A Method for Constructing Knowledge Bases in Early Childhood Education by Integrating Knowledge Graphs and Large Language Models

Abstract

This invention discloses a method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models, belonging to the field of early childhood education method optimization technology. It involves collecting and cleaning preprocessing early childhood education data from multiple channels; extracting entities and relationships using a customized model to construct a structured knowledge graph; fine-tuning the large language model for the early childhood education domain; integrating knowledge from the knowledge graph and model through a gating mechanism with double verification; establishing a monthly update mechanism to achieve incremental learning; furthermore, it achieves personalized knowledge recommendations based on user profiles, optimizes interactive question answering through a combination of retrieval and models, and ensures data security through anonymization and access control. This invention, through the integration of multi-source data and intelligent technologies, constructs a precise and dynamic early childhood education knowledge base, more accurately understands children's needs, generates content that aligns with their cognitive development, achieves personalized learning recommendations, improves the interactive experience and knowledge service quality, while ensuring data security and promoting the digital upgrade of the early childhood education field.

Description

Technical Field

[0001] This invention relates to the field of early childhood education method optimization technology, and in particular to a method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models. Background Technology

[0002] With the rapid development of digital education, the construction of knowledge bases in the field of early childhood education has become a crucial step in improving the quality of early childhood education. However, traditional methods of constructing early childhood education knowledge bases have many limitations and are unable to meet the diverse and personalized needs of early childhood education in the new era.

[0003] Most existing early childhood education knowledge bases are based on static text resources, such as fixed textbooks and lesson plans. Data sources are singular and updates are slow, failing to incorporate the latest educational philosophies and children's cultural content in a timely manner. These knowledge bases often employ simple classification and storage methods, lacking in-depth exploration and structured presentation of knowledge systems. They struggle to systematically demonstrate the connections between knowledge in multiple areas such as language development, cognitive development, and art cultivation, resulting in low knowledge retrieval efficiency and hindering support for personalized teaching and children's independent learning. Furthermore, due to a lack of consideration for children's cognitive characteristics, the presentation of knowledge is often dull and fails to stimulate children's learning interest.

[0004] At the knowledge processing technology level, traditional methods face numerous challenges when applied to early childhood education. During knowledge extraction, due to the colloquial and non-standardized nature of young children's language, conventional named entity recognition and relation extraction techniques struggle to accurately capture key information, easily leading to knowledge omissions or incorrect associations. In terms of knowledge application, rule-based question-answering systems lack flexibility and cannot understand the ambiguous semantics and implicit needs in children's questions; while simple retrieval-based interaction methods lack intelligence, failing to provide vivid and engaging answers based on children's cognitive levels, thus hindering effective human-computer interaction.

[0005] In recent years, large language models and knowledge graph technologies have demonstrated great potential in natural language processing and knowledge management, but their direct application in early childhood education still has significant shortcomings. While large language models possess excellent language generation capabilities, their training data is mostly general corpora, lacking targeted optimization based on early childhood education expertise. The generated content may contain deviations in educational philosophy or expressions that do not align with the cognitive characteristics and psychological needs of young children. Regarding knowledge graph construction, existing technologies often neglect the laws of early childhood cognitive development, with rigid and mechanical designs for knowledge nodes and relationships. They fail to achieve personalized knowledge recommendations and engaging presentations, and struggle to transform abstract knowledge into forms easily understood and accepted by young children, thus failing to fully realize the educational and entertaining value of technology in early childhood education scenarios. Summary of the Invention

[0006] The present invention proposes a method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models, in order to solve the problems mentioned in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models, comprising:

[0008] Data collection steps in the field of early childhood education: Data is collected from early childhood education textbooks, children's picture books, educational videos, and early childhood teachers' lesson plans. The early childhood education textbooks include content in the fields of language, mathematics, and art. The children's picture books are selected from classic award-winning works. The educational videos cover animation and popular science short films. At the same time, unstructured data such as interview records of early childhood education experts and observation logs of kindergarten teaching activities are collected to construct the original dataset.

