Infant teaching resource intelligent searching and recommending system based on AI large model

Abstract

The invention relates to an infant teaching resource intelligent searching and recommending system based on an AI large model. The system comprises a resource acquisition module, a label indexing module, an A I semantic understanding module, a resource screening and adapting module, a recommendation sorting module and a compliance control module. Teachers input teaching requirements through natural languages, voices, teaching plans or pictures, and the system performs semantic analysis and keyword extension based on a large language model and completes multi-modal content retrieval in combination with a resource map. The system introduces a complexity analysis and right age adaptation mechanism, intelligently simplifies and reconstructs contents, and generates a personalized recommendation list based on teacher portraits and a collaborative filtering algorithm. And the resource use process is constrained by the compliance control module to ensure that the content copyright is legal. The system can significantly improve the lesson preparation efficiency of kindergarten teachers and the teaching resource adaptation degree, and is suitable for a digital teaching platform and an intelligent education system.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary applications of artificial intelligence and educational informatization, specifically a multimodal intelligent search and recommendation system for early childhood education resources based on an AI large language model, which aims to serve teachers in the 3- to 6-year-old early childhood education stage in their teaching preparation, resource acquisition, and personalized recommendation tasks. Background Technology

[0002] From a technological development perspective, the emergence of AI-based large language models such as the Transformer architecture (Vaswani et al., 2017) and the BERT model (Devlin et al., 2018) has laid the theoretical foundation for semantic understanding, multimodal processing, and personalized recommendations. Furthermore, the development of multimodal fusion technology (Lu et al., 2019) and knowledge graph augmentation technology (Wang et al., 2014) has also provided technical support for the intelligent processing of educational resources. Simultaneously, research findings in the field of educational informatization, such as Bloom's theory of educational objective classification (Bloom, 1956) and Brusilovsky's adaptive hypermedia method (Brusilovsky, 1996), have provided pedagogical basis for the development of intelligent systems.

[0003] This system innovatively combines AI large-scale model technology with the characteristics of early childhood education, demonstrating significant advantages in semantic understanding, multimodal processing, and personalized recommendations. By integrating knowledge graphs and adaptive learning mechanisms, the system can deeply understand user needs and provide accurate resource matching services, effectively solving the problems of low relevance and poor adaptability in traditional systems.

[0004] In current early childhood education practice, the quality and efficiency of accessing teaching resources directly impact teachers' lesson preparation and classroom performance. Traditional search engines (such as Baidu and Google) often struggle with early childhood education resource searches due to issues like excessive generalization, insufficient semantic recognition capabilities, and a lack of age-appropriate labeling. This leads to lengthy resource acquisition processes and heavy filtering burdens for teachers, severely limiting the availability of teaching resources. Meanwhile, while AI large-scale model technology has enhanced natural language processing and multimodal semantic understanding capabilities, its application in vertical education sectors (especially early childhood education) still lacks systematic integrated solutions. Existing educational intelligence systems have yet to construct a highly adaptable, semantically accurate, and content-structure-diverse intelligent learning resource acquisition system tailored to early childhood teachers. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of poor resource compatibility, weak semantic recognition ability, and poor recommendation effect in traditional search tools and existing educational platforms, and to provide an intelligent search and recommendation system for early childhood teaching resources based on an AI large model.

[0006] To achieve the above objectives, this invention designs a system architecture that integrates multiple functional modules, including resource acquisition and management, semantic understanding and expansion, complexity analysis and content adaptation, collaborative filtering recommendation and interest modeling, and resource compliance control. This system is designed for kindergarten teachers, enabling them to acquire highly adaptable and diverse teaching resources through natural language input, thereby improving teachers' lesson preparation efficiency and the quality of teaching content.

[0007] The technical solution of this system includes the following core modules:

[0008] Resource Acquisition and Structured Content Module: This module acquires multimodal educational content, including text, audio, video, interactive courseware, and AR 3D models, from multiple educational resource libraries through an asynchronous acquisition mechanism and authorized API interfaces. After resource acquisition, the system automatically tags and labels the resources based on teaching themes (such as seasons, animals, and festivals), activity types (language, art, and science), and age-appropriate grading standards (3-4 years, 4-5 years, and 5-6 years), constructing a resource map index database.

[0009] Natural Language Semantic Recognition and Expansion Module: Teachers input query commands through the interface, such as "short videos suitable for middle-grade children to talk about rainbows." The system calls BERT, ChatGLM, or T5 models to encode the input statement into word vectors and calculates semantic similarity in the resource tag word vector space. Through contextual analysis and semantic expansion, the system can automatically expand keywords such as "light refraction," "post-rain phenomena," and "rainbow principle" to enrich the search scope and content diversity.

