameters, while simultaneously performing deep feature extraction and fusion on the input data. In a single inference process, it simultaneously completes three tasks: scene classification, key information extraction, and generation of historical execution action sequences. Afterward, the hidden layer passes the processing results of these three tasks to the output layer. Finally, the output layer integrates the scene classification results, key information extraction results, and historical execution action sequence generation results into structured data, and then vectorizes this structured data before storing it in a vector database, providing data support for subsequent scene memory retrieval.

[0119] In one possible implementation, this application provides a training process for a target model, including: when training the target model, the neural network model is trained using a dynamically weighted multi-task loss function, wherein the dynamically weighted multi-task loss function includes: during model training, a loss function that dynamically adjusts the weights of each task based on task relevance, dynamic weights, and confidence.

[0120] Understandably, task relevance is captured by introducing a task relevance matrix R to show the dependencies between different tasks. The elements in the matrix... Indicates task and tasks The degree of correlation between them.

[0121] For example, assuming a task relevance matrix R is constructed based on 3 sets of samples, then: .

[0122] The first column represents the scene classification task, i.e., task 1; the second column represents the key information extraction task, i.e., task 2; and the third column represents the action sequence generation task, i.e., task 3.

[0123] diagonal elements in a matrix , , , which represents the relevance of the task itself, is usually defined as 1.0, meaning that the task itself is fully relevant.

[0124] Off-diagonal elements in a matrix The degree of correlation between the scene classification task (Task 1) and the key information extraction task (Task 2). , indicating the degree of correlation between the scene classification task (task 1) and the action sequence generation task (task 3).

[0125] ------Formula 1

[0126] Formula 1 represents the precondition function; Indicates an indicator function; when hour, The value is 1, otherwise it is 0; This represents the original loss value for the scene classification task; This is a preset loss threshold used to determine whether the scene recognition task has "made an error" or "has not yet learned well." When When the learning effect of the scene classification task is poor, it is necessary to adjust the loss weight of the associated tasks based on the task relevance and strengthen the joint optimization of the associated tasks.

[0127] -----Formula 2

[0128] Wherein, Formula 2 represents the loss adjustment function. Indicates task and tasks The degree of correlation between them This represents the task loss value for tasks that were not adjusted. The table shows the task loss value after adjustment.

[0129] For example, taking three tasks—scene classification, key information extraction, and action sequence generation—as an example, assuming a preset loss threshold... When the original loss value of the scene classification task When 0.6 is greater than 0.5, we can obtain the result according to Formula 1. The value is 1.

[0130] Furthermore, assuming the task relevance matrix contains... (Indicating the correlation between the scene classification task and the key information extraction task), the original loss value of the key information extraction task. According to Formula 2, the adjusted key information is used to extract the task loss value. This indicates that when the scene classification task fails, the loss weight of the key information extraction task increases, prompting the model to optimize these tasks simultaneously. This avoids a decline in overall performance due to errors in the fundamental scene classification task, and improves the model's comprehensive performance across multiple tasks.

[0131] The dynamic weight adjustment process consists of three sequential steps: calculating the loss decrease rate, calculating the normalized weights, and smoothly updating the weights. These three steps enable dynamic adjustment of task weights based on the loss decrease rate of each task during training.

[0132] The loss descent rate is a quantitative measure of how quickly the loss value decreases for each task during model training. It clarifies the learning dynamics and difficulty differences of each task during training, providing crucial information for subsequent adjustments to task weights. The functional expression for the loss descent rate is:

[0133] -----Formula 3

[0134] in, Represents the task number; It is the window size. For example, setting it to 10 means that the loss changes over 10 consecutive training rounds should be considered. Indicates task In the The loss value during each training round. This formula reflects the decreasing trend of task loss by calculating the average difference between the loss values ​​of adjacent rounds within a window.

[0135] For example, suppose there is a classification task where the loss value is 0.8 in the first round of training (rounds 1-10), the loss value is 0.7 in the second round, and so on. The formula can be used to obtain the rate of loss reduction within 10 rounds.