[0009] Data preprocessing and cleaning steps: The collected text data is segmented using the Jieba segmentation tool combined with a custom thesaurus for early childhood education; duplicate content is removed by calculating text fingerprints using the SimHash algorithm and setting a similarity threshold of 0.9 to filter duplicate text; keyframes are extracted from educational videos using the OpenCV library; for missing data, imputation is used for structured data and missing parts of unstructured text are marked.

[0010] Knowledge extraction and entity recognition steps: Entities, including people, things, and concept categories, are extracted from the data using the Named Entity Recognition (NER) model; the model adopts the BERT-BiLSTM-CRF architecture and is trained on an labeled dataset in the early childhood education domain; simultaneously, relationships between entities are extracted through dependency parsing.

[0011] Knowledge graph construction steps: Extracted entities are used as nodes and relationships as edges. A knowledge graph for early childhood education is constructed using the Neo4j graph database. The node attributes are enriched, and the knowledge graph is mapped to a low-dimensional vector space using the TransE graph embedding algorithm to achieve subsequent integration with a large language model.

[0012] Large Language Model Adaptation and Fine-tuning Steps: The open-source large language model LLaMA is selected as the base model. Supervised fine-tuning is performed using collected early childhood education data, taking into account the characteristics of the early childhood education domain. The cross-entropy loss function is used during the fine-tuning process. Where N is the number of samples, C is the number of categories, yij is the true label (0 or 1) of sample i belonging to category j, and pij is the probability predicted by the model that sample i belongs to category j; at the same time, domain knowledge hints are introduced to enhance the model's understanding of early childhood education scenarios;

[0013] Knowledge fusion and verification steps: The vector representation of the knowledge graph is fused with the input and output of the large language model, using a gating mechanism: G = σ(W1h) kg +W2hllm ), h fusion =G⊙h kg +(1-G)⊙h llm , where h kg For embedding vectors in a knowledge graph, h llm W1 and W2 are the output vectors of the hidden layer of the large language model, W1 and W2 are the learnable weight matrices, σ is the sigmoid function, and ⊙ is the element-wise multiplication. The fused knowledge is verified by a combination of expert review and automated verification. The automated verification uses a contradiction detection algorithm to check the logical consistency of the knowledge.

[0014] Knowledge base update and maintenance steps: Establish a regular update mechanism to collect the latest early childhood education research results and newly published textbook and picture book data every month; use incremental learning methods to update the knowledge graph and large language model. For newly added nodes and edges in the knowledge graph, update the node embedding vectors through graph neural network algorithms; for the large language model, use the parameter fine-tuning technique LoRA.

[0015] Furthermore, it also includes:

[0016] Personalized knowledge recommendation steps: Based on user profile data such as the child's age, learning progress, and interests, calculate the correlation between the user and knowledge points in the knowledge graph: Where S ui Let represent the degree of relevance between user u and knowledge point i, M be the number of relevance factors, wk be the weight of the k-th factor, and rik be the relevance score between user u and knowledge point i under the k-th factor.

[0017] Interactive question-answering optimization steps: When a user asks a question, first use the knowledge graph to search. If a definite answer is found, return it directly; if not found, integrate the question with the knowledge graph information and input it into the large language model; after the model generates an answer, use a text quality assessment model to score it: Score = αP fluency +βP accuracy +γP relevance Where Score is the answer score, P fluency To score language fluency, P accuracy To score for content accuracy, P relevance To determine the relevance score to the question, α, β, and γ are weighting coefficients, and the optimal weighting combination is determined through user testing.

[0018] Furthermore, in the knowledge extraction and entity recognition steps, fuzzy matching and semantic expansion techniques are employed to address the non-standard nature of young children's language expression; a thesaurus is constructed, and the Word2Vec model is used to calculate word semantic similarity. Where vec(w1) and vec(w2) are vector representations of words w1 and w2. When the similarity is greater than 0.7, they are considered synonyms, which improves the ability to recognize entities in children's colloquial expressions.

[0019] Furthermore, in the large language model adaptation and fine-tuning steps, an adversarial training mechanism is introduced; a discriminator D is trained to distinguish between early childhood education texts generated by the fine-tuned model and real texts, while the fine-tuned model G tries to generate texts that will cause the discriminator to misjudge. The loss function is: Where pdata is the distribution of real early childhood education texts and pz is the noise distribution, the realism of the content generated by the model is improved through adversarial training.

