A method and apparatus for switching a reply strategy, an electronic device, and a storage medium

Abstract

A method and device for switching a reply strategy, an electronic device, and a storage medium are disclosed. The method comprises determining target strategy information corresponding to a feedback instruction in a first reply strategy in response to a user input, and modifying the target strategy information based on the feedback instruction to obtain a second reply strategy that is more in line with user needs, thereby achieving quick switching of a large language model reply strategy. The method provided by the present disclosure enables the large language model to generate a personalized reply in line with user intent according to the modified second reply strategy during user-human interaction, thereby enhancing the degree of personalization and user experience of the large language model reply in the intelligent cockpit scenario. At the same time, the user does not need to perform multiple rounds of feedback during each human-computer interaction process, significantly improving the efficiency of human-computer interaction.

Description

Technical Field

[0001] This disclosure relates to the field of large language model technology, and in particular to a method, apparatus, electronic device and storage medium for switching response strategies. Background Technology

[0002] With the development of artificial intelligence, large language models are widely used in the smart cockpits of intelligent vehicles. In the field of smart cockpits, the response strategies of large language models are pre-set by developers. The large models cannot adjust their response strategies based on user feedback. Users need to provide multiple rounds of information feedback to obtain personalized responses that match their intentions, resulting in low efficiency in human-computer interaction. Summary of the Invention

[0003] To address the aforementioned technical problems, this disclosure provides a method, apparatus, electronic device, and storage medium for switching response strategies, thereby resolving the issue of low human-computer interaction efficiency when users generate personalized responses using large language models.

[0004] The first aspect of this disclosure provides a method for switching response strategies, including: In response to the user's input feedback command, the target strategy information corresponding to the feedback command in the first response strategy is determined, wherein the first response strategy is a pre-set response strategy; based on the feedback command, the target strategy information is modified to obtain the second response strategy; the response strategy of the large language model is switched from the first response strategy to the second response strategy.

[0005] A second aspect of this disclosure provides a switching device for a response strategy, comprising: The feedback module is configured to respond to user-input feedback commands and determine the target strategy information corresponding to the feedback command in the first response strategy, wherein the first response strategy is a pre-set response strategy; the strategy modification module is configured to modify the target strategy information based on the feedback command to obtain the second response strategy; the switching module is configured to switch the response strategy of the large language model from the first response strategy to the second response strategy.

[0006] A third aspect of this disclosure provides a computer-readable storage medium storing a computer program for executing the switching method of the response strategy proposed in the first aspect embodiment.

[0007] A fourth aspect of this disclosure provides an electronic device including a processor; a memory for storing processor-executable instructions; and the processor for reading executable instructions from the memory and executing the instructions to implement the switching method of the response strategy proposed in the first aspect embodiment.

[0008] This disclosure provides a method, apparatus, electronic device, and storage medium for switching response strategies. Responding to user-input feedback commands, a target strategy information corresponding to the feedback command is determined from a pre-set first response strategy. Based on the feedback command, the target strategy information is modified to obtain a second response strategy that better suits the user's needs, thereby enabling rapid switching of the large language model's response strategy. The method provided by this disclosure, through modification and switching of response strategies, allows the large language model to generate personalized responses that match the user's intent during human-computer interaction, according to the modified second response strategy. This enhances the personalization of the large language model's responses and the user experience in intelligent cockpit scenarios. Simultaneously, users no longer need to perform multiple rounds of feedback during each human-computer interaction, significantly improving human-computer interaction efficiency. Attached Figure Description

[0009] Figure 1 This is a diagram of an in-vehicle system architecture provided in an exemplary embodiment of the present disclosure; Figure 2 This is a diagram of an in-vehicle system architecture provided in another exemplary embodiment of this disclosure; Figure 3 This is a flowchart illustrating an exemplary embodiment of the present disclosure providing a method for switching response strategies; Figure 4 This is a flowchart of step S301 provided in an exemplary embodiment of this disclosure; Figure 5 This is a flowchart illustrating step S3012 provided in an exemplary embodiment of this disclosure; Figure 6 This is a schematic flowchart of step S302 provided in an exemplary embodiment of this disclosure; Figure 7 This is a schematic flowchart of step S3023 provided in an exemplary embodiment of this disclosure; Figure 8 This is another flowchart illustrating a method for switching response strategies provided in an exemplary embodiment of this disclosure; Figure 9 This is another flowchart illustrating the method for switching response strategies provided in an exemplary embodiment of this disclosure; Figure 10 This is a flowchart illustrating step S901 provided in an exemplary embodiment of this disclosure; Figure 11 This is a flowchart illustrating step S902 provided in an exemplary embodiment of this disclosure; Figure 12 This is a structural diagram of a response strategy switching device provided in an exemplary embodiment of the present disclosure; Figure 13This is a structural diagram of an electronic device provided by an exemplary embodiment of the present disclosure. Detailed Implementation

[0010] To explain this disclosure, exemplary embodiments of the disclosure will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the disclosure, and not all of them. It should be understood that the disclosure is not limited to exemplary embodiments.

[0011] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0012] Application Overview With the rapid development of artificial intelligence technology, Large Language Models (LLMs) are widely used in the intelligent cockpits of smart cars. Leveraging their powerful natural language understanding and generation capabilities, LLMs enable intelligent and personalized interaction within the smart cockpit.

[0013] In the application scenario of smart cockpits, large language models can receive command information input by users through various means such as voice and touch, and analyze and reason according to pre-set response strategies to generate appropriate response content, realize interactive response with users, and improve the user experience of smart cockpits.

[0014] However, the response strategies of the large language models currently used in smart cockpits are all pre-designed and fixed by technical developers during the model deployment phase. In actual operation, the large language models do not have the ability to autonomously adjust their response strategies based on real-time feedback from users.

[0015] In actual human-computer interaction, different users have significant differences in usage habits, interaction needs, and language expression styles. Moreover, the usage scenarios of smart cockpits (such as driving conditions, in-vehicle environment, user status, etc.) are constantly changing, and pre-fixed response strategies are difficult to fully adapt to diverse user needs and dynamic usage scenarios.

[0016] When the response generated by the large language model does not match the user's intent, the user needs to provide multiple rounds of feedback to guide the model to generate a personalized response that aligns with their intent. This process not only increases the user's interaction costs but also extends the interaction cycle and reduces the efficiency of human-computer interaction.

