Knowledge base construction method in preschool education field by fusing knowledge graph and large language model

Abstract

The invention discloses a preschool education field knowledge base construction method fusing a knowledge graph and a large language model, and relates to the technical field of preschool education method optimization. Multi-channel collection of preschool education data and cleaning preprocessing; utilizing a customized model to extract entities and relationships, and constructing a structured knowledge graph; performing fine tuning on the big language model in the preschool education field; fusing atlas and model knowledge through a gating mechanism, and performing dual verification; establishing a monthly updating mechanism to realize incremental learning; in addition, personalized knowledge recommendation is realized based on a user portrait, interactive questions and answers are optimized through combination of retrieval and a model, and data security is guaranteed by adopting anonymization and authority control. Through multi-source data integration and intelligent technology fusion, an accurate and dynamic preschool education knowledge base is constructed, requirements of children are more accurately understood, content fitting cognition is generated, personalized learning recommendation is realized, interaction experience and knowledge service quality are improved, data security is guaranteed, and digital upgrading of the preschool education field is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of early childhood education method optimization, and in particular to a method for constructing a knowledge base in the early childhood education field by integrating a knowledge graph and a large language model. Background Art

[0002] With the rapid development of digital education, the construction of knowledge bases in the field of early childhood education has become an important part of improving the quality of early childhood education. However, traditional methods of constructing early childhood education knowledge bases have many limitations and cannot meet the diverse and personalized needs of early childhood education in the new era.

[0003] Existing early childhood education knowledge bases are mostly based on static text resources, such as fixed textbooks and lesson plans. These data sources are limited and update slowly, making them unable to promptly incorporate the latest educational concepts and children's cultural content. These knowledge bases often employ simple categorized storage methods, lacking in-depth exploration and structured presentation of knowledge systems. This makes it difficult to systematically showcase the connections between knowledge in multiple fields, such as language enlightenment, cognitive development, and artistic cultivation. This results in inefficient knowledge retrieval and makes it difficult to support teachers' personalized teaching and children's independent learning. Furthermore, due to a lack of consideration for children's cognitive characteristics, the knowledge presented is dull and fails to stimulate children's learning interest.

[0004] In terms of knowledge processing technology, traditional methods face numerous challenges when applied to early childhood education. During the knowledge extraction process, due to the colloquial and non-standard nature of children's language, conventional named entity recognition and relationship 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 are unable to understand the ambiguous semantics and implicit requirements of children's questions. Simple search-based interactive methods lack intelligence and are unable to provide vivid and informative responses tailored to children's cognitive levels, making effective human-computer interaction difficult.

[0005] In recent years, large language models and knowledge graph technologies have demonstrated strong potential in natural language processing and knowledge management, but their direct application in early childhood education remains significantly limited. While large language models possess excellent language generation capabilities, their training data is mostly general corpus, lacking targeted optimization of early childhood education expertise. The generated content may exhibit deviations from educational concepts or be expressed in ways that are inconsistent with children's cognitive characteristics and psychological needs. Regarding knowledge graph construction, existing technologies often ignore the laws of early childhood cognitive development, resulting in mechanical and rigid design of knowledge nodes and relationships, making it impossible to achieve personalized recommendations and entertaining presentation of knowledge. Furthermore, it is difficult to transform abstract knowledge into a form that is easily understood and accepted by children, failing to fully leverage the value of technology in integrating education and entertainment into early childhood education. 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 to solve the problems mentioned in the above-mentioned 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 a knowledge graph and a large language model, comprising:

[0008] Data collection steps for early childhood education: Data was collected from early childhood education textbooks, children's picture books, educational videos, and early childhood teacher lesson plans. Early childhood education textbooks included content in the fields of language, mathematics, and art. Children's picture books selected classic award-winning works. Educational videos covered animations and popular science shorts. Unstructured data was also collected from interviews with early childhood education experts and observation logs of kindergarten teaching activities to construct the original dataset.