[0020] Furthermore, a knowledge conflict detection algorithm is employed in the knowledge fusion and verification steps: for the triples (s, p, o) in the knowledge graph and the knowledge (s', p', o') generated by the large language model, the conflict probability is calculated: When P conflict If the value exceeds the threshold of 0.6, a conflict is identified, triggering a manual review process.

[0021] Furthermore, a version management mechanism is adopted in the knowledge base update and maintenance process; each update generates a new knowledge base version, recording the update content, time, and operator information; the update effect is evaluated by comparing the changes in the knowledge graph structure and the performance indicators of the large language model in different versions.

[0022] Furthermore, in the personalized knowledge recommendation step, the user profile data adopts a dynamic update strategy; based on the behavioral data of children using the knowledge base, the profile weights are updated using a sliding window algorithm. Where wk,t is the weight of the k-th factor at time t, and n is the window size, so that the recommended content can be dynamically adjusted according to the child's learning situation.

[0023] Furthermore, in the interactive question-answering optimization step, an evaluation corpus of question-answer pairs in the early childhood education field was constructed; 2000 sets of question-answer pairs were collected from real teaching scenarios, and 5 early childhood education experts were invited to conduct 5-level scoring, calculating the average score of each question-answer pair as a benchmark; the model-generated answers were compared with the answers in the evaluation corpus, and cosine similarity was used for calculation. in These are vector representations of the model's answer and the baseline answer, respectively, and the answer generation strategy is adjusted based on similarity.

[0024] Furthermore, security measures are adopted throughout the entire construction process; the collected data related to children is anonymized using a hash function: H(x) = hash(x + salt) to achieve data privacy, where x is the original data and salt is a randomly generated salt value; at the same time, role-based access control (RBAC) is used to access the knowledge base, assigning different permissions to children, teachers, and administrators.

[0025] Compared with existing technologies, the beneficial effects of this invention are:

[0026] In terms of knowledge resource construction, this method integrates rich data from textbooks, picture books, videos, and expert experience through a multi-channel collection strategy, and establishes a regular update mechanism to ensure that the knowledge base content is comprehensive and keeps pace with the forefront of education. In-depth data preprocessing and professional knowledge extraction techniques, optimized in conjunction with the characteristics of young children's language, significantly improve the accuracy and completeness of knowledge construction, forming a clearly structured and closely interconnected knowledge system.

[0027] The innovative integration of knowledge graphs and large language models is the core advantage of this method. Knowledge graphs present the logical relationships between knowledge points in a structured manner, providing a solid foundation for knowledge reasoning and personalized recommendations; large language models endow the system with powerful natural language understanding and generation capabilities, accurately interpreting children's vague and colloquial questions and generating vivid, engaging answers that match children's cognitive levels. The synergy between the two significantly enhances the intelligence and fun of the interaction, effectively stimulating children's learning interest and desire to explore.

[0028] The personalized knowledge recommendation feature builds a dynamic user profile based on a child's age, learning progress, and interests, enabling the creation of a customized learning path for each child to meet their diverse learning needs and promote personalized development. Furthermore, this feature can adjust the recommendation strategy in real time based on the child's learning behavior, ensuring that the learning content always aligns with the child's cognitive development pace and improves learning effectiveness.

[0029] The introduction of security protection and version management mechanisms provides reliable guarantees for the stable operation of the knowledge base and data security. Anonymized data processing protects the privacy of children's information, while role-based access control policies prevent knowledge leakage and unauthorized tampering. Version management functions facilitate the evaluation and retrospective analysis of knowledge base updates, ensuring the system continuously and stably provides high-quality services.

[0030] Furthermore, the method proposed in this application not only provides young children with scientific and engaging learning tools, but also offers efficient support for early childhood teachers in acquiring teaching resources, designing curricula, and conducting teaching assessments. This helps to promote the transformation of early childhood education from an experience-driven to a data-driven, intelligent decision-making model, and facilitates the overall improvement of digitalization in the field of early childhood education. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models, as proposed in this invention.

[0032] Figure 2 This diagram illustrates the comparison of knowledge extraction and entity recognition accuracy in the early childhood education knowledge base construction method that integrates knowledge graphs and large language models, as proposed in this invention.

