Personalized interactive generation methods, devices and media

Abstract

This application provides a personalized interaction generation method, device, and medium, belonging to the field of intelligent interaction technology. The method includes: acquiring a child's basic profile, toy character parameters, and hierarchical memory information; generating an input comprehension result corresponding to the child's current input; retrieving generation control information based on the input comprehension result and the current interaction scenario; this generation control information includes the child's basic profile, toy character parameters, hierarchical memory information, and / or a child's profile; determining whether the current interaction scenario triggers a special interaction mode based on participation indicators; if so, entering the template skeleton constraint generation path corresponding to the special interaction mode; in the template skeleton constraint generation path, using the generation control information to perform dynamic slot filling operations to form a structured input package; inputting the structured input package into a large language model to generate and output personalized interaction content. This application can solve the problem of insufficient personalized interaction due to the reliance on free generation from a general large model in existing technologies.

Description

Technical Field

[0001] This application belongs to the field of intelligent interaction technology, specifically relating to a personalized interaction generation method, device, and medium. Background Technology

[0002] In the field of intelligent interaction technology, because AI products for children's companionship (such as children's toys, children's machines, or companion assistants) can meet children's personalized companionship and interaction needs, AI products for children's companionship are gradually becoming the development trend in the direction of children's companionship and intelligent interaction.

[0003] While existing AI-powered companionship products for children can answer their questions, they struggle to engage children in sustained interaction, resulting in low willingness to express themselves, short conversation rounds, and a lack of long-term companionship. This is primarily because the dialogue in these products largely relies on a general-purpose language model for free generation. Even solutions with some prompts or rule base configurations only offer minor personalization adjustments to the dialogue content through simple prompts or keyword matching, lacking sufficient interactive engagement. The excessive freedom of interaction through the language model leads to a lack of sustained engagement, and the shallow personalization hinders the formation of a long-term sense of companionship. Engaging interactions depend on the language model's spontaneous generation, making character styles prone to shifting and failing to balance age-appropriateness, fun, and stability.

[0004] In other words, existing personalized interactions for children based on large language models lack structured control links and constraint generation mechanisms, making it difficult for existing AI products for children to stably generate age-appropriate, controllable, and engaging interactive content. Consequently, the retention rate, subscription conversion rate, and repurchase rate of these AI products for children are not high. Summary of the Invention

[0005] This application aims to provide a personalized interaction generation method, device, and medium that can solve the problems of low retention rate, subscription conversion rate, and repurchase rate of existing AI products for children's companionship, which mostly rely on general large language models for free generation, lack personalized interaction, lack the ability to continuously attract children to continue interacting, and are difficult to generate age-appropriate, controllable, and interesting interactive content.

[0006] According to a first aspect of this application, this application provides a method for generating personalized interactions, including: Obtain basic child profiles, toy character design parameters, and layered memory information; Obtain the child's current input, identify and judge the child's current input, and obtain the input comprehension result; Based on the input understanding results and combined with the current interaction scenario, the generated control information is retrieved; the generated control information includes one or more of the following: child basic profile, toy character design parameters, layered memory information, and child portrait. Based on the child's engagement metrics, determine whether the current interactive scenario triggers a preset special interaction mode; If a special interaction mode is triggered, the template skeleton constraint generation path corresponding to the special interaction mode will be entered. In the template skeleton constraint generation path, select the template skeleton corresponding to the special interaction mode, and use the generation control information to perform dynamic slot filling operation on the template skeleton to form a structured input package. The structured input package is fed into the large language model, which then controls the large language model to generate and output personalized interactive content under the constraints of the structured input package.

[0007] Preferably, the above-mentioned personalized interaction generation method, after generating and outputting personalized interactive content, further includes: Based on the current interaction result corresponding to the current interaction scenario, determine whether the current interaction result contains reusable information; among which, reusable information is used to refill the memory of hierarchical memory information; If the information contains reusable information or meets the preset memory retention conditions, the results of this round of interaction are extracted in a structured manner to obtain event cards or structured summaries. Based on the nature of the information, event cards or structured summaries are hierarchically populated into the corresponding permanent memory, short-term memory, and long-term memory in the hierarchical memory information to update the hierarchical memory information; Based on the updated hierarchical memory information, the parameters of the child's portrait or toy character design are updated conditionally or periodically.

[0008] Preferably, the above-mentioned personalized interaction generation method further includes: privacy protection processing of children's data; privacy protection processing of children's data includes memory backfilling of hierarchical memory information or generation and calling of large models: Sensitive information about children in hierarchical memory information is desensitized or labeled; the labeling process includes generating anonymized labels. When inputting structured input packages into a large language model, anonymized labels, structured summaries, and necessary memory fragments are used to replace the original child-sensitive information as model input. Access control, encrypted storage, and encrypted transmission are implemented for toy character design parameters, layered memory information, child profiles, and current interactive content. Access control includes guardian permissions for managing child profiles and layered memory information, as well as disabling memory functions.

[0009] Preferably, the above-mentioned personalized interaction generation method, after determining whether the current interaction scenario triggers a preset special interaction mode based on the child's participation index, further includes: If no special interaction mode is triggered, proceed to the lightweight generation path; In the lightweight generation path, safety constraint information is extracted from the child's basic profile; The input understanding results, child profiles, toy character design parameters, and safety constraint information are assembled into a lightweight prompt input package; The lightweight prompt input package is fed into the large language model, which then controls the large language model to generate and output personalized interactive content under the constraints of the lightweight prompt input package.

[0010] Preferably, in the above-mentioned personalized interaction generation method, in the template skeleton constraint generation path, a template skeleton corresponding to a special interaction mode is selected, and the template skeleton is dynamically filled with slots using generation control information to form a structured input package, including: Based on the specific interaction mode and combined with the generation control information, a template skeleton is selected; the template skeleton is used to constrain the content structure generated by the large language model. Based on the generated control information, personalized fill content is constructed and filled into the corresponding slots of the template skeleton to obtain the slot content; Based on specific interaction modes, the template skeleton and slot content are combined to form a structured input package.

[0011] Preferably, the above-mentioned personalized interaction generation method obtains the child's basic profile, toy character parameters, and layered memory information, including: When interacting with a child for the first time or configuring guardian permissions, build a basic profile of the child and initial toy persona parameters; Based on the child's basic profile, an initial version of the child's portrait is generated; Configure tiered memory information based on the child's basic profile or the guardian's permissions; the tiered memory information includes permanent memory, short-term memory, and long-term memory. After interacting with children, historical interaction information is obtained. Based on the stability and timeliness of the information, the historical interaction information is classified and refilled into permanent memory, short-term memory and long-term memory to update the hierarchical memory information. When the update conditions are met or when periodic updates are required, the child's profile and / or evolutionary toy character parameters are updated based on the child's basic profile and hierarchical memory information.

