Data set generation method, model training method, server and computer medium

Abstract

The invention discloses a data set generation method, a voice interaction model training method, a server and a computer readable storage medium. The generation method comprises the steps of determining prompt words according to a preset data set; according to the prompt word and a pre-trained large language model, determining voice interaction content of the user and the vehicle; and determining a data set according to the voice interaction content. According to the voice interaction method provided by the embodiment of the invention, the cue word is determined by using the voice interaction, the cue word is input into the large language model, the voice interaction content determined according to the voice interaction is obtained, and the data set is determined according to the voice interaction content, so that the data set is generated based on the large language model, and the generation efficiency of the data set is improved.

Description

Technical Field

[0001] The present application relates to the field of voice interaction technology, and more specifically, to a method for generating a data set, a method for training a voice interaction model, a server, and a computer-readable storage medium. Background Art

[0002] Voice interaction is a basic application of most smart devices and has been widely integrated into daily life, covering multiple fields such as smart furniture and in-vehicle interaction. However, after the user issues a command, if the system fails to correctly understand the command or the execution is deviated, it will affect the user experience. In the related technology, there is a lack of methods to detect whether the user's command is executed correctly. Summary of the invention

[0003] Embodiments of the present application provide a method for generating a data set, a method for training a voice interaction model, a server, and a computer-readable storage medium.

[0004] The present application provides a method for generating a data set, the method comprising:

[0005] Determine the prompt word according to the preset data set;

[0006] Determining the content of the voice interaction between the user and the vehicle according to the prompt words and the pre-trained large language model;

[0007] The data set is determined according to the voice interaction content.

[0008] In this way, in the generation method, training method, server and computer-readable storage medium of the implementation mode of the present application, prompt words are determined by using voice interaction, and the prompt words are input into a large language model to obtain voice interaction content determined according to the voice interaction, and a data set is determined according to the voice interaction content, thereby realizing the generation of a data set based on a large language model to improve the efficiency of data set generation.

[0009] In some embodiments, determining the prompt word according to a preset data set includes:

[0010] The prompt word is determined according to the preset data set, preset settings, preset tasks and preset result requirements, wherein the preset settings include settings for users and dialogue scenarios, and the preset result requirements include output format requirements for the voice interaction content.

[0011] In this way, by determining the prompt words according to the preset data set, preset settings, preset tasks and preset result requirements, the prompt words in a standardized format can be obtained so that when the prompt words are input into the large language model, the model can understand and process them.

[0012] In some embodiments, determining the content of the voice interaction between the user and the vehicle based on the prompt word and the pre-trained large language model includes:

[0013] Based on the prompt words and the pre-trained large language model, multiple rounds of voice requests between the user and the vehicle and associated voice request feedback are determined to obtain the voice interaction content.

[0014] In this way, the large language model can output multiple rounds of voice requests and related voice request feedback from multiple rounds of users and vehicles based on the prompt words, which can achieve the rapid generation of multiple rounds of voice interaction content in the same interaction scenario, thereby improving the efficiency of data set generation. The use of a standardized prompt word input format can ensure the quality of the generated voice interaction content and facilitate subsequent data processing and analysis.

[0015] In some implementations, determining the data set according to the voice interaction content includes:

[0016] The data set is obtained according to the result of the annotation processing of the voice interaction content, and the annotation processing result includes the domain of the voice interaction content and the corresponding slot information.

[0017] In this way, by annotating the voice interaction content, a data set including the voice interaction content and its annotation information can be obtained, so that the data set can be used to train or test the voice interaction model.

[0018] In some implementations, obtaining the data set according to the result of the annotation processing of the voice interaction content includes:

[0019] Pre-annotating the voice interaction content to obtain an initial annotation processing result;

[0020] In response to the secondary annotation processing of the initial annotation processing result, the annotation processing result is obtained to obtain the data set.