[0009] Data preprocessing and cleaning steps: The collected text data was segmented using the Jieba word segmentation tool combined with a custom vocabulary in the field of early childhood education; duplicate content was removed, and the SimHash algorithm was used to calculate text fingerprints, with a similarity threshold of 0.9 set to filter out duplicate text; keyframes were extracted from educational videos using the OpenCV library; for missing data, the structured data was filled in, and missing parts of unstructured text were marked;

[0010] Knowledge extraction and entity recognition steps: Use the named entity recognition (NER) model to extract entities from the data, including people, things, and concept categories. The model uses the BERT-BiLSTM-CRF architecture and is trained on a labeled dataset in the early childhood education field. At the same time, dependency parsing is used to extract relationships between entities.

[0011] Knowledge graph construction steps: Using the extracted entities as nodes and relationships as edges, we construct a preschool knowledge graph using the Neo4j graph database. Node attributes are enriched, and the knowledge graph is mapped to a low-dimensional vector space using the graph embedding algorithm TransE, enabling subsequent integration with a large language model.

[0012] Large language model adaptation and fine-tuning steps: We selected the open-source large language model LLaMA as the base model and used the collected early childhood education data for supervised fine-tuning based on the characteristics of the early childhood education field. The cross-entropy loss function was 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 for 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 hint engineering is 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 integrated 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 is the knowledge graph embedding vector, h llm is the output vector of the hidden layer of the large language model, W1 and W2 are learnable weight matrices, σ is the Sigmoid function, and ⊙ is element-wise multiplication. 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.

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

[0015] Furthermore, it also includes:

[0016] Personalized knowledge recommendation steps: Based on user profile data of children's age, learning progress, and interest preferences, calculate the correlation between users and knowledge points in the knowledge graph: Among them S ui is the correlation between user u and knowledge point i, M is the number of correlation factors, wk is the weight of the kth factor, and rik is the correlation score between user u and knowledge point i under the kth factor.

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

[0018] Furthermore, in the knowledge extraction and entity recognition steps, fuzzy matching and semantic expansion techniques are used to address the non-standard characteristics of children's language expression; a synonym library is constructed, and the Word2Vec model is used to calculate the semantic similarity of words: Among them, 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 synonymous, which improves the entity recognition ability of children's spoken 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 the preschool education texts generated by the fine-tuned model and the real texts. The fine-tuned model G tries to generate texts that the discriminator misjudges. The loss function is: Where pdata is the real early childhood education text distribution, and pz is the noise distribution. Adversarial training is used to improve the authenticity of the content generated by the model.

[0020] Furthermore, 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 When the value is greater than the threshold of 0.6, it is determined that there is a conflict and the manual review process is triggered.

[0021] Furthermore, a version management mechanism is adopted in the knowledge base update and maintenance steps; a new knowledge base version is generated for each update, and the update content, time, and operator information are recorded; the update effect is evaluated by comparing the structural changes of the knowledge graphs and the performance indicators of the large language model of different versions.

[0022] Furthermore, in the personalized knowledge recommendation step, the user portrait data adopts a dynamic update strategy; based on the children's behavioral data in the process of using the knowledge base, the sliding window algorithm is used to update the portrait weight: Where wk,t is the weight of the kth factor at the tth moment, 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, we constructed an evaluation corpus of question-answer pairs in the field of early childhood education. We collected 2,000 question-answer pairs from real teaching scenarios, invited five early childhood education experts to perform a five-level rating, and calculated the average score of each question-answer pair as a benchmark. We then compared the answers generated by the model with those in the evaluation corpus using cosine similarity: in are the vector representations of the model answer and the benchmark answer, respectively, and the answer generation strategy is adjusted according to the similarity.

[0024] Furthermore, security measures are adopted throughout the entire construction process; the collected child-related data is anonymized and data privacy is achieved using the hash function: H(x) = hash(x + salt), 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 for knowledge base access, and different permissions are assigned to children, teachers, and administrators.