[0012] Preferably, the above-mentioned personalized interaction generation method updates the child's profile and / or the evolutionary toy character design parameters, including: Update the child's profile based on the child's basic records, short-term memory, and long-term memory; Based on children's interaction preferences, determine the offset of toy character design parameters in multiple preset candidate evolution directions; The toy character parameters are weighted and adjusted based on the offset to ensure that the toy character parameters evolve in a controlled manner in multiple candidate evolution directions.

[0013] Preferably, the above-mentioned personalized interaction generation method determines whether the current interaction scenario triggers a preset special interaction mode based on the child's participation index, including: Special interaction modes include one or more of the following: awkward silence rescue mode, interest enhancement mode, story companionship mode, game interaction mode, and emotional soothing mode; Based on the participation index, when low participation occurs in multiple consecutive rounds of dialogue or the current round number decreases, the "awkward silence rescue mode" is triggered. Based on engagement metrics, when specific interest content in short-term memory or stable preferences in long-term memory appear frequently recently, an interest enhancement mode in a special interaction pattern is triggered. Based on the engagement metrics, when the current child input contains a pattern-triggered instruction, select the special interaction mode corresponding to the pattern-triggered instruction.

[0014] According to a second aspect of this application, this application also provides a personalized interaction generation system for the personalized interaction method provided by any of the above technical solutions; the personalized interaction generation system includes: The parameter acquisition module is used to acquire children's basic profiles, toy character design parameters, and layered memory information; The input acquisition and recognition module is used to acquire the child's current input, recognize and judge the child's current input, and obtain the input comprehension result; The information retrieval module is used to retrieve and generate control information based on the input understanding results and the current interaction scenario; the generated control information includes one or more of the following: child basic profile, toy character design parameters, layered memory information, and child portrait. The mode triggering module is used to determine whether the current interaction scenario triggers a preset special interaction mode based on the child's participation index; if a special interaction mode is triggered, the module will enter the template skeleton constraint generation path corresponding to the special interaction mode. The dynamic slot filling module is used to select the template skeleton corresponding to a special interaction mode in the template skeleton constraint generation path, and use the generation control information to perform dynamic slot filling operation on the template skeleton to form a structured input package. The constraint generation module is used to input structured input packages into the large language model and control the large language model to generate and output personalized interactive content under the constraints of the structured input packages.

[0015] According to a third aspect of this application, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a personalized interactive generation method of any of the above technical solutions.

[0016] According to a fourth aspect of this application, this application also provides a computer storage medium storing a computer program thereon, which, when executed, implements the personalized interactive generation method provided by any of the above technical solutions.

[0017] The personalized interactive generation scheme provided in this application forms a holistic, collaborative closed loop from the child's basic profile, hierarchical memory information, toy persona parameters, to template skeleton triggering, and finally to the generation of personalized interactive content. Furthermore, this application selectively triggers template skeleton constraint generation based on child participation indicators, selectively entering the corresponding generation path. This application features generation control based on hierarchical memory and child profiles, not just using the current input but combining historical information and child profiles for generation control. Additionally, structured parameters based on toy personas, which control interaction states and role consistency, participate in the generation of personalized interactive content, maintaining role consistency and controllability of interaction states. This application's technical solution, from the child's basic profile, toy persona parameters, hierarchical memory information, and child profile, to pattern triggering and structured input packages into the large language model, implements a constrained generation mechanism for the generation and output of personalized interactive content from the large language model. Compared to the direct generation method of general large models, this application's technical solution introduces child profiles, hierarchical memory, toy persona parameters, and pattern judgment before interactive content generation, ensuring more controllable output. Compared to existing personalized processing methods that rely on simple prompts, this application's technical solution can generate generation control information based on the understanding of the child's current input and the current interaction scenario. This information includes one or more of the following: the child's basic profile, toy character parameters, hierarchical memory information, and the child's portrait. This allows for the construction of a structured input package containing personalized content, forming a long-term, updatable, personalized interaction loop for children. Compared to simply stacking historical dialogues, this application improves the proportion of effective information and reduces contextual noise through hierarchical memory. This application maintains character consistency by setting toy character parameters and evolving these parameters based on interaction feedback. Furthermore, compared to existing random fun generation solutions, this application's technical solution enhances product stability and age-appropriateness in specific scenarios through template skeletons and dynamic slot filling.

[0018] In summary, this application introduces basic child profiles, toy character parameters, hierarchical memory information, and basic child portraits, further combining these with the child's current input and engagement metrics to form a structured input package. This package controls the generation of manageable, personalized interactive content by the large language model. Through this structured control chain and the specific constraint generation mechanism based on special interaction patterns and generation control information to form a structured input package, the child companion AI product can stably generate age-appropriate, controllable, and engaging interactive content with a high degree of personalization. This attracts children to continue interacting, fostering a long-term sense of companionship, and ultimately improving the retention rate, subscription conversion rate, and repurchase rate of the child companion AI product. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a personalized interaction generation method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a personalized interactive generation system provided in an embodiment of this application; Figure 3 For the corresponding Figure 2 A flowchart illustrating a personalized interaction generation method according to an embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0022] In this application, unless otherwise expressly specified and limited, the descriptions using terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0023] The existing technology has the following technical problems: Existing AI products for child companionship typically generate responses directly using a general-purpose language model, or only make minor personalized adjustments based on simple prompts. This interaction method, lacking structured control links and constraint generation mechanisms, leads to the following problems: 1. While the responses can provide simple answers to questions, they lack the ability to sustain children's continued interaction.

[0024] 2. The level of personalization is relatively shallow, usually only reaching superficial information such as nicknames and basic preferences, and cannot form a long-term sense of companionship.

[0025] 3. Historical dialogues are often piled up in the context in their entirety, resulting in a low proportion of effective information and unstable generation results.

[0026] 4. Engaging interactions often rely on models to be generated freely in real time, making it easy for styles to drift and difficult to balance age-appropriateness, fun, and stability.

[0027] 5. Toy characters are usually fixed and lack the companionship features that gradually evolve as children use them.

