Generative large model-based stylized interaction method and system, refrigerator and storage medium

Abstract

The invention provides a stylized interaction method and system based on a generative large model, a refrigerator and a storage medium, and the method comprises the steps: firstly, pre-generating a stylized reply through the generative large model, and in the interaction process, processing an interaction text of a user through intention recognition and slot recognition technologies, and extracting the demand intention and related information of the user; and then, according to the identified intention information, the system retrieves and selects a matched template from a preset standard reply verbal skill template library, and according to specific slot position information and account information of the user, personalized filling and adjustment are carried out on the selected template, and stylized reply is carried out. In order to meet the personalized demand of the user for the reply style, before the user initiates the request, the standard reply verbal skill template and the stylized demand of the user are combined in advance, the stylized reply content is created through the generation type large model, through the pre-generation strategy, the system can provide the stylized reply in real time, and the user experience is improved. And the response efficiency and the user satisfaction are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a style interaction method and system based on a generative large model, a refrigerator and a storage medium. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, intelligent dialogue systems such as chat robots and voice assistants have become an indispensable part of daily life. These systems can provide information queries, life assistants, entertainment interactions and other diversified services, greatly facilitating the lives of users. However, for a long time, artificial intelligence technology has been in a relatively early stage, and most enterprises and organizations have mainly focused on basic technical indicators such as reply accuracy and model generalization ability, paying little attention to style replies.

[0003] In the prior art, multiple "translation" models are trained using different data, each model representing a style and being able to translate a standard reply into a specified style, or a reply mapping table is established, i.e. manually writing variant replies of different styles for each standard reply. The above methods can achieve a certain degree of style reply, but the disadvantages are that multiple models need to be maintained, which is cumbersome, and each model corresponds to one style, which is single-dimensional, has few configurable items, and can only choose one style as a whole, but cannot be configured according to different style points. SUMMARY

[0004] The present application relates to the technical field of computer, in particular to a style interaction method and system based on a generative large model, a refrigerator and a storage medium.

[0005] The present application provides a style interaction method based on a generative large model, which includes a style customization process and an interaction process:

[0006] The style customization process includes the following steps:

[0007] Obtaining user desired style requirements;

[0008] Retrieving a preset standard reply script template, splicing the standard reply script template and the desired style requirements to form prompt word data for inputting a generative large model, and generating a style reply corresponding to the standard reply script template through the generative large model;

[0009] Saving the style reply data;

[0010] The interaction process includes the following steps:

[0011] Obtaining user interaction data;

[0012] Performing intent recognition on the user interaction data to obtain a corresponding standard reply script template;

[0013] Based on saving the stylized reply data, the stylized reply data corresponding to the standard reply dialogue template is used to reply.

[0014] As a further improvement of the present application, after saving the stylized reply data, the method further comprises the steps of:

[0015] Periodically inputting the prompt word data spliced from the standard reply dialogue template and the expected stylization requirement into the generative large model to generate new stylized replies;

[0016] The new stylized reply data is saved by overwriting.

[0017] As a further improvement of the present application, the user interaction data is obtained, specifically comprising:

[0018] Obtaining user interaction voice data,

[0019] After cleaning and converting the user interaction voice data, transcribing it into user interaction text.

[0020] As a further improvement of the present application, the user interaction data is subjected to intent recognition to obtain a corresponding standard reply dialogue template, and the stylized reply data corresponding to the standard reply dialogue template is used to reply, specifically comprising:

[0021] Performing intent recognition and slot recognition on the user interaction text to obtain user interaction text intent information and slot information, and obtaining user current account information;

[0022] According to the user interaction text intent information, a corresponding standard reply dialogue template is obtained;

[0023] According to the slot information and the user current account information, the slot information is supplemented to the standard reply dialogue template;

[0024] The stylized reply data corresponding to the standard reply dialogue template is used to reply.

[0025] As a further improvement of the present application, it further comprises:

[0026] The stylization requirement is audited, and the user is prompted to re-input when the stylization requirement does not meet the regulations;

[0027] The stylized reply data is audited, and the stylized reply data is regenerated when there is no slot loss in the stylized reply data.