[0010] In this system, semantic similarity is calculated using the following formula:

[0011]

[0012] Where: Q is the retrieval command vector input by the teacher, T i For the tag word vectors, "·" denotes the vector dot product, ||Q|| and ||T i || represents the modulus of each element. This formula is used to measure the degree of semantic matching between the input intent and the resource tag.

[0013] Resource filtering and adaptation module: The filtering module incorporates metrics such as syntactic complexity, video length, speech rate, and text reading difficulty to exclude content unsuitable for the target age group by setting age-appropriate thresholds. For some more challenging resources, the adaptation module performs intelligent content reconstruction operations, including:

[0014] Simplification of text and illustrations and transformation of picture book style;

[0015] Keyframe extraction and segment reconstruction of video content;

[0016] Segmentation and prompt-based playback conversion of long audio or voice content;

[0017] The user manual for AR models has been generated and the operation process has been simplified.

[0018] To facilitate the quantification of the language difficulty of text resources and assist in the selection of adaptation modules, the system introduces a text complexity estimation function:

[0019] C text =λ1·Avg len +λ2·V diff +λ3·F rate

[0020] Among them: Avg len Indicates the average sentence length; V diff Indicates the percentage of uncommon or technical terms; F rate This represents the proportion of words that appear more frequently than those in the primary school Chinese vocabulary list; λ1, λ2, and λ3 are empirical weights (e.g., 0.3, 0.5, 0.2). If the estimated result exceeds the system threshold, it is automatically marked as a "slightly complex" resource.

[0021] Teacher Behavior Modeling and Recommendation Module: This module records teachers' behavioral characteristics such as resource clicks, favorites, downloads, and usage time, forming personalized interest profiles. The recommendation strategy combines collaborative filtering and interest-weighted scoring mechanisms to personalize and optimize the ranking of new search results, prioritizing the display of content related to the teacher's usage habits and teaching topics, and has the ability to provide real-time feedback and adjustments.

[0022] During the recommendation and ranking process, the system uses the following comprehensive scoring function to prioritize candidate resources:

[0023] R score =α×Sim sem +β×Sim int +γ×Sim age +δ×Sim pop

[0024] Where: Sim sem Indicates semantic similarity; Sim int Indicates interest fit; Sim age Indicates age group suitability; Sim pop This represents the resource popularity score (such as click-through rate, number of favorites, etc.); α, β, γ, and δ are the system parameter tuning weights, satisfying α+β+γ+δ=1.

[0025] Furthermore, the teacher interest profile is dynamically updated using a vector weight mechanism, with the update formula as follows:

[0026]

[0027] Where: i k η represents the interest score of the k-th label; η is the learning rate (e.g., 0.1); f action The behavior type is scored (click = 1, favorite = 2, download = 3). This formula is used to dynamically adjust teachers' interest preferences and improve the personalized response capability of the recommendation system.

[0028] Resource Compliance Control Module: During the content indexing and distribution process, the system automatically verifies the usage authorization type of each resource, allowing only resources with legal sources, clear labeling, and CC licenses or institutional authorization to participate in the recommendation output, thus ensuring the compliance and commercial security of resource use in educational activities.

[0029] The workflow of this system is as follows: Teachers input natural language query commands → AI module recognizes semantics and expands keywords → The system performs fuzzy matching in the structured resource library → Several candidate resources are initially retrieved → Complexity and age-appropriateness screening is performed → The adaptation module reconstructs some content → The recommendation module sorts the remaining resources → The final results display page is returned, and the resources are ranked comprehensively according to factors such as suitability, educational value, and teacher interest.

[0030] Compared with existing educational platforms or search tools, this invention has the following outstanding advantages:

[0031] 1. High semantic understanding capability: It can recognize complex natural language input such as teachers' teaching intentions, students' learning objectives, and content preferences;

[0032] 2. Multimodal resource adaptation capability: The system supports unified retrieval and adaptation of text, audio, video, interactive resources, and AR models;

[0033] 3. Age-appropriate intelligent simplification mechanism: Content complexity assessment and automatic reconstruction technology ensure that the content is highly consistent with children's cognitive abilities;

[0034] 4. Personalized recommendations enhance teaching efficiency: Based on teacher profiles, the system accurately predicts the required content, significantly reducing search and filtering time;

[0035] 5. Compliant, transparent and traceable: Each resource has authorization information and usage scope identification, making it easy for teachers to use it with peace of mind.

[0036] This system is applicable to various use cases, including daily teaching in kindergartens, resource management of educational platforms, lesson preparation support for teaching and research activities, content distribution of educational apps, and family education engine. It has significant educational technology value and promising prospects for widespread application. Attached Figure Description

[0037] Figure 1 This is a diagram showing the overall architecture of the system of the present invention.