[0028] To address the aforementioned technical issues, the following embodiments of this application provide a personalized interaction generation scheme. Based on a child's basic profile, event memories, interests, interaction status, and toy persona, the scheme generates a target skeleton constraint based on the child's participation index. By generating a structured input package, it constrains a large language model to stably generate age-appropriate, controllable, engaging, and sustainably evolving interactive content, thereby enhancing children's willingness to actively express themselves and the number of continuous dialogue rounds.

[0029] To achieve the above objectives, see [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating a personalized interaction generation method provided in an embodiment of this application. Figure 1 As shown, this personalized interaction generation method includes: S110: Obtain basic child profiles, toy character design parameters, and layered memory information; A child's basic profile refers to a set of basic information formed when a child first uses the app or when it is configured by a guardian. This information includes age group, nickname, interests, taboos and preferences, favorite IPs / characters, toy naming relationships, etc.

[0030] Toy character design parameters refer to a set of structured parameters used to control the interaction style and character consistency of children's companion toys, including character design direction weights, tone tags, commonly used word tags, preferred play styles, and prohibited styles.

[0031] Layered memory information refers to a memory management method that divides children's interactive information into permanent memory, short-term memory, and long-term memory according to their stability, timeliness, and future reusability.

[0032] When interacting with a child for the first time or configuring guardian permissions, it is necessary to build a basic profile of the child and initial toy character parameters, and then generate an initial version of the child's portrait based on the basic profile. Layered memory information includes permanent memory, short-term memory, and long-term memory; permanent memory can be obtained from the child's basic profile or configured by the guardian's permissions, while short-term memory and long-term memory are categorized and filled in after interacting with the child using historical interaction information, thereby continuously improving the layered memory information.

[0033] Specifically, as a preferred embodiment, in this personalized interaction generation method, S110: obtaining the child's basic profile, toy character design parameters, and hierarchical memory information, including: S111: When interacting with a child for the first time or configuring guardian permissions, build a basic profile of the child and initial toy persona parameters.

[0034] A child's basic profile includes, but is not limited to, the following: the child's nickname or user identifier; age or age group; toy name (i.e., the name the child gives to the toy); initial interest tags, favorite IPs, characters or themes; taboo preferences or sensitive content, and age-appropriate content restrictions, etc.

[0035] The initial character design of a toy may include: character orientation (e.g., adventurous partner, gentle companion, knowledge explorer, lively and mischievous); tone tags (e.g., encouraging, curious, slightly naughty); preferred play style (e.g., story adventure, quiz challenge, role-playing); prohibited styles (e.g., frightening, adult-oriented, overly didactic); and commonly used word tags, etc.

[0036] S112: Generate an initial version of the child's portrait based on the child's basic profile.

[0037] A child profile refers to structured user identification information based on a child's basic profile, hierarchical memory information, and historical interaction feedback. This includes interests, interaction preferences, activity levels, emotional trends, and taboo preferences. During the initial interaction with a child, since historical interaction feedback is not yet available and hierarchical memory information is limited, an initial version of the child profile can be generated based on the child's basic profile (e.g., age or age group, initial interest tags, and taboo preferences or sensitive content). This child profile is then continuously updated as the child's interactions become more frequent.

[0038] S113: Configure hierarchical memory information based on the child's basic profile or the guardian's permissions.

[0039] In the technical solution provided in this application, the hierarchical information includes permanent memory, short-term memory, and long-term memory. Short-term memory primarily stores recent events, recent states, and short-term valid information; for example, recently occurred events, recently mentioned people or places, and the most recent emotional state. Long-term memory primarily stores information that has been stably maintained after multiple rounds of verification and can continuously improve the quality of future interactions; for example, consistently stable interests, long-term relationships, and long-term preferences for certain gameplay modes. Permanent memory stores long-term stable information such as identity and naming relationships, which can be accessed in each round.

[0040] When initializing a child companion AI product, since there has been no interaction with the child and no historical interaction information with the child has been obtained, it is necessary to first configure layered memory information (mainly permanent memory) based on the child's basic profile or the guardian's permissions. For example, the child's basic profile is a stable source of information in permanent memory; the configured permanent memory includes: the child's basic profile, stable identity information, long-term taboo information, stable relationship information, and other long-term valid facts.

[0041] S114: After interacting with children, obtain historical interaction information, and classify and refill the historical interaction information into permanent memory, short-term memory and long-term memory according to the stability and timeliness of the information, so as to update the hierarchical memory information.

[0042] After obtaining historical interaction information through interaction with children, the historical interaction information can be categorized and backfilled into permanent information, short-term memory, and long-term memory based on the stability and timeliness of the information.

[0043] For example, if a child mentions in a recent conversation that they "drew a dinosaur at kindergarten today," or if the child's current emotional state is "excited," this information is stored in short-term memory and can automatically decay or shift after a certain period (e.g., after several rounds of conversation or a preset time window). After obtaining such events, they can be structured and extracted, and then added back to short-term memory. Additionally, if a child has repeatedly expressed interest in "Tyrannosaurus Rex" over the past month, and each time they mention it with positive emotions, this information can be transferred from short-term memory to long-term memory.

[0044] S115: When the update conditions are met or periodic updates are performed, update the child's profile and / or the character design parameters of the evolving toy based on the child's basic profile and hierarchical memory information.

[0045] Regarding updates to children's portraits, portrait tags can be extracted based on the child's basic profile, layered memory information, and historical interaction feedback. The weights of these tags can then be determined, and the children's portraits can be updated through a structured combination of these tags and their weights. As for toy character design parameters... "Toy character parameters" are reflected in the tone, frequently used words, interaction preferences, and game type preferences of the content subsequently generated by AI products for children's companionship. They are part of the personalized interaction generation control for children and are mainly used to control the output style of the large language model during the generation process. Toy character parameters are preferably stored in a structured form, such as tone style tags; frequently used word tags or word sets; preferred game type tags, etc. Therefore, storing toy character parameters essentially represents the toy character's state as structured tags + weights / probabilities.

[0046] The technical solution provided in this application, by configuring a child's basic profile, toy character parameters, and hierarchical memory information, and generating an initial version of the child's portrait based on the child's basic profile, can perform structured control and personalized constraints on the output of the large language model based on the aforementioned child's basic profile, toy character parameters, hierarchical memory information, and child portrait. This allows the content generated by the large language model to continuously attract children's interaction and possess a high degree of personalization, while also considering age-appropriateness, fun, and stability. Furthermore, by continuously enriching the hierarchical memory information and updating the toy character parameters through historical interaction information with the child after multiple subsequent interactions, the output and interaction of the child companion AI product are continuously optimized.

