Voice interaction methods, servers, and readable storage media

Abstract

This application discloses a voice interaction method, which includes: receiving a current voice request forwarded by a vehicle; determining target input information based on the current voice request and target prompt information; determining, based on a large language model, the slot recognition result and application programming interface (API) prediction result corresponding to the current voice request based on the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information; and outputting an execution result to the vehicle to complete the voice interaction based on the slot recognition result and the API prediction result. Thus, the server in this application enables the large language model to complete tasks such as slot recognition and API prediction, thereby directly processing the voice request to complete the voice interaction. This reduces the need to call related models from different vertical domains, lowers latency, improves the voice request response speed, and ensures the reliability of both the slot recognition result and the API prediction result.

Description

Technical Field

[0001] This application relates to the field of voice interaction technology, and in particular to a voice interaction method, server, and readable storage medium. Background Technology

[0002] In related technologies, in-vehicle voice dialogue systems typically utilize a natural language generation module to parse user statements into machine-understandable semantic tags. A dialogue state tracking module maintains an internal dialogue state as a compact representation of the entire dialogue history. Based on this state, a dialogue strategy module selects appropriate dialogue actions, and finally, the natural language generation module converts these actions into natural language responses. However, in real-world interaction scenarios, due to insufficient prior training, the recognition results in these technologies may be erroneous, failing to extract the expected slot results and thus failing to meet users' voice interaction needs in in-vehicle environments. Summary of the Invention

[0003] This application provides a voice interaction method, a server, and a readable storage medium.

[0004] This application provides a voice interaction method, including:

[0005] Receive the current voice request forwarded by the vehicle;

[0006] Based on the current voice request and target prompt information, target input information is determined, wherein the target input information includes multiple semantic units;

[0007] Based on the large language model, according to the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, the slot recognition result and application interface prediction result corresponding to the current voice request are determined. The large language model is pre-trained and can identify the slot corresponding to the voice request and predict the application interface corresponding to the voice request.

[0008] Based on the slot identification result and the application interface prediction result, the execution result is output and sent to the vehicle to complete the voice interaction.

[0009] In the voice interaction method provided in this application, the server can receive the current voice request forwarded by the vehicle, determine the target input information based on the current voice request and target prompt information, and, based on a pre-trained large language model that can identify the slot corresponding to the voice request and predict the application programming interface (API) corresponding to the voice request, determine the slot recognition result and the API prediction result corresponding to the current voice request based on the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, and output the execution result and send it to the vehicle to complete the voice interaction based on the slot recognition result and the API prediction result corresponding to the current voice request.

[0010] Thus, in this embodiment, the server can complete tasks such as slot identification and application programming interface (API) prediction through a large language model, thereby directly processing voice requests and completing voice interaction. This embodiment, based on a pre-trained large language model, reduces the need to call related models from different domains, effectively reducing latency and improving the response speed to user voice requests. This embodiment enables the large language model to output the slot identification result and API prediction result corresponding to the current voice request based on the target prompt information in the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, ensuring the reliability of both the slot identification result and the API prediction result.

[0011] In some embodiments of this application, the method further includes:

[0012] The target prompt information is determined based on multiple pre-configured application interfaces and the corresponding description information for each application interface.

[0013] Thus, the server in this embodiment enables the large language model to predict the application interface of the current voice request based on the description information of the application interface in the target input information. Therefore, the reliability of the application interface prediction result is guaranteed, and it can avoid to some extent the situation where the large language model creates a new application interface as the prediction result when facing an unknown or rare current voice request.

[0014] In some embodiments of this application, determining the target prompt information based on a plurality of pre-configured application programming interfaces (APIs) and descriptive information corresponding to each API includes:

[0015] Select the target application interface from the plurality of application interfaces;

[0016] The target prompt information is determined based on the interface description information corresponding to the target application interface.

[0017] Thus, the server in this embodiment can determine the target application interface among all application interfaces, and determine the target prompt information and target input information according to the description information corresponding to the target application interface. This can avoid the situation where the target prompt information and target input information are too long to a certain extent, and the slot identification and application interface prediction performed by the large language model based on the target prompt information and target input information can be carried out stably.

[0018] In some embodiments of this application, determining the slot recognition result and application programming interface prediction result corresponding to the current voice request based on the large language model, according to the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, includes:

[0019] Based on the large language model, the encoding result corresponding to the target input information is determined according to the relative distance information and the semantic unit;

[0020] The slot identification result and the application interface prediction result are determined based on the encoding result.

[0021] Thus, the server in this application embodiment can encode the target prompt information based on a large language model, using the relative distance information between the first semantic unit and the second semantic unit in the target prompt information to obtain the encoding result. Then, when slot identification and application interface prediction are performed using the relative distance information indicated by the encoding result, the accuracy of slot identification and the accuracy of application interface prediction can be guaranteed to a certain extent.

[0022] In some embodiments of this application, determining the encoding result corresponding to the target input information based on the large language model, the relative distance information, and the semantic unit includes:

[0023] Based on the large language model, the attention score information corresponding to the target input information is determined according to the relative distance information and the query vector and key vector of each semantic unit.

[0024] The encoding result is determined based on the attention score information.

[0025] Thus, the embodiments of this application enable the large language model to focus on the relative distance between the first semantic unit and the second semantic unit in the target input information based on the encoding result determined by the attention score information, thereby ensuring the reliability of slot recognition and application interface prediction to a certain extent.

[0026] In some embodiments of this application, the training process of the large language model includes:

[0027] The reference model is trained using a predetermined input information sample and the relative distance information of the third semantic unit to the fourth semantic unit in the input information sample to obtain the large language model, wherein the input information sample is determined by the target prompt information and the voice request sample.

[0028] Thus, the server in this embodiment can train a reference model to obtain a large language model through input information samples, so that the trained large language model can be well applied to the processing of input information containing voice requests and prompts, and the robustness of the large language model is guaranteed.

[0029] In some embodiments of this application, determining the slot recognition result and application programming interface prediction result corresponding to the current voice request based on the large language model, according to the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, includes:

[0030] Based on the large language model, the slot identification result, the application programming interface (API) prediction result, and the target parameters corresponding to the API prediction result are determined according to the target input information and the relative distance information.