[0021] In this way, by combining pre-annotation processing with secondary annotation processing, the annotation of voice interaction content can be achieved, which can improve the efficiency and accuracy of annotation, thereby improving the efficiency of data set generation.

[0022] In some embodiments, the generating method further comprises:

[0023] The preset data set is determined according to the user tracking log.

[0024] In this way, based on the first-round voice request and the associated voice request feedback in the user's embedding log, the voice interaction content in the generated data set can be made more in line with the user's real scenario, thereby improving the quality of the data set.

[0025] In some implementations, determining the preset data set according to the user tracking log includes:

[0026] The preset data set is determined according to the first-round voice request and the associated voice request feedback in the user embedding log.

[0027] In this way, the reference model can be trained using the data set to obtain a voice interaction model and realize the voice interaction function.

[0028] The present application provides a method for training a speech interaction model, the training method comprising:

[0029] The reference model is trained based on a data set as in any of the above implementations to obtain the voice interaction model.

[0030] An embodiment of the present application provides a server, including a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, a method as described in any of the above embodiments is implemented.

[0031] An embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed by one or more processors, the method of any of the above embodiments is implemented.

[0032] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0034] Figure 1 It is one of the flow charts of the generation method of certain embodiments of the present application;

[0035] Figure 2 This is the second flow chart of the generation method of certain embodiments of the present application;

[0036] Figure 3 This is the third flow chart of the generation method of certain embodiments of the present application;

[0037] Figure 4 This is a fourth flow chart of a generation method of certain embodiments of the present application;

[0038] Figure 5 This is the fifth flow chart of the generation method of certain embodiments of the present application;

[0039] Figure 6 This is the sixth flow chart of the generation method of certain embodiments of the present application;

[0040] Figure 7 This is the seventh flow chart of the generation method of certain embodiments of the present application;

[0041] Figure 8 This is the eighth flow chart of the generation method of certain embodiments of the present application. DETAILED DESCRIPTION

[0042] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0043] Voice interaction is a basic application of most smart devices and has been widely integrated into daily life, covering multiple fields such as smart furniture and in-vehicle interaction. In order for the voice interaction model to better understand the user's real semantics, it is necessary to use a preset data set to train the model and evaluate the training effect. The higher the quality of the data set, the better the training effect of the model and the more accurate the evaluation of the model performance. In the related technology, the data set is obtained by manual setting and writing, which has low generation efficiency and small coverage of the data set, resulting in low quality of the compiled data set.

[0044] Based on the above unresolved issues, please refer to Figure 1 The present application embodiment provides a method for generating a data set, the generating method comprising:

[0045] 011: Determine the prompt word according to the preset data set;

[0046] 012: Determine the content of the voice interaction between the user and the vehicle based on the prompt words and the pre-trained large language model;

[0047] 013: Determine the data set based on the voice interaction content.

[0048] The embodiment of the present application provides a server, including a memory and a processor, wherein a computer program is stored in the memory and can be executed by the processor. The processor can be used to determine a prompt word according to a preset data set; and to determine the content of the voice interaction between the user and the vehicle according to the prompt word and a pre-trained large language model; and can also be used to determine a data set according to the voice interaction content.

[0049] The embodiment of the present application provides a generating device, which includes a second determining module and a third determining module. The first determining module can be used to determine a prompt word according to a preset data set; the second determining module can be used to determine the content of the voice interaction between the user and the vehicle according to the prompt word and the pre-trained large language model; the third determining module can be used to determine the data set according to the voice interaction content.

[0050] Specifically, the preset data set includes multiple preset voice interaction contents. The preset voice interaction contents include the voice request sent by the user to the vehicle voice interaction module, and the voice request feedback determined by the vehicle for the voice request. Among them, the voice request and the voice request feedback are the interaction contents that the user actually performed previously. Generating voice dialogue contents based on the actual interaction contents to obtain a data set can make the contents of the data set more in line with actual real-life scenarios, improve the coverage of the data set under different user needs, and avoid the subjectivity of manually setting the data set.