[0047] The personalized interaction generation method of this application does not use fixed parameters for the child's profile and toy character design; instead, they dynamically evolve as the interaction progresses. Therefore, as a preferred embodiment, in the above-mentioned personalized interaction generation method, S115: updating the child's profile and / or evolving the toy character design parameters includes: S1151: Update the child profile based on the child's basic profile, short-term memory, and long-term memory.

[0048] The updating of child profiles is based on the child's basic profile and evolves from short-term and long-term memories. Specifically, based on recent interests and emotions in short-term memory, stable preferences in long-term memory, and initial information in the child's basic profile, dimensions such as interest preferences, interaction style preferences, activity level, and emotional trends in the child profile are updated.

[0049] S1152: Based on the child's interaction preferences, determine the offset of the toy character design parameters in multiple preset candidate evolution directions; these multiple preset candidate evolution directions can be obtained from a finite set of evolution directions.

[0050] S1153: Adjust the weights of the toy character parameters based on the offset to make the toy character parameters evolve in a controlled manner in multiple candidate evolution directions.

[0051] In this embodiment, the toy character design parameters evolve based on the child's interaction preferences. Specifically, based on the interaction preferences between the child-accompaniment AI product and the child, the offset of the toy character design parameters in multiple preset candidate evolution directions is determined. Then, the weights of the toy character design parameters are adjusted according to these offsets to ensure controlled evolution in multiple candidate evolution directions. For example, if a child consistently prefers an "adventure-type" interaction style, the weight of the "adventure-type companion" direction in the toy character design parameters can be gradually increased, while the weights of directions such as "gentle companion" can be correspondingly decreased. It should be noted that the weight adjustments are limited to a preset weight range to prevent drastic changes in the toy character design parameters. This controlled evolution mechanism allows the toy character to gradually adapt to the child's personalized preferences while maintaining basic consistency, thereby enhancing the long-term sense of companionship.

[0052] The technical solution provided in this application, compared to existing simple personalized prompts, can continuously enrich layered memory information through interaction with children. This layered memory information is used to continuously update the child's profile and the toy's character parameters, thereby forming a long-term, updatable, personalized interactive loop for children. Compared to the fixed role settings of existing companion AI products, the method provided in this application, through toy character parameters, can maintain role consistency and can be updated based on historical feedback to adapt to the child's continuous growth and generate age-appropriate and stable interactive content.

[0053] Figure 1 The personalized interaction generation method provided in the illustrated embodiment, after step S110: obtaining the child's basic profile, toy character design parameters, and hierarchical memory information, further includes: S120: Obtain the child's current input, identify and judge the child's current input, and obtain the input understanding result.

[0054] When a child activates the AI ​​companion product and inputs voice, the product automatically converts the speech into text using the Automatic Speech Recognition (ASR) module and understands the input to obtain an input comprehension result. This comprehension result includes, but is not limited to: recognizing the child's intent (e.g., asking questions, engaging in small talk, requesting games, asking for stories, and expressing emotions); recognizing emotions or levels of engagement (e.g., positive, negative, hesitant, silent, and low engagement); determining whether the input might contain memorable information; and determining whether it involves sensitive information or inappropriate content.

[0055] For example, when a child inputs the voice message "Tell me a dinosaur story," the child companion AI product converts it into text using ASR (Automatic Speech Recognition). It then identifies the child's intention as "requesting a story," their emotion as "positive," and their level of engagement as "high engagement." The product determines that the input contains memorable information (the interest tag "dinosaur") and does not involve sensitive information or inappropriate content.

[0056] By recognizing and judging the child's current input, the system obtains input comprehension results. Based on these results, it can select and generate control information from the child's basic profile, toy character parameters, hierarchical memory information, and child profile. It also identifies the child's level of participation and determines the interaction mode that should be triggered in the current interaction scenario based on the child's participation index. Finally, it constrains the generated content of the large language model based on the current interaction mode.

[0057] Figure 1 The personalized interaction generation method provided in the illustrated embodiment, after step S120: obtaining the child's current input, recognizing and judging the child's current input, and obtaining the input understanding result, further includes: S130: Based on the input understanding results and combined with the current interaction scenario, retrieve the generation control information; the generation control information includes one or more of the following: the child's basic profile, toy character parameters, hierarchical memory information, and the child's portrait. Additionally, the generation control information may also include taboo preferences and safety constraints to constrain the large language model from generating content that is taboo for children and to improve the safety of dialogue with children.

[0058] The technical solution provided in this application retrieves the necessary information for generation from the child's basic profile, hierarchical memory information (including permanent memory, short-term memory, and long-term memory), toy character design parameters, and child portrait based on the input understanding results and the current interaction scenario. This information serves as the generation control information, which constrains and controls the generated content of the large language model, reducing the problem of character style drift caused by the large language model's improvisation and the difficulty in balancing age-appropriateness, fun, and stability.

[0059] In this embodiment, it is not necessary to retrieve all the acquired information in every round of dialogue. Instead, necessary information is selected based on the input comprehension result of the current input and the current interaction scenario to reduce interference from irrelevant information. For example, when a child asks a factual question, the child companion AI product may only need to retrieve the child's age range (for age-appropriate control) and toy character parameters (for style control); while when a child expresses emotions, the child companion AI product may need to retrieve recent emotional states from short-term memory and stable preferences from long-term memory.

[0060] Figure 1 The personalized interaction generation method provided in the illustrated embodiment, after step S130: retrieving generation control information based on the input understanding result and the current interaction scenario, further includes: S140: Based on the child's participation index, determine whether the current interactive scenario triggers a preset special interactive mode.

[0061] The technical solution provided in this application requires the selection of interaction modes based on children's participation indicators, including special interaction modes corresponding to special interaction scenarios and regular dialogue modes corresponding to ordinary interaction scenarios. The constraint generation paths for the two interaction modes are different; for special interaction modes, stronger structured generation control is required to constrain the large language model to generate more interesting, age-appropriate, and controllable interactive content.

[0062] Specifically, as a preferred embodiment, this personalized interaction generation method, after determining whether the current interaction scenario triggers a preset special interaction mode based on the child's participation index, further includes: if no special interaction mode is triggered, entering a lightweight generation path; in the lightweight generation path, extracting safety constraint information from the child's basic profile; assembling the input understanding result, child profile, toy character parameters, and safety constraint information into a lightweight prompt input package; inputting the lightweight prompt input package into a large language model, and controlling the large language model to generate and output personalized interactive content under the constraints of the lightweight prompt input package.