[0028] As a further improvement of the present application, it further comprises:

[0029] The standard reply template and the desired stylization requirements are combined to form prompt word data for input into the generative large model. A character limit for the reply content is added, and the generative large model generates a stylized reply that meets the character limit requirement and corresponds to the standard reply template.

[0030] This invention also provides a stylized interaction system based on a generative large model, comprising:

[0031] The stylization customization module is configured to obtain the user's desired stylization requirements; retrieve a preset standard reply script template; concatenate the standard reply script template with the desired stylization requirements to form prompt word data for input into the generative big model; and generate a stylized reply corresponding to the standard reply script template through the generative big model.

[0032] A storage module is configured to store the stylized response data;

[0033] The interaction module is configured to acquire user interaction data, perform intent recognition on the user interaction data, acquire the corresponding standard reply script template, and, based on the saved stylized reply data, use the stylized reply data corresponding to the standard reply script template to make a reply.

[0034] As a further improvement of the present invention, an update module is also included, which is configured to periodically input the standard response script template and the prompt word data spliced ​​with the desired stylization requirements into the generative large model to generate new stylized responses; and to overwrite and save the new stylized response data.

[0035] The present invention also provides a refrigerator, comprising:

[0036] Memory, used to store executable instructions;

[0037] When the processor runs the executable instructions stored in the memory, it implements the above-described stylized interaction method based on a generative large model.

[0038] The present invention also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, implement the above-described stylized interaction method based on a generative large model.

[0039] The beneficial effects of this invention are as follows: First, this invention pre-generates stylized responses using a generative large-scale model. During the interaction, intent recognition and slot recognition technologies are used to process the user's interactive text, extracting the user's needs and related information. Then, based on the identified intent information, the system retrieves and selects a matching template from a pre-set standard response template library. According to the user's specific slot information and account information, the selected template is personalized and adjusted to provide a stylized response. To meet users' personalized needs for response style, this solution pre-combines the standard response template with the user's stylized requirements before the user initiates a request, creating stylized response content through a generative large-scale model. Through this pre-generation strategy, the system can provide stylized responses instantly, greatly improving response efficiency and user satisfaction. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the stylization customization process steps of a stylized interaction method based on a generative large model according to one embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the interaction process steps in a stylized interaction method based on a generative large model according to an embodiment of the present invention.

[0042] Figure 3 This is a block diagram of a stylized interactive system based on a generative large model according to one embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0044] Embodiments of the present invention are described in detail below. Examples of these 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 present invention, and should not be construed as limiting the present invention.

[0045] This implementation proposes a method for stylized rewriting of dialogue responses using a generative large model. This effectively addresses the problem of monotonous and rigid response styles in traditional dialogue systems, allowing users to describe their preferences for response styles using natural language, and generating stylized response templates based on these preferences. Simultaneously, it overcomes the limitations of generative large models in terms of long response times and high costs.

[0046] This implementation provides a stylized interaction method based on a generative large model, which includes a stylized customization process and an interaction process:

[0047] like Figure 1 As shown, the stylization customization process includes the following steps:

[0048] S11: Obtain the user's desired stylization requirements.

[0049] S12: Retrieve the preset standard response script template, combine the standard response script template with the desired stylization requirements to form prompt word data for input into the generative big model, and generate a stylized response corresponding to the standard response script template through the generative big model.

[0050] S13: Save stylized response data.

[0051] like Figure 2 As shown, the interaction process includes the following steps:

[0052] S21: Obtain user interaction data.

[0053] S22: Perform intent recognition on user interaction data and obtain the corresponding standard response template.

[0054] S23: Based on the saved stylized reply data, use the stylized reply data corresponding to the standard reply script template to make a reply.

[0055] For the stylization customization process, in step S11, the user's desired stylization requirements are first obtained, which specifically include:

[0056] The system acquires user interaction voice data, cleans and transforms the data, and then transcribes it into user interaction text.

[0057] Users utilize voice interaction, such as the built-in microphones or external microphones of devices like smart speakers, mobile phones, and computers, to express their desired response style through voice assistants, chat interfaces, or other interactive methods. Users can directly express their preference for response style in natural language or input stylistic requirements directly through apps. For example, a user might say, "I want the response to be more formal" or "I like cute responses."

