Intelligent interaction method and system based on hierarchical semantic cross-fusion

Abstract

The application discloses a type intelligent interaction method and system based on hierarchical semantic cross fusion. The method comprises the following steps: acquiring voice or text input information and an interaction scene; analyzing the input information and the interaction scene, and determining basic semantic information, an emotion category and a scene intention; converting the basic semantic information, the emotion category and the scene intention into low-dimensional vectors, and fusing the low-dimensional vectors through a dynamic weight adjustment mode to generate a thinking vector which comprehensively considers semantics, emotions and contexts; creating response content which conforms to language ability and reflects emotional needs according to the thinking vector; performing safety inspection on the response content, and fine-tuning output according to set preferences and restriction conditions to obtain final output content. Through implementation of the method of the application, the communication quality and intimacy between intelligent toys and children are significantly improved, so as to solve the problem of insufficient language adaptation for children in the prior art.

Description

Technical Field

[0001] This invention relates to human-computer interaction methods, and more specifically to intelligent interaction methods and systems based on hierarchical semantic cross-fusion. Background Technology

[0002] With the development of artificial intelligence technology, smart toys and educational tools are playing an increasingly important role in children's growth. However, current technologies still face many challenges in handling language interaction with children. Most general language models are designed primarily for adult written or standard spoken language, lacking sufficient adaptability to children's unique expressions, such as incomplete sentence structures, grammatically incorrect expressions, and emotionally charged language, leading to comprehension biases. For example, the expression "I don't want you to go" implies a child's genuine need for companionship, rather than its literal meaning. Furthermore, the response mechanisms of existing smart toys are mostly based on fixed templates, making the tone of voice stiff and repetitive, failing to provide children with a "friend-like" experience. At the same time, these systems also have problems with the consistency of emotional expression and the age-appropriateness of content; responses are often too adult-like and unsuitable for young children's comprehension abilities.

[0003] Large-scale language models face challenges in toy scenarios, including high resource consumption and long response times, making direct deployment on toy devices virtually impossible. Relying on cloud computing not only increases network latency, impacting user experience, but also shortens battery life due to frequent calls. More importantly, general-purpose model training data contains a large amount of adult contextual information, potentially outputting content unsuitable for children, and lacks specific safety filtering mechanisms for children's scenarios. Parents have high demands for the appropriateness and controllability of content, expecting interpretable and configurable safety controls to ensure content safety. Considering the significant differences in language abilities and interests among children of different ages, existing systems struggle to flexibly adapt to meet each child's unique needs and cannot effectively utilize past interaction records to build lasting emotional connections, lacking stable character settings.

[0004] Therefore, it is necessary to design a new approach to significantly improve the quality and intimacy of communication between smart toys and children, in order to address the problem of insufficient adaptation of existing technologies to children's language, provide a more personalized and emotionally rich interactive experience, and create a more friendly and safe learning and entertainment environment for children. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent interaction method and system based on hierarchical semantic cross-fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a hierarchical semantic cross-fusion-based intelligent interaction method, comprising:

[0007] Acquire voice or text input information and interaction scenarios;

[0008] Analyze the input information and interaction scenario to determine basic semantic information, emotion category, and scenario intent;

[0009] The basic semantic information, emotion category, and scene intent are transformed into low-dimensional vectors, and the low-dimensional vectors are fused by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion, and context.

[0010] Based on the aforementioned thinking vectors, response content is created that is both linguistically competent and reflects emotional needs;

[0011] The response content is subjected to a security check, and the output is fine-tuned according to the set preferences and constraints to obtain the final output content.

[0012] Its further technical solution is as follows: the analysis of the input information and interaction scenario, and the determination of basic semantics, emotion category and scenario intent, includes:

[0013] A lightweight encoder and dependency fragment analysis technique are used, combined with a children's vocabulary list, to extract words, entities and events from the input information to obtain basic semantic information;

[0014] By analyzing the acoustic features and textual cues in the input information, an emotion intensity scale is established to distinguish and quantify fine-grained emotional states in order to obtain emotion categories.

[0015] Based on dialogue history, toy patterns, and time factors, common children's scene templates are used to match the corresponding interaction scene to determine the user's specific needs or intentions in order to obtain the scene intent.

[0016] The further technical solution is as follows: A lightweight encoder and dependency fragment analysis technology are employed, combined with a children's vocabulary list, to extract words, entities, and events from the input information to obtain basic semantic information, including:

[0017] The input information is initially processed by a lightweight encoder. Dependency fragment analysis and fragment matching strategies are used to adapt to the incoherent nature of children's language. In addition, vocabulary recognition and extraction of colloquial expressions, misspellings, and words, entities and events in children's language designed specifically for children are used to obtain basic semantic information.

[0018] The further technical solution is as follows: the acoustic features include pitch, energy and speech rate, and the toy mode includes story time, game time or learning time.

[0019] The further technical solution is as follows: The basic semantic information, emotion category, and scene intent are transformed into low-dimensional vectors, and the low-dimensional vectors are fused by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion, and context, including:

[0020] The basic semantic information, emotion category, and scene intent are converted into low-dimensional vector representations, and irrelevant information is filtered out by introducing a learnable masking mechanism to obtain multiple low-dimensional vectors.

[0021] The low-dimensional vectors are fused into a thinking vector that takes into account semantics, emotion, and context through linear combination. The weight parameters in the linear combination are adjusted according to the specific scenario of the current interaction.

[0022] The further technical solution is as follows: the creation of response content based on the thought vector that conforms to both language ability and emotional needs includes:

[0023] Adjusting the vocabulary difficulty according to the set children's age, replacing complex words with easier-to-understand expressions, and restricting the use of abstract nouns in order to achieve vocabulary control results;

[0024] Based on the results of the vocabulary control, simple short sentences and parallel structures are used, and a mild tone is adopted to generate sentence patterns;

[0025] Based on the emotional information in the thought vector, the sentence structure is infused with emotional expression and multimodal output is triggered to provide soothing and supportive responses to negative emotions, thereby obtaining response content.