[0031] Thus, in the embodiments of this application, the large language model can complete slot identification, application interface prediction, and determine the target parameters for filling the slot identification results into the application interface based on the target input information.

[0032] In some embodiments of this application, the step of outputting the execution result and sending it to the vehicle to complete the voice interaction based on the slot identification result and the application programming interface prediction result includes:

[0033] The slot identification result is filled into the target parameter corresponding to the application interface prediction result to obtain the execution result of the application interface parameter filling;

[0034] The execution result is output and sent to the vehicle to complete the voice interaction.

[0035] Thus, in this embodiment, the slot identification results can be filled into the corresponding target parameters of the application interface based on the slot identification results, the application interface prediction results, and the target operation.

[0036] This application provides a server including a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the above-described voice interaction method.

[0037] This application provides a computer-readable storage medium storing a computer program that, when executed by one or more processors, implements the above-described voice interaction method.

[0038] The server and computer-readable storage medium provided in this application embodiment can perform tasks such as slot recognition and application programming interface (API) prediction through a large language model, thereby directly processing voice requests and completing voice interaction. Based on a pre-trained large language model, this application embodiment reduces the need to call related models in different vertical domains, effectively reducing latency and improving the response speed to user voice requests. This application embodiment enables the large language model to output the slot recognition result and API prediction result corresponding to the current voice request based on the target prompt information in the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, ensuring the reliability of both the slot recognition result and the API prediction result.

[0039] Additional aspects and advantages of embodiments of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of this application. Attached Figure Description

[0040] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein:

[0041] Figure 1 This is a schematic diagram of the architecture of a traditional dialogue system in related technologies;

[0042] Figure 2 This is a schematic diagram of an end-to-end dialogue system architecture.

[0043] Figure 3 This is a flowchart illustrating the voice interaction method in some embodiments of this application;

[0044] Figure 4 This is a schematic diagram of the architecture of an end-to-end dialogue system based on a large language model in some embodiments of this application.

[0045] Figure 5 This is a flowchart illustrating the voice interaction method in some embodiments of this application;

[0046] Figure 6 This is a flowchart illustrating the voice interaction method in some embodiments of this application;

[0047] Figure 7 This is a flowchart illustrating the voice interaction method in some embodiments of this application;

[0048] Figure 8 This is a flowchart illustrating the voice interaction method in some embodiments of this application;

[0049] Figure 9 This is a flowchart illustrating the voice interaction method in some embodiments of this application. Detailed Implementation

[0050] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the embodiments of this application, and should not be construed as limiting the embodiments of this application.

[0051] Please see Figure 1 Traditional in-vehicle voice system architecture is based on a modular strategy, which divides the work among components to realize the entire dialogue process. It typically includes a natural language understanding module, a dialogue state tracking module, a dialogue strategy module, and a natural language generation module. Among them, the dialogue state tracking module and the dialogue strategy module are collectively referred to as the dialogue management module.

[0052] From the perspective of the overall dialogue processing chain, traditional in-vehicle voice system architectures, after acquiring the user's voice input request, first need to use a natural language understanding module to parse the user's voice request into machine-understandable semantic tags, i.e., perform domain classification, intent recognition, and slot recognition. Then, the dialogue state tracking module maintains the dialogue state, and based on this state, the dialogue strategy module selects the appropriate dialogue action. That is, in the dialogue management module, the dialogue state and dialogue strategy are combined to select and execute the application programming interface (API) that meets the user's input requirements. Finally, the natural language generation module converts the dialogue action into a natural language response, which is the system output returned to the user.

[0053] In traditional in-vehicle voice system architectures, each module integrates models and processing logic from different domains. Each module requires training its own model and implementing its own processing logic, necessitating significant development effort and sufficient domain-specific labeled data. Furthermore, interaction with real users is costly, and data feedback cycles are slow, insufficient to support rapid model iteration, thus limiting the scalability of the dialogue system. Additionally, maintaining the rule system requires substantial manpower and time, further restricting its extensibility. Moreover, due to the high latency requirements of voice interaction in in-vehicle scenarios, the timeliness of the dialogue system executing user commands through four modules and multiple models presents a significant challenge.

[0054] Please see Figure 2To address the shortcomings and deficiencies of traditional in-vehicle voice system architectures, a new end-to-end voice system architecture has been proposed, comprising three core algorithm modules: Named Entity Recognition (NER), Action Prediction (AP), and Argument Filling (AF). The NER module identifies entities in the user's voice request. The AP module predicts the application programming interface (API) corresponding to the voice request, which aims to achieve the user's current goal. The AF module selects the parameters from the identified entities in the voice request to execute the API obtained in the previous step.

[0055] In other words, the NER module is used to obtain the entities that need to be called in the application programming interface (API), the AP module determines whether the API called by the user voice input is correct, and the AF module selects which entities to use as parameters for the API.

[0056] Understandably, the end-to-end in-vehicle voice system architecture integrates some modules of the traditional in-vehicle voice system architecture, reducing the relatively independent data requirements of each module and the need for manual rule formulation. Due to its streamlined architecture and lack of domain-specific distinctions, it reduces subsequent module maintenance costs and saves on machine and manpower expenses. Meanwhile, vehicle control requirements in in-vehicle scenarios are highly time-sensitive, and the computing power of the vehicle's hardware differs from the computing power used by developers to train offline models, resulting in high latency requirements for voice interaction in in-vehicle environments. The end-to-end architecture, by streamlining intermediate modules, can effectively reduce the latency of the in-vehicle system.

[0057] From the perspective of the overall dialogue processing chain, the end-to-end in-vehicle voice system architecture, after obtaining the user's voice request, will directly generate system instructions or responses by combining the dialogue history. That is, the NER module obtains slot information, the AP module obtains the API to be executed, and the AF module selects the parameters corresponding to the API execution from the identified slots.

[0058] However, in actual voice interaction scenarios, since the corpus in the early training process cannot cover all possible dialogues in the actual scenario, slot recognition errors and / or API prediction errors may occur. It is even possible to predict non-existent slot information or non-existent APIs, making it difficult to reliably meet the voice interaction needs in the vehicle scenario.