[0063] If the technical solution provided in this application does not trigger a special interaction mode, it automatically enters the lightweight generation path. In this lightweight generation path, this application extracts safety constraint information (such as age-appropriate content constraints and taboo preferences) from the child's basic profile, assembles the input comprehension results, necessary information of the child's profile, toy character parameters, and safety constraint information into a lightweight prompt input package, and then inputs it into the large language model to generate a response. The lightweight generation path is suitable for ordinary interaction scenarios such as simple factual questions and answers, general confirmation, or casual conversation without strong guidance.

[0064] Specifically, as a preferred embodiment, the personalized interaction generation method described above, S140: determining whether the current interaction scenario triggers a preset special interaction mode based on the child's participation index, including: S141: Special interaction modes include one or more of the following: awkward silence rescue mode, interest enhancement mode, story companion mode, game interaction mode, and emotional soothing mode.

[0065] Special interaction modes are often triggered when children’s participation decreases or when they need interaction to soothe them. By triggering special interaction modes, the large language model can be controlled in a structured manner to generate age-appropriate, controllable and interesting interactive content, thereby improving children’s personalized interaction ability and frequency.

[0066] S142: Based on the participation index, when low participation occurs in multiple consecutive rounds of dialogue or the current round number decreases, the "awkward silence rescue mode" is triggered.

[0067] S143: Based on the participation index, when a specific interest content in short-term memory or a stable preference exists in long-term memory that appears frequently recently, the interest enhancement mode in the special interaction mode is triggered.

[0068] S144: Based on the engagement metrics, when the current child input contains a pattern-triggered instruction, select the special interaction mode corresponding to the pattern-triggered instruction.

[0069] Engagement metrics include, but are not limited to, whether low-engagement expressions (such as "I don't know," "I'm bored," or "whatever") appear in consecutive rounds of dialogue, whether the current number of dialogue rounds is decreasing, whether specific interest content appears frequently in short-term memory, whether stable preferences exist in long-term memory, and whether the current input from the child contains clear pattern-triggered instructions.

[0070] For example, when a child exhibits low participation in multiple consecutive rounds of conversation (e.g., answering "I don't know" or "whatever" for three consecutive rounds) or when the current number of conversation rounds is significantly lower than the recent average, an awkward silence rescue mode is automatically triggered. In this mode, the child-accompaniment AI product can proactively propose a topic or game invitation related to the child's interests to re-engage the child's participation.

[0071] When a particular interest appears frequently in short-term memory recently (e.g., the child mentions "dinosaurs" three times in the past five rounds of conversation), or when a stable preference exists in long-term memory (e.g., the "dinosaur" interest tag consistently has a high weight in the child's profile), the interest enhancement mode is automatically triggered. In interest enhancement mode, the child companion AI product can engage in in-depth conversations or generate related stories around the topic of interest.

[0072] When a child inputs a mode-triggered instruction that includes a special interaction mode (such as "Tell me a story" or "Let's play a game"), the special interaction mode corresponding to that instruction is automatically selected (such as story companion mode or game interaction mode).

[0073] The technical solution provided in this application uses the child's participation index to determine when the current interactive scenario triggers a preset special interactive mode, and then triggers the corresponding special interactive mode. This allows the child companion AI product to output more interesting and targeted interactive content based on the child's current state, thereby enhancing personalized interaction with children and improving the retention rate, subscription conversion rate, and repurchase rate of the child companion AI product.

[0074] Figure 1 The technical solution provided in the illustrated embodiment, after S140: determining whether the current interactive scenario triggers a preset special interactive mode based on the child's participation index, further includes: S150: If a special interaction mode is triggered, the template skeleton constraint generation path corresponding to the special interaction mode will be entered.

[0075] In the technical solution provided in this application embodiment, the corresponding generation path is selected based on whether a special interaction mode is triggered. This generation path corresponds to the current interaction scenario. For example, in a normal scenario, a lightweight generation path is triggered, while in a special interaction scenario, the feature triangle lake mode is entered, and the corresponding template skeleton constraint generation path is selected.

[0076] If a special interaction mode is triggered, the template skeleton constraint generation path will be automatically entered. The template skeleton constraint generation path is an enhanced control layer for special scenarios, used for scenarios that require enhanced companionship and controllability, such as rescuing from awkward silences, enhancing interest, story games, emotional companionship, and proactive requests. By using the target estimation constraint generation path, the generated content of the large language model can be constrained in a more stable and structured way, enabling the AI ​​for child companionship to stably generate age-appropriate, controllable, and engaging interactive content.

[0077] S160: In the template skeleton constraint generation path, select the template skeleton corresponding to the special interaction mode, and use the generation control information to perform dynamic slot filling operation on the template skeleton to form a structured input package.

[0078] In specific interaction scenarios, embodiments of this application can select a specific template skeleton based on the current interaction mode, age group, child profile, short-term memory, long-term memory, toy character parameters, and prohibited content safety constraints, and dynamically fill slots using the aforementioned generation control information, thereby forming a structured input package. The structured input package may include: age group, current interaction mode, recent available memory, long-term preference tags, child profile tags, toy character parameters, template skeleton structure, slot filling content, prohibited content, and safety constraints. A large language model generates natural language response text based on the structured input package. The generated text undergoes content safety and age-appropriateness checks before speech synthesis.

[0079] The structured input package not only includes the content to be generated, but it is also obtained through a structured template skeleton and slot filling. Therefore, the content of the structured input package is structured, which also allows for controllable and stable personalized interactive content generated by the large language model. In addition, by introducing personalized generation information, this structured input package can make the interactive content generated by children's companion AI products more interesting.

[0080] Specifically, in a preferred embodiment, step S160 of the personalized interaction generation method involves selecting a template skeleton corresponding to a specific interaction mode within the template skeleton constraint generation path, and using generation control information to dynamically fill slots in the template skeleton to form a structured input package. This specifically includes: S161: Select a template skeleton based on the special interaction mode and the generation control information; wherein, the template skeleton is used to constrain the content structure generated by the large language model.

[0081] S162: Construct personalized fill content based on the generated control information, and fill the personalized fill content into the corresponding slots in the template skeleton to obtain the slot content. Personalized fill content can be obtained through web searches, preset discourse libraries, and preset content templates.

[0082] S163: Based on the special interaction mode, combine the template skeleton and slot content to form a structured input package.