[0033] Figure 3 This diagram illustrates a satisfaction survey of personalized knowledge recommendation based on the method for constructing a knowledge base in the early childhood education field that integrates knowledge graphs and large language models, as proposed in this invention. Detailed Implementation

[0034] 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 embodiments of the present invention, and not all embodiments. 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.

[0035] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0037] Reference Figures 1 to 3 : A specific implementation method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models

[0038] 1. Data collection steps in the field of early childhood education

[0039] Data was collected from multiple sources, including early childhood education textbooks, children's picture books, educational videos, and lesson plans from kindergarten teachers. The textbooks covered eight major areas, including language, mathematics, and art. Over 200 classic, award-winning children's picture books were selected. Educational videos encompassed animation, short science videos, and other types, totaling over 500 hours. Additionally, unstructured data such as interview records with early childhood education experts and observation logs of kindergarten teaching activities were collected to construct the original dataset.

[0040] This method collects data from the early childhood education field through multiple channels, covering both structured and unstructured data types. Regarding textbooks and picture books, it collaborates with 10 well-known early childhood education publishers to acquire over 1200 electronic textbooks across eight major areas, including language, mathematics, and art, and simultaneously collects over 300 classic children's picture books. Educational video data is crawled from mainstream educational platforms, including science animations and craft tutorials, totaling over 600 hours, and video subtitles are obtained via API interfaces. Simultaneously, the team conducted on-site surveys of 50 kindergartens, collecting over 800 teacher lesson plans, 1200 teaching activity observation logs, and recording 200 hours of expert interview videos, which were then transcribed into text. All data is stored in the HDFS distributed file system, and a data index directory is established for easy subsequent retrieval and access.

[0041] 2. Data Preprocessing and Cleaning Steps

[0042] The collected text data underwent word segmentation using the Jieba word segmentation tool combined with a custom vocabulary for early childhood education to identify specialized terms such as "sensitive period for young children" and "Montessori teaching method." Duplicate content was removed by calculating text fingerprints using the SimHash algorithm and setting a similarity threshold of 0.9 to filter duplicate text, which, according to tests, reduced data redundancy by approximately 18%. Keyframe extraction was performed on educational videos, extracting one frame per second using the OpenCV library, followed by audio extraction and conversion to text format using FFmpeg. For missing values ​​in structured data, a multiple imputation method based on random forest was used; missing parts of unstructured text were marked as "[missing]" and subsequently filled in manually or through knowledge reasoning.

[0043] 3. Knowledge Extraction and Entity Recognition Steps

[0044] Named Entity Recognition (NER) is used to extract entities from the data, including categories such as people (e.g., "Sun Wukong" and "Andersen"), objects (e.g., "building blocks" and "piano"), and concepts (e.g., "color recognition" and "digital literacy"). The model adopts the BERT-BiLSTM-CRF architecture and is trained on an labeled dataset in the early childhood education field, achieving an F1 score of 92% for entity recognition. Simultaneously, dependency parsing is used to extract relationships between entities, such as "Snow White - belongs to - fairy tales" and "building blocks - used for - shape recognition".

[0045] The named entity recognition model uses the BERT-BiLSTM-CRF architecture and is trained on a self-built corpus containing 100,000 labeled data points in the early childhood education domain. The model's input layer converts text into character vectors, extracts features using a BERT pre-trained model, and then performs sequence labeling using a bidirectional LSTM and CRF layer. Ultimately, it identifies six major entity categories, including people, things, and concepts, achieving an F1 score of 92.3% in testing. In the relation extraction stage, Stanford CoreNLP is used for dependency parsing, defining 20 common relation types in the early childhood education domain (such as "belongs to," "used for," and "cultivates"). Relationships between entities are extracted through a combination of rule matching and machine learning. For example, from the sentence "Building blocks can help young children with shape recognition," the relation triple (building blocks, used for, shape recognition) can be extracted.