[0026] Its further technical solution is as follows: based on the emotional information in the thought vector, the sentence structure is infused with emotional expression and multimodal output is triggered to provide soothing and supportive responses to negative emotions, so as to obtain response content, including:

[0027] Based on the emotional information in the thought vector, the sentence structure is infused with emotional expression. When the required emotional intensity is detected, a multimodal output instruction is triggered. For negative emotions, comforting and supportive responses are generated to obtain the response content.

[0028] The further technical solution is as follows: The process of performing a security check on the response content and fine-tuning the output according to set preferences and constraints to obtain the final output content includes:

[0029] The response content is subjected to sensitivity checks, and the output is adjusted according to the style preferences and disabling rules set by the parents to obtain the final output content.

[0030] This invention also provides an intelligent interaction system based on hierarchical semantic cross-fusion, comprising:

[0031] The acquisition unit is used to acquire voice or text input information and interaction scenarios;

[0032] The analysis unit is used to analyze the input information and interaction scenario, and determine basic semantic information, emotion category and scenario intent;

[0033] The transformation and fusion unit is used to transform the basic semantic information, emotion category and scene intent into low-dimensional vectors, and fuse the low-dimensional vectors by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion and context.

[0034] The content creation unit is used to create response content that is both linguistically competent and reflects emotional needs based on the thought vector.

[0035] The adjustment unit is used to perform security checks on the response content and fine-tune the output according to the set preferences and constraints to obtain the final output content.

[0036] The further technical solution is as follows: the analysis unit includes:

[0037] The semantic extraction subunit is used to extract words, entities and events from the input information by using a lightweight encoder and dependency fragment analysis technology, combined with a children's vocabulary, to obtain basic semantic information.

[0038] The emotion determination subunit is used to establish an emotion intensity scale by analyzing the acoustic features and textual cues in the input information, so as to distinguish and quantify fine-grained emotional states and obtain emotion categories.

[0039] The intent determination subunit is used to determine the user's specific needs or intents by matching the relevant interaction scenario with common children's scenario templates based on dialogue history, toy patterns, and time factors, in order to obtain the scenario intent.

[0040] The advantages of this invention compared to existing technologies are as follows: By acquiring and analyzing children's voice or text input information and their interaction scenarios, this invention accurately extracts basic semantics, emotion categories, and scene intentions, and transforms them into low-dimensional vectors. These vectors are then dynamically weighted and fused into a thinking vector, thereby creating response content that matches children's language abilities and reflects their emotional needs. This method not only considers the accuracy of semantic understanding but also incorporates emotion perception and contextual association, ensuring that the response accurately matches the child's language level while providing appropriate emotional support. Furthermore, by conducting security checks and personalized fine-tuning of the response content, the safety and appropriateness of the interaction are further guaranteed. This method significantly improves the communication quality and intimacy between smart toys and children, solves the shortcomings of existing technologies in adapting to children's language, provides a more personalized and emotionally rich interactive experience, and creates a more friendly and safe learning and entertainment environment for children.

[0041] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating the intelligent interaction method based on hierarchical semantic cross-fusion provided in an embodiment of the present invention;

[0044] Figure 2 A schematic block diagram of an intelligent interactive system based on hierarchical semantic cross-fusion provided in an embodiment of the present invention;

[0045] Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0048] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0049] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0050] Please see Figure 1 , Figure 1 This is a flowchart illustrating the hierarchical semantic cross-fusion-based intelligent interaction method provided in this invention. This method, applied to a server, acquires voice or text input and interaction scenarios, analyzes and determines basic semantic information, emotion categories, and scenario intentions, and transforms these into low-dimensional vectors for dynamic fusion to generate thought vectors. This results in responses that are both linguistically appropriate and reflect emotional needs. Specifically designed for children, this method employs a lightweight encoder and dependency fragment analysis technology combined with a children's vocabulary to extract words, entities, and events. It utilizes acoustic features to analyze emotional states and determines specific needs or intentions based on dialogue history, ensuring a high degree of adaptation to children's language characteristics. By adjusting vocabulary difficulty to suit children's age, using simple sentences and a gentler tone, and incorporating emotional expressions, the method significantly improves the quality and intimacy of communication with children, providing a personalized and emotionally rich interactive experience. Simultaneously, by fine-tuning the output content through security checks and parental preference settings, the method ensures the safety and suitability of the communication environment, creating a friendly and educational learning and entertainment space for children.

[0051] Figure 1 This is a flowchart illustrating the intelligent interaction method based on hierarchical semantic cross-fusion provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S150.

[0052] S110, acquiring voice or text input information and interaction scenarios.

[0053] In this embodiment, the voice input information refers to the child's voice data captured by the microphone equipped with the smart toy. This voice data may include, but is not limited to, daily conversations, questions, emotional expressions (such as happiness or sadness), and requests for storytelling. Considering the characteristics of children's language, such as inaccurate pronunciation, incomplete sentence structure, and simple and colloquial vocabulary, the system needs to be specially designed to adapt to and correctly parse this type of input.

[0054] Text input: In some cases, if the smart toy supports text input (e.g., via a touchscreen or physical keyboard), this input method is also acceptable. Text input can be questions that children type directly, shared story snippets, or anything they want to communicate with the toy.

[0055] Whether it's voice or text input, the system first preprocesses the raw data, such as endpoint detection (determining the start and end times of speech), noise suppression (improving the signal-to-noise ratio for more accurate subsequent analysis), and speech rate estimation, providing a high-quality data foundation for the next step of basic semantic extraction.

[0056] Interactive scenarios refer to:

[0057] Current environmental conditions: This includes time (specific times of day, such as morning, noon, and night), location (family room, outdoors, etc.), and the device's operating mode (such as story time, play mode, and learning mode). Different times and locations can influence a child's behavior and emotions, so understanding this information helps in a more accurate understanding of a child's needs and intentions.

