An education resource sharing method and system based on an education cloud platform

Abstract

This invention discloses a method and system for sharing educational resources based on an educational cloud platform, comprising: acquiring a resource dataset from a service cloud platform, wherein the service cloud platform resource dataset contains the original content of educational resources to be processed; inputting the service cloud platform resource dataset into a pre-trained resource extraction model, wherein the resource extraction model performs deep semantic understanding and information extraction on the resource content, and simultaneously outputs a structured recognition result containing resource summary, global semantics of the resource, types of resource elements, and keywords of the resource, wherein the pre-trained resource extraction model is a lightweight large language model fine-tuned in the field of education; based on the structured recognition result, filtering out invalid interference content to obtain a denoised semantic representation of the resource; and responding to the semantic representation of the resource, generating and exporting a general resource template that matches the resource content.

Description

Technical Field

[0001] This invention relates to the field of educational resource technology, and in particular to a method and system for sharing educational resources based on an educational cloud platform. Background Technology

[0002] Traditional educational resource sharing models suffer from uneven distribution of high-quality resources, fragmented sharing channels, delayed updates, restricted access, and low utilization efficiency. These issues hinder the effective narrowing of educational gaps between urban and rural areas and across regions, and fail to meet the urgent needs of the new era of educational informatization for high resource mobility, high utilization rates, and high adaptability. The aim is to build a resource-sharing ecosystem through an educational cloud platform, integrating standardized uploading, categorized storage, intelligent retrieval, on-demand distribution, dynamic updates, and multi-terminal collaboration. This will enable teachers, students, educational institutions, and other users to access diverse resources such as text, audio, video, courseware, and question banks. The convenient access, efficient utilization, and secondary creation of educational resources break down geographical and temporal barriers, promoting the widespread dissemination and precise matching of high-quality resources. This is significant not only in advancing educational equity from "equal opportunity" to "equal resources," improving teaching quality and learning efficiency, and supporting personalized learning and the construction of a lifelong education system, but also in enhancing the utilization efficiency and reproduction capacity of educational resources through data-driven resource optimization and intelligent recommendation algorithms. This provides a scientific basis for educational decision-making and promotes the transformation of educational informatization from "infrastructure construction" to "ecosystem building."

[0003] In existing technologies, resource summaries are extracted based on summarization algorithms. Specifically, the resource dataset to be identified is extracted, and the dataset is gradually decomposed to obtain segmented and sentence-based resource sets. The sentences are then segmented using the jieba word segmentation tool to obtain the segmentation results. The word segmentation ranking is calculated using the TextRank algorithm to obtain the top-ranked words as resource summaries. However, TextRank relies on word frequency and co-occurrence relationships to construct a graph model, which cannot identify synonyms, polysemous words, and implicit semantic connections. For example, in interdisciplinary resources, the potential connection between quantum computing and cryptography may be overlooked, resulting in the summary omitting key information. Therefore, this paper proposes an educational resource sharing method and system based on an educational cloud platform. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for sharing educational resources based on an educational cloud platform.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for sharing educational resources based on an educational cloud platform includes: Obtain the service cloud platform resource dataset, which contains the original content of the educational resources to be processed; The resource dataset of the service cloud platform is input into a pre-trained resource extraction model, which performs deep semantic understanding and information extraction on the resource content and outputs a structured recognition result containing resource summary, global semantics of the resource, types of resource elements and keywords. The pre-trained resource extraction model is a lightweight large language model that has been fine-tuned in the field of education. Based on the structured recognition results, invalid interference content is filtered out to obtain the denoised semantic representation of the resources. In response to the resource semantic representation, a general resource template matching the resource content is generated and exported.