[0017] Exemplary System Figure 1 This is a diagram of an in-vehicle system architecture provided in an exemplary embodiment of this disclosure. (As follows) Figure 1As shown, the vehicle-mounted system 100 of this disclosure may include a processor 110, a memory 120, and a vehicle-mounted processing module 130. The processor 110 is used to execute program instructions to achieve coordinated control and data processing of various modules (devices) in the vehicle-mounted system. The processor 110 is communicatively connected to the memory 120 and the vehicle-mounted processing module 130 to realize instruction transmission and data transmission.

[0018] The memory 120 is used to store the large language model 140, thereby providing the necessary computing resources and data cache space for the large language model 140. The memory 120 is also used to store the operating system, applications, and various data resources, providing an operating environment and data support for other functions of the intelligent cockpit.

[0019] The in-vehicle processing module 130 includes a data acquisition device 131 and an intent recognition device 132. The data acquisition device 131 is used to acquire multimodal data, such as user-inputted voice commands, text commands, or real-time acquired images. For example, the data acquisition device 131 may include an audio input device, such as a microphone, for acquiring user-inputted voice commands. The data acquisition device 131 may include a touch sensor, which can acquire touch operation commands input by the user based on the central control screen of the smart cockpit, such as text commands input via touch. The data acquisition device 131 may include an image acquisition device, such as a camera, for acquiring road condition images outside the smart cockpit to determine the vehicle's driving status, and / or acquiring user images inside the smart cockpit to determine the user's activity status within the smart cockpit.

[0020] The data acquisition device 131 may also include environmental sensors, such as light sensors and temperature sensors, for collecting environmental status information within the smart cockpit. Furthermore, the data acquisition device 131 can also collect user command information via input devices such as steering wheel buttons and physical knobs.

[0021] The intent recognition device 132 is used to identify the intent type of multimodal data, wherein the multimodal data includes three intent types: functional interaction intent, feedback intent, and irrelevant topic.

[0022] When the intent type is a functional interaction intent, the multimodal data is used to trigger operations related to the business functions of the smart cockpit and generate relevant response content. For example, when a user inputs a function command into the smart cockpit via voice or touch, the in-vehicle processing module 130 preprocesses the input command. For instance, for voice-input function commands, the in-vehicle processing module 130 can perform noise reduction enhancement, text conversion, and semantic parsing. Then, the in-vehicle processing module 130 sends the preprocessed structured data to the processor 110. The processor 110 executes the function command based on the structured data and calls the large language model 140 deployed in the memory 120 to generate corresponding response content based on the structured data and response strategy. This response content is then output to the user via the in-vehicle processing module 130 in the form of voice broadcast, screen display, or a multimodal combination.

[0023] When the intent type is feedback intent, the multimodal data is used to represent the user's satisfaction with the current response and their adjustment needs. In this case, the multimodal data may include feedback instructions. For example, when a user inputs a feedback instruction to the response generated by the large language model 140, the vehicle processing module 130 analyzes the user's voice tone characteristics, touch operation behavior, facial expression changes, or text evaluation of the response based on the feedback instruction to identify the user's intent information, namely, the user's satisfaction with the current response and their adjustment needs. The vehicle processing module 130 can send the user intent information to the processor 110, thereby instructing the processor 110 to dynamically adjust the response strategy of the large language model 140 so that the response strategy of the large language model 140 can match the user's feedback intent.

[0024] If the intent type is irrelevant to the topic, indicating that the multimodal data is invalid, the processor 110 can generate a prompt message to guide the user to re-enter the information.

[0025] Figure 2 This is a diagram of an in-vehicle system architecture provided in another exemplary embodiment of this disclosure. (See diagram for example.) Figure 2 As shown, the in-vehicle system 100 in this disclosure may further include a communicator 150. The communicator 150 is used to establish a communication connection with the server 160. The large language model 140 can be deployed on the server 160 to generate response content on the server 160, reducing the computing power consumption of the smart cockpit. The communicator 150 can upload the multimodal data collected by the in-vehicle processing module 130 to the server 160 through the communication connection, thereby generating response content or adjusting the response strategy through the large language model 140, and returning the response content or the adjusted response strategy to the in-vehicle system 100 through the communication connection.

[0026] Exemplary methods Figure 3 This is a flowchart illustrating a method for switching response strategies provided in an exemplary embodiment of this disclosure. This embodiment can be applied to electronic devices, such as... Figure 3 As shown, it includes the following steps: S301: In response to a user-inputted feedback instruction, determine the target policy information corresponding to the feedback instruction in the first response policy.

[0027] Feedback instructions are the user's input expressing satisfaction and adjustment needs after receiving the first response content generated by the large language model based on the first response strategy. The first response strategy is a pre-set response strategy for the large language model, designed and solidified by the technical developers during the model deployment phase.

[0028] After receiving user input, the electronic device can feed the input into a large language model. The large language model can then perform semantic understanding on the input and generate a first response based on the semantic understanding results, according to a first response strategy.

[0029] The target strategy information is the strategy information that needs to be modified in the first response strategy as indicated by the feedback instruction. It is used to represent the direction in which the user is dissatisfied with the content of the first response or expects adjustments.

[0030] In some examples, electronic devices can perform semantic parsing on feedback instructions to extract the policy adjustment methods explicitly or implicitly expressed in the feedback instructions, thereby determining the target policy information in the first response policy based on the semantic parsing results.

[0031] In some examples, electronic devices can use a large language model to recognize the intent of user-input instructions. After the instructions are input into the large language model, it performs semantic encoding to obtain instruction codes. Then, it calculates the similarity between the instruction code and the code corresponding to each intent type to determine the intent type of the instruction code. When the similarity between the instruction code and the code corresponding to the feedback intent is greater than or equal to a preset similarity threshold, the instruction can be determined to be a feedback instruction.

[0032] S302: Based on the feedback instruction, modify the target strategy information to obtain the second response strategy.

[0033] Electronic devices can adjust one or more parameters or style settings in the target strategy information based on feedback instructions to generate a second response strategy that adapts to the user's current needs.

[0034] It should be noted that the second response strategy retains other strategy information from the first response strategy, and only modifies the target strategy information based on user feedback, thereby maintaining consistency in responses while meeting users' personalized needs.

[0035] S303: Switch the response strategy of the large language model from the first response strategy to the second response strategy.

[0036] After receiving the second response strategy, the electronic device switches the response strategy of the large language model from the first response strategy to the second response strategy, enabling the large language model to generate response content based on the second response strategy. After the switch is complete, when the user enters new command information again, the large language model will generate a second response content that meets the user's expectations based on the second response strategy.