[0136] Normalized weights are applied by standardizing the weights of each task based on the loss descent rate, ensuring that the sum of all task weights equals 1. The aim is to allocate model resources fairly and reasonably in multi-task learning, ensuring that the impact of each task on model training is within an appropriate proportion. The functional expression for normalized weights is:

[0137] -------Formula 4

[0138] in, Indicates task In the Normalized weights during round training; It is a task In the The rate of loss reduction during each training round, calculated using a specific formula, reflects the task... The rate at which the loss value decreases.

[0139] For example, suppose there are three tasks, and the loss decrease rate of task 1 is... The rate of decrease in losses for Task 2 The rate of decrease in losses for Task 3 Then, by calculating the reciprocal of the rate of decrease of loss for each task, and then normalizing the calculation according to the formula, we can obtain the result for each task on the [missing information]. The normalized weights of the wheel.

[0140] Smooth weight updates aim to make the adjustment of task weights more stable and continuous, avoiding significant fluctuations in weights between adjacent rounds that could affect the stability of model training. This is achieved by introducing an update rate, which weights the current round's weights and the new weights calculated based on the current loss decrease rate. The functional expression for smooth weight updates is:

[0141] -------Formula 5

[0142] in, This indicates the update speed, such as setting it to 0.1. When the weights are larger, they can respond more quickly to changes in the rate of loss decrease; When the weights are smaller, the weight updates are more stable, retaining more weight information from the previous round; these are the tasks. In the Wheel and the first Weights during training rounds.

[0143] For example, setting If the weight of task 1 in the current round is By calculating the new weights based on the current rate of loss decrease using the previous method, and then performing a weighted calculation according to the formula, the smooth update weights for the next round of Task 1 can be obtained. .

[0144] Confidence-weighted loss function is a method used for specific tasks (such as scene classification) where the loss function is adjusted based on the model's confidence level in predicting the results for that task. This allows the model to focus more on samples with high prediction difficulty and low confidence during the learning process, while avoiding overfitting to simple samples, thus improving the model's generalization ability. The confidence-weighted loss function is expressed as follows:

[0145] ---Formula 6

[0146] in, This is the original loss value for the scene classification task. This represents the confidence level of the model's classification result for that scene, with a value ranging from 0 to 1. It is the loss value of the scene classification task after confidence weighting.

[0147] For low-confidence samples: when the confidence level of the model output... When smaller, The value will be relatively large, then the weighted loss value will be large. The loss weight will be relatively high. This means that the model assigns higher loss weights to low-confidence samples (usually samples that are difficult to classify accurately, i.e., hard samples). During model training, in order to reduce the overall loss value, the model will pay more attention to these hard samples, investing more resources in learning how to correctly classify them, thereby improving the model's ability to handle complex situations.

[0148] High-confidence samples: When the confidence level of the model output is high. When it is large, The value is small, and the weighted loss value is small. Therefore, the generalization rate is relatively low. This approach avoids the model from focusing too much on simple samples that it can already classify well, preventing the model from overfitting to these simple samples and giving the model better generalization ability.

[0149] Based on this, the final loss function is provided, and its expression is:

[0150] ---Formula 7

[0151] in, It is the confidence-weighted loss value for the scene classification task mentioned earlier; and These are the loss values ​​of the key information extraction task and the action sequence generation task after being weighted by scene loss correlation; , , During training, the weights are adjusted using dynamic weight adjustment methods (such as the previously mentioned method of dynamically adjusting weights based on the rate of loss descent). The weights of each task determined during the training round.