[0059] Based on the issues mentioned above, please refer to Figure 3 This application provides a voice interaction method, including:

[0060] 01: Receive the current voice request forwarded by the vehicle;

[0061] 02: Based on the current voice request and target prompt information, determine the target input information, which includes multiple semantic units;

[0062] 03: Based on the large language model, the slot recognition result and application interface prediction result corresponding to the current voice request are determined according to the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information. The large language model is pre-trained and can recognize the slot corresponding to the voice request and predict the application interface corresponding to the voice request.

[0063] 04: Based on the slot recognition results and the application programming interface prediction results, output the execution results and send them to the vehicle to complete the voice interaction.

[0064] This application also provides a server, which includes a memory and a processor. The voice interaction method of this application can be implemented by the server of this application. Specifically, the memory stores a computer program, the processor is used to receive the current voice request forwarded by the vehicle, and to determine the target input information based on the current voice request and target prompt information. The target input information includes multiple semantic units. The processor is used to determine the slot recognition result and application interface prediction result corresponding to the current voice request based on a large language model, according to the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information. The large language model is pre-trained and can recognize the slot corresponding to the voice request and predict the application interface corresponding to the voice request. The processor is used to output the execution result and send it to the vehicle to complete the voice interaction based on the slot recognition result and the application interface prediction result.

[0065] This application provides a voice interaction device. The voice interaction method of this application can be implemented by the voice interaction device of this application. Specifically, the voice interaction device includes a receiving module, an information determination module, a result determination module, and an interaction module. The receiving module receives a current voice request forwarded by a vehicle. The information determination module determines target input information, including multiple semantic units, based on the current voice request and target prompt information. The result determination module determines the slot recognition result and application programming interface (API) prediction result corresponding to the current voice request based on a large language model, according to the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information. The large language model is pre-trained and can recognize the slot corresponding to the voice request and predict the corresponding API. The interaction module outputs an execution result and sends it to the vehicle to complete the voice interaction based on the slot recognition result and the API prediction result.

[0066] Specifically, with the rapid development of Large Language Models (LLMs), related technologies have made significant progress and have been applied in various fields. Advances in deep learning and big data technologies have increased the scale and complexity of LLMs, thereby improving their performance and expressive capabilities. LLMs can possess hundreds of billions or even trillions of parameters, enabling them to better understand and generate human language, covering a wider range of semantics and contexts. Consequently, LLMs have achieved performance comparable to human capabilities in various natural language processing tasks such as natural language understanding, natural language generation, contextual understanding, text summarization, text translation, and dialogue systems.

[0067] Therefore, this application provides an in-vehicle speech recognition solution based on a large language model, which can be found in the following details. Figure 4 , Figure 4 This diagram illustrates the architecture of an end-to-end dialogue system based on a large language model (LLM) in certain embodiments of this application. Specifically, this embodiment leverages the superior capabilities of the LLM model in natural language processing tasks, and trains it accordingly based on the task at hand. This embodiment eliminates three modules (NER module, AP module, and AF module) in the end-to-end speech system architecture, meaning independent modules are no longer required. The language understanding, generation, and contextual understanding capabilities of the LLM model are directly utilized to generate the API and AF tags corresponding to the current speech request. This reduces the burden on developers and saves time and costs.

[0068] Furthermore, due to the fact that in such cases Figure 4In the dialogue system shown, the system no longer needs to distinguish between vertical domains, making the large language model responsible for predicting APIs for all vertical domains. Compared to a single natural language processing model in a dialogue system that divides vertical domains, the large language model needs to process and recognize a significantly larger number of APIs.

[0069] Furthermore, in traditional end-to-end systems, since natural language processing models with fewer parameters, such as the BERT model with around 100 million parameters, can be used, a relatively large amount of training data can be used for model training. Moreover, even if the model needs to process a large number of APIs, because the model has fewer parameters, it can still complete model updates with a large amount of training data and a short training time, thus enabling the model to achieve high API recognition accuracy.

[0070] Conversely, in such Figure 4 In the system shown, since the number of parameters in an LLM ranges from 3 billion to trillions, the sheer volume of parameters means that using a large amount of training data to update the parameters of a large language model would significantly increase training time and thus raise training costs. Therefore, to reduce the training time and cost of large language models, for a large number of APIs, the training data for each API may be limited to a few hundred to a few thousand data points, thereby limiting the total amount of training data.

[0071] Furthermore, compared to models in architectures that divide into vertical domains, large language models use less training data during training, which may lead to a decrease in API recognition accuracy.

[0072] Based on the above background, in this embodiment, when a user wants to control the vehicle to perform a certain operation at the current moment, thus triggering a voice request, and the vehicle receives and forwards the current voice request to the server, the server can receive the forwarded current voice request from the vehicle. The server can determine the target input information, which includes multiple semantic units, based on the current voice request and pre-configured target prompt information. The server can also call a pre-trained large language model that can identify the slot corresponding to the voice request and predict the application programming interface (API) that implements the voice request, to identify the slot of the current voice request and predict the corresponding API based on the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, thereby obtaining the slot identification result and the API prediction result. The server can also generate an execution command based on the slot identification result and the API prediction result and send the execution command to the vehicle so that the vehicle performs the corresponding operation according to the execution command, thereby completing the interaction with the user.

[0073] Understandably, given the limited total amount of training data for the large language model, the server in this application converts the target prompt information and the current voice request into target input information, and inputs the target input information into the large language model. This allows the large language model to complete the processing of the current voice request based on the prompts or assistance of the target prompt information. Therefore, even when the total amount of training data is limited, and the training data for each application interface may be limited to a few hundred to a few thousand records, the large language model can still accurately or reliably complete the prediction work of the application interface based on the target prompt information.

[0074] It is also understandable that the target prompt information in the embodiments of this application is content that can be set according to the actual situation.

[0075] For example, in some embodiments of this application, the target prompt information is as follows: "The following is a conversation taking place in the car, with participants including a passenger, a driver, a smart voice assistant, and an Agent. The passenger and driver will chat with each other in the car and also send commands to the smart voice assistant; the smart voice assistant will listen to all the conversations between the passenger and driver, and when it determines that a command is being sent to itself, it needs to understand the content of the command, break down the task, and provide a series of APIs that can be called and their corresponding parameters; the Agent will receive the APIs and their corresponding parameters provided by the voice assistant, execute them, and then provide the execution results. The name of each API is composed of multiple English words and their abbreviations to express its meaning. I will give a command received by the system, please provide the response that the smart voice assistant should give, the response includes two parts, ACTIONS indicating the API to be called, and ARGUMENTS indicating the parameters corresponding to this API."