[0046] 4. Knowledge Graph Construction Steps

[0047] Extracted entities are used as nodes, and relationships as edges. A knowledge graph for early childhood education is constructed using the Neo4j graph database. Node attributes are enriched; for example, character nodes include attributes such as "character appearance" and "works featured," while object nodes include attributes such as "age of use" and "educational goals." A graph embedding algorithm (such as TransE) is used to map the knowledge graph to a low-dimensional vector space, facilitating subsequent integration with large language models. The vector dimension is set to 200 dimensions.

[0048] A knowledge graph for early childhood education was constructed using the Neo4j graph database, storing extracted entities as nodes and relationships as edges. Each node has rich attributes; for example, character nodes include attributes such as "character description," "list of works appearing," and "educational significance," while object nodes include attributes such as "applicable age range," "usage scenario," and "educational goal." The TransE algorithm was used to map the knowledge graph to a 200-dimensional vector space, with 500 training epochs and a learning rate of 0.01. The translation loss function was minimized: L = ∑ (h，rt)∈S ∑ (h′r′t′)∈S′ max[0, d(h+r, t)] 2 -d(h′+r′,t′) 2+γ] where S is the set of positive sample triples, S' is the set of negative sample triples, d(x,y) is the Euclidean distance between vectors x and y, and γ is the margin hyperparameter (set to 1.0). Through training, entities with similar semantics are brought closer together in the vector space.

[0049] 5. Large Language Model Adaptation and Fine-tuning Steps

[0050] We selected an open-source large language model (such as LLaMA) as the base model and performed supervised fine-tuning using collected early childhood education data, tailored to the specific characteristics of the early childhood education field. During the fine-tuning process, the cross-entropy loss function was employed. Where N is the number of samples, C is the number of categories, yij is the true label (0 or 1) of sample i belonging to category j, and pij is the probability predicted by the model that sample i belongs to category j. Simultaneously, a domain knowledge prompting engineering is introduced, adding prompts such as "in early childhood math education" and "for 3-4 year old children" before the input text to enhance the model's understanding of early childhood education scenarios.

[0051] We chose an open-source LLaMA model as the base and performed supervised fine-tuning using 500,000 collected datasets from the early childhood education domain. The fine-tuning process employed the PyTorch framework, setting the batch size to 32, the number of training epochs to 10, and using the AdamW optimizer with a learning rate of 5×10⁻⁶. -5 The loss function used is cross-entropy loss: Where N is the number of samples, C is the number of categories, yij is the true label (0 or 1) of sample i belonging to category j, and pij is the probability predicted by the model that sample i belongs to category j. Simultaneously, a prompting engineering library containing 500 prompt word templates is constructed, such as "When 3-4 year old children learn colors, they can be taught through [specific methods]", adding relevant prompt words before the input text to enhance the model's understanding of early childhood education scenarios.

[0052] 6. Knowledge Integration and Verification Steps

[0053] The vector representation of the knowledge graph is fused with the input and output of the large language model, using a gating mechanism: G = σ(W1h) kg +W2h llm ),h fusion =G⊙h kg +(1-G)⊙h llm , where h kg For embedding vectors in a knowledge graph, h llm W1 and W2 are the output vectors of the hidden layer of the large language model, respectively. σ is the sigmoid function, and ⊙ represents element-wise multiplication. The fused knowledge is validated through a combination of expert review and automated verification. The automated verification uses a contradiction detection algorithm to check the logical consistency of the knowledge.

[0054] A gating mechanism is employed to fuse information from a knowledge graph and a large language model. First, the embedding vector hkg of the knowledge graph nodes and the hidden layer output vector hllm of the large language model are input into the gating unit: G = σ(W1h) kg +W2h llm ), h fusion =G⊙h kg +(1-G)⊙h llm Wherein, W1 and W2 are learnable weight matrices with dimensions of 200x200 and 2048x200 respectively (assuming the LLaMA hidden layer dimension is 2048), σ is the Sigmoid function, and ⊙ represents element-wise multiplication. The fused knowledge is verified through a dual verification mechanism: on the one hand, a review panel of 10 early childhood education experts is invited to manually review 1000 randomly selected knowledge items; on the other hand, an automated verification algorithm is used to detect logical contradictions in the knowledge, such as checking for conflicting statements like "an activity is suitable for 3-4 year old children" and "the difficulty of this activity is only suitable for 5-6 year old children." When the probability of conflict, Pconflict, is greater than the threshold of 0.6, manual review is triggered.