[0152] During model training, to more rationally adjust the loss function for each task weight, multiple factors such as task relevance, dynamic weights, and confidence levels can be considered. First, based on task relevance, tasks like key information extraction and action sequence generation are weighted according to the loss of scene classification tasks, reflecting the impact of the correlation between different tasks on the loss. Next, according to the dynamic weight adjustment mechanism, the weights are dynamically determined based on the rate of loss decrease for each task during training. Tasks with slower loss decrease have increased weights, making the model focus more on them, while tasks that have converged have decreased weights to avoid overfitting. Simultaneously, for scene classification tasks, the loss is also weighted according to the confidence level of the model output. Low-confidence samples have higher loss weights after weighting, allowing the model to focus more on difficult samples, while high-confidence samples have lower loss weights to prevent overfitting to simple samples. Finally, the total loss function is calculated by combining the task losses weighted by task relevance and confidence with the task weights obtained from dynamic weight adjustment. This guides the model to allocate resources rationally during training, improving overall performance.

[0153] In one possible implementation, the data in the electronic device provided in this application includes: vehicle bus data, sensor data, application layer data, user interaction data, context data, and user profile data.

[0154] As is understandable, vehicle bus data refers to the collection of various data transmitted through the vehicle's onboard bus (such as CAN bus, LIN bus, Ethernet) that reflects the vehicle's own operating status and control commands. Vehicle bus data includes, but is not limited to: engine speed, coolant temperature, brake pedal status, door open / closed status, transmission gear position, and vehicle speed.

[0155] Sensor data refers to the raw data collected by various onboard physical sensors used to perceive the vehicle's surrounding environment and its own status. Examples include obstacle distance and speed collected by millimeter-wave radar, lane line positions and traffic sign information collected by cameras, 3D point cloud data collected by lidar, and tire pressure data collected by tire pressure sensors.

[0156] Application layer data refers to the business data generated during the operation of various applications in in-vehicle intelligent terminals. Examples include route planning information, waypoint coordinates, and estimated arrival time for navigation applications; music playback and volume settings for entertainment applications; and air conditioning temperature and seat adjustment positions for vehicle control applications.

[0157] Dialogue data refers to the interactive data generated between users and in-vehicle intelligent interaction systems through voice, text, and other means. Examples include the user's voice command text, the system's response content, the dialogue initiation time, and the dialogue duration.

[0158] Contextual information includes, but is not limited to: the location of the event, the vehicle's location, and real-time weather conditions. For example, when a user issues a voice request to the vehicle's in-vehicle system to "navigate to XX Road XX Intersection," the specific contextual information corresponding to this interaction event is: current time 18:00, real-time vehicle location XX Road XX Intersection, and real-time weather light rain.

[0159] User profile data refers to tagged data built upon users' historical behavior and preferences to characterize user traits. Examples include a user's frequent travel routes, preferred air conditioning temperature range, favorite music genres, and driving habits.

[0160] The request processing method provided in this application first obtains a first request, then identifies the user intent, which includes query intent and / or execution intent, based on the request, and finally determines the target strategy by combining the first request, the user intent, and pre-acquired structured data. This method, by identifying the user intent and combining the first request, the user intent, and the structured data, solves the problem that the system cannot support users in efficiently querying their own historical behavior data, nor can it automatically execute corresponding operations based on past user behavior, thus failing to meet users' scenario-based service needs. It can provide users with efficient, personalized, and scenario-based services, thereby improving user satisfaction.

[0161] Figure 2 The flow of the request processing method provided in this embodiment Figure 2 .like Figure 2 As shown. This embodiment is... Figure 1 Based on the embodiments, the implementation process of the request processing method is described in detail. The request processing method provided in this embodiment includes:

[0162] S201, Obtain the first request.

[0163] The explanation of step S201 is the same as that in the above embodiments, and will not be repeated here.

[0164] S202. Identify user intent based on the first request, whereby user intent includes query intent and / or execution intent.

[0165] The explanation of step S202 is the same as that in the above embodiments, and will not be repeated here.

[0166] S203, Based on the text information included in the first request, determine the target scenario.