[0076] Furthermore, based on the target prompt information shown above, the server can generate the target input information as follows: "The following is a dialogue taking place inside the car, involving passengers, the driver, the intelligent voice assistant, and the Agent. The passenger and driver will chat with each other and send commands to the intelligent voice assistant; the intelligent voice assistant will listen to all the conversations between the passenger and the driver, and when it determines that a command is being sent to itself, it needs to understand the command content, break down the task, and provide a series of APIs that can be called and their corresponding parameters; the Agent will receive the APIs and their corresponding parameters from the voice assistant, execute them, and then provide the execution results. The name of each API is composed of multiple English words and their abbreviations to express its meaning. I will give a command received by the system, and ask the intelligent voice assistant to give a response. The response contains two parts: ACTIONS indicates the API to be called, and ARGUMENTS indicates the parameters corresponding to this API. Voice command: 'Second row right side sliding door pauses movement'."

[0077] Furthermore, after inputting the aforementioned target input information into the large language model in this application embodiment, the large language model can complete slot identification and application interface prediction based on the reasoning or thinking methods indicated by the target prompt information, thereby outputting the corresponding slot identification results and application interface prediction results. The server can generate an execution result based on the slot identification results and application interface prediction results. In some embodiments, the server can generate an execution result in JSON format, such as "{'ACTIONS':'ControlSet','ARGUMENTS':[('DEVICE','Sliding Door'),('SET_TYPE','Pause'),('POSITION','Second Row Right Side')]}".

[0078] It's also understandable that during the pre-training phase or other training phases of a large language model, the information input to the model might only include voice request samples. Therefore, after training, the large language model is typically used to process voice requests with shorter information lengths, or in other words, voice requests with smaller data sizes. For example, the length of the input information received by the large language model during the pre-training phase is typically 2K, or 2048 characters.

[0079] Therefore, in this embodiment, the large language model processes target input information that is relatively long (or large in data size) and consists of the current voice request and target prompt information. The large language model may not have learned how to process target input information during the pre-training phase, or in other words, it may not have learned how to process information with a length greater than the voice request sample. Furthermore, retraining the large language model is costly.

[0080] Therefore, in this embodiment, the large language model completes the reasoning of the target input information and the current voice request based on the relative distance information between the first semantic unit and the second semantic unit in the target input information. Here, a semantic unit can be understood as a token. The first semantic unit and the second semantic unit can be understood as any two semantic units in the target input information. The relative distance information between the first semantic unit and the second semantic unit can be understood as the distance between one semantic unit and another semantic unit in the target input information, or as the difference in the sequence number between one semantic unit and another semantic unit in the target input information.

[0081] For example, let the target input information be S. We perform semantic unit partitioning on S to obtain the semantic unit partitioning result corresponding to S, namely {Token1, Token2, ..., Token...}. n-1 ,Token n Furthermore, the first semantic unit is Token2, and the second semantic unit is Token.n-1 In this case, the relative distance between the first semantic unit and the second semantic unit can be n-1-2.

[0082] It is understandable that relative distance information can, to some extent, characterize the correlation between the first semantic unit and the second semantic unit in the target speech request.

[0083] For example, the closer the first semantic unit is to the second semantic unit, the shorter the relative distance between the first semantic unit and the second semantic unit, and therefore the higher the correlation between the first semantic unit and the second semantic unit can be considered. Conversely, the less close the first semantic unit is to the second semantic unit, the longer the relative distance between the first semantic unit and the second semantic unit, and therefore the lower the correlation between the first semantic unit and the second semantic unit can be considered.

[0084] Therefore, the embodiments of this application can enable the large language model to refer to or consider the relative distance information between any two semantic units in the target input information, and understand the target input information, thereby reliably completing tasks such as slot identification and application programming interface prediction.

[0085] Based on this, the implementation method of this application enables the large language model to understand target input information with a length greater than 2K, such as 8K to 16K, based on the relative distance information between the first semantic unit and the second semantic unit in the target input information. This allows for a reliable understanding of the target prompt information in the target input information, thereby enabling tasks such as current voice request slot recognition and application programming interface (API) prediction. At the same time, it avoids the need to retrain the large language model to make it suitable for target input information processing. The length of input information that the large language model can process can be easily and efficiently expanded from 2K to 8K to 16K, thus ensuring that the model input can contain longer information. When the large language model predicts APIs, it can be based on longer or richer input information. Therefore, API prediction can be based on rich input information, and the prediction accuracy is guaranteed.

[0086] In summary, in this embodiment, the server can complete tasks such as slot identification and application programming interface (API) prediction using a large language model, thereby directly processing voice requests and completing voice interaction. This embodiment, based on a pre-trained large language model, reduces the need to call related models from different domains, effectively reducing latency and improving the response speed to user voice requests. This embodiment enables the large language model to output the slot identification result and API prediction result corresponding to the current voice request based on the target prompt information in the target input information and the relative distance information between the first semantic unit and the second semantic unit in the target input information, ensuring the reliability of both the slot identification result and the API prediction result.

[0087] Please see Figure 5 In some embodiments of this application, the voice interaction method further includes:

[0088] 05: Determine the target prompt information based on multiple pre-configured application interfaces and the corresponding description information for each application interface.

[0089] The voice interaction device according to this application embodiment further includes a prompt information determination module. The prompt information determination module is used to determine the target prompt information based on a plurality of pre-configured application programming interfaces and the description information corresponding to each application programming interface.

[0090] The processor in this embodiment is also used to determine target prompt information based on a plurality of pre-configured application programming interfaces and description information corresponding to each application programming interface.

[0091] It should be noted that in end-to-end dialogue systems based on large language models without vertical domain segmentation, the total amount of training data used to train the large language model is relatively small due to the high training time and cost. Consequently, because the large language model needs to process a large number of APIs, and each API has limited training data, the large language model may predict incorrect APIs when faced with unknown or rare voice requests, or even create a new API as output, leading to API recognition errors.