[0055] 7. Knowledge Base Update and Maintenance Steps

[0056] A regular update mechanism is established to collect the latest early childhood education research findings, newly published textbooks and picture books, and other data monthly. Incremental learning methods are used to update the knowledge graph and the large language model. For newly added nodes and edges in the knowledge graph, the node embedding vectors are updated using graph neural network algorithms. For the large language model, parameter-efficient fine-tuning techniques (such as LoRA) are used to update only a small number of adapter parameters, reducing computational resource consumption.

[0057] A monthly update mechanism was established, with the data collection module activated at the beginning of each month to acquire the latest early childhood education research findings, newly published textbooks and picture books, and other data. Incremental learning was employed to update the knowledge graph and the large language model: for the knowledge graph, the GraphSAGE algorithm was used to update the embedding vectors of newly added nodes, with the aggregator type set to "mean," and training was performed for three epochs; for the large language model, LoRA technology was used, freezing most of the parameters of the pre-trained model and training only the newly added adapter parameters (approximately 0.1% of the total parameters), using the RMSProp optimizer with a learning rate set to 1×10⁻⁶. -4 The training process was repeated for 5 rounds. Simultaneously, version information for each update was recorded, including update time, update content, and operator details. A version comparison report was generated, and the update effectiveness was evaluated by comparing the number of knowledge graph nodes, relationship change rate, and the perplexity and accuracy of the large language model on the test set for different versions.

[0058] 8. Personalized Knowledge Recommendation Steps

[0059] Based on user profile data such as children's age, learning progress, and interests, the correlation between users and knowledge points in the knowledge graph is calculated: Where S ui Let M represent the relevance between user u and knowledge point i, M be the number of relevance factors, wk be the weight of the k-th factor, and rik be the relevance score between user u and knowledge point i under the k-th factor. Personalized learning content is recommended to preschoolers based on the relevance ranking; for example, "Children's Drawing Introduction" related knowledge is prioritized for 4-year-olds interested in drawing.

[0060] Dynamic user profiles are built based on data such as children's age, learning progress, and interests. User profile data is stored in a MongoDB database, and profile weights are updated using a sliding window algorithm (window size n is set to 30 days). Where wk,t is the weight of the k-th factor at time t, and rik is the relevance score of user u to the k-th factor at time i. Calculate the association degree between the user and knowledge points in the knowledge graph: Where M represents the number of correlation factors (set to 5, including age matching degree, learning progress matching degree, interest relevance, knowledge difficulty matching degree, and educational goal matching degree). Knowledge points are ranked according to their relevance, and the top 10 relevant knowledge contents are recommended to children, displayed as animated cards through a visual interface.

[0061] 9. Steps to optimize interactive question answering

[0062] When a user asks a question, the system first uses a knowledge graph for retrieval. If a definitive answer is found, it is returned directly. If not, the question is fused with relevant information from the knowledge graph and then input into the large language model. After the model generates an answer, a text quality assessment model is used to score it: Score = αP fluency +βP accuracy +γP relevance Where Score is the answer score, P fluency To score language fluency, P accuracy To score for content accuracy, P relevance To determine the relevance score to the question, α, β, and γ are weighting coefficients. The optimal weighting combination is determined through multiple rounds of user testing to ensure that the answers meet the needs of early childhood education scenarios.

[0063] When a user asks a question, the system first searches the knowledge graph. If a definitive answer is found, it is returned directly. If not, the question is fused with information from five relevant nodes in the knowledge graph (filtered using cosine similarity calculation) and then input into the large language model. After the model generates an answer, a text quality assessment model is used to score it: Score = αP fluency +βP accuracy +γP relevanceWhere Score is the score for the answer, P fluency For the language fluency score (calculated based on the BLEU-4 index), P accuracy To score the accuracy of the content (calculated by comparison with an authoritative knowledge base), P relevance The relevance score to the question (calculated using the BM25 algorithm) uses α, β, and γ as weighting coefficients. The optimal weighting combination was determined through machine learning training on 1000 manually labeled question-and-answer pairs: α = 0.3, β = 0.5, and γ = 0.2. If the score is below the threshold of 70, a new answer is generated or the user is prompted to provide additional information.