[0167] Understandably, the text information in the first request is the direct carrier of user needs. The keywords and sentence logic contained therein can intuitively reflect the user's interaction purpose. For example, if the text information includes the keyword "music" in "play yesterday's music", it can directly point to the "entertainment" scenario. Compared with other non-text input forms, text information has the advantages of clear semantics and easy feature extraction. Determining the target scenario based on this text information can ensure the accuracy and efficiency of scenario determination and avoid subsequent processing deviations due to information ambiguity.

[0168] This step determines the target scene in a way that can be achieved, for example, through a large language model fine-tuned using LoRA technology. The specific process is as follows: First, the text information of the first request is input into the input layer of the large language model, where the input layer segments and encodes the text, converting the unstructured text content into a numerical vector representation that the model can compute. Second, the hidden layer of the model receives the numerical vector passed from the input layer and, relying on the scene classification knowledge injected through LoRA fine-tuning, uses a self-attention mechanism to deeply mine the scene keywords, semantic features, and demand logic in the text, matching them with a preset scene classification system. Subsequently, the hidden layer passes the scene matching result to the output layer. Finally, the output layer decodes and normalizes the processing result of the hidden layer, outputting the target scene with the highest matching degree to the current text information.

[0169] S204. Obtain the sequence of target execution actions corresponding to the target scene from the structured data.

[0170] Understandably, structured data pre-stores a clear correspondence between scenario information and historical action sequences. These historical action sequences represent the solutions electronic devices have used to handle similar scenarios in the past. Therefore, obtaining the target action sequence based on this data not only avoids the computational waste caused by the system repeatedly generating action sequences, but also ensures that the response solution remains consistent with the user's historical interaction habits, thereby significantly improving the response speed of electronic devices to user requests.

[0171] This step of determining the target execution action sequence can be achieved, for example, through vector retrieval. The specific process is as follows: First, the identified target scene is used as the search keyword; second, the structured data in the built-in vector database of the electronic device is retrieved, and vector retrieval technology is used to match the scene information entries with the highest relevance to the target scene; finally, the historical execution action sequence bound to the matched entries is identified and determined as the target execution action sequence.

[0172] Assuming the target scenario is the "music-playback" scenario within the "entertainment" scenario, firstly, the "music-playback" scenario is used as the search keyword to retrieve structured data stored in the vector database; secondly, entries corresponding to the "music-playback" scenario are matched in the structured data, and the context information associated with these entries is "user prefers to play music listened to yesterday," and the bound historical execution action sequence is "wake up the in-car music module → retrieve the user's playback records from yesterday → play music in the historical playback order"; finally, this historical execution action sequence is determined as the target execution action sequence.

[0173] In one possible implementation, where pre-acquired structured data is stored in an electronic device, this application provides a method for updating the structured data, including: updating the structured data, wherein the update includes: retaining important data in the structured data, and / or deleting unimportant data after a preset time period has elapsed; important data includes data in the structured data that has been used based on user requests, and unimportant data includes data in the structured data that has not been used due to user requests.

[0174] Understandably, the preset duration can be, for example, 30 days or 20 days.

[0175] With structured data stored locally on the electronic device, the structured data is first divided into two categories, important data and unimportant data, based on whether the data has been requested by the user. Then, an update operation is performed to retain important data and delete unimportant data after a preset time. This can prevent local storage space from being excessively occupied by invalid data without affecting the user's core usage needs.

[0176] S205. Verify the validity of each execution action parameter in the target execution action sequence.

[0177] Understandably, the target execution action sequence is obtained based on historical structured data matching. The parameters it carries may become invalid due to environmental changes or system status updates. For example, the "gas station location parameter" in historical navigation actions may become invalid if the station is closed. Verifying the validity of these parameters can promptly remove invalid information, prevent electronic devices from executing incorrect actions, and at the same time ensure the accuracy and security of action execution, thereby improving the user interaction experience.

[0178] In verifying the validity of each execution action parameter in the target execution action sequence, the core contents to be verified include, but are not limited to: whether the destination is still valid, whether the waypoints are still valid, and whether the playlist exists.