[0037] In some examples, the strategy switch can take effect immediately, meaning that the electronic device completes the switching operation immediately after generating the second response strategy, ensuring that the user can directly generate the corresponding response content according to the second response strategy in the next interaction, without the need for multiple rounds of information feedback, thus improving interaction efficiency.

[0038] In other examples, electronic devices can also send policy switching confirmation messages to users, such as displaying a prompt like "The response policy has been modified based on your feedback," so that users are clearly aware of the current status of the effective response policy.

[0039] In addition, electronic devices can associate and store the second response strategy with the user's identifier so that when the user uses the large language model later, the user's preferred response strategy can be automatically loaded, avoiding the user having to repeatedly adjust the strategy and improving the consistency of human-computer interaction and user experience.

[0040] The method provided in this disclosure modifies the target strategy information in the first response strategy in real time in response to user input feedback instructions to obtain a second response strategy, and switches the large language model's response strategy from the first response strategy to the second response strategy. This allows the large language model to autonomously adjust its response strategy based on real-time user feedback, without requiring multiple rounds of user feedback to guide the large language model to generate responses that match the user's intent. This effectively reduces the user's interaction operation cost and interaction cycle, and improves the efficiency of human-computer interaction.

[0041] Figure 4 This is a flowchart of step S301 provided in an exemplary embodiment of this disclosure.

[0042] like Figure 4 As shown above, in the above Figure 3 Based on the illustrated embodiment, step S301 may include the following steps: S3011: In response to a user input feedback command, perform intent recognition on the content of the feedback command to obtain the user intent information of the feedback command.

[0043] In some examples, after receiving a user's feedback instruction, the electronic device can extract the instruction content from the feedback instruction and perform intent recognition on the instruction content to obtain the user's intent information. For instance, the large language model generates the first response content based on the "standard tone" in the first response strategy. If the user finds the tone of the first response content to be harsh, they will input the feedback instruction "I hope your tone is more humorous."

[0044] Electronic devices can convert feedback instructions into text, resulting in the instruction "I hope your tone is more humorous." Then, they can perform intent recognition on the instruction to obtain the user intent information "Adjust your tone style to a humorous tone."

[0045] S3012: Based on the user intent information of the feedback instruction, determine the target strategy information corresponding to the feedback instruction in the first response strategy.

[0046] After determining the user's intent information, the dimensions can be adjusted according to the strategy specified by the user's intent information, thereby locating the target strategy information corresponding to the feedback instruction in the first response strategy. Continuing the previous example, after obtaining the user intent information of "adjust the tone style to a humorous tone," the electronic device can determine that the "tone style" of the first response strategy needs to be adjusted.

[0047] Based on this, the electronic device can find the currently effective "tone style", i.e., "standard tone", in the first response strategy as the target strategy information, so as to perform subsequent modifications to the target strategy information based on the user's intent information.

[0048] The method provided in this disclosure obtains user intent information by recognizing the execution intent of feedback instructions, and then determines target policy information based on the user intent information. This can accurately locate the modification requirements of feedback instructions, avoid deviations in policy positioning due to semantic ambiguity of instructions, improve the accuracy of matching target policy information in the first response policy, ensure that subsequent policy modifications meet the actual needs of users, and improve the accuracy and effectiveness of large language model response policy modifications.

[0049] Figure 5 This is a flowchart illustrating step S3012 provided in an exemplary embodiment of this disclosure.

[0050] like Figure 5 As shown above, in the above Figure 4 Based on the illustrated embodiment, step S3012 may include the following steps: S30121: Obtain the functional keywords corresponding to the user intent information of the feedback command.

[0051] In some examples, after obtaining user intent information, electronic devices can perform word segmentation on the user intent information to extract functional keywords that can represent the dimensions of strategy adjustment. Continuing the previous example, if the user intent information is "adjust the tone and style to humorous", after word segmentation, functional keywords such as "tone" and "style" can be extracted.

[0052] S30122: Among the multiple business functions corresponding to the first response strategy, determine the target function corresponding to the functional keyword.

[0053] In some instances, the first response strategy may include multiple business functions, such as driving functions, response functions, seat adjustment, and air conditioning control. Electronic devices can match extracted functional keywords with the various business functions in the first response strategy to determine the target function indicated by the user's intent information.

[0054] Continuing with the previous example, the user intent information contains functional keywords such as "tone" and "style". These functional keywords are all related to the response content generated by the large language model. It can be determined that the user intent information needs to make strategic adjustments to the tone, style and emotional expression intensity of the large language model. Therefore, the electronic device can determine that the target function is the response function.

[0055] S30123: Determine the target policy information corresponding to the feedback instruction in the policy document corresponding to the target function.

[0056] In some examples, electronic devices can query the policy document corresponding to the target function within the first response policy. In the first response policy, each business function corresponds to a policy document, which is used to control the large language model to generate response content related to that business function according to the policy document when the user inputs instructions related to that specific business function.

[0057] Continuing with the previous example, once the electronic device determines that the target function is the reply function, it can query the strategy document corresponding to the reply function in the first reply strategy. For example, the strategy document for the reply function records strategy information related to the reply function, such as the currently effective "tone style", "emotional expression intensity parameter", "reply length parameter", "reply detail", "expression method", and "empathy level".

[0058] Since the user's intent information clearly points to the need for tone style adjustment, electronic devices can define "standard tone" in "tone style" as the target policy information corresponding to the feedback instruction in the policy document of the response strategy.

[0059] The method provided in this disclosure extracts functional keywords from user intent information, accurately locates the target function among multiple business functions of the first response strategy, and determines the target strategy information in the strategy document corresponding to the target function. This realizes the mapping of ambiguous feedback instructions to specific target strategy information, significantly improving the efficiency and accuracy of strategy information determination and providing reliable basic data support for subsequent strategy modifications.

[0060] Figure 6 This is a flowchart illustrating step S302 provided in an exemplary embodiment of this disclosure.

[0061] like Figure 6 As shown above, in the above Figure 3 Based on the illustrated embodiment, step S302 may include the following steps: S3021: Obtain evaluative keywords corresponding to the user intent information of the feedback instruction.

[0062] In some examples, during the keyword extraction process of user intent information by electronic devices, evaluative keywords can also be obtained from the user intent information. These evaluative keywords are used to indicate a need for modification of the strategy information in the area to be modified, such as descriptive words like "more humorous," "too abrupt," or "would like more detail."

[0063] Continuing with the previous example, the user intent information is "adjust the tone style to humorous". Through keyword extraction, we can obtain "adjust" and "humorous" as evaluative keywords. These keywords clearly express that the user wants to adjust the "tone style" of the large language model to "humorous tone".