[0038] Figure 2 This is a flowchart of the resource acquisition and structuring module.

[0039] Figure 3 This is a structural diagram of the Natural Language Semantic Recognition and Extension Module.

[0040] Figure 4 This is a flowchart of the resource complexity analysis and adaptation module.

[0041] Figure 5 This is a flowchart of the teaching video editing and structural reconstruction module.

[0042] Figure 6 A schematic diagram of the module for modeling teacher interests and updating weights for personalized recommendations.

[0043] Figure 7 Flowchart for automatic detection and tag addition in the resource compliance control module. Detailed Implementation

[0044] To more clearly illustrate the technical solution of this invention, the following detailed description of the system's composition and operation process, combined with the system structure diagram and the workflow diagram of each module, is provided. The intelligent search and recommendation system for early childhood teaching resources based on an AI large-scale model described in this invention mainly consists of six functional modules: a resource acquisition and structuring module, a natural language semantic recognition and expansion module, a resource filtering and adaptation module, a teacher behavior modeling and recommendation module, a resource compliance control module, and a user interface and interaction module. These modules form a complete closed loop through data sharing and communication protocols, and are uniformly connected to a multimodal educational resource database and a teacher profile behavior database, supporting online retrieval, recommendation, and lesson preparation operations.

[0045] Example

[0046] Please see Figure 1 The system adopts a distributed service architecture, with each core module deployed as a functional sub-service on the cloud platform, supporting elastic invocation and horizontal scaling. The system as a whole is triggered by a front-end query command, which calls an AI large model for semantic parsing and completes semantic vector retrieval and matching in a structured tag graph. Subsequently, it goes through content adaptation, interest recommendation, and compliance screening modules, and finally returns a structured list of educational resources to the user interface.

[0047] (I) Resource Acquisition and Structured Module

[0048] The resource acquisition module is used to extract multimodal early childhood education resources, including text, images, videos, audio, picture books, courseware, and AR 3D models, from multiple mainstream educational content platforms and authorized content providers. This module consists of three parts: an asynchronous acquisition unit, a resource cleaning submodule, and a structured annotation system.

[0049] The data collection methods are divided into open API crawling and a self-managed crawler scheduling mechanism. The authorized platform provides structured resource call interfaces through the API protocol, and the system performs data synchronization updates at fixed intervals. Unstructured web page content is executed by an asynchronous crawler system on a scheduled basis, ensuring high-coverage content acquisition.

[0050] The collected resources first undergo a cleaning process. The system automatically removes advertising content, platform identification information, and files with incorrect formats. It then performs OCR to extract text from images, extracts descriptive keywords from video frames, and performs ASR transcription on audio files. After format standardization, all resources enter the structured annotation process.

[0051] Please see Figure 2 The structured annotation module labels each resource with a five-dimensional tagging system based on its teaching theme, applicable age group, activity type, resource format, and content complexity. For example, a video teaching resource for middle-grade students about the butterfly life cycle will be tagged with: "middle grade", "science", "video", "insect theme", and "illustrated + narrated".

[0052] All tag information is entered into the resource graph database in the form of triples (resource-attribute-value). This graph provides retrieval paths and semantic weights for subsequent semantic query and recommendation algorithms.

[0053] (II) Natural Language Semantic Recognition and Extension Module

[0054] Please see Figure 3 This module is one of the core technology modules of this system. Teachers can enter natural language search commands through the search bar, such as "video courseware suitable for middle-grade students to talk about the water cycle". The system first calls the AI ​​semantic modeling submodule to encode the input command into word vectors. The AI ​​model used is a fine-tuned language model that integrates the BERT structure and pre-trained with Chinese basic education corpora, and has strong semantic generalization ability in the education field.

[0055] Through syntactic analysis, named entity recognition, and dependency parsing, the system identifies the main teaching intentions, target audience, formal requirements, and core themes from the input statements. For example, the above input content will be parsed as: "Target audience: middle class", "Teaching objective: understanding the water cycle phenomenon", "Content format: video", "Keywords: water cycle, evaporation, precipitation", etc.

[0056] In addition, the system introduces a semantic expansion mechanism to generate keyword expansion groups based on the synonyms, near-synonyms and hierarchical relationships of existing tags in the resource graph. For example, "water cycle" can be expanded to "rainwater formation", "evaporation and condensation", "weather change", etc., in order to improve resource recall rate and matching quality.

[0057] Each time a teacher queries, the resulting instruction vector is compared with the existing tag word vectors in the system. The Top-N relevant tags are obtained through the cosine similarity algorithm, and a multi-path retrieval sequence is constructed. This sequence is used to evoke a set of candidate resources in the resource database that match the tags.