[0006] The above technical solution further includes: Furthermore, obtain the service cloud platform resource dataset, specifically including: By using the resource crawling engine configured in the education cloud platform, based on predefined resource types, subject tags and update cycle strategies, the platform actively scans and discovers newly released or updated educational resource datasets within the education cloud platform. The resource types include at least documents, courseware, videos, audio, and interactive simulation software packages. For the original content of educational resources in different modalities and formats, the corresponding standardized adapter is called for unified access. The adapter includes a document parser, a video keyframe and subtitle extractor, an audio-to-text engine, and a software package metadata reader, so as to initially convert the unstructured or semi-structured original data into an intermediate representation that can be processed by subsequent models. The intermediate representation of the data is cleaned, deduplicated, and formatted. Core metadata related to the resource is automatically extracted or supplemented. The core metadata includes the resource title, creator, grade level and subject, file format, and basic descriptive information, forming a preliminarily organized multimodal educational resource dataset. By scanning the dataset with pre-defined filtering rules and content security models, the system identifies and filters out content containing sensitive information, copyright disputes, or content that does not comply with educational standards, ensuring the compliance and security of the input data.

[0007] Furthermore, the pre-trained resource extraction model is a lightweight large language model fine-tuned in the education domain. Specifically, the pre-trained resource extraction model is a model built on the Transformer architecture and obtained by domain-adaptive pre-training or supervised fine-tuning on a massive educational professional corpus. It is deployed locally on the service nodes or edge devices of the education cloud platform to ensure that educational resource data is processed locally and does not flow out of the private domain.

[0008] Furthermore, the resource extraction model performs deep semantic understanding and information extraction on the resource content, and simultaneously outputs structured recognition results, specifically including: By designing specific prompt word engineering, the lightweight large language model is guided to generate a structured output containing the resource summary, the global semantics of the resource, the types of resource elements, and the keywords of the resource in one go; The resource keywords and their semantic relationships are represented in the form of a semantic network or graph structure, rather than an independent list of keywords.

[0009] Furthermore, the resource extraction model incorporates an educational terminology knowledge graph for reinforcement learning during the training phase, specifically as follows: Establish a terminology knowledge graph that includes core concepts in the field of education and their relationships; During the training or inference process of the resource extraction model, when a word in the input text is identified as matching a standard node in the terminology knowledge graph, the corresponding semantic representation within the resource extraction model is weighted and strengthened or linked to nodes to correct general semantic biases and enhance the accuracy of domain concepts.

[0010] Furthermore, during the training of the resource extraction model, a strong regularization strategy for text data was adopted, including: The training data was augmented using synonym replacement, sentence back-translation, and random deletion techniques based on an education-related terminology database. During model training, Dropout, weight decay, and layer normalization techniques are applied. In addition, an early stopping method is used to terminate training based on validation set performance monitoring to prevent the model from overfitting on a finite-sized educational dataset.

[0011] Furthermore, the generation and export of a generic resource template that matches the resource content specifically includes: Using the denoised resource semantic representation as input, the resource extraction model or a dedicated template generation module is called again to generate template creation instructions that describe the target resource framework and filling requirements. Based on the template creation instructions, an initial template framework is matched or dynamically created from the template library; The initial template framework is iteratively adjusted and verified by importing standardized sample data, and a general resource template that meets the requirements is output.

[0012] An educational resource sharing system based on an educational cloud platform includes: Resource dataset acquisition module: Acquires the service cloud platform resource dataset to be processed from the education cloud platform; Deep semantic understanding and extraction module: It has a built-in pre-trained resource extraction model for deep analysis of resource datasets and outputting structured recognition results; Semantic denoising and representation module: Cleans and refines the structured recognition results to obtain a pure resource semantic representation; Resource template generation module: Based on the resource semantic representation, it generates and exports a concrete general resource template.

[0013] Furthermore, the deep semantic understanding and extraction module includes a localized model deployment unit and a structured output parsing unit. The localized model deployment unit deploys the lightweight large language model on a private server or edge computing device. The structured output parsing unit parses and verifies the structured data output by the model and converts it into a unified semantic representation format within the system.

[0014] Furthermore, the resource template generation module includes an instruction generation unit, a template adaptation and adjustment unit, and a template export unit. The instruction generation unit generates executable template construction instructions based on semantic representation. The template adaptation and adjustment unit matches or initializes the template according to the instructions and performs multiple rounds of optimization using standard data. The template export unit outputs the finally determined general resource template to the cloud platform shared resource library in a standardized format.