[0179] The validity verification of the action parameters in this step can be achieved in the following way, with the specific process including: First, extract the core parameters of each action in the target action sequence and clarify the validity judgment rules for each parameter; Second, call the real-time status data of the electronic device and the real-time service data of the network side, and compare and verify them with the extracted core parameters; Finally, mark the valid parameters and invalid parameters according to the verification results, and correct or remove the actions corresponding to invalid parameters to ensure that the parameters of the remaining action sequence meet the execution requirements.

[0180] For example, the specific process for validating the parameters of the target action sequence "wake up the car music module → retrieve yesterday's playback history → play music in historical playback order" is as follows: First, extract the core parameters of each action in the sequence and clarify the validity judgment rules for each parameter. The core parameter for "wake up the car music module" is the running status of the car music module, and the judgment rule is whether the module is in a normal state where it is not occupied and can be woken up; the core parameter for "retrieve yesterday's playback history" is the storage path and data integrity of yesterday's playback history, and the judgment rule is whether the record file exists and is not corrupted; the core parameter for "play music in historical playback order" is the access link of the audio resources in the playback history, and the judgment rule is whether the link is valid and whether the audio can be played normally.

[0181] Secondly, the system calls the real-time status data of the electronic device to detect whether the in-vehicle music module is currently occupied by other applications. At the same time, it retrieves the user's playback record file from yesterday's local storage to verify its integrity, and then verifies the validity of each audio resource link in the record through the network-side service interface.

[0182] Finally, based on the verification results, valid and invalid parameters are marked, and actions corresponding to invalid parameters are corrected or eliminated. For example, if an audio resource link is detected to be invalid, "play music in historical playback order" is corrected to "skip invalid audio and play the remaining valid audio", ensuring that the adjusted action sequence parameters all meet the execution requirements.

[0183] S206. If the user intent is a query intent and the parameters of each execution action in the target execution action sequence are valid, output the information of the target execution action sequence.

[0184] Understandably, firstly, when the user's intent is to query, it means that the user's need is to obtain information rather than to directly perform an operation; secondly, the validity of the target action sequence parameters is a prerequisite for information output. If the parameters are invalid, the output action sequence information has no practical reference value and may even mislead the user.

[0185] Therefore, if the user's intent is determined to be a query intent and all parameters of the target execution action sequence are valid, the information of the target execution action sequence can be output, for example, through voice broadcasting via the TTS speech synthesis module built into the electronic device, or through display via the display module of the electronic device.

[0186] S207. If the user intent is an execution intent and the parameters of each execution action in the target execution action sequence are valid, output the information of the target execution action sequence.

[0187] Understandably, firstly, when the user's intent is to perform an action, it means that the user's need is for the electronic device to complete a specific operation, rather than to obtain information; secondly, the validity of the target action sequence parameters is a prerequisite for information output. If the parameters are invalid, the output action sequence information has no practical reference value and may even mislead the user.

[0188] Therefore, when the user's intent is to execute, and the parameters of each execution action in the target execution action sequence are valid, the information of the target execution action sequence is output to confirm with the user, for example, by voice broadcasting through the TTS speech synthesis module built into the electronic device, or by displaying through the display module of the electronic device.

[0189] S208. Based on the user's confirmation of the target execution action sequence, control the electronic device to sequentially execute the actions in the target execution action sequence.

[0190] Understandably, the user's confirmation action on the sequence of actions to be performed can verify whether the sequence fully matches the user's actual needs, effectively avoiding unexpected operations caused by the system due to misjudgment of the scenario, parameter deviation, or other issues.

[0191] Therefore, after confirming the user's confirmation action for the target action sequence, it is necessary to control the electronic device to execute the actions in the target action sequence sequentially.

[0192] This step determines the user's confirmation action for the target's action sequence. This can be obtained through multimodal interaction methods supported by the electronic device. For example, the complete information of the target's action sequence can be provided to the user through the electronic device's display module or TTS speech synthesis module. Then, the confirmation command from the user is received through any interaction method such as voice command, touch operation, or physical button, thereby completing the acquisition of the user's confirmation action.