[0058] The system receives users' stylistic requests through speech recognition technology or text input. If it is speech input, the system converts the speech into text data in real time for subsequent processing.

[0059] The acquired real-time speech data is processed, including noise reduction, echo removal, reverberation removal, and data format conversion and storage. Techniques such as spectral subtraction, Wiener filtering, and deep learning models are used to analyze the spectral characteristics of the speech signal, distinguishing between noise and valid speech, thereby removing or reducing noise components, minimizing background noise interference, and improving speech clarity. Echo components in the speech signal are eliminated using techniques such as adaptive filters and nonlinear processing. The reverberation component is estimated and compensated for by analyzing the time and frequency domain characteristics of the speech signal to reduce or eliminate reverberation effects. These steps effectively improve the quality of the speech data, providing clearer and more accurate input for subsequent speech recognition and semantic understanding.

[0060] After the audio data is processed, the raw audio data needs to be converted to a unified audio format. Simultaneously, sampling rate adjustments and bit depth conversions can be performed to match the needs of subsequent processing. The converted data can be compressed to save storage space and stored in a local or cloud database for quick retrieval and access.

[0061] Speech-to-text technology is used to transcribe cleaned speech data into text data, transforming unstructured speech input into structured text output. This can be achieved through methods such as the STT model. Afterwards, the user's stylistic requirements are analyzed to identify desired style keywords or descriptions, such as "formal," "cute," or "humorous."

[0062] Following step S11, the following is also included:

[0063] Styling requirements are reviewed, and users are prompted to re-enter their input if the stylizing requirements do not meet the requirements.

[0064] The system automatically reviews users' stylistic requests using preset rules and algorithms, including checking for sensitive words, inappropriate content, and information that violates laws and regulations. Automated review can quickly process large amounts of data and provide timely feedback. After the review is complete, if the stylistic request fails, the system will provide feedback to the user explaining the reasons for the failure and prompting the user to re-enter their stylistic request according to the guidelines.

[0065] In step S12, after obtaining the user's desired stylistic requirements, a preset standard response template is retrieved, such as "{num} portions of {food} have been added," "{num} portions of {food} have been deleted," or "No {food} found in the refrigerator." The standard response template is a predefined series of fixed-format response texts containing standardized statements designed to answer common user questions or complete specific tasks, aiming to provide fast, accurate, and consistent responses. The standard response template has a fixed structure and format, containing replaceable variable parts (such as {num}, {food}, etc.), which are replaced by specific information during actual use to adapt to different contexts.

[0066] Next, the user's desired stylistic requirements are combined with the standard response template to form a clear prompt word. For example, if a user requests a "cute" style, and the standard response template is "{num} servings of {food} have been added," the resulting prompt word might be "Please rephrase in a cute way: {num} servings of {food} have been added." This concatenated prompt word data is then input into a generative large-scale model. By learning from a large amount of text data, the model can understand the task requirements based on the provided prompt words and generate text that meets these requirements—that is, a stylized response. The generative large-scale model generates response text that corresponds to the standard response template and also conforms to the user's stylistic requirements based on the input prompt word data.

[0067] For example, "{num} servings of {food} have been added" will generate the message "Cute little master, {num} servings of {food} have been added to the menu~".

[0068] The generated stylized response data needs to be reviewed again to check for any missing slots. If any slots are missing, the stylized response data needs to be regenerated.

[0069] Furthermore, step S12 also includes: adding a character limit to the reply content, and generating a stylized reply that meets the character limit requirement and corresponds to the standard reply script template through a generative large model.

[0070] The character limit for replies is added because generative models typically generate long texts, while voice assistant replies need to be as concise as possible. Users often want to quickly obtain the information they need, and overly long replies may increase the user's burden and negatively impact the experience. By limiting the number of characters in the reply, the generated content is forced to focus more on core information and reduce redundant expressions. For example, the character limit can be set to 30 characters or less, and this can be explicitly stated when generating prompts.

[0071] In practical applications, adjustments can be made based on user feedback regarding the number of characters in the reply to find the optimal balance point that conveys sufficient information without being verbose.