[0092] Based on the above background, the embodiments of this application add API description information to the target prompt information, that is, the description information corresponding to the API, so that the large language model can complete the slot recognition and application interface prediction of the current voice request based on the description information of the API in the target prompt information.

[0093] For example, in one example, an API is named "NavigationSetRoutePrefer". The description of this API includes: "NavigationSetRoutePrefer is composed of four words: Navigation, Set, Route, and Prefer, which represent navigation, settings, routes, and preferences, respectively. Therefore, the meaning of the NavigationSetRoutePrefer API is to set navigation route preferences, such as setting preferences for expressways and avoiding toll roads."

[0094] In another example, an API is named "NavigationSetMapMode". The description for this API includes the following: "NavigationSetMapMode is composed of the three words Navigation, Set, and MapMode, which represent navigation, settings, and map mode, respectively. Therefore, this API means setting the map mode for navigation. Specifically, it is used to control the mode in which the map is displayed in the navigation system, such as switching the map mode from 2D to 3D."

[0095] In another example, an API is named "VoiceSet". The description for this API includes the following: "The VoiceSet API consists of two words: Voice and Set. It represents settings related to in-car sound. It is typically used to control the sound output and effects of car audio systems, air conditioning, and other devices. The VoiceSet API can be used to set parameters such as volume, sound quality, and mode, and is commonly used to adjust the volume and sound quality of the in-car sound. Similar to VoiceOpen and VoiceClose, VoiceSet is also a commonly used API, typically used to control the sound output and effects of car audio systems, air conditioning, and other devices."

[0096] Furthermore, in some embodiments of this application, the target prompt information generated by the server based on the above API description information may include: "The following is a dialogue that takes place in the car, with participants including passengers, drivers, intelligent voice assistants, and Agents. Passengers and drivers will chat with each other in the car and also send commands to the intelligent voice assistant; the intelligent voice assistant will listen to all the conversations between passengers and drivers, and when it determines that a command is being sent to itself, it needs to understand the content of the command and then break down the task, providing a series of APIs that can be called and corresponding parameters; the Agent will receive the APIs and corresponding parameters provided by the voice assistant, execute them, and then provide the execution result. Each API name is composed of multiple English words and their abbreviations to express its meaning. Next, I will give you a few APIs and their meanings. Please read the API definitions, understand their constituent meanings, and understand the relationship between the API definitions and API names. API Definition: API1: NavigationSetRoutePrefer Meaning: NavigationSetRoutePrefer is composed of the four words Navigation, Set, Route, and Prefer, which represent navigation, settings, routes, and preferences, respectively. Therefore, the meaning of the NavigationSetRoutePrefer API is to set navigation route preferences, such as setting..." Preferring expressways and avoiding toll roads, etc. API 2: NavigationSetMapMode Meaning: NavigationSetMapMode is composed of the words Navigation, Set, and MapMode, representing navigation, settings, and map mode respectively. Therefore, this API means setting the map mode for navigation. Specifically, it controls the map display mode in the navigation system, such as switching the map mode from 2D to 3D. API 3: VoiceSet Meaning: The VoiceSet API consists of two words, Voice and Set, representing the control over the in-car audio. Related settings operations. This is typically used to control the sound output and effects of devices such as car audio systems and air conditioning. The VoiceSet API can be used to set parameters such as sound volume, sound quality, and mode, and is typically used to adjust the volume and sound quality inside the car. Similar to VoiceOpen and VoiceClose, VoiceSet is also a commonly used API, typically used to control the sound output and effects of devices such as car audio systems and air conditioning. Finally, I will provide a command received by the system, asking the intelligent voice assistant to provide a response. The response consists of two parts: ACTIONS indicating the API to be called, and ARGUMENTS indicating the parameters corresponding to this API.

[0097] It should be noted that the above example shows a scenario where the target prompt information includes descriptions of 3 APIs. However, it is understood that the number of API descriptions that the target prompt information may include is a number that can be set according to the actual situation, and 3 is only an illustrative number.

[0098] It should also be understood that when the target prompt information and target input information include descriptions of all application programming interfaces (APIs), if the large language model faces an unknown or rare current voice request, it can select the corresponding API from all APIs as the prediction result for the current voice request based on the descriptions of all APIs. This can avoid the situation where the large language model creates a non-existent API as the prediction result.

[0099] Thus, the server in this embodiment enables the large language model to predict the application interface of the current voice request based on the description information of the application interface in the target input information. Therefore, the reliability of the application interface prediction result is guaranteed, and it can avoid to some extent the situation where the large language model creates a new application interface as the prediction result when facing an unknown or rare current voice request.

[0100] Please see Figure 6 In some embodiments of this application, step 05 includes:

[0101] 050: Select the target application interface from multiple application interfaces;

[0102] 051: Determine the target prompt information based on the interface description information corresponding to the target application interface.

[0103] The prompt information determination module in this application embodiment is further configured to select a target application interface from multiple application interfaces, and to determine the target prompt information based on the interface description information corresponding to the target application interface.

[0104] The processor in this application embodiment is further configured to select a target application interface from a plurality of application interfaces, and to determine target prompt information based on the interface description information corresponding to the target application interface.

[0105] Specifically, without segmentation by vertical domain, a large language model needs to process all APIs within each vertical domain. Furthermore, if the number of APIs within each vertical domain is large—for example, in a large-scale, multi-functional end-to-end system—the number of APIs could reach hundreds. Simultaneously, to ensure the recognition performance of the large language model, the API descriptions are quite detailed, resulting in long information lengths (or, larger data sizes).

[0106] Furthermore, in this embodiment, when the target input information is long due to the inclusion of descriptions for multiple APIs, the large language model can understand and process the target prompt information based on the relative distance information between the first semantic unit and the second semantic unit in the target input information. However, if the length of the target input information exceeds the upper limit of the input information length of the large language model due to too many APIs, the large language model may not be able to correctly complete the API prediction and slot recognition for the current voice request.

[0107] Based on the above background, the server in the embodiments of this application can extract a certain number of APIs from a pre-configured plurality of APIs as target APIs, and then use the description information of the target APIs to complete the generation of target prompt information and target input information.