[0025] Compared with the existing technology, the beneficial effects of the present invention are:

[0026] In terms of knowledge resource development, this approach utilizes a multi-channel acquisition strategy, integrating rich data from textbooks, picture books, videos, and expert experience. A regular update mechanism is established to ensure the knowledge base is comprehensive and up-to-date with educational trends. In-depth data preprocessing and specialized knowledge extraction techniques, optimized based on the specific language characteristics of young children, significantly improve the accuracy and completeness of knowledge construction, forming a clearly structured and closely connected knowledge system.

[0027] The innovative integration of a knowledge graph and a large language model is the core advantage of this approach. The knowledge graph presents the logical relationships between knowledge in a structured manner, providing a solid foundation for knowledge reasoning and personalized recommendations. The large language model empowers the system with powerful natural language understanding and generation capabilities, enabling it to accurately interpret children's ambiguous and colloquial questions and generate vivid, engaging responses tailored to their cognitive level. Working together, these two significantly enhance the intelligence and engaging nature of interactions, effectively stimulating children's interest in learning and their desire to explore.

[0028] The personalized knowledge recommendation function creates a dynamic user profile based on a child's age, learning progress, and interests. It can customize a unique learning path for each child, meeting their differentiated learning needs and promoting their individual development. Furthermore, this function can adjust recommendation strategies in real time based on the child's learning behavior, ensuring that learning content always aligns with the child's cognitive development rhythm and improves learning outcomes.

[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 children's privacy, and role-based access control prevents knowledge leakage and unauthorized tampering. Version management facilitates the evaluation and review of knowledge base updates, ensuring the system continues to provide high-quality services.

[0030] In addition, this application method not only provides young children with scientific and interesting learning tools, but also provides efficient support for preschool teachers in acquiring teaching resources, curriculum design and teaching evaluation. It helps to promote the transformation of preschool education from an experience-driven to a data-driven, intelligent decision-making model, and promote the overall improvement of the digital level in the field of preschool education. BRIEF DESCRIPTION OF THE DRAWINGS

[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 proposed in the present invention;

[0032] Figure 2 This is a schematic diagram comparing the accuracy of knowledge extraction and entity recognition of the method for building a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models proposed in the present invention;

[0033] Figure 3 Schematic diagram of the personalized knowledge recommendation satisfaction survey for the method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models proposed in the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0036] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below 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 by integrating knowledge graphs and large language models

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

[0039] Data was collected from multiple sources, including preschool textbooks, children's picture books, educational videos, and preschool teacher lesson plans. The textbooks cover eight major areas, including language, mathematics, and art. The children's picture books include over 200 classic and award-winning works. The educational videos range from animations to popular science shorts, totaling over 500 hours. Unstructured data, such as interviews with preschool education experts and observation logs of kindergarten teaching activities, was also collected to construct the original dataset.

[0040] This method collects data in the field of early childhood education through multiple channels, covering both structured and unstructured data types. In terms of teaching materials and picture book resources, we cooperate with 10 well-known early childhood education publishers to obtain more than 1,200 electronic teaching materials in eight major fields such as language, mathematics, and art, and simultaneously collect more than 300 classic children's picture books; educational video data is crawled from mainstream education platforms, including popular science animations, hand-made teaching and other types, with a cumulative duration of more than 600 hours, and video subtitles are obtained through the API interface. At the same time, the team organized field surveys in 50 kindergartens, collected more than 800 teacher lesson plans, 1,200 teaching activity observation logs, recorded 200 hours of expert interview videos, and transcribed them into text. All data is stored in the distributed file system HDFS, and a data index directory is established to facilitate subsequent retrieval and calling.