[0072] In step S13, the generated stylized response data is saved. Stylized responses are pre-generated instead of being generated in real-time using a generative large model because generative large models have high computational requirements, especially when processing a large number of requests, which can lead to significant delays. For voice assistants or chatbots that expect instant responses, any delay exceeding a few seconds can negatively impact the user experience. By pre-generating and storing stylized responses, the system can provide a response immediately upon receiving a user request, significantly improving response speed and thus enhancing the user experience. Furthermore, pre-generating stylized responses allows the system to more thoroughly review and optimize the generated content, ensuring the quality and compliance of the responses and avoiding potential errors or inappropriate content that may occur with real-time generation.

[0073] Furthermore, after saving the stylized response data, the following steps are also included:

[0074] Periodically input the standard response template and the desired stylization prompt data into the generative large model to generate new stylized responses; then save the new stylized response data overwrites the existing data.

[0075] Regularly updating stylized response data can maintain the freshness, relevance, and user experience of the responses. By periodically reassembling standard response templates with users' desired stylization needs and inputting them into a generative model to generate new stylized responses, the system can continuously adapt to changes in user needs, updates in socio-cultural trends, and improvements in model performance.

[0076] An appropriate update cycle can be determined based on the frequency of user interaction, the pace of cultural and social events, and the system's processing capacity; such as weekly, monthly, or triggered by specific events. Regularly review and analyze users' stylization needs, including emerging style preferences or changes in existing needs, to ensure that generated stylized responses still align with users' current preferences. Newly approved stylized responses overwrite stored data, updating older response data in the database.

[0077] In the interaction process, user interaction data is obtained. Similar to the process of obtaining user expected stylistic requirements mentioned above, after collecting user voice or text information, the voice data is processed and transcribed into text data.

[0078] Step S22 specifically includes:

[0079] S221: Perform intent recognition and slot recognition on user interaction text, obtain user interaction text intent information and slot information, and obtain the user's current account information.

[0080] S222: Obtain the corresponding standard response template based on the user's interactive text intent information.

[0081] S223: Supplement the standard reply script template with slot information based on slot information and the user's current account information.

[0082] S224: Reply using stylized reply data corresponding to the standard reply script template.

[0083] Through natural language processing (NLP) technology, the system analyzes the main purpose or request of user interaction text, extracting core actions or objectives from user inquiries or instructions. Simultaneously with intent recognition, the system also needs to identify "slot information" in the text, i.e., the specific data required to execute the intent. For example, in "Add 3 apples to the shopping cart," "3" and "apples" are the slot information for quantity and item, respectively. The system obtains the user's current account information, such as the refrigerator's current inventory, to supplement the response information. Based on the identified intent information, the system retrieves the template that best matches the user's request from a pre-set library of standard response templates. Using the slot information extracted from the user's text and relevant user account information, the selected standard response template is filled and modified. Finally, the system selects appropriate stylized response data to reply based on the user's style preferences.

[0084] Furthermore, the final stylized responses are output as voice broadcasts. To maintain consistency, the voice style and emotion match the stylized text responses. If the text response adopts a lighthearted and humorous style, the corresponding voice should also express the same emotional tone. By converting stylized text responses into voice broadcasts, a more natural and convenient interaction method is provided.

[0085] This embodiment also provides a stylized interaction system based on a generative large model, which includes:

[0086] Styling customization module 1 is configured to obtain the user's expected stylization requirements; retrieve a preset standard reply script template; concatenate the standard reply script template and the expected stylization requirements to form prompt word data for input into the generative large model; and generate a stylized reply corresponding to the standard reply script template through the generative large model.

[0087] Storage module 2 is configured to store the stylized response data;

[0088] Interaction module 3 is configured to acquire user interaction data, perform intent recognition on the user interaction data, acquire the corresponding standard reply script template, and, based on the saved stylized reply data, use the stylized reply data corresponding to the standard reply script template to make a reply.

[0089] Furthermore, the stylized interaction system based on the generative large model also includes an update module 4, which is configured to periodically input the standard response script template and the prompt word data spliced ​​with the desired stylization requirements into the generative large model to generate new stylized responses; and to overwrite and save the new stylized response data.

[0090] This embodiment also provides a refrigerator, including:

[0091] Memory, used to store executable instructions;

[0092] When the processor runs the executable instructions stored in the memory, it implements the above-described stylized interaction method based on a generative large model.