[0015] The present invention has the following beneficial effects: In this invention, an educational terminology knowledge graph is used as external prior knowledge. It is dynamically integrated with the internal reasoning process of a lightweight large language model through weighted reinforcement or node linking mechanisms, thereby realizing the internalization of domain knowledge and significantly improving the accuracy of deep understanding of educational professional terms and cross-concept semantic associations. Based on this structured result, after noise reduction and semantic representation, the system can automatically generate and export a highly matched general resource template. Attached Figure Description

[0016] Figure 1 This is a flowchart of an educational resource sharing method based on an educational cloud platform proposed in this invention; Figure 2 This is a system block diagram of an educational resource sharing system based on an educational cloud platform proposed in this invention. Detailed Implementation

[0017] 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.

[0018] Please see Figures 1-2 As shown, this invention is a method for sharing educational resources based on an educational cloud platform, comprising: Obtain the service cloud platform resource dataset, which contains the original content of the educational resources to be processed; The resource dataset of the service cloud platform is input into a pre-trained resource extraction model, which performs deep semantic understanding and information extraction on the resource content and outputs a structured recognition result containing resource summary, global semantics of the resource, types of resource elements and keywords. The pre-trained resource extraction model is a lightweight large language model that has been fine-tuned in the field of education. Based on the structured recognition results, invalid interference content is filtered out to obtain the denoised semantic representation of the resources. In response to the resource semantic representation, a general resource template matching the resource content is generated and exported.

[0019] In one embodiment, obtaining the service cloud platform resource dataset specifically includes: By using the resource crawling engine configured in the education cloud platform, based on predefined resource types, subject tags and update cycle strategies, the platform actively scans and discovers newly released or updated educational resource datasets within the education cloud platform. The resource types include at least documents, courseware, videos, audio, and interactive simulation software packages. For the original content of educational resources in different modalities and formats, the corresponding standardized adapter is called for unified access. The adapter includes a document parser, a video keyframe and subtitle extractor, an audio-to-text engine, and a software package metadata reader, so as to initially convert the unstructured or semi-structured original data into an intermediate representation that can be processed by subsequent models. The intermediate representation of the data is cleaned, deduplicated, and formatted. Core metadata related to the resource is automatically extracted or supplemented. The core metadata includes the resource title, creator, grade level and subject, file format, and basic descriptive information, forming a preliminarily organized multimodal educational resource dataset. By scanning the dataset with pre-defined filtering rules and content security models, the system identifies and filters out content containing sensitive information, copyright disputes, or content that does not comply with educational standards, ensuring the compliance and security of the input data.

[0020] In one embodiment, the pre-trained resource extraction model is a lightweight large language model fine-tuned in the education domain. Specifically, the pre-trained resource extraction model is a model built on the Transformer architecture and obtained by domain-adaptive pre-training or supervised fine-tuning on a massive educational professional corpus. It is deployed locally on the service nodes or edge devices of the education cloud platform to ensure that educational resource data is processed locally and does not flow out of the private domain.

[0021] More specifically: The basic lightweight large language model is fine-tuned for the education field. The training corpus used for fine-tuning includes curriculum standard documents, professional textbook texts, academic papers, and teaching design schemes, so that the basic lightweight large language model can gain a deep understanding of professional terminology, concept system and teaching logic in the education field. The fine-tuning in the education field specifically involves: using a command-based fine-tuning approach to organize training data pairs containing task instances from educational scenarios. These task instances include, but are not limited to, generating summaries of key teaching points from long texts, extracting core concepts and definitions from teaching materials, and inferring the teaching sequence or logical relationship between concepts. The educational resource dataset to be parsed is input into the fine-tuned lightweight large language model, and prompt words containing specific task instructions are constructed. These prompt words are used to guide the model to perform a composite task of generative summarization, entity recognition, and relation extraction. The constructed specific task instruction prompts explicitly require the lightweight large language model to distinguish and identify different types of entities and relationships in its output; The entity types include at least subject knowledge points, teaching objectives, teaching methods, and teaching resources, and the relation types include at least "presupposes", "applies to", "belongs to", and "compares to". The system receives and parses the output generated by the lightweight large language model based on the prompt words, and parses the output into a predefined, machine-readable structured semantic information format. The structured semantic information format includes a JSON or XML data structure for encapsulating the resource summary, the key entities, and the relationships. The generated structured semantic information has its inherent relationships organized and represented in the form of a graph structure or attribute graph, where the key entities are used as nodes and the semantic or logical relationships between entities are used as edges, thereby forming a lightweight semantic network that represents the core content of educational resources. This semantic relationship network is used to represent the membership, causal, or analogical relationships between concepts.