[0058] To measure the matching degree between input semantics and resource tags, the system uses the following semantic similarity calculation formula:

[0059]

[0060] Where Q is the natural language retrieval vector input by the teacher, and Ti is the tag word vector; "·" denotes the dot product, ||Q|| and ||T||. i || represents the respective modulus.

[0061] Example 1:

[0062] Let Q = [0.4, 0.3, 0.6], T i =[0.5,0.2,0.7]

[0063] The dot product = 0.4 × 0.5 + 0.3 × 0.2 + 0.6 × 0.7 = 0.2 + 0.06 + 0.42 = 0.68

[0064]

[0065] Sim sem ≈0.68 / (0.781×0.883)≈0.987

[0066] (III) Resource Selection and Adaptation Module

[0067] The main function of this module is to further determine whether the candidate resources, which are initially retrieved and are related to the query intent, meet the cognitive abilities, classroom time constraints, and teaching format requirements of the target learners, thereby improving the teaching usability and age-appropriateness of the resources.

[0068] Please see Figure 4 The resource filtering module consists of a complexity discrimination unit and a content adaptation unit. The complexity discrimination unit extracts features from multiple dimensions such as text, audio, video, and images for each candidate resource and compares them with the system's built-in cognitive threshold standard. Specifically, this includes:

[0069] Text Resources: The system analyzes the average sentence length, keyword semantic level, and frequency of professional terms. If a large number of adult science terms (such as "reflectivity" and "water vapor content") are found in the text, it is marked as "slightly complex".

[0070] To assess whether the text content is suitable for young children's cognitive level, the system sets the following text complexity function:

[0071] C text =λ1·Avg len +λ2·V diff +λ3·F rate

[0072] Among them, Avg len V represents the average sentence length. diff For the proportion of uncommon words, F rate λ1, λ2, and λ3 represent the frequency of uncommon words; λ1, λ2, and λ3 are weight parameters, typically with values ​​of 0.3, 0.5, and 0.2, respectively.

[0073] Example 2:

[0074] Avg len =18, V diff =0.32, F rate =0.27

[0075] C text =0.3×18+0.5×0.32+0.2×0.27=5.4+0.16+0.054=5.614

[0076] Video Resources: The system analyzes the total video length (recommended not to exceed 6 minutes), speech rate (recommended not to exceed 180 words / minute), subtitle synchronization, and the appropriate frequency of scene transitions. If the scene changes frequently, the narration is unclear, or there is a lack of subtitles, it is recommended not to recommend the video or to proceed with the adaptation process.

[0077] Audio resources: This mainly analyzes whether the speech clarity, sentence structure, intonation, and rhythm of the language are acceptable to children. If the rhythm is monotonous or the content is lengthy, the system will automatically suggest "segmented playback" or a combination of text and images for supplementary information.

[0078] Image resources and interactive courseware: Analyze the page information density, whether elements have visual guide lines, and whether the images and text form a consistent semantic field. If the content in the image is abstract or lacks guiding text, the adaptation module will add supplementary explanatory text or restructure the layout.

[0079] The content adaptation unit is responsible for restructuring the structures of the resources marked as "relatively complex." Adaptation methods include, but are not limited to:

[0080] Simplification of text and image resources: The BART model, which is fine-tuned based on picture book corpus, is used to abbreviate the text and replace words, transforming the adult language style into a child-friendly expression style, and adding facial expressions, interjections, onomatopoeia, etc.

[0081] Video content editing: Identifying core teaching segments (e.g., using a SceneDetect-based LSTM keyframe extraction algorithm) Figure 5 The experimental segments (such as evaporation and condensation) shown in the video were compressed into an animated short film of no more than 3 minutes, with added narration.

[0082] Audio content graphic culture: For lengthy voice broadcasts, the system transcribes them into short text segments and matches them with illustrations to form printable "sound and voice synchronization" graphic materials.

[0083] AR model simplification: Complex 3D structures will be replaced with cartoon models with fewer nodes, while adding child-friendly operation commands and voice prompts such as "one-click rotation" and "click to display information".

[0084] All reconstructed resources will be labeled "AI Simplified Version," while the original version is also available for teachers to evaluate and choose from. Appropriate resources will be prioritized in search results and will include an "Age-Appropriateness Index."

[0085] (iv) Teacher Behavior Modeling and Recommendation Module

[0086] The teacher behavior modeling module dynamically updates its interest profile model based on various behavioral records of teachers on the platform (such as search terms, click records, saved resources, download frequency, browsing time, etc.). This module uses a combination of vector space modeling and collaborative filtering, which can both construct interest vectors based on teachers' historical preferences and supplement recommendations by referring to the behavior of similar user groups.