[0058] Toy Mode: Depending on the toy's designated mode, such as educational mode, entertainment mode, or soothing mode, the system will adjust its response strategy. For example, in soothing mode, the toy may pay more attention to the child's emotional changes and provide more comforting and supportive responses; while in learning mode, it focuses on knowledge transfer and intellectual development.

[0059] Dialogue history: This includes the content and context of recent interactions, which is crucial for maintaining consistency and coherence across multiple rounds of dialogue. For example, if dinosaurs were discussed previously, toys can enhance the realism and intimacy of interactions when similar topics are mentioned in subsequent conversations by recalling previous conversations.

[0060] In conclusion, acquiring voice or text input information and interaction scenarios is the first step in building an intelligent interactive system that can understand and respond to children's needs. This involves not only technical implementation but also how to better adapt to children's language characteristics and emotional needs, thereby improving the quality and intimacy of communication.

[0061] S120. Analyze the input information and interaction scenario, and determine the basic semantic information, emotion category and scenario intent.

[0062] In this embodiment, basic semantic information refers to:

[0063] Vocabulary, Entities, and Events: Through a lightweight encoder and dependency parsing (DP) technology, combined with a child-specific vocabulary list, the system identifies keywords, entities (such as names of people and locations), and event relationships (such as what happened) from the input information. For example, if a child says "I went to the park with Xiaoming," the system can identify "Xiaoming" as an entity, "park" as a location, and "play" as an event.

[0064] Sentence completeness assessment: The system evaluates sentence completeness based on extracted information, taking into account that children may use incomplete sentence structures. For example, for the sentence "It's hot today, I want to eat ice cream," the system should be able to understand that even without a clear subject, the sentence expresses a desire to eat ice cream.

[0065] Emotion categories refer to:

[0066] Emotional State and Intensity: By analyzing acoustic features (such as pitch, energy, and speech rate) and textual cues (such as interjections and repetitions) in the input information, an emotional intensity scale is established. For example, descriptions such as "a little sad" or "very angry" can be quantified into specific numerical values, thereby distinguishing and quantifying fine-grained emotional states. This helps to more accurately capture children's true emotions, such as identifying whether it is "whining" or "genuine distress."

[0067] Scene intent refers to:

[0068] The system determines a user's specific needs or intentions based on dialogue history, the toy's current mode (story / game / learning), and the timing of the interaction. It uses common child scenario templates to match the relevant interaction scenario. For example, if a child says "I want to hear a story" before bedtime, the system's scenario intent might be "requesting a bedtime story." Similarly, if a child says "This math problem is so hard" after experiencing frustration, the scenario intent might be interpreted as "seeking help to solve a difficult problem."

[0069] In one embodiment, step S120 described above may include steps S121 to S123.

[0070] S121. Using a lightweight encoder and dependency fragment analysis technology, combined with a children's vocabulary list, words, entities and events are extracted from the input information to obtain basic semantic information.

[0071] This step focuses on parsing the input content, ensuring that even incoherent or colloquial expressions are correctly understood and processed. At the same time, a vocabulary list designed specifically for children's language habits helps the system better adapt to children's language characteristics.

[0072] Specifically, the input information is initially processed by a lightweight encoder, and dependency fragment analysis and fragment matching strategies are used to adapt to the incoherent nature of children's language. In addition, vocabulary recognition and extraction of colloquial expressions, misspellings and words, entities and events in children's language designed specifically for children are combined to obtain basic semantic information.

[0073] Among them, dependency fragment analysis and fragment matching strategy is a technical approach for processing incoherent or colloquial language. After initially processing the input information using a lightweight encoder, it uses dependency grammar theory to identify lexical dependencies in the sentence and segments it into multiple semantically self-contained fragments for independent analysis. Subsequently, based on predefined child language patterns and vocabulary lists, these fragments are matched with known patterns to accurately extract basic semantic information such as words, entities, and events, thereby effectively understanding and processing children's unique non-standard language expressions. This method is particularly suitable for parsing children's language input because it can adapt to incomplete syntactic structures and diverse colloquial expressions.

[0074] This process enables the system to understand the core content of children's speech, even if the speech contains non-standard grammatical structures or vocabulary.

[0075] S122. By analyzing the acoustic features and textual cues in the input information, an emotion intensity scale is established to distinguish and quantify fine-grained emotional states in order to obtain emotion categories.

[0076] This step focuses on understanding the child's emotional state. The acoustic features include pitch, energy, and speech rate, while textual cues may include interjections and repeated words. By comprehensively considering vocal characteristics and textual content, the system can not only identify basic emotion types (such as happiness or sadness) but also accurately measure their intensity, which is crucial for generating appropriate responses. For example, the system can identify whether a child is "somewhat sad" or "very angry," which helps adjust the response accordingly.

[0077] S123. Based on dialogue history, toy patterns, and time factors, use common children's scene templates to match the corresponding interaction scene, determine the user's specific needs or intentions, and obtain the scene intent.

[0078] This step aims to gain a comprehensive understanding of the current interaction context in order to make the most appropriate response. Using predefined scenario templates, such as story time, playtime, or learning time, and taking into account conversation history and temporal factors (such as time of day), the system can dynamically adjust its behavior based on context, providing more personalized and relevant answers. For example, in the evening after a child has just experienced a failed attempt, the system might choose to respond with encouragement and support rather than simply providing knowledge. This scenario-based understanding mechanism improves the quality and relevance of the interaction, making smart toys true friends and guides for children.

[0079] S130. The basic semantic information, emotion category and scene intent are transformed into low-dimensional vectors, and the low-dimensional vectors are fused by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion and context.

[0080] The basic semantic information, emotion category, and scene intent extracted from the child's input are converted into low-dimensional vectors. These vectors are then fused by dynamically adjusting their weights to generate a thinking vector that comprehensively considers semantics, emotion, and context. This process is a key step in achieving highly human-like interaction, allowing smart toys to understand and respond to children by considering not only the content spoken (i.e., basic semantics) but also the child's emotional state (i.e., emotion category) and the context or purpose of the dialogue (i.e., scene intent).