[0108] For example, in some embodiments of this application, the server stores 10 pre-configured APIs and description information for each API. The server can then randomly select 3 APIs from the 10 APIs to obtain 3 target APIs. Furthermore, the server can generate target prompt information based on the description information of these 3 target APIs.

[0109] Optionally, in some embodiments of this application, the server can determine, based on prior knowledge, that the upper limit of the input information length for a "large language model capable of slot identification and API prediction based on relative distance information" is 16K, meaning the information input to a large language model can contain a maximum of 16×1024 characters. Combined with the pre-calculated average information length of the API descriptions, the server can ultimately determine that the target prompt information can include descriptions of up to 80 APIs. Therefore, when the total number of pre-configured APIs exceeds 80, the server can select 80 APIs from all APIs to obtain the target API. When the number of pre-configured APIs does not exceed 80, the server can use all APIs as the target API.

[0110] Furthermore, it is understandable that, compared to large language models that do not complete tasks such as slot identification and API prediction through relative distance information, the large language model in this application has a longer input information length, and thus can complete the API prediction task based on the description information of more APIs, thereby ensuring the accuracy of API prediction.

[0111] Furthermore, it is understood that the specific method for selecting the target API is configurable based on the circumstances, and random selection is only one of the feasible implementation methods of this application. In another implementation method of this application, the server can classify or predict the API of the current voice request based on a pre-trained classification model with a small number of parameters to obtain the API classification result, and recall a certain number of APIs from all APIs based on this API classification result to obtain the target API.

[0112] In addition, it is also understood that when the target prompt information and target input information include API description information, the implementation method of this application can be applied to various vertical domains such as music, navigation, encyclopedia, weather, telephone, etc. without relying on vertical domain expertise and manual intervention, thus not being limited to the vehicle domain in the above example.

[0113] For example, when the target prompt and target input information include description information for an API named NavigationSearchOnwayPoi, even if the large language model has not processed the voice request "find charging operators along the way", the large language model can still predict that NavigationSearchOnwayPoi is the API corresponding to "find charging operators along the way" based on the API description information in the target prompt and target input information.

[0114] Thus, the server in this embodiment can determine the target application interface among all application interfaces, and determine the target prompt information and target input information according to the description information corresponding to the target application interface. This can avoid the situation where the target prompt information and target input information are too long to a certain extent, and the slot identification and application interface prediction performed by the large language model based on the target prompt information and target input information can be carried out stably.

[0115] Please see Figure 7 In some embodiments of this application, step 03 includes:

[0116] 030: Based on a large language model, determine the encoding result corresponding to the target input information according to relative distance information and semantic units;

[0117] 031: Determine the slot identification result and application interface prediction result based on the encoding result.

[0118] The result determination module in this application implementation is also used to determine the encoding result corresponding to the target input information based on the large language model, relative distance information and semantic units, and to determine the slot identification result and application interface prediction result based on the encoding result.

[0119] The processor in this embodiment is also used to determine the encoding result corresponding to the target input information based on a large language model, relative distance information, and semantic units, and to determine the slot identification result and application interface prediction result based on the encoding result.

[0120] Specifically, in order to ensure that the large language model can reasonably complete slot identification and application interface prediction through target input information, the implementation method of this application enables the large language model to encode the target input information based on the relative distance information between any two speech units in the target prompt information and in combination with all semantic units in the target prompt information. Thus, the large language model can infer the slot identification result and the application interface prediction result based on the encoding result.

[0121] It is understandable that the specific process of determining the encoding structure based on relative distance information and semantic units is customary. For example, in some embodiments of this application, the server can determine the weight of the first semantic unit based on the relative distance information between the first semantic unit and the second semantic unit, and determine the weighted expression of the first semantic unit using the weight of the first semantic unit. Then, if the second semantic unit is fixed to a specific semantic unit, and each semantic unit in the target input information is sequentially used as the first semantic unit, the server can complete the encoding of the target input information based on the weighted expression of each semantic unit in the target input information to obtain the encoding result.

[0122] It is also understandable that, since the encoding result of the target input information is determined based on relative distance information, the encoding result can reflect the relative distance between any first semantic unit and the second semantic unit in the target input information to a certain extent. Therefore, when the large language model performs inference, it can take into account the relative distance between two semantic units in the target input information.

[0123] It is understandable that the closer the relative distance between two semantic units, the higher the correlation between them. Therefore, the implementation method of this application enables large language models to reliably complete slot identification and application programming interface prediction based on the encoding results determined by the relative distance information.

[0124] Thus, the server in this application embodiment can encode the target prompt information based on a large language model, using the relative distance information between the first semantic unit and the second semantic unit in the target prompt information to obtain the encoding result. Then, when slot identification and application interface prediction are performed using the relative distance information indicated by the encoding result, the accuracy of slot identification and the accuracy of application interface prediction can be guaranteed to a certain extent.

[0125] Please see Figure 8 In some embodiments of this application, step 030 includes:

[0126] 0300: Based on a large language model, attention score information corresponding to the target input information is determined according to relative distance information and query vector and key vector of each semantic unit;

[0127] 0301: Determine the encoding result based on the attention score information.

[0128] The result determination module in this application implementation is also used to determine the attention score information corresponding to the target input information based on the large language model, according to the relative distance information and the query vector and key vector of each semantic unit, and to determine the encoding result based on the attention score information.

[0129] The processor in this embodiment is also used to determine the attention score information corresponding to the target input information based on the large language model, according to the relative distance information and the query vector and key vector of each semantic unit, and to determine the encoding result based on the attention score information.

[0130] Specifically, the implementation of this application can complete the encoding of target input information through an attention mechanism. That is, by performing query vector mapping and key vector mapping on each semantic unit in the target input information, the query vector and key vector of each semantic unit are obtained.

[0131] Furthermore, based on the query vector and key vector of each semantic unit, the relative distance information between the first semantic unit and the second semantic unit in the target input information is combined, or in other words, the relative distance information between any two semantic units in the target input information is combined, thus generating the attention score information of the target input information.