[0193] The request processing method provided in this application first obtains the user's first request, then identifies the query intent or execution intent based on the request, and determines the target scenario based on the text information in the request; subsequently, it retrieves the target execution action sequence corresponding to the target scenario from the structured data and verifies the validity of each execution action parameter in the sequence; for query intent, if the parameters are valid, it directly outputs the target execution action sequence information; for execution intent, if the parameters are valid, it first outputs the target execution action sequence information, and then controls the electronic device to execute the actions in the sequence sequentially based on the user's confirmation command.

[0194] This method solves the technical problem that the system cannot support users to efficiently query their own historical behavior data, nor can it automatically execute corresponding operations based on users' past behavior, thus failing to meet users' contextual service needs. It achieves intelligent and efficient response of electronic devices to user requests by matching user requests with scene action sequences in historical structured data and adding parameter validity verification and user confirmation steps.

[0195] Figure 3 This is a schematic diagram of the structure of the request processing device provided in the embodiments of this application, such as... Figure 3 As shown in the figure, this application provides a request processing apparatus applied to an electronic device, the apparatus comprising:

[0196] Module 301 is used to obtain the first request;

[0197] The identification module 302 is used to identify the user intent based on the first request, wherein the user intent includes query intent and / or execution intent;

[0198] The determination module 303 is used to determine the target strategy based on the first request, user intent, and pre-acquired structured data;

[0199] The structured data includes the following correspondences: scene information, context information, and the historical execution sequence of the electronic device; the historical execution sequence is used to indicate the content and order of actions performed by the electronic device in the past.

[0200] The target strategy includes: outputting information about the target execution action sequence, and / or executing the target execution action sequence; the target execution action sequence is a historical execution action sequence in structured data that is adapted to the first request.

[0201] In one possible implementation, the scene information includes: first-level scene information and one or more second-level scene information corresponding to the first-level scene information. The first-level scene information is used to describe the user's behavior on the electronic device, and the second-level scene information is a subdivision of the first-level scene information.

[0202] In one possible implementation, the structured data is obtained by processing data in the electronic device based on a target model; the target model is used to classify the data in the electronic device into scenarios, extract key information, and generate action sequences; the key information includes contextual information.

[0203] In one possible implementation, the target model is trained using a dynamically weighted multi-task loss function. This dynamically weighted multi-task loss function includes a loss function that dynamically adjusts the weights of each task based on task relevance, dynamic weights, and confidence during model training.

[0204] In one possible implementation, the data in the electronic device includes: vehicle bus data, sensor data, application layer data, user interaction data, context data, and user profile data.

[0205] In one possible implementation, the determining module 303 is further configured to determine the target scene based on the text information included in the first request;

[0206] The acquisition module 301 is specifically used to acquire the sequence of target execution actions corresponding to the target scene from structured data;

[0207] The device also includes: a verification module 304;

[0208] The verification module 304 is used to verify the validity of each execution action parameter in the target execution action sequence;

[0209] The device also includes: output module 305;

[0210] The output module 305 is specifically used to output information about the target execution action sequence when the user intent is a query intent and the parameters of each execution action in the target execution action sequence are valid.

[0211] In one possible implementation, the output module 305 is specifically used to output information about the target execution action sequence when the user intent is an execution intent and the parameters of each execution action in the target execution action sequence are valid;

[0212] The device also includes: execution module 306;

[0213] The execution module 306 is used to control the electronic device to sequentially execute the actions in the target execution action sequence based on the user's confirmation action of the target execution action sequence.

[0214] In one possible implementation, the device further includes: an update module 307;

[0215] The update module 307 is used to update structured data, wherein the update includes: retaining important data in the structured data, and / or deleting unimportant data after a preset time period has elapsed; important data includes data in the structured data that has been used based on user requests, and unimportant data includes data in the structured data that has not been used due to user requests.