[0064] S3022: Generate modification requests based on evaluative keywords.

[0065] In some examples, electronic devices can match evaluative keywords with a pre-defined rule base to generate structured modification requests. These requests can include the target policy information to be modified and the direction of the modification.

[0066] Continuing with the previous example, the electronic device can generate a modification request based on the evaluative keyword "humor" and the target strategy information "tone and style". The specific content of the modification request could be "Change the 'tone and style' in the first response strategy from 'standard tone' to 'humorous tone'", thereby transforming the vague user intent of "adjusting the tone and style to a humorous tone" into a more specific modification request, improving the accuracy of modifying the first response strategy.

[0067] S3023: In response to the modification request, modify the target policy information to obtain the second response policy.

[0068] In some examples, after generating a modification request, the electronic device can respond by modifying the target policy information in the policy document of the target function. This replaces the current parameter value in the target policy information with the target parameter value specified in the modification request, while keeping other policy information in the first response policy unchanged, thus generating a second response policy. Continuing with the previous example, the electronic device can modify the "tone and style" in the policy document of the response function from "standard tone" to "humorous tone," while keeping other policy information in the policy document of the response function unchanged, ultimately generating a second response policy that includes the modified humorous tone.

[0069] In some examples, electronic devices can also extract quantification terms from feedback instructions. Quantification terms are used to indicate the magnitude of modification to the target policy information, such as "slightly," "a little," "more," "especially," "very," "reject," "don't," etc.

[0070] In some examples, electronic devices can adaptively adjust the magnitude of modifications to the target strategy information by incorporating quantified words in feedback instructions. For instance, if the feedback instruction is "I hope your tone is slightly more humorous," the electronic device can recognize the quantified words "slightly" and "a little," indicating that the user wants a mild modification to the tone style of the large language model. In this case, the "tone style" in the strategy document for the response function can be changed from "standard tone" to "mildly humorous tone" to reflect the user's precise needs for tone style adjustment and avoid excessive modifications that would cause the modified first response strategy to deviate from the user's expected effect.

[0071] In some examples, if the feedback instruction is "I hope your tone is particularly humorous", the electronic device recognizes the quantitative word "particular" to indicate that the user wants a high degree of modification to the tone style of the large language model. In this case, the "tone style" in the strategy document of the response function can be changed from "standard tone" to "highly humorous tone" to fully meet the user's expectations for the intensity of humorous expression.

[0072] In some examples, the electronic device can also synchronously adjust relevant strategy information based on the modified target strategy information to maintain coordination and consistency among parameters within the strategy document. For instance, when the electronic device changes the "tone style" from "standard tone" to "highly humorous tone," it can simultaneously adjust the "emotional expression intensity parameter" in the response function from "neutral" to "positive" to ensure that the second response strategy matches each other in both tone style and emotional expression intensity, avoiding stylistic disjointedness in the response content due to a single parameter adjustment.

[0073] In some examples, users can input feedback commands based on the first response. Therefore, after the electronic device switches the response strategy of the large language model from the first response strategy to the second response strategy, the electronic device can re-input the command information corresponding to the first response content into the large language model. The large language model will then regenerate the first response content corresponding to the command information based on the second response strategy and output the regenerated first response content to the user. This allows the user to intuitively perceive the difference in the effect of the large language model after switching from the first response strategy to the second response strategy, improving the user's satisfaction with the strategy adjustment results.

[0074] The method provided in this disclosure generates a modification request by extracting evaluative keywords from user intent information, and performs precise modifications to the target policy information based on the modification request. This transforms ambiguous natural language feedback into specific, executable policy adjustment operations, effectively improving the automation and accuracy of policy modification and ensuring that the second response policy accurately reflects the user's personalized needs.

[0075] Figure 7 This is a flowchart illustrating step S3023 provided in an exemplary embodiment of this disclosure.

[0076] like Figure 7 As shown above, in the above Figure 6 Based on the illustrated embodiment, step S3023 may include the following steps: S30231: In response to a modification request, determine the type of operation to modify the target strategy information based on evaluative keywords.

[0077] In some examples, the operation types include add, replace, and delete. To differentiate between different operation types, electronic devices can pre-establish a mapping table between evaluative keywords and operation types. For example, evaluative keywords such as "add," "join," and "increase" correspond to add operations; evaluative keywords such as "change to," "adjust to," and "replace with" correspond to replace operations; and evaluative keywords such as "delete," "remove," and "remove" correspond to delete operations.

[0078] After generating a modification request, the electronic device can look up the operation type corresponding to the evaluative keywords in the user intent information based on the mapping table between evaluative keywords and operation types, thereby determining what modification operation to perform on the target policy information.

[0079] S30232: In response to the add operation, add new information based on evaluative keywords to the preset position of the target strategy information to obtain the second response strategy.

[0080] In some examples, when the operation type is an add operation, the electronic device can respond to the add operation by determining the preset location of the target policy information. The preset location refers to a specific field or parameter node in the policy document that allows the insertion of new information, such as the extended parameter area at the end of the policy document, or supplementary descriptions after specific policy information.

[0081] Electronic devices can add new information based on evaluative keywords at preset locations to obtain a second response strategy. For example, if the modification request is "add a blessing at the end of the reply", and the evaluative keywords are "add" and "blessing", the electronic device can add a "blessing template" as a new information under "end processing strategy parameters" in the strategy document of the reply function, and set the content of the "blessing template" to a blessing generated based on user intent information, such as "wishing you a happy life", thereby generating a second response strategy that includes the strategy of adding a blessing.

[0082] S30233: In response to the replacement operation, the target strategy information is replaced with strategy information generated based on evaluative keywords to obtain the second response strategy.

[0083] In some examples, when the operation type is a replacement operation, the electronic device can respond to the replacement operation by directly replacing the target policy information with policy information generated based on evaluative keywords, thereby obtaining a second response policy.

[0084] For example, continuing the previous example, if the modification request is "to change the 'tone style' in the first response strategy from 'standard tone' to 'humorous tone'", then the target strategy information is "standard tone" in "tone style", and the strategy information generated by the electronic device based on the evaluative keywords is "humorous tone" in "tone style". The electronic device can replace the "tone style" in the strategy document of the response function from "standard tone" to "humorous tone", while keeping other strategy information unchanged, to generate a second response strategy with "humorous tone" as the "tone style".

[0085] S30234: In response to the deletion operation, delete at least a portion of the policy information in the target policy information to obtain a second response policy.