[0064] 10. Security Protection and Version Management

[0065] The collected data related to young children is anonymized using a salted hash function: H(x) = hash(x + salt), where x is the original data and salt is a 16-bit random string, stored in a separate key management system. A role-based access control (RBAC) model is adopted, assigning different permissions to three roles: children, teachers, and administrators. Children can only access personalized learning content; teachers can view teaching resources and manage class student learning data; administrators have the highest privileges, including knowledge base updates and user management.

[0066] The version control system records detailed information for each update and generates version comparison reports. The following is a comparison of the effects of a particular update:

[0067]

[0068] The data above shows that this update has improved the knowledge richness of the knowledge graph, and the increase in the number of nodes and relationships signifies the integration of more educational knowledge. The decrease in the perplexity of the large language model indicates reduced uncertainty in the generated content, while the improved accuracy of responses directly reflects improved interaction quality. This verifies the effectiveness of the knowledge base update and maintenance mechanism in this method, enabling continuous optimization of knowledge base performance and providing higher-quality knowledge services for the early childhood education field. Simultaneously, security measures ensure data security during updates and use, while version management provides reliable assurance for the stable evolution of the knowledge base.

[0069] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge base in the field of early childhood education by fusing a knowledge graph and a large language model, characterized in that, include: Data collection steps in the field of early childhood education: Data is collected from early childhood education textbooks, children's picture books, educational videos, and early childhood teachers' lesson plans. The early childhood education textbooks include content in the fields of language, mathematics, and art. The children's picture books are selected from classic award-winning works. The educational videos include animation and popular science short films. At the same time, unstructured data such as interview records of early childhood education experts and observation logs of kindergarten teaching activities are collected to construct the original dataset. Data preprocessing and cleaning steps: The collected text data is segmented using the Jieba segmentation tool combined with a custom thesaurus for early childhood education; duplicate content is removed by calculating text fingerprints using the SimHash algorithm and setting a similarity threshold of 0.9 to filter duplicate text; keyframes are extracted from educational videos using the OpenCV library; for missing data, imputation is used for structured data and missing parts of unstructured text are marked. Knowledge extraction and entity recognition steps: Entities, including people, things, and concept categories, are extracted from the data using the Named Entity Recognition (NER) model; the model adopts the BERT-BiLSTM-CRF architecture and is trained on an labeled dataset in the early childhood education domain; simultaneously, relationships between entities are extracted through dependency parsing. Knowledge graph construction steps: Extracted entities are used as nodes and relationships as edges. A knowledge graph for early childhood education is constructed using the Neo4j graph database. The node attributes are enriched, and the knowledge graph is mapped to a low-dimensional vector space using the TransE graph embedding algorithm to achieve subsequent integration with a large language model. The large language model adaptation and fine-tuning steps: Select the open source large language model LLaMA as the base model, and use the collected early childhood education data for supervised fine-tuning according to the characteristics of early childhood education; Cross-entropy loss function is used in the fine-tuning process: Where N is the number of samples, C is the number of categories, yij is the true label 0 or 1 of sample i belonging to category j, pij is the probability of the model predicting that sample i belongs to category j; At the same time, the field knowledge prompt engineering is introduced to enhance the understanding of the model to the early childhood education scene; Knowledge fusion and verification steps: The vector representation of the knowledge graph is fused with the input and output of the large language model, using a gating mechanism. , where h kg For embedding vectors in a knowledge graph, h llm W1 and W2 are the output vectors of the hidden layer of a large language model, and W1 and W2 are the learnable weight matrices. For the Sigmoid function, The product is multiplied by elements; the fused knowledge is verified through a combination of expert review and automated verification. The automated verification uses a contradiction detection algorithm to check the logical consistency of the knowledge. Knowledge base update and maintenance steps: Establish a regular update mechanism to collect the latest early childhood education research results and newly published textbook and picture book data every month; use incremental learning methods to update the knowledge graph and large language model. For newly added nodes and edges in the knowledge graph, update the node embedding vectors through graph neural network algorithms; for the large language model, use the parameter fine-tuning technique LoRA. Personalized knowledge recommendation steps: Based on user profile data such as the child's age, learning progress, and interests, calculate the correlation between the user and knowledge points in the knowledge graph. S ui Let M be the degree of relevance between user u and knowledge point i, M be the number of relevance factors, wk be the weight of the kth factor, and rik be the relevance score between user u and knowledge point i under the kth factor. In the knowledge extraction and entity recognition steps, fuzzy matching and semantic expansion techniques are employed to address the non-standard language expression characteristics of young children; a thesaurus is constructed, and the Word2Vec model is used to calculate word semantic similarity. , where vec(w1) and vec(w2) are the vector representations of words w1 and w2. When the similarity is greater than 0.7, they are considered synonyms, which improves the entity recognition ability of children's colloquial expressions.