[0087] Please see Figure 6 The recommendation module workflow includes the following steps:

[0088] Teacher interest vector generation: The system calculates the frequency of user behavior on various tags (such as "science videos", "kindergarten activities" and "nature themes") by weighting, and forms a set of interest weight vectors.

[0089] Similar user identification: The system uses KNN (k-nearest neighbor algorithm) to find similar teacher groups in a multi-dimensional interest space and build a collaborative reference pool.

[0090] Recommendation candidate generation: For initially matched resources, the system calculates the "semantic relevance score (Sim_sem)," "interest suitability score (Sim_int)," "age suitability score (Sim_age)," and "click popularity score (Sim_pop)," and generates a recommendation ranking index by weighting them according to the following function:

[0091] R score =α×Sim sem +β×Sim int +γ×Sim age +δ×Sim pop

[0092] Where α, β, γ, and δ are empirical parameter tuning coefficients, and satisfy α+β+γ+δ=1.

[0093] The default configuration is α=0.4, β=0.3, γ=0.2, δ=0.1, and teachers can make personalized adjustments in the settings.

[0094] Example 3:

[0095] Sim sem =0.88, Sim int =0.90, Sim age =0.80, Sim pop =0.60

[0096] α=0.4, β=0.3, γ=0.2, δ=0.1

[0097] R score =0.4×0.88+0.3×0.90+0.2×0.80+0.1×0.60=0.352+0.27+0.16+0.06

[0098] =0.842

[0099] Results sorting and display: The top-N resources will be displayed on the interface, prioritizing resources tagged with "frequently clicked by teachers", "highly relevant to course objectives", and "updated this week".

[0100] The recommendation results have self-learning capabilities. If a teacher clicks on or saves a certain type of resource a high number of times, the system will automatically increase the ranking priority of that type of content in the results in the next search, forming an "intelligent preference memory".

[0101] To support dynamic updates of teacher profiles, the system is designed with the following interest weight update function:

[0102]

[0103] Among them, i k Let f be the interest value for the k-th class label, η be the learning rate (e.g., 0.1), and f be the learning rate. action This represents the behavioral feedback value (click = 1, favorite = 2, download = 3).

[0104] Example 4:

[0105]

[0106] Based on the increase in interest value, the system prioritizes recommending resources with this type of tag, forming a dynamic learning feedback loop, and is used to adjust personalized ranking strategies in real time.

[0107] Meanwhile, the system supports a "feedback mechanism." If a teacher explicitly marks a resource as "useless," "inappropriate," or "too difficult," the system will automatically filter that type of resource in the next search, further optimizing the user experience.

[0108] (V) Resource Compliance Control Module

[0109] This module is used to ensure that the system's output resources comply with national education policies and teaching platform standards in terms of content copyright, scope of use authorization, and educational legality, and to prevent copyright disputes from arising during teachers' use in the classroom.

[0110] Please see Figure 7 The system initiates an authorization verification process during the resource entry stage, conducting preliminary screening based on the authorization agreements recorded in the resource metadata (such as "CC-BY", "CC-BY-NC", "Educational Use Authorization", etc.). For content without an indicated authorization source or with unclear copyright ownership, the system will automatically label it "Not Recommended" and exclude it from the recommendation pool.

[0111] During the resource display phase, the system will prominently display a usage license summary below each resource, such as: "This resource can be used for educational purposes and non-commercial sharing," "This video is for preview only and cannot be downloaded," "The resource belongs to XX Publishing House, and the copyright has been clearly authorized," etc., so that teachers can determine whether it can be included in formal teaching materials.

[0112] The system also includes a manual review interface, where platform administrators periodically conduct manual inspections of the collected content to confirm there are no obvious copyright infringement risks or repeated hotlinking. For disputed resources, the system can automatically disable them and notify the content provider to update the authorization status.

[0113] (vi) User Interface and Interaction Module

[0114] The user interaction module serves as the core bridge connecting teachers and the system. Its goal is to simplify the complex AI search and multimodal recommendation process into a low-cognitive-burden process of "input-confirmation-retrieval," thereby improving usability and acceptance.

[0115] Teachers can access resources in the following four ways:

[0116] Natural language text input: such as "Find a picture book suitable for a small class to talk about insect metamorphosis";

[0117] Voice input: The system connects to Whisper or a speech recognition API to transcribe commands into text and recognize keywords in real time;

[0118] Lesson plan upload: Supports .docx or .pdf format. The system automatically parses the title, chapter, and teaching objectives to generate recommended keywords.

[0119] Image input: Upload photos of teaching aids, picture books, or cards. The system generates label vectors for querying through image recognition and semantic back-inference (such as the CLIP model).