[0022] In one embodiment, the process of performing deep semantic understanding and information extraction on resource content by the resource extraction model, and simultaneously outputting structured recognition results, specifically includes: By designing specific prompt word engineering, the lightweight large language model is guided to generate a structured output containing the resource summary, the global semantics of the resource, the types of resource elements, and the keywords of the resource in one go; The resource keywords and their semantic relationships are represented in the form of a semantic network or graph structure, rather than an independent list of keywords.

[0023] More specifically: Constructing a domain-adaptive cue word template: Based on the characteristics of educational resources, a structured cue word template is pre-designed. This template explicitly instructs the lightweight large language model to perform the following composite tasks: Summarize the core content of the original text and generate a resource summary. Analyze and explain the overall intent and core ideas of the resource to form a global semantic description of the resource; Identify and classify the types of components contained in the resources, such as text, test questions, courseware, video scripts, etc., and determine the types of resource elements; Extract core terms and analyze the semantic, logical, or pedagogical connections between these terms; The prompt word template requires the model to organize all the above output items in a uniform, machine-readable structured format, such as JSON or XML; Injecting resource content and calling model inference: The original text content in the service cloud platform resource dataset to be processed is filled into the specified position in the constructed prompt word template. Then, the assembled complete prompt is input into the locally deployed lightweight large language model, triggering the model to perform a forward inference calculation. Based on the parameters fine-tuned on the corpus in the education field, the injected resource content is deeply analyzed and understood. The combined output of the parsing and verification model: Receives and captures the original text response generated by the lightweight large language model based on the prompt word instructions, parses the response using a predefined parser, and separates and extracts the corresponding resource summary, global semantic description of the resource, list of resource element types, and set of resource keywords and their relationship descriptions according to the agreed structured format; Constructing a keyword semantic network graph: For the parsed resource keywords and their relational descriptions, a network construction process is performed, treating each resource keyword as an independent node in the graph. Then, based on the relational descriptions output by the model (such as "belongs to", "premise is", "applies to", "compared to"), directed or undirected edges are established between the corresponding keyword nodes, and the relational type is used as the edge label. Finally, a data representation of a semantic network or graph structure with resource keywords as nodes and semantic associations as edges is generated, thereby intuitively representing the complex, interconnected, and non-isolated relationships between keywords. The final structured recognition result is integrated and packaged with the parsed resource summary, global resource semantics, and list of resource element types, together with the constructed resource keyword semantic network graph. This results form a complete, machine-readable structured recognition result, which will serve as direct input for subsequent processes such as resource denoising and template generation. In one embodiment, the resource extraction model incorporates an educational terminology knowledge graph for reinforcement learning during the training phase, specifically: Establish a terminology knowledge graph that includes core concepts in the field of education and their relationships; During the training or inference process of the resource extraction model, when a word in the input text is identified as matching a standard node in the terminology knowledge graph, the corresponding semantic representation within the resource extraction model is weighted and strengthened or linked to nodes to correct general semantic biases and enhance the accuracy of domain concepts.

[0024] In one embodiment, a strong regularization strategy for text data is employed during the training of the resource extraction model, including: The training data was augmented using synonym replacement, sentence back-translation, and random deletion techniques based on an education-related terminology database. During model training, Dropout, weight decay, and layer normalization techniques are applied. In addition, an early stopping method is used to terminate training based on validation set performance monitoring to prevent the model from overfitting on a finite-sized educational dataset.