[0086] In some examples, when the operation type is a delete operation, the electronic device can respond to the delete operation by determining the policy information to be deleted from the target policy information. The policy information to be deleted can be a portion of the target policy information or all of the target policy information.

[0087] The electronic device can remove the policy information to be deleted from the target policy information to obtain a second response policy. For example, if the modification request is "delete the technical term explanation in the response", and the evaluative keywords are "delete" and "technical term explanation", the electronic device can determine that the target policy information is the "terminology explanation policy parameter" in the response function policy document, and delete the parameter and its related terminology explanation template from the policy document to generate a second response policy that does not contain technical term explanations, making the response content of the large language model more concise and direct.

[0088] In some examples, before performing a deletion operation, the electronic device can also detect other policy information that is related to the policy information to be deleted in the first response strategy. If a strong dependency relationship is detected between the policy information to be deleted and other policy information, the electronic device can generate a risk warning message to remind the user that deleting the policy information may cause the associated function to fail or the response quality to decrease. The deletion operation will only be performed after obtaining user feedback confirmation, so as to avoid accidentally deleting the target policy information and affecting the normal response capability of the large language model.

[0089] It should be noted that the embodiments provided above in this disclosure determine one or more steps in S30232-S30234 to be executed based on the modification type corresponding to the modification request. This disclosure does not impose specific restrictions on the execution order of steps S30232-S30234.

[0090] In some examples, electronic devices can also perform a validity check on the modification operation before executing it. The validity check includes verifying whether the modified policy document conforms to preset policy rules. If the check passes, the modification operation is executed; if the check fails, a modification failure message is returned to the user, explaining the reason for the failure, such as "The large language model does not support this tone style" or "The modified policy information conflicts with other existing policy information," guiding the user to re-enter the feedback command.

[0091] The method provided in this disclosure determines the type of modification operation (addition, replacement, deletion) through evaluative keywords, and executes the corresponding modification operation to generate a second response strategy. This enables diverse and flexible modifications to the first response strategy, meeting users' strategy adjustment needs across different dimensions. In this way, the effective content of the original response strategy can be retained, while targeted modifications can be made to certain strategy information based on user feedback, making the second response strategy more aligned with users' personalized interaction habits and improving the flexibility and completeness of strategy adjustments.

[0092] Figure 8 This is another flowchart illustrating an exemplary embodiment of the present disclosure providing a method for switching response strategies.

[0093] like Figure 8As shown above, in the above Figure 3 Based on the illustrated embodiment, the following steps may be included after step S303: S304: Save the strategy version information of the first response strategy.

[0094] In some examples, before modifying the first response strategy, the electronic device can also save the strategy version information of the first response strategy to a pre-created strategy database in storage as historical version information of the response strategy. For example, the electronic device can store the strategy version information of the first response strategy in a structured format, such as JSON or XML, to facilitate subsequent querying and analysis.

[0095] To facilitate subsequent rollback of the first response strategy, before storing the strategy version information of the first response strategy in the strategy database, the electronic device can create a strategy index corresponding to the first response strategy based on the strategy version information of the first response strategy, and store the strategy index in association with the strategy version information of the first response strategy, so that the electronic device can quickly retrieve and locate the first response strategy in the strategy database based on the strategy index.

[0096] The policy database stores the response policies adopted by the large language model at different historical points in time. Each response policy in the policy database corresponds to a policy version information, which may include the policy identifier, policy creation (modification) timestamp, policy effective period, policy documentation, etc., providing a data foundation for subsequent policy rollback.

[0097] In some examples, to save storage space in the policy database, electronic devices can calculate the policy creation timestamp corresponding to each policy version information and calculate the time difference with the current time. This allows them to delete response policies whose time difference is greater than a preset time threshold and which are no longer used for rollback from the policy database, thereby freeing up storage resources.

[0098] In some examples, the time threshold can be set according to the storage requirements of the policy database and the policy rollback cycle, such as 30 days, 90 days, or 180 days. After saving the policy version information of the first response policy to the policy database, the electronic device can prompt the user that the first response policy will be automatically deleted after the policy rollback cycle expires and can no longer be rolled back, so that the user is aware of the rollback time limit of the first response policy.

[0099] In another example, for policy version information marked as important or having special functional significance, electronic devices can set the corresponding reply policy to be permanently retained to avoid the loss of critical policy information due to automatic cleanup mechanisms.

[0100] S305: In response to a rollback instruction based on policy version information input, the response policy of the large language model is rolled back from the second response policy to the first response policy.

[0101] In some examples, when it is subsequently detected that the actual application effect of the second response strategy does not meet the user's expectations, the electronic device can respond to the rollback command, obtain the policy index corresponding to the policy version information of the first response strategy, and then query the policy version information of the first response strategy in the policy database according to the policy index to locate the first response strategy.

[0102] After locating the first response strategy, the electronic device can roll back the response strategy of the large language model from the second response strategy to the first response strategy, so as to regenerate the response content according to the first response strategy. After completing the strategy rollback, the electronic device can store the second response strategy in the strategy database in the same way. The method of storing the second response strategy can refer to the method of storing the first response strategy in the strategy database, which will not be described in detail here.

[0103] In some examples, policy rollback can include manual rollback and automatic rollback. Manual rollback involves the user actively inputting a rollback command to revert the response policy of the large language model to a specified version. Automatic rollback, on the other hand, involves the electronic device automatically triggering the policy rollback operation based on preset rollback conditions. For example, after switching from a first response policy to a second response policy, the electronic device can monitor the system's operating status in real time. If a system fault or anomaly is detected, the electronic device can automatically trigger the rollback mechanism, thereby reverting the second response policy back to the first response policy. This restores the response policy to the first response policy without user intervention, ensuring the stability and reliability of the large language model.

[0104] The method provided in this disclosure can save the policy version information of the first response policy after switching the response policy, and can roll back the response policy to the first response policy in response to the rollback command. This can realize the rollback switching of the large language model, improve the stability and security of the operation of the large language model response policy, and ensure the continuous reliability of human-computer interaction in the intelligent cockpit.

[0105] Figure 9 This is another flowchart illustrating the method for switching response strategies provided in an exemplary embodiment of this disclosure.

[0106] like Figure 9 As shown above, in the above Figure 3 Based on the illustrated embodiment, before step S301, the following steps may also be included: S901: Obtain multimodal data for generating response content.

[0107] In some examples, electronic devices can acquire multimodal data from inside the smart cockpit through acquisition devices. This multimodal data can include at least one of text data, voice data, image data, video data, and sensor data. For example, in a smart cockpit scenario, the acquisition device can include audio input devices such as microphones; image acquisition devices such as in-vehicle cameras and dashcams; and various devices such as touchscreens and in-vehicle sensors, used to collect user voice commands, facial expressions, gestures, and vehicle status information, respectively.