[0041] 2. Data preprocessing and cleaning steps

[0042] The collected text data was segmented using the Jieba word segmentation tool combined with a custom vocabulary for early childhood education to identify specialized terms such as "children's sensitive period" and "Montessori teaching method." Duplicate content was removed by calculating text fingerprints using the SimHash algorithm, with a similarity threshold of 0.9 set to filter out duplicate text. Testing showed this reduced data redundancy by approximately 18%. Keyframe extraction was performed on educational videos at a rate of one frame per second using the OpenCV library. The audio was then extracted and converted to text using the FFmpeg tool. A random forest-based multi-pass imputation method was used to address missing values ​​in structured data. Missing portions of unstructured text were marked as "[missing]" and subsequently filled in through manual completion or knowledge inference.

[0043] 3. Knowledge extraction and entity recognition steps

[0044] A named entity recognition (NER) model was used to extract entities from the data, including categories such as people (e.g., "Sun Wukong" and "Hans Christian Andersen"), things (e.g., "building blocks" and "piano"), and concepts (e.g., "color recognition" and "digital enlightenment"). The model, using the BERT-BiLSTM-CRF architecture and trained on a labeled dataset for early childhood education, achieved an F1 score of 92% for entity recognition. Furthermore, dependency parsing was used to extract relationships between entities, such as "Snow White - belongs to - fairy tales" and "building blocks - are used for - shape recognition."

[0045] The named entity recognition model uses the BERT-BiLSTM-CRF architecture and is trained on a self-built early childhood education corpus containing 100,000 annotated data items. The model's input layer converts text into word vectors, extracts features using the BERT pre-trained model, and then performs sequence annotation using bidirectional LSTM and CRF layers. Ultimately, it identifies six major entity categories, including people, things, and concepts, achieving an F1 score of 92.3%. During the relationship extraction phase, the Stanford Core NLP tool performs dependency parsing, defining 20 common relationship types in the early childhood education field (such as "belongs to," "used for," and "cultivate"). 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 recognize shapes," the relationship triple (building blocks, used for, shape recognition) can be extracted.

[0046] 4. Knowledge graph construction steps

[0047] Using the Neo4j graph database, we constructed a preschool education knowledge graph, using extracted entities as nodes and relationships as edges. We enriched node attributes, such as "character image" and "appearance in works" for character nodes and "age of use" and "educational goals" for object nodes. We mapped the knowledge graph into a low-dimensional vector space using graph embedding algorithms (such as TransE) to facilitate subsequent integration with large language models. The vector dimension was set to 200.

[0048] We used the Neo4j graph database to build a preschool education knowledge graph, storing extracted entities as nodes and relationships as edges. Each node was assigned a variety of attributes, such as "character description," "list of works in which the text appears," and "educational significance" for character nodes, and "applicable age range," "usage scenarios," and "educational objectives" for object nodes. We used the TransE algorithm to map the knowledge graph to a 200-dimensional vector space, setting the number of training rounds to 500 and the learning rate to 0.01. We minimized the translation loss function: 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 triplets, S' is the set of negative sample triplets, d(x,y) is the Euclidean distance between vectors x and y, and γ is the interval hyperparameter (set to 1.0). Through training, entities with similar semantics are brought closer 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 used the collected early childhood education data for supervised fine-tuning based on the characteristics of the early childhood education field. During the fine-tuning process, we used the cross-entropy loss function: Where N is the number of samples, C is the number of categories, yij is the true label (0 or 1) for sample i belonging to category j, and pij is the model's predicted probability that sample i belongs to category j. Furthermore, a domain knowledge prompting project is introduced, adding prompts such as "in early childhood mathematics enlightenment" and "for children aged 3-4" before the input text to enhance the model's understanding of early childhood education scenarios.

[0051] We selected the open-source LLaMA model as the basis and used 500,000 early childhood education data collected for supervised fine-tuning. The fine-tuning process used the PyTorch framework, set the batch size to 32, the number of training rounds to 10, and used the AdamW optimizer with a learning rate of 5×10 -5 The loss function uses cross entropy loss: Where N is the number of samples, C is the number of categories, yij is the true label (0 or 1) for sample i belonging to category j, and pij is the model's predicted probability that sample i belongs to category j. A prompt engineering library containing 500 prompt word templates was also constructed, such as "When 3-4-year-old children learn colors, [specific method] can be used for teaching." Relevant prompt words are added before input text to enhance the model's understanding of early childhood education scenarios.