[0083] The technical solution provided in this application allows for the selection or construction of a template skeleton in the template skeleton constraint generation path, based on factors such as the triggered special interaction mode, the child's age group, child profile, short-term memory, long-term memory, toy character design parameters, and safety constraints. A template skeleton is a constrained structural framework pre-built before content generation in a special interaction mode. It specifies the content structure of the large language model's response content, rather than directly fixing the final dialogue. For example, for the story-accompaniment mode, the template skeleton can specify the content structure as "story beginning (including character introduction) → story development (including conflict or adventure plot) → story ending (including moral or interactive questions)"; for the awkward silence rescue mode, the template skeleton can specify the content structure of the response content as "empathic expression → interest-awakening questions → open-ended interactive invitations".

[0084] The template skeleton can be constructed using methods such as manual pre-setting, rule base configuration, configuration file definition, or model-assisted generation. This application embodiment can maintain multiple template skeleton libraries according to different special interaction modes, selecting the appropriate template skeleton from them when the corresponding mode is triggered.

[0085] After selecting a template skeleton, personalized fill content can be built based on the generated control information and then filled into the corresponding slots of the template skeleton to obtain the slot content. For example, for the template skeleton of the story companion mode, the slots to be filled include: "Character Name" (fill in the name the child gives to the toy or the IP character the child likes), "Story Theme" (fill in the child's recent high-frequency interest tags), "Interactive Questions" (fill in personalized questions generated based on the child's profile), etc.

[0086] Subsequently, based on the current specific interaction mode, the template skeleton and slot content are combined to form a structured input package. The structured input package may include: age group, current interaction mode, recent available memories, long-term preference tags, child profile tags, toy character design parameters, template skeleton structure, slot filling content, prohibited content, and safety constraints, etc.

[0087] Through the above process, structured input is generated by using template skeletons, personalized fill content, fill slots, and slot content. This allows for structured constraints on the large language model, controlling the generation of age-appropriate, interesting, and stable interactive content.

[0088] Figure 1 The technical solution provided in the illustrated embodiment, in step S160: after selecting the template skeleton corresponding to the special interaction mode in the template skeleton constraint generation path, and using the generation control information to perform dynamic slot filling operation on the template skeleton to form a structured input package, further includes: S170: Input the structured input package into the large language model, and control the large language model to generate and output personalized interactive content under the constraints of the structured input package.

[0089] The technical solution provided in this application involves inputting a structured input package into a large language model, enabling the model to generate natural language response text under the constraints of the structured input package. Because the template skeleton in the structured input package constrains the structural framework of the content generated by the large language model, the slot-filling content provides personalized materials, the toy character parameters constrain the tone and style, and safety constraints and prohibited content ensure the age-appropriateness and safety of the content. In this way, the large language model no longer generates content freely, but rather completes content generation within a controlled, structured space, effectively avoiding problems such as style drift and age-inappropriate content.

[0090] In addition, after generating personalized interactive content using the large language model, the generated text needs to undergo content safety and age-appropriateness checks before being input into the relevant speech synthesis module and played back to children in speech form. The output personalized interactive content should conform to the child's age group, current interaction mode, child's interests and preferences, and toy character design requirements.

[0091] In addition, as a preferred embodiment, the above-described personalized interaction generation method further includes, after step S170: generating and outputting personalized interaction content, the following steps: S180: Recording and backfilling of the results of this round of interaction.

[0092] The record of the results of this round of interaction includes, but is not limited to: a summary of the input and output of this round, the number of dialogue rounds, the length of the child's speech, whether the interaction continued, whether there was positive feedback, whether new interests were generated, event or emotional information, and the use of this record to determine whether to perform memory filling later.

[0093] Specifically, step S180 above: the steps of recording and memory backfilling the results of this round of interaction, specifically includes: S181: Based on the current interaction result corresponding to the current interaction scenario, determine whether the current interaction result contains reusable information; wherein, the reusable information is used to fill in the memory of the hierarchical memory information.

[0094] S182: If the information contains reusable information or meets the preset memory retention conditions, the results of this round of interaction are extracted in a structured manner to obtain event cards or structured summaries.

[0095] S183: Based on the nature of the information, the event cards or structured summaries are hierarchically backfilled into the corresponding permanent memory, short-term memory, and long-term memory in the hierarchical memory information to update the hierarchical memory information.

[0096] S184: Based on the updated hierarchical memory information, conditionally or periodically update the parameters of the child's portrait or toy character design.

[0097] The technical solution provided in this application embodiment performs subsequent processing on the interaction results after the current round of response output to determine whether memory refilling is required, and simultaneously implements child privacy protection.

[0098] This application's embodiments do not require the mandatory generation of event cards or updating of profiles for each round of dialogue. Instead, it first determines whether the current round contains reusable information. This reusable information includes: newly expressed interests, preferences, or taboos of the child; recent events; recent emotional state; important people or relationships; feedback on interactive methods such as games, stories, and Q&A; and data on the effectiveness of this round of interaction, such as the number of dialogue rounds, speech length, and willingness to continue interacting. When reusable information that can be stored in hierarchical memory exists in this round, it is first extracted into event cards or structured summaries. Event cards may include event summaries, event types, time attributes, emotional tendencies, mentioned objects, interest tags, confidence levels, and the number of times they are mentioned.

[0099] Subsequently, based on the nature of the content, a layered backfilling operation is performed on the permanent memory, short-term memory, and long-term memory within the layered memory information. The layered backfilling includes: Information such as stable identity, naming relationships, and long-term taboos is backfilled into permanent memory; Recent events, recent interests, recent emotions, and recently mentioned objects are filled into short-term memory; Recurring, emotionally resonant, and interactive information is refilled into long-term memory.

[0100] The parameters of children's portraits and toy character designs can be updated conditionally or periodically based on updated memories, and are not required to be updated every round.

[0101] It should be noted that the parameters of the child's portrait and toy character design can be updated conditionally or periodically based on the updated memories, and are not required to be updated every round.

[0102] In addition, as a preferred embodiment, the above-mentioned personalized interaction generation method further includes: S190: Privacy protection processing of children's data; This privacy protection processing of children's data includes the following steps during the memory backfilling of hierarchical memory information or the generation and calling of large models: S191: Desensitize or label sensitive information about children in hierarchical memory information; among which, labeling includes generating anonymized labels.

[0103] S192: When inputting structured input packages into a large language model, use anonymized labels, structured summaries, and necessary memory fragments to replace the original child-sensitive information as model input.