[0051] The prompt has a fixed format, including a setting module, an input module, and a task module, and different modules are separated by separators. Among them, the setting module is used to determine the identities and dialogue scenarios of the two parties in the voice interaction content. For example, the two parties in the dialogue are the user and the voice assistant. The setting module sets the user's personal information and the capabilities and personality of the voice assistant. The input module is used to input the preset voice interaction content in the preset data set so that the large language model can determine the current conversation background and topic. The task module is used to determine the content of the task, for example, the number of rounds to generate multiple rounds of voice interaction content, the number of words in the voice request issued by the user in each round, the format of the output voice interaction content, etc.

[0052] In some embodiments, a delimiter corresponding to each module is set at the beginning and end of each module to facilitate the large speech model to recognize each module. For example, the delimiter of the module is set to <instruction>, the delimiter for the input module is<current_dialogue> , the delimiter of the task module is <task>.

[0053] For another example, the input module can be "<current_dialogue> [User]:{query}\n[Voice Assistant]:{answer}< / current_dialogue> ”

[0054] in,<current_dialogue> is the separator of the input module, {query} is the voice request made by the user in the preset data set, and {answer} is the voice request feedback associated with the voice request in the preset data set.

[0055] By inputting Prompt into the large language model, the output voice interaction content can be obtained. The large language model can be a trained generative large language model, such as ChatGPT, BERT, T5, etc. The output voice interaction content is determined based on the preset voice interaction content in the preset data set and is in the same context as the preset voice interaction content.

[0056] The voice interaction content output by the large language model is screened and sorted to process erroneous, abnormal or repeated voice interaction content to obtain a data set.

[0057] Furthermore, the data set may be a training set, a validation set or a test set. The training set and the validation set may be used to train the voice interaction model. The test set may be used to test the trained voice interaction model to evaluate the processing capability of the trained voice interaction model for unknown voice data.

[0058] In this way, by using voice interaction to determine the prompt words, and inputting the prompt words into the large language model, the voice interaction content determined according to the voice interaction is obtained, and the data set is determined according to the voice interaction content, thereby realizing the generation of the data set based on the large language model to improve the efficiency of data set generation. And by using the powerful language generation ability of the large language model, it is possible to quickly generate a variety of dialogue scenarios and improve the quality of the data set.

[0059] See also Figure 2 In some embodiments, step 011 (determining a prompt word according to a preset data set) includes:

[0060] 0111: Determine the prompt words based on the preset data set, preset settings, preset tasks and preset result requirements. The preset settings include the settings for users and dialogue scenarios. The preset result requirements include the output format requirements for voice interaction content.

[0061] In some embodiments, the processor may be configured to determine a prompt word based on a preset data set, preset settings, preset tasks, and preset result requirements.

[0062] In some embodiments, the first determination module includes a first determination submodule. The first determination submodule can be used to determine the prompt word according to a preset data set, preset settings, preset tasks, and preset result requirements.

[0063] Specifically, the preset data set, preset settings, preset tasks and preset result requirements are respectively filled into the modules corresponding to the Prompt to complete the setting of the Prompt. Among them, the preset settings can be filled into the setting module, the preset voice interaction content in the preset data set is filled into the input module, and the preset tasks and preset result requirements are filled into the task module.

[0064] The preset setting is the setting of the voice interaction scene and the identities of the two parties in the conversation. The preset setting may include user settings and voice assistant settings, and the corresponding separators are used before and after the settings to distinguish them. In one embodiment, the setting module is: <instruction> Suppose you are a [voice assistant] chatting with a [user].<user_info> [The user] may be the driver or the passenger.< / user_info><bot_info> [Voice Assistant] is an in-vehicle voice assistant equipped with large model capabilities. It can accurately identify user needs, make voice replies, support vehicle function operations, and also support chat, encyclopedia, etc. It has a gentle personality, humor, and extensive knowledge.< / bot_info>< / instruction> ”

[0065] The default setting is to assume that you are a voice assistant chatting with the user. The user may be the driver or the passenger. The voice assistant is an in-vehicle voice assistant with large model capabilities. It can accurately identify user needs, make voice replies, support vehicle function operations, and also support chat, encyclopedia and other Q&A. Gentle, humorous, and knowledgeable.< / task> < / instruction> To set the module separator,< / user_info> The delimiter set by the user.< / bot_info> The separator character set for the voice assistant.