[0132] Optionally, in some embodiments of this application, the attention score information corresponding to the target input information can be represented by the following formula:

[0133] softmax(q i K T +m·[-(i-1],…,-2,-1,0])

[0134] In the formula, softmax() can be understood as an activation function, used to map the parameters within the parentheses to the interval [0,1]. i K represents the query vector for the i-th semantic unit. T This can be understood as a matrix composed of the key vectors of each semantic unit in the target input information. m is a pre-defined fixed constant. [-(i-1],…,-2,-1,0] can be understood as q i The relative distance to the key vector of each semantic unit.

[0135] It is understandable that the relative distance information between two semantic units also exists in the key vectors and query vectors corresponding to the two semantic units respectively. For example, if the relative distance between the i-th semantic unit and the j-th semantic unit is ij (or ji), then the query vector q of the i-th semantic unit... i The key vector k relative to the j-th semantic unit j The relative distance is also ij (or ji).

[0136] It is also understood that the attention score information calculated by the above formula can be interpreted as the attention score information obtained based on the multi-head attention mechanism. Therefore, m in the above formula can be understood as a parameter related to the "head" in the multi-head attention mechanism. In some embodiments of this application, when the multi-head attention mechanism includes 8 "heads", the value of m is 1 / 2, 1 / 4, 1 / 8...1 / 2. 8 .

[0137] What is understandable is that, regarding q i In terms of q i The closer the relative distance between semantic units, the greater the contribution of their key vectors to the attention score; the farther the relative distance between semantic units, the lower the contribution of their key vectors to the attention score. In other words, attention score information can be penalized based on relative distance.

[0138] Furthermore, since attention score information is penalized by relative distance, for any semantic unit in the target input, the large language model will focus on semantic units that are relatively close to the semantic unit. Consequently, based on the encoding results determined by the attention score information, the large language model can reliably perform slot identification and application interface prediction based on phase distance.

[0139] Furthermore, it is understandable that, compared to the traditional attention mechanism's method of mapping position vectors, the implementation method of this application avoids the calculation of position vectors because it is based on the predetermined relative distance between each semantic unit. At the same time, the calculation of position vectors in the traditional attention mechanism depends on input information of a fixed length, while the implementation method of this application uses the relative distance between each semantic unit, thus enabling attention calculation and encoding processing of input information of arbitrary length.

[0140] Thus, the embodiments of this application enable the large language model to focus on the relative distance between the first semantic unit and the second semantic unit in the target input information based on the encoding result determined by the attention score information, thereby ensuring the reliability of slot recognition and application interface prediction to a certain extent.

[0141] In some embodiments of this application, the training process of the large language model includes:

[0142] A reference model is trained using pre-determined input information samples and the relative distance information between the third semantic unit and the fourth semantic unit in the input information samples to obtain a large language model. The input information samples are determined by target prompt information and voice request samples.

[0143] The voice interaction device according to the embodiments of this application further includes a training module. The training module trains a reference model using predetermined input information samples and the relative distance information of the third semantic unit to the fourth semantic unit in the input information samples to obtain a large language model, wherein the input information samples are determined by target prompt information and voice request samples.

[0144] The processor in this embodiment is further configured to train a reference model using predetermined input information samples and the relative distance information between the third semantic unit and the fourth semantic unit in the input information samples, to obtain a large language model, wherein the input information samples are determined by target prompt information and voice request samples.

[0145] Specifically, in order to ensure that the large language model can be well applied to slot identification and application programming interface prediction of target input information, the implementation method of this application can fine-tune or supervised fine-tune the large language model using pre-set input information samples before the large language model is applied. Then, after fine-tuning, the large language model can output more accurate or reliable results based on the target input information.

[0146] Understandably, during fine-tuning training, because the large language model can be trained based on the relative distance information between the third semantic unit and the fourth semantic unit in the input information sample, it can learn the knowledge or ability to understand the input information sample through relative distance information. Here, the third and fourth semantic units can be understood as any two semantic units in the input information sample.

[0147] Optionally, in some embodiments of this application, the reference model for fine-tuning the training process can determine the attention score information based on the formula above, that is:

[0148] softmax(q i K T +m·[-(i-1],…,-2,-1,0])

[0149] Understandably, when determining attention scores based on this formula, the greater the relative distance between two semantic units, the lower the attention score; conversely, the closer the relative distance, the higher the attention score, thus achieving attention score penalty based on relative distance. Furthermore, this attention score penalty mechanism exhibits good robustness and has relatively lenient requirements on the value of m. Moreover, this attention score penalty mechanism is unlikely to significantly impact the fundamental language capabilities of the reference model or the large language model. Therefore, the reference model and the large language model can reliably learn how to process input information samples through fine-tuning training, thereby stably completing the prediction of the application programming interface corresponding to the target input information.

[0150] Optionally, in some embodiments of this application, the target prompt information includes descriptive information about the application programming interface (API). Furthermore, during the fine-tuning process of the reference model and the large language model, the model can learn the knowledge or ability to predict the API based on its descriptive information, thereby ensuring the accuracy of API recognition.

[0151] Furthermore, even when the target prompt information includes descriptions of multiple application interfaces, and each application interface description is lengthy, the reference model and the large language model can understand the target prompt information and input information samples based on the relative distance information between the third semantic unit and the fourth semantic unit. This avoids the situation where the reference model cannot process input information samples because it has not learned how to process them before fine-tuning the training.

[0152] It is understood that the reference model in the embodiments of this application can be understood as a base model with general natural language processing capabilities, trained using samples with relatively short information lengths. It can also be understood as a large language model that has been infused with knowledge of a specific vertical domain, such as the vehicle domain, and is applicable to task processing within that specific vertical domain.

[0153] Thus, the server in this embodiment can train a reference model to obtain a large language model through input information samples, so that the trained large language model can be well applied to the processing of input information containing voice requests and prompts, and the robustness of the large language model is guaranteed.

[0154] In some embodiments of this application, step 03 includes:

[0155] Based on the large language model, the slot identification result, application interface prediction result, and target parameters corresponding to the application interface prediction result are determined according to the target input information and relative distance information.

[0156] The result determination module in this embodiment is also used to determine the slot identification result, the application interface prediction result, and the target parameters corresponding to the application interface prediction result based on the large language model, target input information, and relative distance information.