[0081] In this embodiment, the thinking vector refers to a unified vector representation that comprehensively considers semantics, emotion, and context by transforming basic semantic information, emotion category, and scene intent into low-dimensional vectors and then dynamically adjusting the weights according to the specific interaction scenario for linear combination, so as to guide the smart toy to make appropriate responses.

[0082] In one embodiment, step S130 described above may include steps S131 to S132.

[0083] S131. The basic semantic information, emotion category and scene intent are converted into low-dimensional vector representations respectively, and irrelevant information is filtered out by introducing a learnable masking mechanism to obtain multiple low-dimensional vectors.

[0084] In this step, the system first needs to encode three independent information streams—basic semantic information, emotion category, and scene intent—into low-dimensional vector representations. This step involves using appropriate machine learning models to capture the unique features of each type of data and transforming them into easily tractable mathematical forms. Specifically:

[0085] Basic semantic vector (h) sem): This represents the core information extracted from the child's words, such as who did what, when they did it, and where they did it.

[0086] Emotion Vector (h) emo (): This reflects the emotional tone of the speech, including emotions such as happiness, sadness, and anger, and their intensity.

[0087] Context vector (h) ctx ): This includes the specific context and intention of the conversation, such as whether the child is seeking help, sharing happiness, or looking for comfort.

[0088] To improve the quality of vectors, the system introduces a learnable masking mechanism to filter out information irrelevant to the current task, ensuring that each vector accurately reflects the key characteristics of its corresponding dimension.

[0089] S132. The low-dimensional vector is fused into a thinking vector that takes into account semantics, emotion and context through linear combination. The weight parameters in the linear combination are adjusted according to the specific scenario of the current interaction.

[0090] Next, in step S132, these three low-dimensional vectors are fused into a unified thinking vector through linear combination. This process involves weight parameters (α, β, γ), which are dynamically adjusted according to the specific context of the current interaction. For example:

[0091] In the educational model, the value of α may be increased in order to place greater emphasis on the accurate transmission of knowledge content;

[0092] When in soothing mode, the weight of β may be increased, making the response more focused on emotional empathy;

[0093] When playing multiple rounds of a story or game, the weight of γ will increase to maintain plot consistency and character continuity.

[0094] This flexible weighting strategy ensures that the final generated response accurately addresses the child's words, appropriately reflects their emotional state, and takes into account the overall context of the conversation, thus providing a smart and warm interactive experience.

[0095] In summary, through the above two steps, S130 not only achieves effective integration of various types of information, but also ensures that the smart toy can make the most appropriate response according to different situations, greatly enhancing the quality and depth of interaction with children.

[0096] S140. Based on the thought vector, create response content that is both linguistically competent and reflects emotional needs.

[0097] In this embodiment, the response content refers to generating an answer that is both suitable for the child's comprehension level and meets their emotional needs by integrating basic semantic information, emotional categories, and contextual intent. This answer not only accurately conveys the information but also expresses it in a warm and encouraging manner, thereby enhancing the emotional connection with the child.

[0098] In one embodiment, step S140 described above may include steps S141 to S143.

[0099] S141. Adjust the vocabulary difficulty according to the set child's age, replace complex words with easier-to-understand expressions, and restrict the use of abstract nouns to obtain vocabulary control results.

[0100] In this embodiment, vocabulary control refers to ensuring that the vocabulary and concepts used in the output text are understandable and acceptable to children of the target age group by filtering and replacing words. For example, complex scientific terms or abstract concepts are transformed into more easily understood figurative metaphors or concrete examples (such as changing "dinosaur era" to "a long, long time ago"), while reducing the use of long sentences and complex structures to make sentences more concise and clear.

[0101] S142. Based on the results of the vocabulary control, simple short sentences and parallel structures are used, and a mild tone is adopted to generate sentence patterns.

[0102] In this embodiment, sentence structure refers to language structures specifically designed to facilitate children's understanding and acceptance. It emphasizes the use of short sentences and parallel structures to organize information, avoiding the use of multiple clauses or multi-layered nested sentence structures that may cause confusion. In addition, it focuses on using a gentle and encouraging tone, especially when correction or negation is needed, prioritizing more positive expressions such as "think about it again" or "let's try another way" to protect children's self-confidence and curiosity.

[0103] S143. Based on the emotional information in the thought vector, the sentence structure is infused with emotional expression and multimodal output is triggered to provide soothing and supportive responses to negative emotions, so as to obtain response content.

[0104] In this embodiment, based on the emotional information in the thought vector, the sentence structure is infused with an emotional expression. When a required emotional intensity is detected, a multimodal output instruction is triggered. For negative emotions, comforting and supportive responses are generated to obtain the response content.

[0105] This means enriching the response with emotional information contained in the thought vector, making it not only a transmission of information but also an expression of emotional resonance and support. The specific steps are as follows:

[0106] Emotional expression: The system adds corresponding empathetic remarks to the generated response text based on the intensity and type of emotion extracted from the thought vector (such as happiness, sadness, anger, etc.). For example, when the system detects that a child expresses "a little regret," it may generate a response like, "I also feel a little regretful, but we can think of other interesting things together."

[0107] Emotional adjustment and tone selection: The system selects different tones and sentence styles for different emotional states. For example, when facing negative emotions, the system tends to use gentle and encouraging language; while in positive emotional scenarios, it may adopt a more cheerful and energetic tone.

[0108] Enhanced Multisensory Experience: For inputs with high emotional intensity, especially negative emotions, the system not only provides verbal comfort but also triggers additional multimodal output commands, such as facial animations, voice changes, or background music. This helps strengthen emotional connections and provides children with more comprehensive support. For example, when a child is scared, the smart toy's facial expression screen can display a warm smile while playing a soft voice saying, "Don't worry, I'm here with you."