[0066] The preset tasks are the tasks that the large language model is expected to perform, such as limiting the number of rounds of speech interaction content generation, word count, context reference specifications, etc. The preset result requirements are the format requirements for the speech interaction content output by the large language model, so as to facilitate the subsequent data cleaning of the speech interaction content. For example, the speech interaction content is required to be data in Json, dict, or xml formats.

[0067] In another embodiment, the task module is: <task> Next, please continue to complete this conversation based on the above content: - The number of conversation rounds should be controlled within 4 to 6 rounds. - Each round of [user] content should be controlled within 20 words. - [User] content can remove the subject and add context reference. Please return the conversation content in JSON format.< / task> ”

[0068] The preset task is "Next, please continue to complete this conversation based on the above content: the number of conversation rounds is controlled within 4 to 6. The user content in each round is controlled within 20 words. The user content can remove the subject and add context reference." The preset result requirement is "Please return the conversation content in JSON format." is the delimiter of the task module.

[0069] In this way, by determining the prompt words according to the preset data set, preset settings, preset tasks and preset result requirements, the prompt words in a standardized format can be obtained so that when the prompt words are input into the large language model, the model can understand and process them.

[0070] See also Figure 3 In some embodiments, step 012 (determining the content of the voice interaction between the user and the vehicle based on the prompt word and the pre-trained large language model) includes:

[0071] 0121: Based on the prompt words and the pre-trained large language model, determine multiple rounds of voice requests between the user and the vehicle and the associated voice request feedback to obtain the voice interaction content.

[0072] In certain embodiments, the processor may be used to determine multiple rounds of voice requests between the user and the vehicle and associated voice request feedback based on the prompt words and a pre-trained large language model to obtain voice interaction content.

[0073] In some embodiments, the second determination module includes a second determination submodule. The second determination submodule can be used to determine multiple rounds of voice requests between the user and the vehicle and associated voice request feedback based on the prompt word and the pre-trained large language model to obtain voice interaction content.

[0074] Specifically, a pair of voice requests and their associated feedback in a preset data set can be input into Prompt each time, and the preset task requires the output of multiple rounds of voice interaction content, so that after Prompt is input into the large language model, the large language model outputs multiple rounds of voice requests and associated voice request feedback in the same scenario as the input voice request to obtain voice interaction content.

[0075] In one embodiment, the preset voice request content of the preset data set is "turn on the width light", and the associated voice request feedback is "the width light is turned on", that is, the input module of Prompt is:<current_dialogue> [User]: "Turn on the width lights"\n[Voice Assistant]: "The width lights are on"< / current_dialogue> The output of the large language model is: "dialogue":[{"user":"Can you adjust the temperature in the car?"},{"voice assistant":"Of course, what temperature do you want to adjust it to?"},{"user":"Set it to 22 degrees."},{"voice assistant":"The temperature is set to 22 degrees, wait a moment."},{"user":"How long will it take to get to the destination?"},{"voice assistant":"It will take about 15 minutes to arrive."}]. Understandably, the voice interaction content output by the large language model and the input voice request both belong to the scenario when the user initially gets on the car.

[0076] In another embodiment, the preset voice request content of the preset data set is "Why does one plus one equal two?", and the associated voice request feedback is "One plus one equals two. This is the basic arithmetic rule in mathematics and the axiom in the mathematical system. In human daily life, this is also a common cognition and experience. In mathematics, one plus one equals two is defined as an axiom and is the foundation of the mathematical system. This axiom is set to make the mathematical system self-consistent and orderly. In human daily life, we can also observe the phenomenon of one plus one equals two. For example, two apples put together are still two apples. The power of one person plus the power of another person can accomplish more things."