[0108] In some examples, electronic devices can preprocess multimodal data, such as noise filtering and feature extraction, to eliminate differences between different modalities and improve the accuracy of the first response content generated by the subsequent large language model.

[0109] S902: Utilize multimodal data and a first-response strategy based on a large language model to generate the first-response content.

[0110] In some examples, electronic devices can input multimodal data into a large language model, which then analyzes and processes the multimodal data based on a preset first response strategy to generate the corresponding first response content.

[0111] To facilitate the processing of multimodal data by large language models, electronic devices can perform vector transformation on multimodal data, mapping data from different modalities to a unified high-dimensional vector space to form a structured vector representation. For example, text data can be converted into text vectors through word embedding models, speech data can be converted into speaker vectors based on acoustic features, image data can be converted into image feature vectors through visual encoders, video data can be decomposed into temporal frame feature sequences, and sensor data can be converted into numerical state vectors.

[0112] Electronic devices can fuse vectors from different modalities according to a preset weighting relationship to obtain multimodal vectors, which are then input into a large language model. This allows the model to analyze and process the multimodal vectors to generate the corresponding initial response. It's important to note that the vectors from different modalities must maintain a consistent dimensionality to facilitate subsequent multimodal feature fusion. The weighting relationship represents the degree of influence of different modal features on the generated response; features with higher influence receive higher weights, and vice versa.

[0113] The method provided in this disclosure can acquire multimodal data before determining target strategy information, thereby generating a first response content that conforms to the actual interaction scenario and interaction intent of the smart cockpit by combining the multi-dimensional information of the multimodal data. At the same time, it provides an accurate reference basis for subsequent user feedback and strategy adjustment, and improves the coherence and rationality of the overall interaction process.

[0114] Figure 10 This is a flowchart illustrating step S901 provided in an exemplary embodiment of this disclosure.

[0115] like Figure 10 As shown above, in the above Figure 9 Based on the illustrated embodiment, step S901 may include the following steps: S9011: In response to user input instructions, determine the first time point.

[0116] In some examples, when a user inputs instructions, the acquisition device can adopt a corresponding acquisition method based on the user's input method. For example, for user-inputted voice instructions, the acquisition device can acquire them through a microphone; for user-inputted text instructions, it can acquire them through a touchscreen. The electronic device can determine the first time point based on the timestamp of the acquired instruction information after it has been acquired.

[0117] S9012: Acquire the target image acquired by the image acquisition device based on the first time point.

[0118] In some examples, electronic devices can acquire image data of the interior of a smart cockpit in real time via image acquisition devices to observe the environmental conditions and driver status within the cockpit. For instance, the image acquisition device can acquire image data at intervals, such as every 10 seconds.

[0119] In some examples, to maintain the temporal consistency of multimodal data, after acquiring user-input instructions, the electronic device can obtain a target image corresponding to a first time point from the image data acquired by the image acquisition device. For example, the electronic device can acquire image data identical to the first time point as the target image. If no image data identical to the first time point exists, the electronic device can acquire image data with the smallest time difference from the first time point as the target image, or the electronic device can acquire image data within a preset time range before and after the first time point as the target image; this application does not limit this.

[0120] S9013: Based on the instruction information and the target image, acquire multimodal data.

[0121] In some examples, electronic devices can integrate instruction information and target images to form multimodal data containing both textual semantics and visual information. Specifically, electronic devices can extract textual or speech features from instruction information and image features from the target image, aligning the two types of features at a first time point to obtain multimodal data. This multimodal data retains the user's explicit interaction intent while also including the real-time environmental state within the smart cockpit, providing complete information input for subsequent context-aware response generation by a large language model.

[0122] The method provided in this disclosure, by determining the first time point to acquire the target image collected by the image acquisition device and combining it with instruction information to generate multimodal data, can accurately collect real-time image information inside the intelligent cockpit, making the multimodal data more consistent with the current scene state, improving the authenticity and timeliness of the multimodal data, and providing more accurate data support for subsequent response content generation and strategy adjustment.

[0123] Figure 11 This is a flowchart illustrating step S902 provided in an exemplary embodiment of this disclosure.

[0124] like Figure 11 As shown above, in the above Figure 9 Based on the illustrated embodiment, step S902 may include the following steps: S9021: Acquire contextual information collected by multiple sensors.

[0125] In some examples, electronic devices can collect contextual information through multiple sensors. This contextual information includes the smart cockpit's status, environment, navigation, and surrounding facilities information. For example, the smart cockpit's status information may include vehicle interior operating parameters such as air conditioning temperature, window open / closed status, door lock status, and lighting mode; environmental information may include external environmental parameters such as outside temperature, humidity, light intensity, and weather conditions; navigation information may include trip-related data such as the current route, destination location, estimated arrival time, and real-time traffic conditions; and surrounding facilities information may include the distribution of points of interest such as nearby gas stations, parking lots, restaurants, and repair shops.

[0126] For example, multiple sensors may include temperature sensors, humidity sensors, light sensors, GPS positioning modules, radar sensors, vehicle network communication modules, etc., each used to collect different types of situational information.

[0127] S9022: Using multimodal data, determine the first response template corresponding to the multimodal data in the first response strategy.

[0128] In some examples, semantic recognition of multimodal data instructions can be performed using a large language model to determine the user's interaction intent. Based on this intent, a matching response template can be queried in the first response strategy to obtain the first response template corresponding to the multimodal data. For example, if a user inputs the instruction "I'm a little tired," semantic recognition can determine that "tired" is a word related to driving status. Therefore, the user's interaction intent can be determined as "I need to rest." The electronic device can then, based on the "need to rest" interaction intent, query a matching response template in the first response strategy, such as "We have detected that you have driven for XX hours. We suggest you rest at a nearby service area," and mark this response template as the first response template.

[0129] It should be noted that the first response template may include a fixed wording structure and replaceable variables to adapt to personalized expressions in different situations. This application does not impose specific limitations on the template structure of the first response template.

[0130] S9023: In the context information, determine the target information corresponding to the multimodal data.

[0131] In some examples, electronic devices calculate the relevance between contextual information and instruction information corresponding to multimodal data, thereby extracting target information corresponding to the multimodal data from the contextual information based on the relevance. For example, contextual information with a relevance greater than or equal to a relevance threshold is extracted as target information.