[0093] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, implement the above-described stylized interaction method based on a generative large model.

[0094] In summary, this implementation first pre-generates stylized responses using a generative large-scale model. During the interaction, intent recognition and slot recognition technologies process the user's interactive text to accurately extract the user's needs and related information. Then, based on the identified intent information, the system retrieves and selects a matching template from a pre-set standard response template library. According to the user's specific slot information and account information, the selected template is personalized and adjusted to provide a stylized response. To meet users' personalized needs for response style, this solution pre-combines standard response templates with the user's stylized requirements before the user initiates a request, creating stylized response content through a generative large-scale model. This pre-generation strategy allows the system to provide stylized responses instantly, greatly improving response efficiency and user satisfaction.

[0095] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0096] The detailed descriptions listed above are merely specific descriptions of feasible implementations of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent implementations or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A stylized interaction method based on a generative large model, characterized in that, This includes the stylization customization process and the interaction process: The stylistic customization process includes the following steps: Obtain the user's desired stylization requirements; Retrieve a preset standard response script template, and combine the standard response script template with the desired stylization requirement to form prompt word data for input into the generative large model. Generate a stylized response corresponding to the standard response script template through the generative large model. Save the stylized response data; The interaction process includes step: Obtain user interaction data; The user interaction data is subjected to intent recognition to obtain the corresponding standard response template; Based on the saved stylized response data, a response is made using the stylized response data corresponding to the standard response script template.

2. The stylized interaction method based on a generative large model according to claim 1, characterized in that, After saving the stylized response data, the method further includes the following steps: The standard response template and the prompt word data, which are spliced ​​together with the desired stylization requirements, are periodically input into the generative large model to generate new stylized responses; Overwrite and save the new stylized response data.

3. The stylized interaction method based on a generative large model according to claim 1, characterized in that, The acquisition of user interaction data specifically includes: Obtain user interaction voice data. After cleaning and transforming the user interaction voice data, it is transcribed into user interaction text.

4. The stylized interaction method based on a generative large model according to claim 3, characterized in that, The user interaction data is subjected to intent recognition to obtain the corresponding standard response script template; Respond using the stylized response data corresponding to the standard response template, specifically including: The user interaction text is subjected to intent recognition and slot recognition to obtain user interaction text intent information and slot information, and the user's current account information is obtained. Obtain the corresponding standard response template based on the user interaction text intent information; The slot information is supplemented for the standard reply script template based on the slot information and the user's current account information; Reply using the stylized reply data corresponding to the standard reply template.

5. The stylized interaction method based on a generative large model according to claim 1, characterized in that, Also includes: The stylization requirements are reviewed, and if the stylization requirements do not meet the requirements, the user is prompted to re-enter the information. The stylized response data is reviewed, and if there are no missing slots in the stylized response data, the stylized response data is regenerated.

6. The stylized interaction method based on generative large models according to claim 1, characterized in that, Also includes: The standard reply template and the desired stylization requirements are concatenated to form prompt word data for input into the generative large model. A character limit for the reply content is added, and the generative large model generates a stylized reply that meets the character limit requirement and corresponds to the standard reply template.

7. A stylized interaction system based on a generative large model, characterized in that, include: The stylization customization module is configured to obtain the user's desired stylization requirements; Retrieve a preset standard response script template, and combine the standard response script template with the desired stylization requirement to form prompt word data for input into the generative large model. Generate a stylized response corresponding to the standard response script template through the generative large model. A storage module is configured to store the stylized response data; The interaction module is configured to acquire user interaction data, perform intent recognition on the user interaction data, acquire the corresponding standard reply script template, and, based on the saved stylized reply data, use the stylized reply data corresponding to the standard reply script template to make a reply.

8. The stylized interaction system based on a generative large model according to claim 7, characterized in that, It also includes an update module, which is configured to periodically input the standard response script template and the prompt word data spliced ​​with the desired stylization requirements into the generative large model to generate new stylized responses; and to overwrite and save the new stylized response data.

9. A refrigerator, comprising: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the stylized interaction method based on a generative large model as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing executable instructions that, when executed by a processor, implement the stylized interaction method based on a generative large model as described in any one of claims 1 to 6.

Images