[0120] Teachers can manually filter the recommended results, such as switching the sorting logic to "video priority", "image and text priority" or "this week's update priority", and participate in the system feedback mechanism by scoring, collecting, and reporting, so as to promote the dynamic fine-tuning of the model.

[0121] The search results support one-click export, automatic organization into lesson preparation packages, PPT template filling, or generation of QR codes for family sharing. For example, in the "Spring-themed Activities" search, teachers can choose to generate an integrated package that includes a PPT with illustrations and text, audio of spring poems, printable sticker templates, and a 3D AR model for classroom teaching and home-school collaboration.

[0122] (VII) Usage Procedures and Application Examples in Teaching Scenarios

[0123] To further illustrate the working logic of this system, a typical teaching process is given below:

[0124] Teacher input: "Animated videos about butterfly growth suitable for middle-grade children";

[0125] The system identifies keywords such as "middle class", "butterfly", and "animated video", expands them into tags such as "caterpillar", "breaking out of the cocoon", and "metamorphosis", converts them into semantic vectors Q, and compares their similarity with the tag vectors Ti of each resource in the tag library.

[0126] The system uses the following semantic similarity calculation formula:

[0127]

[0128] If Q=[0.4,0.3,0.6], T i =[0.5,0.2,0.7]

[0129] Then Q·T i =0.4×0.5+0.3×0.2+0.6×0.7=0.68

[0130]

[0131] Sim sem≈0.68 / (0.781×0.883)≈0.986

[0132] This indicates that the query semantics highly match resources related to "butterfly metamorphosis";

[0133] The system retrieves picture book videos, teaching demonstration clips, interactive courseware, etc. that match the tags from the graph;

[0134] The system performs content complexity analysis on the recalled resource text, and uses the following function:

[0135] C text =λ1·Avg len +λ2·V diff +λ3·F rate

[0136] If the text has Avg len =18, V diff =0.32, F rate =0.27

[0137] If the empirical weights are λ1 = 0.3, λ2 = 0.5, and λ3 = 0.2, then...

[0138] C text =0.3×18+0.5×0.32+0.2×0.27=5.4+0.16+0.054=5.614

[0139] If the system determines that the content exceeds the age-appropriate recommendation limit (e.g., 5.0), it will trigger content editing to generate a 3-minute abridged version.

[0140] The sorting module performs a comprehensive score on streamlined resources using the following function:

[0141] R score =α×Sim sem +β×Sim int +γ×Sim age +δ×Sim pop

[0142] If the resource rating is Sim sem =0.92, Sim int =0.85, Sim age =0.90, Sim pop =0.70

[0143] If the weights are set as α = 0.4, β = 0.3, γ = 0.2, and δ = 0.1, then...

[0144] R score

[0145] =0.4×0.92+0.3×0.85+0.2×0.90+0.1×0.75=0.368+0.255+0.18+0.075=0.878

[0146] The system determines that the resource's score exceeds the threshold and places it in the Top recommendation slot;

[0147] When a teacher clicks to play and saves the resource, the system records the behavior tag "butterfly metamorphosis" and updates the interest vector I_user;

[0148] The interest weight update function is:

[0149]

[0150] If the original interest value i k =0.55, collection behavior f action =2, learning rate η = 0.1, then

[0151]

[0152] The system has completed the update of interest profiles and will prioritize displaying related topics such as "insect life cycle" in subsequent recommendations;

[0153] Teachers select "Add to Course Package," and the system generates a PPT presentation of the lecture notes.

[0154] Simultaneously generate a QR code for home-school education, which parents can scan to review, forming a closed loop between the classroom and the family.

[0155] Test feedback shows that the system helps teachers reduce lesson preparation time by an average of over 60%, while improving the adaptability of classroom resources and interaction rates. Particularly in thematic monthly lesson design, the system automatically collects past resources and packages them into "Spring Theme Library" and "Festival Customs Material Collection," greatly enhancing teachers' content coordination capabilities.

[0156] In other use cases, such as holiday teaching (Spring Festival, Mid-Autumn Festival), natural phenomena (four seasons, precipitation), scientific knowledge (magnets, buoyancy), and art activities (paper cutting, color matching), the system can complete accurate recommendations and content adaptation through semantic analysis and recommendation scoring mechanisms.

[0157] (viii) System Deployment and Scalability Design

[0158] To ensure the platform's sustainable operation and high availability, the system of this invention supports the following three deployment modes:

[0159] Cloud platform hosting mode: suitable for regional education bureaus and education platforms, and supports integration with existing government cloud infrastructure;

[0160] Local deployment mode: Suitable for standalone kindergartens and educational institutions, providing a simplified containerized service installation;

[0161] SaaS model: Supports monthly rental or individual teacher subscription operation, lightweight front-end + middle platform service separation architecture, and clear interface documentation.