[0052] 6. Knowledge fusion and verification steps

[0053] The vector representation of the knowledge graph is integrated with the input and output of the large language model, and a gating mechanism is used: G = σ(W1h kg +W2h llm ),h fusion =G⊙h kg +(1-G)⊙h llm , where h kg is the knowledge graph embedding vector, h llm is the output vector of the hidden layer of the large language model, W1 and W2 are learnable weight matrices, σ is the Sigmoid function, and ⊙ is element-wise multiplication. 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.

[0054] The gating mechanism is used to fuse the information of the knowledge graph and the large language model. First, the embedding vector hkg of the knowledge graph node 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 Here, 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 ⊙ is element-wise multiplication. The integrated knowledge is verified through a dual validation mechanism: first, a review panel of 10 early childhood education experts manually reviews 1,000 randomly selected pieces of knowledge; second, an automated validation algorithm detects logical contradictions in the knowledge. For example, by checking for conflicting statements such as "a certain activity is suitable for children aged 3-4" and "the difficulty of this activity is only suitable for children aged 5-6," manual review is triggered when the conflict probability Pconflict exceeds a threshold of 0.6.

[0055] 7. Knowledge base update and maintenance steps

[0056] A regular update mechanism is established to collect data on the latest early childhood education research results, newly published textbooks, and picture books every month. Knowledge graphs and large language models are updated using incremental learning methods. For new nodes and edges in the knowledge graph, node embedding vectors are updated using a graph neural network algorithm. For the large language model, efficient parameter fine-tuning techniques (such as LoRA) are used to update only a small number of adapter parameters, reducing computing resource consumption.

[0057] A monthly update mechanism was established, with the data collection module launched at the beginning of each month to obtain the latest early childhood education research results, newly published textbooks, and picture books. The knowledge graph and large language model were updated using incremental learning: 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 trained for three rounds. For the large language model, LoRA technology was used, freezing most of the parameters of the pre-trained model and only training the newly added adapter parameters (approximately 0.1% of the total parameters). The RMSProp optimizer was used with a learning rate of 1×10 -4 , and train for 5 rounds. At the same time, record the version information of each update, including update time, update content, operator, etc., generate a version comparison report, and evaluate the update effect by comparing the number of knowledge graph nodes, relationship change rate, and perplexity and accuracy of the large language model on the test set of different versions.

[0058] 8. Personalized knowledge recommendation steps

[0059] Based on user profile data such as children's age, learning progress, and interest preferences, the correlation between users and knowledge points in the knowledge graph is calculated: Among them S ui is the correlation between user u and knowledge point i, M is the number of correlation factors, wk is the weight of the kth factor, and rik is the relevance score between user u and knowledge point i under the kth factor. We recommend personalized learning content to children based on correlation ranking. For example, for a 4-year-old child interested in painting, we prioritize "Children's Painting Enlightenment" content.

[0060] Dynamic user profiles are constructed based on the child's age, learning progress, interests, and other data. User profile data is stored in a MongoDB database, and a sliding window algorithm (with a window size n set to 30 days) is used to update the profile weights: Where wk,t is the weight of the kth factor at time t, and rik is the correlation score between user u and the kth factor at time i. Calculate the correlation between the user and the knowledge points in the knowledge graph: Where M is the number of correlation factors (set to 5, including age matching, learning progress matching, interest relevance, knowledge difficulty matching, and educational goal matching). Knowledge points are sorted according to the correlation, and the top 10 relevant knowledge contents are recommended to children. These are displayed in the form of animated cards through a visual interface.