[0109] Personalized Interaction Design: Recognizing that each child's reactions and preferences may differ, the system can customize personalized multimodal responses based on interest tags and preference settings accumulated from previous interactions. For example, if it knows a child particularly likes dinosaurs, the system might suggest, "Let's imagine ourselves as brave little dinosaurs and go on an adventure together!"

[0110] Negative Emotion Processing Mechanism: Once the system identifies a negative emotion (such as sadness, anger, fear, etc.), it prioritizes generating soothing and supportive responses rather than directly providing rational explanations. This strategy aims to first alleviate the child's emotional stress and then gradually guide them to solve the problem or shift their attention.

[0111] Positive guidance and educational opportunities: While dealing with negative emotions, the system also seeks opportunities to transform them into educational opportunities. For example, when a child feels frustrated because of failure, in addition to providing emotional support, they can be encouraged to try new methods or the importance of perseverance can be emphasized.

[0112] In summary, step S143 integrates emotional information into language output and combines it with a multimodal feedback mechanism, enabling smart toys to more effectively understand and respond to children's emotional needs, thereby establishing a deeper emotional connection while ensuring the safety and appropriateness of the interaction process. This approach not only enhances the user experience but also promotes children's mental health development.

[0113] S150. Perform a security check on the response content and fine-tune the output according to the set preferences and restrictions to obtain the final output content.

[0114] In this embodiment, the response content is subjected to sensitivity checks, and the output is adjusted according to the style preferences and disabling rules set by the parents to obtain the final output content.

[0115] This process ensures that the output of smart toys not only meets children's cognitive level and psychological needs, but also adheres to the safety and style preferences set by parents. Specifically, the S150 steps can be divided into the following key stages:

[0116] Sensitivity test:

[0117] Sensitive Word Scan: The system first performs a comprehensive scan of the generated candidate responses using a maintained children-specific sensitive word database. These sensitive words cover topics such as violence, self-harm, and terror, as well as any content that may cause fear or anxiety. Once a sensitive word or topic is detected, the system automatically initiates a rewriting mechanism to transform the relevant content into more positive or educational expressions.

[0118] Negative Emotion Processing: For input containing negative emotions (such as sadness and anger), the system prioritizes generating soothing and supportive responses. For example, if a child expresses "I hate math," the system doesn't simply correct this attitude, but instead tries to understand the underlying emotion and offer encouragement: "I know math can be a little difficult sometimes, but every time you solve a problem, you get smarter!"

[0119] Style preferences and disabling rules adjustments:

[0120] Application of parent-defined style preferences: The system reads parameters pre-set by parents through a configuration interface, such as their style preference for "focusing on learning" or "focusing on companionship." For example, in a family environment that emphasizes learning, the system may be more inclined to provide knowledge-based answers; while in a family that values ​​emotional companionship, it will adopt a gentler, more empathetic language style.

[0121] Personalized rule enforcement: Parents can also set specific topics or keywords as prohibited items. For example, if parents do not want to discuss certain types of horror stories, the system will avoid mentioning these topics when generating responses. In addition, parents can also set different language difficulty levels for children of different ages to suit their children's language development levels.

[0122] Output fine-tuning and review:

[0123] Vocabulary and Sentence Structure Adjustment: Based on the above checks and adjustments, the system will further fine-tune the initially generated answers. This includes replacing overly complex words or words that are beyond a child's cognitive range, simplifying long sentences and multiple clauses, and replacing them with short sentences and parallel structures to make the language more suitable for children to understand and accept.

[0124] Multimodal output planning: In addition to textual adjustments, the system also incorporates text-to-speech (TTS), facial animations, lighting effects, and other sensory experiences to enhance interactivity. For example, when expressing comfort, in addition to gentle words, it can be accompanied by warm smile animations or soft background music to create a more intimate and friendly atmosphere.

[0125] Final review and confirmation: All revised responses must undergo a final security review to ensure no sensitive information is omitted and that they fully comply with all requirements set by the parents. Only outputs that pass this step will be officially sent to the child.

[0126] In summary, the S150 process, through meticulous safety and compliance checks on the response content and personalized adjustments based on parents' specific preferences and constraints, ensures that the smart toy's output meets children's learning and emotional needs while maintaining suitability and safety within diverse family value environments. This approach not only enhances the user experience but also creates a healthy and positive growth environment for children.

[0127] The method in this embodiment belongs to the field of natural language understanding and human-computer interaction technology. It particularly focuses on corpus modeling and pattern recognition techniques for children's language features, semantic extraction and representation learning algorithms for multi-dimensional information, and emotion recognition, tone analysis, and safe expression constraint mechanisms. It is mainly applied to intelligent toy interaction systems, aiming to support the deployment of local semantic understanding modules on resource-constrained toy terminals and achieve anthropomorphic, continuous, and personalized language response behaviors, while ensuring good recognition capabilities for the incomplete language and colloquial expressions of young children.

[0128] The method described in this embodiment has a wide range of applications, including but not limited to children's voice-interactive smart toys, educational toys and learning machines, and toys for comfort and emotional support. By engaging in daily conversations and Q&A with children through voice and facial expressions, smart toys can not only continuously accompany children, such as by telling stories, greeting them, and giving game commands, but also maintain consistency in identity and style across multiple rounds of dialogue. Furthermore, in interactive Q&A, it can dynamically adjust the depth of explanation based on the child's language level, packaging knowledge-based answers into stories or metaphors suitable for children's understanding, and providing encouraging feedback in conjunction with emotional states to reduce frustration. For emotionally sensitive scenarios, this invention can identify emotional signals in children's speech and provide comforting and empathetic responses through gentle tone and appropriate content, avoiding irritating or age-inappropriate expressions.