[0162] All modules support microservice encapsulation and API exposure. Educational platforms can call upon the system's search and recommendation capabilities to embed into existing teaching platforms, enabling the injection of "resources as a service" capabilities. In the future, this system can also be extended to more scenarios such as lower elementary grades and special education.

[0163] The system supports the generation of data visualization reports and the export of behavioral statistics, providing regional education management platforms with auxiliary decision-making basis such as "teacher usage heatmap", "resource usage frequency analysis" and "recommendation effect evaluation".

[0164] The system supports language model replacement and knowledge graph update interfaces, facilitating access to next-generation model capabilities such as GPT and GLM4 or synchronous textbook version upgrades.

[0165] (ix) Implementation Results and Application Prospects

[0166] Based on comprehensive functional assessments and teacher survey feedback, this system possesses the following significant advantages:

[0167] Search accuracy: Query intent recognition accuracy exceeds 92%, and effective matching rate is improved by more than 70%;

[0168] Satisfaction with teaching resources: Teachers' average score for the "age-appropriateness" and "classroom applicability" of the recommended content was 4.7 / 5;

[0169] Compliance coverage: Content authorization clarity reaches over 97%, and all recommended resources have a basis for use;

[0170] Scalability: The graph updates and interest models can be continuously iterated online, resulting in low platform operation and maintenance costs;

[0171] Teacher dependency: The reuse rate exceeds 85%, and active teachers initiate searches an average of 3.5 times per week.

[0172] This system is particularly suitable for business scenarios such as regional smart education cloud platforms, kindergarten digital teaching and research systems, and AI-assisted teaching platforms. It not only effectively improves teacher resource acquisition and teaching efficiency but also provides a complete, replicable, and engineering-feasible paradigm for the informatization of early childhood education. The system's core modules have been service-encapsulated, possessing SaaS deployment, API integration, and local embedding capabilities, facilitating its widespread adoption in preschool education institutions, teaching software platforms, and family education scenarios. It demonstrates significant engineering implementability, market transformation prospects, and international promotion potential.

Claims

1. A smart search and recommendation system for early childhood education resources based on an AI large-scale model, characterized in that, The system includes: The resource acquisition module is used to collect multimodal teaching resources such as images, text, audio, video, and AR models from multiple authorized educational platforms and public educational resource libraries, and to perform data standardization and preliminary format conversion on the resource content; The resource labeling and indexing module is used to label the collected multimodal resources with structured tags according to multiple dimensions such as educational theme, target age group, and activity type, and to build a multi-level index relationship table to support rapid semantic-level retrieval. The AI ​​semantic understanding module is built on a large-scale pre-trained language model. After receiving natural language input from teachers, it combines contextual analysis and semantic graph label structure, and uses word vector expansion and similarity matching technology to realize teaching keyword expansion, semantic intent recognition and vector retrieval representation generation. The resource filtering module is used to filter candidate resources that are closest to the user's intent from the resource database based on semantic matching degree and resource tag relationship, and to filter inappropriate content by combining content complexity and age-appropriate judgment model. The content adaptation module is used to reconstruct the format and content of candidate resources, including but not limited to automatic simplification of text content, editing of essential video clips, conversion of text and image content into picture books for younger audiences, and generation of voice broadcast styles. The compliance control module is used to determine the copyright status of resources during the resource crawling and display process, and only allows resources with CC licenses, open source licenses or platform authorizations to appear in search results, and prompts information on resource usage restrictions; The adaptation scoring mechanism module is used to score and rank candidate resources based on the matching degree function R_score, wherein the scoring function is: R score =α×Sim sem +β×Sim int +γ×Sim age +δ×Sim pop Where: Sim sem Indicates semantic similarity; Sim int Indicates interest fit; Sim age Indicates age group suitability; Sim pop This represents the resource popularity score (such as click-through rate, number of favorites, etc.); α, β, γ, and δ are the system parameter tuning weights, satisfying α+β+γ+δ=1; User-based collaborative filtering algorithm, using the formula: User similarity calculation: in: u and v represent two users, r ui R represents the rating of user u for item i. vi This represents user v's rating of item i; Predictive score calculation: in: P(u,i) is the predicted rating of user u for item i, R u R is the average rating for user u. v The average rating for user v; Nearest neighbor selection calculation: N(u) = TopK(Argmax(Sim(u,v))) Select the K neighbors that are most similar to the target user; The recommendation and ranking module, based on user historical behavior data and teacher profile models, combines collaborative filtering algorithms with an adaptation system scoring mechanism to perform weighted ranking of the final resource list; The feedback optimization module collects user data in real time and optimizes system performance.

2. The system according to claim 1, characterized in that, The resource acquisition module collects resources in real time through an asynchronous crawler scheduling mechanism and API data interface, and performs resource cleaning, desensitization processing, and metadata standardization.