[0216] The request processing apparatus provided in this application embodiment can be used to execute the technical solution of the request processing method in any of the above embodiments of this application. Its implementation principle and technical effect are similar, and will not be described again here.

[0217] Figure 4 A schematic diagram of the structure of the electronic device provided in this application. Figure 4 As shown, the electronic device 400 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the device 400 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0218] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0219] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0220] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0221] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0222] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0223] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0224] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0225] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0226] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0227] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0228] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0229] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0230] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0231] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0232] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A request processing method, characterized in that, Applied to electronic devices, the method includes: Get the first request; Based on the first request, the user intent is identified, including query intent and / or execution intent; Based on the first request, the user intent, and the pre-acquired structured data, a target strategy is determined; The structured data includes the following correspondences: scene information, context information, and the historical execution action sequence of the electronic device; the historical execution action sequence is used to indicate the content and order of actions historically performed by the electronic device. The target strategy includes: outputting information about the target execution action sequence, and / or executing the target execution action sequence; the target execution action sequence is a historical execution action sequence in the structured data that is adapted to the first request.

2. The method according to claim 1, characterized in that, The scene information includes: first-level scene information and one or more second-level scene information corresponding to the first-level scene information. The first-level scene information is used to describe the user's behavior on the electronic device, and the second-level scene information is a subdivision of the first-level scene information.

3. The method according to claim 2, characterized in that, The structured data is obtained by processing the data in the electronic device based on the target model; the target model is used to classify the data in the electronic device into scenarios, extract key information, and generate action sequences; the key information includes the context information.

4. The method according to claim 3, characterized in that, The target model is trained by a dynamically weighted multi-task loss function, which includes a loss function that dynamically adjusts the weights of each task based on task relevance, dynamic weights, and confidence during model training.

5. The method according to claim 3 or 4, characterized in that, The data in the electronic device includes: vehicle bus data, sensor data, application layer data, user interaction data, context data, and user profile data.

6. The method according to any one of claims 1-4, characterized in that, The step of determining the target strategy based on the first request, the user intent, and pre-acquired structured data includes: Based on the text information included in the first request, the target scenario is determined; Obtain the target execution action sequence corresponding to the target scene from the structured data; Verify the validity of each execution action parameter in the target execution action sequence; If the user intent is the query intent and the parameters of each execution action in the target execution action sequence are valid, output the information of the target execution action sequence.

7. The method according to claim 6, characterized in that, The step of determining the target strategy based on the first request, the user intent, and pre-acquired structured data further includes: If the user intent is the execution intent and each execution action parameter in the target execution action sequence is valid, output the information of the target execution action sequence; Based on the user's confirmation of the target execution action sequence, the electronic device is controlled to sequentially execute the actions in the target execution action sequence.

8. The method according to any one of claims 1-4, characterized in that, The pre-acquired structured data is stored in the electronic device, and the method further includes: The structured data is updated, wherein the update includes: retaining important data in the structured data, and / or deleting unimportant data after a preset time period has elapsed; the important data includes data in the structured data that has been used based on user requests, and the unimportant data includes data in the structured data that has not been used due to user requests.

9. A request processing apparatus, characterized in that, Applied to electronic devices, the device includes: The acquisition module is used to acquire the first request; The identification module is configured to identify user intent based on the first request, wherein the user intent includes query intent and / or execution intent; The determination module is used to determine the target strategy based on the first request, the user intent, and pre-acquired structured data; The structured data includes the following correspondences: scene information, context information, and the historical execution action sequence of the electronic device; the historical execution action sequence is used to indicate the content and order of actions historically performed by the electronic device. The target strategy includes: outputting information about the target execution action sequence, and / or executing the target execution action sequence; the target execution action sequence is a historical execution action sequence in the structured data that is adapted to the first request.

10. An electronic device, characterized in that, include: Memory and processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-8.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.

Images