[0129] To achieve the above objectives, the method in this embodiment first establishes a multi-layered semantic structure representation system, breaking down children's sentences into a basic semantic layer (core information such as who / what / when / where), an emotional tone layer (happy, sad, angry, etc.), and a scene intention layer (asking for help, sharing, resistance, curiosity, etc.), and realizes calculable representation of incomplete sentences, emotional expressions, and imaginative narratives. Secondly, by implementing a biomimetic thinking architecture that achieves semantic cross-fusion, it simulates the hierarchical reasoning sequence of humans: "first understand the content, then feel the emotion, and then consider how to answer in the context," fusing semantic information from different levels into a unified "thinking vector," enabling toys to communicate accurately and emotionally when responding. Finally, a child-safe and child-friendly output mechanism is constructed, limiting vocabulary difficulty and sentence complexity to ensure that children of different ages can understand it. Expressions that may cause fear, anxiety, or conflict are rewritten or weakened, and the system automatically guides sensitive topics such as health and dangerous behaviors towards strategies for positive advice and informing parents.

[0130] These designs significantly improve the accuracy of semantic understanding, enhance anthropomorphism and a sense of companionship, systematically ensure child-friendliness and safety, and balance deployment costs and scalability. Specifically, through multi-layer semantic extraction and vector fusion, this invention can accurately capture key information in incomplete sentences, colloquial expressions, and fantasy narratives; the biomimetic thinking architecture allows the toy to undergo a process of "understanding—feeling—choosing expression" internally, resulting in outputs that naturally reflect human-like thought processes; combined with emotion enhancement and sentence softening strategies, the dialogue style remains gentle, friendly, and encouraging. Furthermore, the modular design facilitates hardware simplification or replacement with lighter implementations, and the thinking vector h serves as a unified interface, enabling future integration with more powerful cloud-based models or local upgrades. This not only meets children's learning and emotional needs but also creates a healthy and positive growth environment for them.

[0131] Specifically, the method in this embodiment starts with multi-dimensional input acquisition and semantic extraction, then proceeds through semantic vector encoding and biomimetic thinking vector fusion, to the generation of child-friendly output based on thinking vectors, and finally ensures the safety and suitability of the final output through safety constraint verification and personalized parental control.

[0132] First, in the perception and understanding stage, the system needs to acquire data on the child's original speech stream (or converted text) and the current interactive scenario. In basic semantic parsing, a lightweight encoder encodes the input and employs dependency segment analysis and segment matching strategies to adapt to the disjointed nature of children's speech. Simultaneously, a child-specific vocabulary list is used to identify colloquial expressions, misspellings, and preschool slang to extract basic semantic information such as vocabulary, entities, and events. In the emotion and tone determination stage, acoustic features (such as pitch, energy, and speech rate) and textual cues (such as interjections and repeated words) are extracted in parallel. A joint determination is used to establish an emotion intensity scale, distinguishing fine-grained emotions. Scene intent inference utilizes dialogue history, toy patterns, and timely tags, combined with common child scene templates, to establish the scene intent of the current dialogue.

[0133] Next, in the core decision-making stage, the second step, the system transforms the basic semantics, emotion categories, and scene intentions obtained in the first step into low-dimensional vector representations and introduces a learnable masking mechanism to filter noise. Then, by linearly fusing these vectors of different dimensions, a unified "bionic thinking vector" is generated. A dynamic weight adjustment strategy automatically adjusts the weights of each vector according to the current scene, ensuring that the importance of the corresponding emotional or knowledge content is prioritized in different modes.

[0134] The third step is the expression generation stage. Based on the unified thinking vector generated in the previous step and the configured parameters for the child's age group, the system performs vocabulary difficulty control, sentence softening, and emotion enhancement. The decoder filters and replaces the generated content according to the set child's age group and language level to ensure that the vocabulary is suitable for children's understanding. The generation engine avoids using long sentences and multiple clauses, especially in negation or error correction scenarios, prioritizing the use of milder sentence structures to protect children's confidence. In addition, based on the emotional intensity in the thinking vector, the system adds anthropomorphic expressions to the text and maps high emotional intensity to multimodal output instructions.

[0135] The final step is the review and output stage, which includes security constraint verification and personalized control for parents. The system scans candidate responses using a sensitive word database, and immediately initiates a rewriting mechanism if a sensitive topic is triggered, guiding the conversation towards positive values. The system also reads the preference settings set by parents and fine-tunes the output style to ensure that the final output is both safe and in line with family education principles.

[0136] The child-friendly semantic understanding process begins with the input analysis stage, where children's speech first undergoes preprocessing steps such as endpoint detection, noise suppression, and speech rate estimation to generate candidate text and acoustic features. Subsequently, this textual information is fed into a lightweight encoder for basic semantic extraction, identifying keywords, entities, and event relationships, and evaluating sentence completeness. Simultaneously, the system uses acoustic features and textual cues to infer emotion categories, intensity, and dialogue intent labels, forming multi-layered vectors, including a semantic vector h. sem Emotion vector h emo and context vector h ctx These vectors are merged into a thinking vector h based on the current mode configuration.

[0137] Once in the biomimetic thinking and response planning phase, the system determines the response type based on the semantic component of the thinking vector, such as directly answering a question, asking a counter-question, or telling a story. It also sets the response tone based on the emotional dimension, such as soothing, encouraging, or celebratory. Furthermore, the system considers the context to determine whether to reference dialogue history, ensuring continuity and personalized interaction. During this process, it simultaneously plans multimodal outputs such as facial animations, lighting changes, and background sound effects to enrich the interactive experience.

[0138] The output generation and safety constraint phases ensure the safety and appropriateness of the final output. First, the draft answer is sent to the child-friendly rewriting module to adjust vocabulary difficulty and sentence length. Then, a safety filter checks for sensitive topics, and if necessary, a safe rewriting strategy is activated or the answer is simplified. Finally, the generated text-driven TTS module reads the content aloud in a child-friendly voice, simultaneously triggering facial expressions, gestures, and lighting effects.

[0139] Parameter settings and personalized learning mechanisms allow for the automatic loading of different vocabulary difficulty tables, sentence templates, and story material libraries based on age level and difficulty mode, with fine-tuning of difficulty parameters as the number of interactions increases. Long-term memory and interest preference modeling maintain a limited capacity of interest tags; when specific topics appear frequently, the weight of related tags increases, and they are proactively introduced in subsequent interactions. The key event recording function enables the system to show care and concern for children at appropriate times.