[0132] For example, in response to the aforementioned interactive intent of "needing a rest", the electronic device can extract vehicle status information, navigation information and surrounding facility information from contextual information as target information.

[0133] S9024: Embed content parameters corresponding to the target information at a specified position in the first reply template to obtain the second reply template.

[0134] In some examples, the content parameters corresponding to the target information are used to represent the specific numerical value or state description of the target information. For example, continuing the previous example, the target information can be the continuous driving time of the vehicle, the available rest areas near the vehicle, and the distance between the vehicle and the available rest areas, etc., and the corresponding content parameters can be "have been driving continuously for 3 hours", "the nearest service area is 3 kilometers ahead", "expected to arrive in 6 minutes", etc.

[0135] The electronic device can embed the content parameters corresponding to the target information into the first response template at once. For example, "You have been driving continuously for 3 hours" can be embedded into "You have been driving for XX hours" in the first response template, and "The nearest service area is 3 kilometers ahead" and "We expect to arrive in 6 minutes" can be embedded after "We suggest you rest at the nearby service area" in the first response template to obtain the second response template. For example, the content of the second response template can be "We have detected that you have been driving continuously for 3 hours. We suggest you rest at the nearby service area. The nearest service area is 3 kilometers ahead and we expect to arrive in 6 minutes." S9025: Using a large language model, generate the content of the first reply based on the second reply template.

[0136] In some examples, electronic devices can utilize a large language model to convert a second response template into a first response that is linguistically coherent and contextually appropriate, according to a first response strategy. For instance, for the aforementioned second response template, the first response generated by the large language model could be: "It has been detected that you have been driving continuously for 3 hours. We suggest you take a short break at the service area 3 kilometers ahead. You will arrive in about 6 minutes. Please drive safely." In another example, the large language model can also generate corresponding control commands based on the first response. The electronic device can then respond to these control commands and execute the corresponding functional interactions.

[0137] For example, when a user inputs the command "It's too hot" while driving, the electronic device can use a large language model to perform semantic recognition on the voice command to determine the user's interaction intent. For instance, for the voice command "It's too hot," the large language model can use semantic recognition to determine that the user wants to lower the temperature inside the car. While performing semantic recognition on the voice command, the large language model also recognizes the driver's facial expressions and posture information in the image data. For example, if the image data shows the driver sweating and appearing slightly agitated, it can be determined that the driver is in a high-temperature environment. Based on the recognized interaction intent, the large language model can generate a first response message, "The air conditioning has been turned on to 23 degrees Celsius for you," based on a first response strategy.

[0138] After generating the first response message "The air conditioner has been turned on to 23 degrees Celsius for you", the large language model can generate a command to start the air conditioner function, so that the electronic device can start the air conditioner function and adjust the cooling parameters to 23 degrees Celsius.

[0139] In conjunction with the foregoing embodiments, the electronic device can receive feedback instructions input by the user based on the first response content. For example, if the feedback instruction is "I'm cold, please adjust to 26 degrees Celsius," the electronic device can respond to this feedback instruction by determining the "air conditioning control" policy document in the first response strategy, obtaining the target policy information, i.e., the cooling parameters of "air conditioning control," and adjusting the default cooling parameters from 23 degrees Celsius to the user-specified 26 degrees Celsius, thus obtaining the second response strategy. In this way, when the user subsequently inputs the instruction "It's too hot," the large language model can directly generate the response content "The air conditioning has been turned on to 26 degrees Celsius for you" according to the second response strategy, and generate a start instruction to instruct the air conditioning function to cool at 26 degrees Celsius. This ensures that both the functional interaction and the response content conform to the user's feedback intent, improving the efficiency and convenience of human-computer interaction.

[0140] The method provided in this disclosure obtains a second response template by acquiring contextual information, matching a first response template, embedding content parameters, and then generating a first response content by combining multimodal data. This method can closely integrate contextual information with the response template, making the first response content adaptable to the real-time status of the smart cockpit and the user's command information, thereby improving the accuracy and adaptability of the first response content generated by the large language model.

[0141] Exemplary device Figure 12 This is a structural diagram of a response strategy switching device provided in an exemplary embodiment of this disclosure.

[0142] See Figure 12 The response strategy switching device 1200 includes a feedback module 1210, a strategy modification module 1220, and a switching module 1230.

[0143] The response strategy switching device 1200 can be used to implement the corresponding method embodiments of this disclosure.

[0144] In some examples, the feedback module 1210 is used to respond to a user-input feedback command and determine the target strategy information corresponding to the feedback command in the first response strategy, wherein the first response strategy is a pre-set response strategy. The strategy modification module 1220 is used to modify the target strategy information based on the feedback command to obtain a second response strategy. The switching module 1230 is used to switch the response strategy of the large language model from the first response strategy to the second response strategy.

[0145] In some examples, the feedback module 1210 is used to respond to a user-input feedback instruction, perform intent recognition on the instruction content of the feedback instruction, and obtain user intent information of the feedback instruction. Based on the user intent information of the feedback instruction, the target strategy information corresponding to the feedback instruction in the first response strategy is determined.

[0146] In some examples, the feedback module 1210 is used to obtain functional keywords corresponding to the user intent information of the feedback instruction. Among multiple business functions corresponding to the first response strategy, the target function corresponding to the functional keyword is determined. The target strategy information corresponding to the feedback instruction is determined in the strategy document corresponding to the target function.

[0147] In some examples, the policy modification module 1220 is used to obtain evaluative keywords corresponding to the user intent information of the feedback instruction, wherein the evaluative keywords represent the modification requirement of the policy information in the area to be modified. A modification request is generated based on the evaluative keywords. In response to the modification request, the target policy information is modified to obtain a second response policy.

[0148] In some examples, the strategy modification module 1220, in response to a modification request, determines the type of operation to modify the target strategy information based on evaluative keywords. The operation types include add, replace, and delete. In response to an add operation, new information generated based on the evaluative keywords is added to a preset position in the target strategy information, resulting in a second response strategy. In response to a replace operation, the target strategy information is replaced with strategy information generated based on the evaluative keywords, resulting in a second response strategy. In response to a delete operation, at least a portion of the strategy information in the target strategy information is deleted, resulting in a second response strategy.

[0149] In some examples, the response strategy switching device 1200 also includes a rollback module for saving the strategy version information of the first response strategy. In response to a rollback command input based on the strategy version information, the response strategy of the large language model is rolled back from the second response strategy to the first response strategy.

[0150] In some examples, the response strategy switching device 1200 also includes a content generation module for acquiring multimodal data for generating response content. Using the multimodal data, a first response is generated based on a first response strategy using a large language model.