[0104] S193: Implement access control, encrypted storage, and encrypted transmission of toy character design parameters, layered memory information, child portraits, and current interactive content; among which, access control includes guardian permissions for managing child portraits and layered memory information, as well as disabling memory functions.

[0105] For example, when a child mentions in a conversation, "My name is Wang Xiaoming, and I live in Building 3 of Sunshine Garden Community," the system will label "Wang Xiaoming" as "child_name_001" and "Building 3 of Sunshine Garden Community" as "home_address_001" during memory backfilling. Only the labeled information will be stored in the hierarchical memory, and the original sensitive information will not be stored for a long time.

[0106] In the technical solution provided in this application embodiment, privacy protection processing is performed on children's data during memory backfilling and subsequent generation and retrieval processes. Privacy protection processing includes desensitization and minimization, specifically including one or more of the following methods: Sensitive information such as children's names, schools, addresses, contact information, and family members' names is anonymized or tagged to generate anonymous tags; the complete original dialogue is not stored as memory for a long time, but is stored in the form of structured summaries; when inputting into the large language model, anonymized tags, structured summaries, and necessary memory fragments (usually the necessary memory fragments themselves also need to be anonymized) are given priority to replace the original sensitive information as input to the large language model; access control, encrypted storage, and encrypted transmission are implemented for current child input such as children's voice text, hierarchical memory information, child profiles, and toy personification parameters; guardians can view, modify, and delete child profiles and memory content, or turn off / disable some memory functions, etc.

[0107] By protecting children's privacy, such as through desensitization, minimization of data storage, structured summaries, and access control, the exposure of children's sensitive information can be reduced, thereby mitigating the risks associated with children's data.

[0108] Compared to the direct generation of general large models, the technical solution provided in the above embodiments of this application introduces child profiles, hierarchical memory, toy character parameters, and pattern judgment before generating interactive content, making the model output more controllable. Compared to the existing personalized processing method of simple prompts, the technical solution of this application can generate generation control information, including one or more of the following, based on the understanding of the child's current input and the current interaction scenario: the child's basic profile, toy character parameters, hierarchical memory information, and child profile. This allows the construction of a structured input package including personalized content, forming a long-term updatable personalized interaction loop for children. Compared to the full-text stacking of historical dialogues, this application can improve the proportion of effective information and reduce contextual noise through hierarchical memory. This application maintains character consistency by setting toy character parameters and evolving these parameters based on interaction feedback. Furthermore, compared to existing random fun generation schemes, the technical solution of this application can improve the stability and age-appropriateness of the product in special scenarios through template skeletons and dynamic slot filling. In summary, the embodiments described above in this application introduce a child's basic profile, toy character parameters, hierarchical memory information, and a child's basic portrait. Further combining this with the child's current input and engagement metrics, a structured input package is formed to control the generation of controllable personalized interactive content by the large language model. Through this structured control link and the specific constraint generation mechanism that forms the structured input package based on special interaction patterns and generation control information, the child companion AI product can stably generate age-appropriate, controllable, and engaging interactive content with a high degree of personalization. This attracts children to continue interacting, fostering a long-term sense of companionship, and ultimately improving the retention rate, subscription conversion rate, and repurchase rate of the child companion AI product.

[0109] In addition, the following embodiments of this application provide product embodiments, the beneficial effects of which are the same as those of the personalized interaction method provided in the above embodiments, and other technical features in the product embodiments are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0110] See Figure 2 , Figure 2 This application provides a schematic diagram of a personalized interaction generation system, used in any of the personalized interaction methods provided in the above embodiments; the personalized interaction generation system includes: The parameter acquisition module 110 is used to acquire children's basic profiles, toy character design parameters, and layered memory information; The input acquisition and recognition module 120 is used to acquire the child's current input, recognize and judge the child's current input, and obtain the input understanding result; The information retrieval module 130 is used to retrieve and generate control information based on the input understanding results and the current interaction scenario; wherein, the generated control information includes one or more of the following: child basic profile, toy character design parameters, hierarchical memory information, and child portrait. The mode triggering module 140 is used to determine whether the current interaction scenario triggers a preset special interaction mode based on the child's participation index; if a special interaction mode is triggered, the template skeleton constraint generation path corresponding to the special interaction mode is entered. The dynamic slot filling module 150 is used to select the template skeleton corresponding to the special interaction mode in the template skeleton constraint generation path, and use the generation control information to perform dynamic slot filling operation on the template skeleton to form a structured input package. The constraint generation module 160 is used to input the structured input package into the large language model and control the large language model to generate and output personalized interactive content under the constraints of the structured input package.

[0111] correspond Figure 2 The personalized interactive generation system shown combines Figure 3 The personalized interaction generation method shown includes: S201: Children input voice / text.

[0112] S202: Current input understanding; including identifying intent, identifying emotion / engagement, and identifying reusable information.

[0113] S203: Retrieve generation control information; including hierarchical memory, child profile, toy character design parameters, and safety and age-appropriate parameters.

[0114] S204: Determine whether a special interaction mode has been triggered; if yes, proceed to step S205; otherwise, proceed to step S206.

[0115] S205: Template skeleton constraint generation.

[0116] S206: Lightweight generation for normal scenes.

[0117] S207: Generate a response using a large model.

[0118] S208: Content safety and age-appropriateness check.

[0119] S209: Voice output for children.

[0120] S210: Determine whether it contains valid sedimentation information; if not, proceed to step S211; if yes, proceed to step S212.

[0121] S211: Keep only the basic logs or discard them.

[0122] S212: Structured Extraction.

[0123] S213: Layered memory backfilling.

[0124] S214: Update child portrait / toy character design parameters.

[0125] In summary, the personalized interaction solution provided in this application forms a holistic collaborative closed loop through the child's basic profile, hierarchical memory, child portrait, toy character parameters, pattern triggering, template skeleton generation, and memory backfilling. Compared to the direct generation of a general large model, this method introduces child portrait, hierarchical memory information, toy character parameters, and interaction mode judgment before generation, making the output of the large language model more controllable.

[0126] Compared to simple personalized prompts, this application's technical solution can form a long-term, updatable, personalized interactive loop for children. Compared to simply stacking historical dialogues, this application's technical solution improves the proportion of effective information and reduces contextual noise through hierarchical memory and structured summarization. Compared to fixed character settings, this application's technical solution maintains character consistency through toy character parameters and can adjust conditions based on historical feedback. Compared to random fun generation, this application's technical solution improves generation stability and age-appropriateness in specific scenarios through template skeletons and dynamic slot filling. Compared to long-term storage of complete original dialogues, this application's technical solution reduces the risk to children's data through selective backfilling and privacy protection.