[0140] This invention provides emotional support and guidance for imaginative play for 3-5 year olds in free-flowing conversation scenarios, and emotional support and problem-solving guidance for 6-8 year olds in homework assistance scenarios. Comparative experimental results show that, compared to traditional template-based systems, this invention significantly improves children's subjective sense of being understood and their desire to continue talking to the toys. It also demonstrates higher accuracy in recognizing the intent behind incomplete sentences, onomatopoeia, and emotional expressions, and effectively reduces the output of age-inappropriate content through safe rewriting and parental configuration mechanisms.

[0141] Therefore, the method in this embodiment not only significantly improves the accuracy of semantic understanding and enhances the sense of anthropomorphism and companionship, but also systematically ensures child-friendliness and safety, while balancing deployment costs and scalability. Through practical implementation and comparative verification, the effectiveness of this solution in improving children's subjective "feeling of being understood" and "wanting to continue talking to the toy," increasing the accuracy of intent recognition, and reducing the probability of age-inappropriate content output is demonstrated.

[0142] The aforementioned hierarchical semantic cross-fusion-based intelligent interaction method acquires and analyzes children's voice or text input information and their interaction scenarios. It accurately extracts basic semantics, emotion categories, and scene intentions, transforming these into low-dimensional vectors and dynamically adjusting their weights to fuse them into a thinking vector. This creates response content that matches children's language abilities and reflects their emotional needs. This method not only considers the accuracy of semantic understanding but also incorporates emotion perception and contextual association, ensuring that responses accurately match children's language levels while providing appropriate emotional support. Furthermore, security checks and personalized fine-tuning of the response content further guarantee the safety and appropriateness of the interaction. This method significantly improves the communication quality and intimacy between smart toys and children, addresses the shortcomings of existing technologies in adapting to children's language, provides a more personalized and emotionally rich interactive experience, and creates a more friendly and safe learning and entertainment environment for children.

[0143] Figure 2 This is a schematic block diagram of an intelligent interactive system 300 based on hierarchical semantic cross-fusion provided in an embodiment of the present invention. Figure 2 As shown, corresponding to the above-described intelligent interaction method based on hierarchical semantic cross-fusion, the present invention also provides an intelligent interaction system 300 based on hierarchical semantic cross-fusion. This intelligent interaction system 300 includes a unit for executing the above-described intelligent interaction method based on hierarchical semantic cross-fusion, and the system can be configured in a server. Specifically, please refer to... Figure 2 The hierarchical semantic cross-fusion-based intelligent interactive system 300 includes an acquisition unit 301, an analysis unit 302, a transformation and fusion unit 303, a content creation unit 304, and an adjustment unit 305.

[0144] The acquisition unit 301 is used to acquire voice or text input information and interaction scenarios; the analysis unit 302 is used to analyze the input information and interaction scenarios, and determine basic semantic information, emotion categories, and scenario intentions; the transformation and fusion unit 303 is used to transform the basic semantic information, emotion categories, and scenario intentions into low-dimensional vectors, and fuse the low-dimensional vectors by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion, and context; the content creation unit 304 is used to create response content that conforms to both language ability and emotional needs based on the thinking vector; and the adjustment unit 305 is used to perform a security check on the response content and fine-tune the output according to set preferences and constraints to obtain the final output content.

[0145] In one embodiment, the analysis unit 302 includes:

[0146] The semantic extraction subunit uses a lightweight encoder and dependency fragment analysis technology, combined with a children's vocabulary, to extract words, entities, and events from the input information to obtain basic semantic information. The emotion determination subunit analyzes the acoustic features and textual cues in the input information to establish an emotion intensity scale, distinguishing and quantifying fine-grained emotional states to obtain emotion categories. The intent determination subunit uses dialogue history, toy patterns, and time factors to match common children's scene templates with the corresponding interaction scene to determine the user's specific needs or intents to obtain scene intents.

[0147] In one embodiment, the semantic extraction subunit is used to perform preliminary processing on the input information through a lightweight encoder, using dependency fragment analysis and fragment matching strategies to adapt to the incoherent nature of children's language, and combining vocabulary recognition and extraction of colloquial expressions, misspellings, and words, entities, and events in children's language designed specifically for children, in order to obtain basic semantic information.

[0148] In one embodiment, the conversion and fusion unit 303 includes:

[0149] The transformation subunit is used to convert the basic semantic information, emotion category, and scene intent into low-dimensional vector representations respectively, and to filter irrelevant information by introducing a learnable masking mechanism to obtain multiple low-dimensional vectors; the fusion subunit is used to fuse the low-dimensional vectors into a thinking vector that comprehensively considers semantics, emotion, and context through linear combination, wherein the weight parameters in the linear combination method are adjusted according to the specific scenario of the current interaction.

[0150] In one embodiment, the content creation unit 304 includes:

[0151] The vocabulary control subunit is used to adjust the vocabulary difficulty according to the set child's age, replace complex words with easier-to-understand expressions, and restrict the use of abstract nouns to obtain the vocabulary control result; the sentence generation subunit is used to generate sentence patterns using simple short sentences and parallel structures based on the vocabulary control result, and adopt a soothing tone; the response subunit is used to incorporate emotional expressions into the sentence patterns based on the emotional information in the thought vector and trigger multimodal output, providing soothing and supportive responses to negative emotions to obtain response content.

[0152] In one embodiment, the response subunit is configured to incorporate emotional expressions into the sentence structure based on the emotional information in the thought vector, trigger a multimodal output instruction when a required emotional intensity is detected, and generate comforting and supportive responses for negative emotions to obtain response content.

[0153] In one embodiment, the adjustment unit 305 is used to perform a sensitivity check on the response content and adjust the output according to the style preferences and disabling rules set by the parents to obtain the final output content.