[0077] The output of the large language model is: "dialogue":[{"user":"What is zero plus zero?"},{"voice assistant":"Zero plus zero equals zero, this is the basic principle of arithmetic."},{"user":"Why can't it equal one?"},{"voice assistant":"Because any number plus zero equals the original number, zero does not increase."},{"user":"Math is really interesting!"},{"voice assistant":"Yes, math is like a window to explore the world!"}]. Understandably, the voice interaction content output by the large language model and the input voice request both belong to the scenario of discussing mathematics.

[0078] In this way, the large language model can output multiple rounds of voice requests and related voice request feedback from multiple rounds of users and vehicles based on the prompt words, which can achieve the rapid generation of multiple rounds of voice interaction content in the same interaction scenario, thereby improving the efficiency of data set generation. The use of a standardized prompt word input format can ensure the quality of the generated voice interaction content and facilitate subsequent data processing and analysis.

[0079] See also Figure 4In some implementations, step 013 (determining a data set based on the voice interaction content) includes:

[0080] 0131: According to the labeling processing result of the voice interaction content, a data set is obtained, and the labeling processing result includes the domain of the voice interaction content and the corresponding slot information.

[0081] In some embodiments, the processor may be used to obtain a data set based on a result of annotating the voice interaction content, where the result of the annotation includes the domain of the voice interaction content and corresponding slot information.

[0082] In some embodiments, the third determination module includes a third determination submodule. The third determination submodule can be used to obtain a data set according to a result of annotating the voice interaction content, where the result of the annotation processing includes the domain of the voice interaction content and corresponding slot information.

[0083] Specifically, the generated multiple rounds of voice interaction content are annotated to mark their domains and slot information, thereby completing the generation of the data set. For example, the domain can be navigation, vehicle control, chatting, encyclopedia, weather, etc.

[0084] In addition, the annotated content also includes the application programming interface (API) that meets the voice request. For example, the user's voice request is "I want to go to Guangzhou", the annotated API is "NavigationSearchPoi", and the annotated slot information is "{"poi_name": "Guangzhou"}".

[0085] In one embodiment, the preset voice interaction content of the preset data set and the voice interaction content output by the large language model are shown in Table 1. Among them, round 1 is the preset voice interaction content, and rounds 2 and 3 are the voice interaction content output by the large language model.

[0086]

[0087] Table 1

[0088] In this way, by annotating the voice interaction content, a data set including the voice interaction content and its annotation information can be obtained, so that the data set can be used to train or test the voice interaction model.

[0089] See also Figure 5 In some implementations, step 0131 (obtaining a data set based on the result of the annotation processing of the voice interaction content) includes:

[0090] 01311: Pre-annotate the voice interaction content to obtain the initial annotation processing result;

[0091] 01312: In response to the secondary annotation processing of the initial annotation processing result, an annotation processing result is obtained to obtain a data set.

[0092] In some embodiments, the processor can be used to pre-annotate the voice interaction content to obtain an initial annotation processing result; it can also be used to respond to a secondary annotation processing of the initial annotation processing result to obtain an annotation processing result to obtain a data set.

[0093] In some embodiments, the third determination submodule includes a labeling unit and a response unit. The labeling unit can be used to perform pre-labeling processing on the voice interaction content to obtain an initial labeling processing result; the response unit can be used to respond to the secondary labeling processing of the initial labeling processing result to obtain the labeling processing result to obtain the data set.

[0094] Specifically, the pre-annotation process can be performed by a natural language understanding (NLU) model. The voice interaction content output by the large language model is input into the NLU model to obtain the pre-annotated label (apiner label) output by the NLU model to obtain the initial annotation processing result of the voice interaction content.