[0151] In some examples, the content generation module is used to determine a first time point in response to user input instructions. The image acquisition device then acquires the target image based on this first time point. Based on the instruction information and the target image, multimodal data is acquired.

[0152] In some examples, the content generation module acquires contextual information from multiple sensors, including the smart cockpit's status, environment, navigation, and surrounding facilities. Using the multimodal data, a first response template corresponding to the multimodal data is determined in the first response strategy. Target information corresponding to the multimodal data is determined from the contextual information. Content parameters corresponding to the target information are embedded at a specified location in the first response template to obtain a second response template. Using a large language model, the first response content is generated based on the second response template.

[0153] The beneficial technical effects corresponding to the exemplary embodiments of this device can be found in the corresponding beneficial technical effects of the exemplary method section above, and will not be repeated here.

[0154] Exemplary electronic devices Figure 13 This is a structural diagram of an electronic device provided by an exemplary embodiment of the present disclosure.

[0155] like Figure 13 As shown, the electronic device 1300 includes at least one processor 1301 and a memory 1302.

[0156] The processor 1301 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 1300 to perform desired functions.

[0157] The memory 1302 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 1301 may execute one or more computer program instructions to implement the switching methods of the response strategies and / or other desired functions of the various embodiments of this disclosure described above.

[0158] In one example, the electronic device 1300 may also include an input device 1303 and an output device 1304, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0159] The input device 1303 may also include, for example, a keyboard, a mouse, etc.

[0160] The output device 1304 can output various information to the outside, including, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices.

[0161] Of course, for the sake of simplicity, Figure 13 Only some of the components of the electronic device 1300 relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device 1300 may include any other suitable components depending on the specific application.

[0162] Exemplary computer program products and computer-readable storage media In addition to the methods and apparatus described above, embodiments of this disclosure may also provide a computer program product, including computer program instructions, which, when executed by a processor, cause the processor to perform steps in the switching methods of the response strategies described in the various embodiments of this disclosure in the "Exemplary Methods" section above.

[0163] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of embodiments of this disclosure. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0164] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform steps in the switching methods of the response strategies of the various embodiments of this disclosure described in the "Exemplary Methods" section above.

[0165] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may include, but is not limited to, systems, apparatuses, or devices that are electrical, magnetic, optical, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0166] The basic principles of this disclosure have been described above with reference to specific embodiments. However, the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0167] Various modifications and variations can be made to this disclosure without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A method for switching response strategies, comprising: In response to a user-inputted feedback command, the target strategy information corresponding to the feedback command in the first response strategy is determined, wherein the first response strategy is a pre-set response strategy; Based on the feedback instruction, the target strategy information is modified to obtain a second response strategy; The response strategy of the large language model is switched from the first response strategy to the second response strategy.

2. The method according to claim 1, wherein, The step of responding to a user-input feedback command and determining the target strategy information corresponding to the feedback command in the first response strategy includes: In response to a user-inputted feedback instruction, intent recognition is performed on the instruction content of the feedback instruction to obtain the user intent information of the feedback instruction; Based on the user intent information of the feedback instruction, the target strategy information corresponding to the feedback instruction in the first response strategy is determined.

3. The method according to claim 2, wherein, The step of determining the target strategy information corresponding to the feedback instruction in the first response strategy based on the user intent information of the feedback instruction includes: Obtain the functional keywords corresponding to the user intent information of the feedback instruction; Among the multiple business functions corresponding to the first response strategy, determine the target function corresponding to the functional keyword; The target policy information corresponding to the feedback instruction is determined in the policy document corresponding to the target function.

4. The method according to claim 2, wherein, Based on the feedback instruction, the target strategy information is modified to obtain a second response strategy, including: Obtain evaluative keywords corresponding to the user intent information of the feedback instruction, wherein the evaluative keywords are used to indicate the need to modify the strategy information in the area to be modified; A modification request is generated based on the evaluative keywords; In response to the modification request, the target policy information is modified to obtain a second response policy.

5. The method according to claim 4, wherein, In response to the modification request, the target policy information is modified to obtain a second response policy, including: In response to the modification request, the operation type for modifying the target strategy information is determined based on evaluative keywords. The operation type includes add, replace, and delete operations. In response to the addition operation, new information generated based on the evaluative keywords is added to the preset position of the target strategy information to obtain the second response strategy; In response to the replacement operation, the target strategy information is replaced with strategy information generated based on the evaluative keywords to obtain the second response strategy; In response to the deletion operation, at least a portion of the policy information in the target policy information is deleted to obtain the second response policy.

6. The method according to claim 1, wherein, After switching the response strategy of the large language model from the first response strategy to the second response strategy, the method further includes: Save the strategy version information of the first response strategy; In response to a rollback command input based on the strategy version information, the response strategy of the large language model is rolled back from the second response strategy to the first response strategy.

7. The method according to claim 1, wherein, Before determining the target strategy information corresponding to the feedback instruction in the first response strategy in response to the user input feedback instruction, the method further includes: Obtain multimodal data for generating response content; Using the multimodal data, a first response is generated based on a first response strategy of a large language model.

8. The method according to claim 7, wherein, The acquisition of multimodal data for generating response content includes: In response to user input commands, determine the first point in time; Acquire the target image captured by the image acquisition device based on the first time point; Based on the instruction information and the target image, the multimodal data is obtained.

9. The method according to claim 7, wherein, The step of generating the first response content using the multimodal data and based on the first response strategy of the large language model includes: The system acquires contextual information collected by multiple sensors, including the status information, environmental information, navigation information, and surrounding facility information of the intelligent cockpit. Using the multimodal data, a first response template corresponding to the multimodal data is determined in the first response strategy; In the context information, target information corresponding to the multimodal data is determined; By embedding content parameters corresponding to the target information at a specified position in the first reply template, a second reply template is obtained; Using the large language model, the first reply content is generated based on the second reply template.

10. A switching device for a response strategy, the device comprising: The feedback module is configured to respond to a user-inputted feedback command and determine the target strategy information corresponding to the feedback command in the first response strategy, wherein the first response strategy is a pre-set response strategy; The strategy modification module is configured to modify the target strategy information based on the feedback instruction to obtain a second response strategy; The switching module is configured to switch the response strategy of the large language model from the first response strategy to the second response strategy.

11. A computer-readable storage medium storing a computer program for performing the switching method of the response strategy according to any one of claims 1-9.

12. An electronic device, comprising: processor; Memory used to store the processor's executable instructions; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the switching method of the response strategy as described in any one of claims 1-9.

Images