3. The system according to claim 1, characterized in that, The AI ​​semantic understanding module uses pre-trained language models such as BERT, ChatGLM3, or T5 to vectorize the natural language input by teachers, and combines semantic expansion algorithms to generate a set of related keywords, thereby improving the educational relevance and content contextual completeness of the search. The semantic similarity Sim sem The calculation method is as follows: Where: Q is the word vector representation of the teacher's input statement, Ti is the label vector, the symbol "·" represents the vector dot product, and ||Q|| and ||T|| are also present. i || represents the vector magnitude, and this similarity value is used to measure the semantic relevance between natural language intent and instructional labels.

4. The system according to claim 1, characterized in that, The content adaptation module includes: The complexity analysis unit is used to determine the sentence structure, vocabulary difficulty, and length of text resources, or the duration, speaking speed, and image density of video resources. The content reconstruction unit is used to generate a younger-adapted version based on the complexity analysis results. This includes operations such as: rewriting text and image summaries, generating video clips, generating voice readings, and simplifying AR model instructions. The style transfer unit is used to transfer the style of resource presentation according to the preferences of the target age group, such as changing the style of adult science videos to cartoon demonstration style; The video summarization unit combines an LSTM-CNN structure to generate essential summaries of teaching videos, which can be distributed in different formats such as "classroom version" and "parent version". The text complexity estimation formula is as follows: C text =λ1·Avg len +λ2·V diff +λ3·F rate Among them: Avg len Indicates the average sentence length; V diff Indicates the percentage of uncommon or technical terms; F rate This indicates the proportion of words that appear more frequently than those in the primary school Chinese vocabulary list; λ1, λ2, and λ3 are empirical weights, which can be set to 0.3, 0.5, and 0.2, respectively.

5. The system according to claim 1, characterized in that, The compliance control module includes: The content security submodule is used for sensitive word filtering, semantic analysis, and risk level assessment. The access control submodule is used for authentication, access control, operation auditing, and data isolation. The privacy protection module is used for key management, user preference settings, and privacy management operations. The compliance review submodule is used for compliance checks, resource audits, and data statistical analysis. The risk control submodule is used for risk matrix analysis, crisis management, vulnerability scanning, and patch management. The operations support submodule is used for operations such as operation log recording, system status monitoring, notification alerts, and trend analysis.

6. The system according to claim 1, characterized in that, The recommendation ranking module includes: The behavioral data analysis unit is used to record and analyze teachers' clicks, downloads, favorites, and sharing actions, and to build behavioral vector models. The collaborative recommendation submodule is used to recommend potential high-quality resources based on the similarity of resource preferences among multiple users, using cosine similarity calculation and matrix factorization techniques. An interest weight update mechanism is used to dynamically update teachers' interest profiles and resource recommendation weights when their resource usage behavior changes significantly. The formula for updating interest weights is as follows: Where: i k η represents the interest score of the k-th label; η is the learning rate (e.g., 0.1); f action User behavior is scored: clicks are 1, favorites are 2, and downloads are 3. The recommendation display control module is used to customize the recommendation list based on the Top-K sorting strategy and the priority of topics specified by the teacher.

7. The system according to claim 1, characterized in that, The system provides a voice input interface, an image reverse search interface, and a course plan embedding interface, allowing teachers to obtain intelligent recommendation results through voice commands, uploaded images, or course schedule files. The interfaces are designed using RESTful principles, support JSON format, and have a response latency of no more than 500ms.

8. A method for retrieving and recommending early childhood education resources using the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1. Teachers input their natural language teaching needs; S2, the AI ​​semantic understanding module parses the input content and generates a keyword set and a retrieval vector; S3. The system performs semantic retrieval in the structured multimodal resource database based on vector matching to obtain a preliminary candidate resource set; S4. The resource filtering module filters out resource content that is not suitable for early childhood education based on tag matching relationship and complexity judgment rules. S5. The content adaptation module reconstructs complex resources to generate resource versions with higher adaptability. S6. The recommendation and sorting module calculates the final resource list based on the teacher user profile and other user behaviors, and combines the recommendation weights. S7. The system outputs a multimodal resource list sorted by matching degree, which supports teachers to preview, favorite, and export PPT.

9. The method according to claim 8, characterized in that, The system can display the age level, educational goal type (cognitive, language, art, etc.), suggested usage time, and interaction methods of resource content in the search results to help teachers quickly assess its suitability for teaching and its operability in the classroom.

10. The method according to claim 8, characterized in that, The system automatically records user feedback after each recommendation, which is used to train the weights of the recommendation model for the next round, thereby improving the system's iteration efficiency and user satisfaction.

Images