[0127] See Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the personalized interaction generation method of any of the above embodiments.

[0128] The following is for reference. Figure 4 The diagram illustrates a structural schematic of an electronic device suitable for implementing the embodiments of this application. The electronic devices in the embodiments of this application can include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, personal digital assistants (PDAs), tablet computers (PADs), portable multimedia players (PMPs), in-vehicle terminals, etc., as well as fixed terminals such as digital TVs, desktop computers, etc. Figure 4 The device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0129] like Figure 4As shown, the electronic device can include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory ROM 1002 or a program loaded from a storage device 1003 into a random access memory RAM 1004. The RAM 1004 also stores various programs and data required for the operation of the electronic device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. The communication device 1009 can operate the electronic device to exchange data with other devices via wireless or wired communication. Although the diagram shows a model building device with various systems, it should be understood that it is not required to implement or have all of the systems shown. It is possible to implement or have more or fewer systems alternatively.

[0130] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram can represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0131] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0132] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0133] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating personalized interactions, characterized in that, include: Obtain basic child profiles, toy character design parameters, and layered memory information; Obtain the child's current input, identify and judge the child's current input, and obtain the input understanding result; Based on the input understanding results and combined with the current interaction scenario, the generated control information is retrieved; wherein, the generated control information includes one or more of the following: the child's basic profile, the toy character design parameters, the hierarchical memory information, and the child's portrait; Based on the child's engagement metrics, determine whether the current interactive scenario triggers a preset special interaction mode; If the special interaction mode is triggered, the template skeleton constraint generation path corresponding to the special interaction mode will be entered. In the template skeleton constraint generation path, the template skeleton corresponding to the special interaction mode is selected, and the template skeleton is dynamically filled with slots using the generation control information to form a structured input package. The structured input package is input into a large language model, and the large language model is controlled to generate and output personalized interactive content under the constraints of the structured input package.

2. The method as described in claim 1, characterized in that, After generating and outputting personalized interactive content, it also includes: Based on the current interaction result corresponding to the current interaction scenario, determine whether the current interaction result contains reusable information; wherein, the reusable information is used to refill the hierarchical memory information. If the reusable information is included or the preset memory retention conditions are met, the results of this round of interaction are extracted in a structured manner to obtain event cards or structured summaries. Based on the nature of the information, event cards or structured summaries are hierarchically populated into the corresponding permanent memory, short-term memory, and long-term memory in the hierarchical memory information to update the hierarchical memory information; Based on the updated hierarchical memory information, the parameters of the child's portrait or toy character design are updated conditionally or periodically.

3. The method as described in claim 2, characterized in that, This also includes privacy protection processing for children's data, which includes memory backfilling of hierarchical memory information or generation and calling of large models: Sensitive information about children in the hierarchical memory information is desensitized or tagged; wherein, the tagging process includes generating anonymized tags; When the structured input package is input into the large language model, the anonymized tags, the structured summary, and the necessary memory fragments are used to replace the original sensitive information of children as model input; The toy character design parameters, the layered memory information, the child portrait, and the current interactive content are subject to access control, encrypted storage, and encrypted transmission; wherein, the access control includes guardian permissions for managing the child portrait and layered memory information, as well as disabling the memory function.

4. The method as described in claim 1, characterized in that, The step of determining whether a preset special interaction mode is triggered in the current interaction scenario based on the child's participation index includes: The special interaction modes include one or more of the following: awkward silence rescue mode, interest enhancement mode, story companionship mode, game interaction mode, and emotional soothing mode; Based on the participation index, when low participation occurs in multiple consecutive rounds of dialogue or the current round number decreases, the awkward silence rescue mode is triggered. Based on the engagement index, when a specific interest content in short-term memory appears frequently recently or a stable preference exists in long-term memory, the interest enhancement mode in the special interaction mode is triggered. Based on the engagement metrics, when the current child input contains a pattern-triggered instruction, the special interaction mode corresponding to the pattern-triggered instruction is selected.

5. The method as described in claim 1, characterized in that, After determining whether the current interactive scenario triggers a preset special interaction mode based on the child's participation index, the method further includes: If the special interaction mode is not triggered, proceed to the lightweight generation path; In the lightweight generation path, safety constraint information is extracted from the child's basic profile; The input understanding results, the child profile, the toy character design parameters, and the safety constraint information are assembled into a lightweight prompt input package; The lightweight prompt input package is input into the large language model, and the large language model is controlled to generate and output personalized interactive content under the constraints of the lightweight prompt input package.

6. The method as described in claim 1, characterized in that, In the template skeleton constraint generation path, the template skeleton corresponding to the special interaction mode is selected, and the template skeleton is dynamically filled with slots using the generation control information to form a structured input package, including: Based on the specific interaction mode and the generation control information, a template skeleton is selected; wherein, the template skeleton is used to constrain the content structure generated by the large language model. Personalized fill content is constructed based on the generated control information, and the personalized fill content is filled into the corresponding slot of the template skeleton to obtain the slot content; Based on the special interaction mode, the template skeleton and the slot content are combined to form the structured input package.

7. The method as described in claim 1, characterized in that, The acquisition of children's basic profiles, toy character design parameters, and layered memory information includes: When interacting with a child for the first time or configuring guardian permissions, build a basic profile of the child and initial toy persona parameters; Based on the child's basic profile, an initial version of the child's portrait is generated; Configure the hierarchical memory information according to the child's basic profile or the guardian's permission settings; wherein, the hierarchical memory information includes permanent memory, short-term memory and long-term memory; After interacting with children, historical interaction information is obtained. Based on the stability and timeliness of the information, the historical interaction information is classified and backfilled into the permanent memory, short-term memory and long-term memory to update the hierarchical memory information. When the update conditions are met or a periodic update is performed, the child's profile and / or the toy character design parameters are updated based on the child's basic profile and the hierarchical memory information.

8. The method as described in claim 7, characterized in that, The updating of the child's portrait and / or the evolution of the toy character parameters includes: The child profile is updated based on the child's basic profile, short-term memory, and long-term memory; Based on children's interaction preferences, the offset of the toy character design parameters in multiple preset candidate evolution directions is determined; The toy character parameters are weighted and adjusted according to the offset to control the evolution of the toy character parameters in the plurality of candidate evolution directions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the personalized interaction generation method as described in any one of claims 1 to 8.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the personalized interaction generation method as described in any one of claims 1 to 8.

Images