[0095] After the initial labeling results are obtained, the labels are manually screened and corrected for secondary labeling. When performing manual labeling, the voice requests issued by users in the previous round and the associated voice request feedback, as well as the voice requests in the current round, are referred to to label the fields and slot information that should be recalled for the voice requests in the current round, thereby achieving the labeling of the voice interaction content, obtaining the labeling results (NLU labels), and forming a data set.

[0096] In this way, by combining pre-annotation processing with secondary annotation processing, the annotation of voice interaction content can be achieved, which can improve the efficiency and accuracy of annotation, thereby improving the efficiency of data set generation.

[0097] See also Figure 6 In some embodiments, the generating method further comprises:

[0098] 014: Determine the preset data set based on the user tracking log.

[0099] In some embodiments, the processor may be used to determine a preset data set based on the user's tracking log.

[0100] In some embodiments, the generating device further includes a fourth determining module. The fourth determining module can be used to determine a preset data set according to the user embedding log.

[0101] Specifically, the preset voice request content in the preset data set can be obtained from the user embedding log. The user embedding log is stored in the server, and the user embedding log includes the voice request actually issued by the online user and the TTS (Text To Speech) response made by the vehicle computer to the voice request.

[0102] By extracting the real voice interaction content from the user's embedded logs and obtaining a preset data set, the large language model can generate multiple rounds of voice interaction content in the same scenario based on the user's real voice interaction content, thereby making the scenario of the voice interaction content in the data set more realistic, avoiding the subjectivity of manually written voice interaction content, and improving the coverage of the data set.

[0103] In this way, by determining the preset data set from the user embedding log and generating conversation scenarios that meet the actual user needs, the coverage of the generated data set can be improved, and manual intervention can be reduced, thereby avoiding the subjectivity of manually setting the preset data set.

[0104] See also Figure 7 In some implementations, step 014 (determining a preset data set based on the user's embedding log) includes:

[0105] 0141: Determine the preset data set based on the first-round voice request and the associated voice request feedback in the user tracking log.

[0106] In some embodiments, the processor may be used to determine a preset data set based on a first-round voice request and associated voice request feedback in a user embedding log.

[0107] Specifically, the first-round voice request refers to the first-round voice request issued in a voice interaction scenario in the user embedding log. For example, in the user embedding log, the user voice request and the associated voice request feedback in the dialogStatus=start state are the first-round voice request and the associated voice request feedback.

[0108] Among them, the associated voice request feedback refers to the feedback made by the server to the user based on the first-round voice request. In addition, the preset data set also includes the fields corresponding to each preset voice interaction content. For example, the first-round voice request is "Exit navigation", the associated voice request feedback is "Help you turn off the navigation", and the corresponding field is "NavigationExit"; or, the first-round voice request is "Play Faye Wong's ambiguity", and the associated voice request feedback is "Okay, let's listen to the song together", and the corresponding field is "MusicSearchPlay"; or, the first-round voice request is "Are you a man or a woman", and the associated voice request feedback is "Oh, as an artificial intelligence, I have no gender, but you can treat me as a boy or a girl, it's okay", and the corresponding field is "ChatGeneral".

[0109] By extracting the first-round voice request and the related voice request feedback from the user's embedded log, it can be ensured that the voice interaction content in the preset data set is the first-round voice interaction content in a certain scenario. The voice interaction content generated by the large language model based on the preset data set is in the same scenario as the corresponding first-round voice request, which is convenient for simulating the user's real voice interaction scenario and improving the authenticity, reliability and coverage of the generated data set.

[0110] In this way, based on the first-round voice request and the associated voice request feedback in the user's embedding log, the voice interaction content in the generated data set can be made more in line with the user's real scenario, thereby improving the quality of the data set.

[0111] In some embodiments, the fourth determination module includes a fourth determination submodule. The fourth determination submodule can be used to determine a preset data set according to the first-round voice request and the associated voice request feedback in the user embedding log.