[0061] 9. Interactive Q&A Optimization Steps

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

[0063] When a user asks a question, the system first searches the knowledge graph and returns the answer if it finds one. If not, the system fuses the question with the five relevant nodes in the knowledge graph (filtered by cosine similarity) and then inputs them into the large language model. After the model generates an answer, it uses the text quality assessment model to score it: Score = αP fluency +βP accuracy +γP relevanceAmong them, Score is the answer score, P fluency is the language fluency score (calculated based on the BLEU-4 index), P accuracy is the content accuracy score (calculated by comparing with the authoritative knowledge base), P relevance α is the relevance score to the question (calculated using the BM25 algorithm), and α, β, and γ are weight coefficients. Machine learning training on 1,000 manually annotated question-answer pairs determined the optimal weight combination to be α = 0.3, β = 0.5, and γ = 0.2. If the score falls 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 child-related data 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 used to assign 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; and administrators have the highest permissions, allowing them to perform operations such as knowledge base updates and user management.

[0066] The version management system records the detailed information of each update and generates a version comparison report. The following is the effect comparison data of a certain update:

[0067]

[0068] The above data demonstrates that this update has enhanced the knowledge richness of the knowledge graph. The increase in the number of nodes and relationships signifies the incorporation of more educational knowledge. The decrease in the perplexity of the large language model indicates a reduction in the uncertainty of the model-generated content, while the increase in answer accuracy directly reflects the improved quality of interaction. This validates the effectiveness of the knowledge base update and maintenance mechanism in this method, which can continuously optimize knowledge base performance and provide higher-quality knowledge services for the early childhood education sector. Furthermore, security measures ensure the security of data during updates and use, and version management functions provide reliable guarantees for the stable evolution of the knowledge base.

[0069] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge base in the field of early childhood education that integrates knowledge graphs and large language models, characterized by: include: Data collection steps for early childhood education: Data was collected from early childhood education textbooks, children's picture books, educational videos, and early childhood teacher lesson plans. Early childhood education textbooks included content in the fields of language, mathematics, and art. Children's picture books selected classic award-winning works. Educational videos included animations and popular science shorts. Unstructured data was also collected from interviews with early childhood education experts and observation logs of kindergarten teaching activities to construct the original dataset. Data preprocessing and cleaning steps: The collected text data was segmented using the Jieba word segmentation tool combined with a custom vocabulary in the field of early childhood education; duplicate content was removed, and the SimHash algorithm was used to calculate text fingerprints, with a similarity threshold of 0.9 set to filter out duplicate text; keyframes were extracted from educational videos using the OpenCV library; for missing data, the structured data was filled in, and missing parts of unstructured text were marked; Knowledge extraction and entity recognition steps: Use the named entity recognition (NER) model to extract entities from the data, including people, things, and concept categories. The model uses the BERT-BiLSTM-CRF architecture and is trained on a labeled dataset in the early childhood education field. At the same time, dependency parsing is used to extract relationships between entities. Knowledge graph construction steps: Using the extracted entities as nodes and relationships as edges, we construct a preschool knowledge graph using the Neo4j graph database. Node attributes are enriched, and the knowledge graph is mapped to a low-dimensional vector space using the graph embedding algorithm TransE, enabling subsequent integration with a large language model. Large language model adaptation and fine-tuning steps: We selected the open-source large language model LLaMA as the base model and used the collected early childhood education data for supervised fine-tuning based on the characteristics of the early childhood education field. The cross-entropy loss function was 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 for sample i belonging to category j, and pij is the model's prediction of the probability that sample i belongs to category j. At the same time, domain knowledge hint engineering is introduced to enhance the model's understanding of early childhood education scenarios.

2. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 1 is characterized in that: Also includes: Knowledge fusion and verification steps: The vector representation of the knowledge graph is integrated 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 is the knowledge graph embedding vector, h llm is the output vector of the hidden layer of the large language model, W1 and W2 are learnable weight matrices, σ is the Sigmoid function, and ⊙ is element-wise multiplication. 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: Set up a regular update mechanism to collect the latest early childhood education research results and newly published textbook picture book data every month; use incremental learning methods to update the knowledge graph and large language model, add nodes and edges to the knowledge graph, and update the node embedding vectors through the graph neural network algorithm; for the large language model, use the parameter fine-tuning technology LoRA; Personalized knowledge recommendation steps: Based on user profile data of children's age, learning progress, and interest preferences, calculate the correlation between users and knowledge points in the knowledge graph: Among them S ui is the correlation between user u and knowledge point i, M is the number of correlation factors, wk is the weight of the kth factor, and rik is the correlation score between user u and knowledge point i under the kth factor.

3. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 1 is characterized in that: Also includes: Interactive Q&A optimization steps: When a user asks a question, first use the knowledge graph to search, and directly return if a definite answer is found; If not found, the question is fused with the knowledge graph information and input into the large language model; after the model generates an answer, the text quality assessment model is used to score it: Score = αP fluency +βP accuracy +γP relevance , where Score is the answer score, P fluency is the language fluency score, P accuracy For content accuracy score, P relevance is the score of relevance to the question, α, β, and γ are weight coefficients, and the optimal weight combination is determined through user testing.

4. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 1 is characterized in that: In the knowledge extraction and entity recognition steps, fuzzy matching and semantic expansion techniques are used to address the irregularities in children's language expression. A synonym database is constructed, and the Word2Vec model is used to calculate the semantic similarity of words. Among them, 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 synonymous, which improves the entity recognition ability of children's spoken expressions.

5. The method for constructing a knowledge base in the field of early childhood education that integrates a knowledge graph and a large language model according to claim 1 is characterized in that: During the large language model adaptation and fine-tuning steps, an adversarial training mechanism is introduced. A discriminator D is trained to distinguish between preschool education texts generated by the fine-tuned model and real texts. The fine-tuned model G tries to generate texts that the discriminator misjudges. The loss function is: Where pdata is the real early childhood education text distribution, and pz is the noise distribution. Adversarial training is used to improve the authenticity of the content generated by the model.

6. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 2 is characterized in that: In the knowledge fusion and verification steps, a knowledge conflict detection algorithm is used: 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 When the value is greater than the threshold of 0.6, it is determined that there is a conflict and the manual review process is triggered.

7. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 2 is characterized in that: In the knowledge base update and maintenance steps, a version management mechanism is adopted; a new knowledge base version is generated for each update, and the update content, time, and operator information are recorded; the update effect is evaluated by comparing the changes in the knowledge graph structure and the performance indicators of the large language model of different versions.

8. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 2 is characterized in that: In the personalized knowledge recommendation step, user portrait data adopts a dynamic update strategy; Based on the behavioral data of young children in the process of using the knowledge base, the sliding window algorithm is used to update the portrait weight: Where wk,t is the weight of the kth factor at the tth moment, and n is the window size, so that the recommended content can be dynamically adjusted according to the child’s learning situation.

9. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 3 is characterized in that: In the interactive question-answering optimization step, we constructed an evaluation corpus of question-answer pairs in the field of early childhood education. We collected 2,000 question-answer pairs from real-world teaching scenarios and invited five early childhood education experts to perform a 5-level rating. We calculated the average score of each question-answer pair as a benchmark. We then compared the answers generated by the model with those in the evaluation corpus using cosine similarity: in are the vector representations of the model answer and the benchmark answer, respectively, and the answer generation strategy is adjusted according to the similarity.

10. The method for constructing a knowledge base in the field of early childhood education by integrating a knowledge graph and a large language model according to claim 1, characterized in that: Security protection measures are adopted throughout the entire construction process; the collected child-related data is anonymized and data privacy is achieved using the hash function: H(x) = hash(x + salt), 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 for knowledge base access, assigning different permissions to children, teachers, and administrators.

Images