[0154] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned hierarchical semantic cross-fusion intelligent interactive system 300 and its various units can be referred to the corresponding descriptions in the foregoing method embodiments. For the sake of convenience and brevity, these details will not be repeated here.

[0155] The aforementioned hierarchical semantic cross-fusion-based intelligent interactive system 300 can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the computer device shown.

[0156] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0157] See Figure 3 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0158] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a hierarchical semantic cross-fusion-based intelligent interaction method.

[0159] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0160] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a type of intelligent interaction method based on hierarchical semantic cross-fusion.

[0161] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0162] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the hierarchical semantic cross-fusion type intelligent interaction method.

[0163] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0164] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0165] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all steps of the hierarchical semantic cross-fusion-based intelligent interaction method.

[0166] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0167] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0168] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0169] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0170] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0171] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hierarchical semantic cross-fusion-based intelligent interaction method, characterized in that: include: Acquire voice or text input information and interaction scenarios; Analyze the input information and interaction scenario to determine basic semantic information, emotion category, and scenario intent; The basic semantic information, emotion category, and scene intent are transformed into low-dimensional vectors, and the low-dimensional vectors are fused by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion, and context. Based on the aforementioned thinking vectors, response content is created that is both linguistically competent and reflects emotional needs; The response content is subjected to a security check, and the output is fine-tuned according to the set preferences and constraints to obtain the final output content.

2. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 1, characterized in that, The analysis of the input information and interaction scenario, and the determination of basic semantics, emotion categories, and scenario intent, includes: A lightweight encoder and dependency fragment analysis technique are used, combined with a children's vocabulary list, to extract words, entities and events from the input information to obtain basic semantic information; By analyzing the acoustic features and textual cues in the input information, an emotion intensity scale is established to distinguish and quantify fine-grained emotional states in order to obtain emotion categories. Based on dialogue history, toy patterns, and time factors, common children's scene templates are used to match the corresponding interaction scene to determine the user's specific needs or intentions in order to obtain the scene intent.

3. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 2, characterized in that, The process employs a lightweight encoder and dependency fragment analysis technology, combined with a children's vocabulary list, to extract words, entities, and events from the input information to obtain basic semantic information, including: The input information is initially processed by a lightweight encoder. Dependency fragment analysis and fragment matching strategies are used to adapt to the incoherent nature of children's language. In addition, vocabulary recognition and extraction of colloquial expressions, misspellings, and words, entities and events in children's language designed specifically for children are used to obtain basic semantic information.

4. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 2, characterized in that, The acoustic features include pitch, energy, and speech rate, and the toy modes include story time, game time, or learning time.

5. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 1, characterized in that, The process of converting the basic semantic information, emotion category, and scene intent into low-dimensional vectors, and fusing these low-dimensional vectors by dynamically adjusting weights to generate a thinking vector that comprehensively considers semantics, emotion, and context, includes: The basic semantic information, emotion category, and scene intent are converted into low-dimensional vector representations, and irrelevant information is filtered out by introducing a learnable masking mechanism to obtain multiple low-dimensional vectors. The low-dimensional vectors are fused into a thinking vector that takes into account semantics, emotion, and context through linear combination. The weight parameters in the linear combination are adjusted according to the specific scenario of the current interaction.

6. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 1, characterized in that, The creation of response content based on the thought vector that conforms to both language ability and emotional needs includes: Adjusting the vocabulary difficulty according to the set children's age, replacing complex words with easier-to-understand expressions, and restricting the use of abstract nouns in order to achieve vocabulary control results; Based on the results of the vocabulary control, simple short sentences and parallel structures are used, and a mild tone is adopted to generate sentence patterns; Based on the emotional information in the thought vector, the sentence structure is infused with emotional expression and multimodal output is triggered to provide soothing and supportive responses to negative emotions, thereby obtaining response content.

7. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 6, characterized in that, The process involves incorporating emotional expressions into the sentence structure based on the emotional information in the thought vector and triggering multimodal output to provide soothing and supportive responses to negative emotions, resulting in response content including: Based on the emotional information in the thought vector, the sentence structure is infused with emotional expression. When the required emotional intensity is detected, a multimodal output instruction is triggered. For negative emotions, comforting and supportive responses are generated to obtain the response content.

8. The intelligent interaction method based on hierarchical semantic cross-fusion according to claim 1, characterized in that, The process of performing a security check on the response content and fine-tuning the output according to set preferences and constraints to obtain the final output content includes: The response content is subjected to sensitivity checks, and the output is adjusted according to the style preferences and disabling rules set by the parents to obtain the final output content.

9. A hierarchical semantic cross-fusion-based intelligent interactive system, characterized in that: include: The acquisition unit is used to acquire voice or text input information and interaction scenarios; The analysis unit is used to analyze the input information and interaction scenario, and determine basic semantic information, emotion category and scenario intent; The transformation and fusion unit is used to transform the basic semantic information, emotion category and scene intent into low-dimensional vectors, and fuse the low-dimensional vectors by dynamically adjusting the weights to generate a thinking vector that comprehensively considers semantics, emotion and context. The content creation unit is used to create response content that is both linguistically competent and reflects emotional needs based on the thought vector. The adjustment unit is used to perform security checks on the response content and fine-tune the output according to the set preferences and constraints to obtain the final output content.

10. The intelligent interactive system based on hierarchical semantic cross-fusion according to claim 9, characterized in that, The analysis unit includes: The semantic extraction subunit is used to extract words, entities and events from the input information by using a lightweight encoder and dependency fragment analysis technology, combined with a children's vocabulary, to obtain basic semantic information. The emotion determination subunit is used to establish an emotion intensity scale by analyzing the acoustic features and textual cues in the input information, so as to distinguish and quantify fine-grained emotional states and obtain emotion categories. The intent determination subunit is used to determine the user's specific needs or intents by matching the relevant interaction scenario with common children's scenario templates based on dialogue history, toy patterns, and time factors, in order to obtain the scenario intent.

Images