[0112] See also Figure 8 , the embodiment of the present application provides a training method for a voice interaction model, the training method comprising:

[0113] 021: Train the reference model based on a data set as in any of the above implementations to obtain a voice interaction model.

[0114] Specifically, the data set can be used as a training set and a validation set to train the reference model to obtain a voice interaction model. Since the data set can be generated based on a preset data set obtained according to the user's embedded point log, the user voice interaction content therein is more in line with the user's real scene, so that the voice interaction model trained with the data set generated by the generation method according to the embodiment of the present application can better process the voice requests issued by the user.

[0115] Through the trained voice interaction model, users can make voice requests to the car computer. After analysis and processing by the voice interaction model, they will receive corresponding voice request feedback to execute the functions that the user expects to achieve.

[0116] In addition, the data set can also be used as a test set to test the trained voice interaction model to evaluate the training effect of the voice interaction model. Using the data set generated by the generation method of the embodiment of the present application to test the voice interaction model can more accurately evaluate the performance of the voice interaction model.

[0117] In this way, the reference model can be trained using the data set to obtain a voice interaction model and realize the voice interaction function.

[0118] In some embodiments, the processor may be used to train a reference model based on a data set such as any of the above embodiments to obtain a voice interaction model.

[0119] The embodiment of the present application provides a training device, which includes a training module. The training module can be used to train a reference model based on a data set as in any of the above embodiments to obtain a voice interaction model.

[0120] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the method of any of the above embodiments is implemented.

[0121] It is understood that a computer program includes computer program code. The computer program code may be in source code form, object code form, executable file or some intermediate form. Computer readable storage media may include: any entity or device capable of carrying computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution medium.

[0122] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0123] In addition, the term "connection" should be understood in a broad sense, for example, it can include fixed connection, detachable connection, or integral connection; it can include direct connection, indirect connection through an intermediate medium, and internal communication between two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0124] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0125] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0126] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in the field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for generating a data set, characterized in that: The generation method comprises: Determine the prompt word according to the preset data set; Determining the content of the voice interaction between the user and the vehicle according to the prompt words and the pre-trained large language model; The data set is determined according to the voice interaction content.

2. The generation method according to claim 1, characterized in that: The step of determining the prompt word according to the preset data set includes: The prompt word is determined according to the preset data set, preset settings, preset tasks and preset result requirements, wherein the preset settings include settings for users and dialogue scenarios, and the preset result requirements include output format requirements for the voice interaction content.

3. The generation method according to claim 1, characterized in that: The determining of the voice interaction content between the user and the vehicle according to the prompt word and the pre-trained large language model includes: Based on the prompt words and the pre-trained large language model, multiple rounds of voice requests between the user and the vehicle and associated voice request feedback are determined to obtain the voice interaction content.

4. The generation method according to claim 1, characterized in that: The determining the data set according to the voice interaction content includes: The data set is obtained according to the result of the annotation processing of the voice interaction content, and the annotation processing result includes the domain of the voice interaction content and the corresponding slot information.

5. The generation method according to claim 1, characterized in that: The step of obtaining the data set according to the result of the annotation processing of the voice interaction content includes: Pre-annotating the voice interaction content to obtain an initial annotation processing result; In response to the secondary annotation processing of the initial annotation processing result, the annotation processing result is obtained to obtain the data set.

6. The generation method according to claim 1, characterized in that: The generation method further includes: The preset data set is determined according to the user tracking log.

7. The generation method according to claim 5, characterized in that: The step of determining the preset data set according to the user embedding log includes: The preset data set is determined according to the first-round voice request and the associated voice request feedback in the user embedding log.

8. A method for training a speech interaction model, characterized in that: The training method comprises: The reference model is trained based on the data set described in any one of claims 1 to 7 to obtain the voice interaction model.

9. A server, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the method described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by one or more processors, implements the method described in any one of claims 1 to 8.

Images