2. The method for constructing a knowledge base in the early childhood education domain that integrates knowledge graphs and large language models according to claim 1, characterized in that, Also includes: Interactive question-and-answer optimization steps: When a user asks a question, first use the knowledge graph to search; if a definite answer is found, return it directly. If not found, the question is fused with knowledge graph information and input into the large language model; after the model generates an answer, a text quality assessment model is used for scoring. Where Score is the answer score, P fluency To score language fluency, P accuracy To score for content accuracy, P relevance To score the relevance to the question, α, β, These are weighting coefficients; the optimal weight combination is determined through user testing.

3. The method for constructing a knowledge base in the early childhood education domain that integrates knowledge graphs and large language models according to claim 1, characterized in that, In the large language model adaptation and fine-tuning steps, an adversarial training mechanism is introduced; a discriminator D is trained to distinguish between early childhood education texts generated by the fine-tuned model and real texts, while the fine-tuned model G tries to generate texts that the discriminator will misjudge. The loss function is: , Where pdata is the distribution of real early childhood education texts and pz is the noise distribution, the authenticity of the content generated by the model is improved through adversarial training.

4. The method for constructing a knowledge base in the early childhood education domain that integrates knowledge graphs and large language models according to claim 1, characterized in that, A knowledge conflict detection algorithm is used in the knowledge fusion and verification steps: for the triples (s, p, o) in the knowledge graph and the knowledge (s', p', o') generated by the large language model, the conflict probability is calculated: When P conflict If the value exceeds the threshold of 0.6, a conflict is identified, triggering a manual review process.

5. The method for constructing a knowledge base in the early childhood education domain that integrates knowledge graphs and large language models according to claim 1, characterized in that, In the knowledge base update and maintenance process, a version management mechanism is adopted; each update generates a new knowledge base version, and records the update content, time, and operator information; the update effect is evaluated by comparing the changes in the knowledge graph structure and the performance indicators of the large language model in different versions.

6. The method for constructing a knowledge base in the early childhood education domain that integrates knowledge graphs and large language models according to claim 1, characterized in that, In the personalized knowledge recommendation step, user profile data adopts a dynamic update strategy; Based on behavioral data of children using the knowledge base, the sliding window algorithm is used to update the portrait weights: , where wk,t is the weight of the k-th factor at time t, and n is the window size, so that the recommended content is dynamically adjusted according to the child's learning situation.

7. The method for constructing a knowledge base in the early childhood education field that integrates knowledge graphs and large language models according to claim 2, characterized in that, In the interactive question-answering optimization step, an evaluation corpus of question-answer pairs in the field of early childhood education was constructed; 2000 sets of question-answer pairs were collected from real teaching scenarios, and 5 early childhood education experts were invited to conduct 5-level scoring, and the average score of each question-answer pair was calculated as a benchmark; the model-generated answers were compared with the answers in the evaluation corpus, and cosine similarity was used for calculation. ,in , These are vector representations of the model's answer and the baseline answer, respectively, and the answer generation strategy is adjusted based on similarity.

8. The method for constructing a knowledge base in the early childhood education domain that integrates knowledge graphs and large language models according to claim 1, characterized in that, Security measures were implemented throughout the entire construction process; the collected data related to young children was anonymized using a hash function. To achieve data privacy, x represents the original data and salt is a randomly generated salt value; meanwhile, role-based access control (RBAC) is used to access the knowledge base, assigning different permissions to children, teachers, and administrators.

Images