[0157] The processor in this embodiment is also used to determine the slot identification result, the application interface prediction result, and the target parameters corresponding to the application interface prediction result based on the large language model, according to the target input information and the relative distance information.

[0158] Specifically, in the embodiments of this application, the large language model can perform slot identification, application interface prediction, and determine the target parameters that need to be filled in the slot identification results based on the understanding of the target input information and the reasoning method or auxiliary information indicated by the target prompt information.

[0159] In one example, the received voice request is "The sliding door on the right side of the second row is paused." Based on the prompt information written in the example above, the large language model can identify entities such as "right side of the second row," "sliding door," and "pause" from the voice request, predict the application interface 'ControlSet,' and fill the corresponding entities into the parameters of this interface, including 'DEVICE,' 'SET_TYPE,' and 'POSITION.'

[0160] Thus, in the embodiments of this application, the large language model can complete slot identification, application interface prediction, and determine the target parameters for filling the slot identification results into the application interface based on the target input information.

[0161] Please see Figure 9 In some embodiments of this application, step 04 includes:

[0162] The interaction module in this embodiment is also used to fill the slot recognition result into the target parameter corresponding to the application interface prediction result, to obtain the execution result of the application interface parameter filling, and to output the execution result to the vehicle to complete the voice interaction.

[0163] The processor in this embodiment is also used to fill the slot identification result into the target parameter corresponding to the application interface prediction result, to obtain the execution result of the application interface parameter filling, and to output the execution result to the vehicle to complete the voice interaction.

[0164] Specifically, in the case where the large language module can perform slot recognition, application interface prediction and determine the target parameters that need to be filled in based on the understanding of the target input information and the reasoning method or auxiliary information indicated by the target prompt information, the implementation of this application can also fill the target parameters of the application interface with the slot recognition results to obtain the execution results that can be output to the vehicle side.

[0165] In one example, the received voice request is "The sliding door on the right side of the second row will stop moving." Based on the prompt information written in the example above, the large language model can identify entities such as "right side of the second row," "sliding door," and "pause" from the voice request, predict the application programming interface 'ControlSet,' and fill the corresponding entities into the parameters of this interface, including 'DEVICE', 'SET_TYPE', and 'POSITION'. The result of parameter filling can be represented as: [('DEVICE', 'sliding door'), ('SET_TYPE', 'pause'), ('POSITION', 'right side of the second row')].

[0166] It is understandable that the slot identification result filling can be performed by the large language model based on the code of the built-in program. That is, after the large language model outputs the slot identification result, the application interface prediction result and the target parameter, the server completes the slot identification result filling operation through pre-configured code or program.

[0167] Thus, in this embodiment, the slot identification results can be filled into the corresponding target parameters of the application interface based on the slot identification results, the application interface prediction results, and the target operation.

[0168] This application also provides a computer-readable storage medium storing a computer program that, when executed by one or more processors, implements the above-described voice interaction method.

[0169] In this specification, the terms "specifically," "furthermore," "particularly," "understandably," etc., refer to specific features, structures, materials, or characteristics described in connection with embodiments or examples that are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0170] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0171] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A voice interaction method, characterized in that, The method comprises: receiving a current voice request forwarded by a vehicle; determining target input information according to the current voice request and target prompt information, wherein the target input information comprises a plurality of semantic units, the target prompt information is pre-configured, and the target prompt information comprises API description information; based on a large language model, determining a slot recognition result and an application program interface prediction result corresponding to the current voice request according to the target input information and relative distance information of a first semantic unit with respect to a second semantic unit in the target input information, wherein the large language model is pre-trained and can identify a slot corresponding to a voice request according to an inference manner indicated by the target prompt information, and can predict an application program interface corresponding to the voice request; outputting an execution result according to the slot recognition result and the application program interface prediction result, and issuing the execution result to the vehicle to complete voice interaction.

2. The method of claim 1, wherein, The method further comprises: determining the target prompt information based on a plurality of pre-configured application program interfaces and corresponding description information of each application program interface.

3. The method of claim 2, wherein, The determination of the target prompt information based on the plurality of pre-configured application program interfaces and the corresponding description information of each application program interface comprises: selecting a target application program interface from the plurality of application program interfaces; determining the target prompt information according to interface description information corresponding to the target application program interface.

4. The method of claim 1, wherein, The determination of the slot recognition result and the application program interface prediction result corresponding to the current voice request based on the large language model according to the target input information and the relative distance information of the first semantic unit with respect to the second semantic unit in the target input information comprises: determining an encoding result corresponding to the target input information based on the large language model according to the relative distance information and the semantic units; determining the slot recognition result and the application program interface prediction result according to the encoding result.

5. The method of claim 4, wherein, The determination of the encoding result corresponding to the target input information based on the large language model according to the relative distance information and the semantic units comprises: determining attention score information corresponding to the target input information based on the large language model according to the relative distance information and a query vector and a key vector of each semantic unit; determining the encoding result according to the attention score information.

6. The method of claim 1, wherein, The training process of the large language model comprises: training a reference model using a pre-determined input information sample and relative distance information of a third semantic unit with respect to a fourth semantic unit in the input information sample to obtain the large language model, wherein the input information sample is determined by the target prompt information and a voice request sample.

7. The method of claim 1, wherein, The determination of the slot recognition result and the application program interface prediction result corresponding to the current voice request based on the large language model according to the target input information and the relative distance information of the first semantic unit with respect to the second semantic unit in the target input information comprises: Based on the large language model, the slot recognition result, the application program interface prediction result and the target parameter corresponding to the application program interface prediction result are determined according to the target input information and the relative distance information.

8. The method of claim 7, wherein, The execution result is outputted according to the slot recognition result and the application program interface prediction result, and is issued to the vehicle to complete the voice interaction, and the execution result comprises: The slot recognition result is filled into the target parameter corresponding to the application program interface prediction result, and the execution result filled with the application program interface parameter is obtained; The execution result is outputted and issued to the vehicle to complete the voice interaction.

9. A server, characterized by The computer readable storage medium stores a computer program, and when the computer program is executed by one or more processors, the method of any one of claims 1-8 is implemented.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and when the computer program is executed by one or more processors, the method of any one of claims 1-8 is implemented.

Images