[0025] More specifically: Obtain the original training text set in the education domain; load the pre-built education domain terminology library, which contains core education concepts, terms, and their synonym mappings; perform at least one of the following augmentation operations on the sentences in the training text: Synonym replacement involves identifying words in a sentence that belong to the terminology database and randomly replacing them with synonyms with a preset probability based on the synonym mapping relationship in the database, thereby increasing vocabulary diversity. Sentence back-translation involves translating a sentence into an intermediate language and then translating it back into the original language to generate new sentences that are semantically consistent but express different ideas, which are used to expand the corpus. Random deletion involves randomly deleting non-key words or clauses from a sentence with a preset probability to improve the model's robustness to incomplete information. Model training integrating multiple regularization techniques: The enhanced training dataset is divided into a training set and a validation set; Construct or load the initial network parameters of the resource extraction model; During the model training iteration, the following regularization technique is applied simultaneously: Dropout, which randomly sets the output of a portion of neurons to zero during the forward propagation process before the fully connected layer or attention layer of the model, with a dropout probability set to 0.1 to 0.5. Weight decay, implemented in the optimizer (AdamW) with L2 regularization, penalizes the magnitude of the weights by adding the product of the sum of squares of the model weights and the decay coefficient to the loss function. The formula is: Total Loss = Original Loss + λ Σ(θ²), where λ is the decay coefficient and θ is the model weight parameter; Layer normalization, applied after each sublayer of the Transformer encoder, normalizes the output of all feature dimensions for a single sample to stabilize training and accelerate convergence. Early stopping based on validation set performance monitoring terminates training: After each training epoch, the current model is evaluated using the validation set, and the validation set loss and key task metrics (such as the ROUGE score generated by the summary) are calculated and recorded. The validation set performance is continuously monitored. When the validation set loss no longer decreases within N consecutive training epochs, or the key task metrics no longer improve, early stopping is triggered to terminate the training process and roll back to the model parameter snapshot with the best validation set performance, where N is a preset positive integer.

[0026] In one embodiment, generating and exporting a generic resource template that matches the resource content specifically includes: Using the denoised resource semantic representation as input, the resource extraction model or a dedicated template generation module is called again to generate template creation instructions that describe the target resource framework and filling requirements. Based on the template creation instructions, an initial template framework is matched or dynamically created from the template library; The initial template framework is iteratively adjusted and verified by importing standardized sample data, and a general resource template that meets the requirements is output.

[0027] An educational resource sharing system based on an educational cloud platform includes: Resource dataset acquisition module: Acquires the service cloud platform resource dataset to be processed from the education cloud platform; Deep semantic understanding and extraction module: It has a built-in pre-trained resource extraction model for deep analysis of resource datasets and outputting structured recognition results; Semantic denoising and representation module: Cleans and refines the structured recognition results to obtain a pure resource semantic representation; Resource template generation module: Based on the resource semantic representation, it generates and exports a concrete general resource template.

[0028] In one embodiment, the deep semantic understanding and extraction module includes a localized model deployment unit and a structured output parsing unit. The localized model deployment unit deploys the lightweight large language model on a private server or edge computing device. The structured output parsing unit parses and verifies the structured data output by the model and converts it into a unified semantic representation format within the system.

[0029] In one embodiment, the resource template generation module includes an instruction generation unit, a template adaptation and adjustment unit, and a template export unit. The instruction generation unit generates executable template construction instructions based on semantic representation. The template adaptation and adjustment unit matches or initializes the template according to the instructions and performs multiple rounds of optimization using standard data. The template export unit outputs the finally determined general resource template to the cloud platform shared resource library in a standardized format.

[0030] All data obtained in this invention has been authorized by the user.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An education resource sharing method based on an education cloud platform, characterized in that, Includes the following steps: Obtain the service cloud platform resource dataset, which contains the original content of the educational resources to be processed; The resource dataset of the service cloud platform is input into a pre-trained resource extraction model, which performs deep semantic understanding and information extraction on the resource content and outputs a structured recognition result containing resource summary, global semantics of the resource, types of resource elements and keywords. The pre-trained resource extraction model is a lightweight large language model that has been fine-tuned in the field of education. Based on the structured recognition results, invalid interference content is filtered out to obtain the denoised semantic representation of the resources. In response to the resource semantic representation, a general resource template matching the resource content is generated and exported.

2. The method of claim 1, wherein, Obtain the service cloud platform resource dataset, specifically including: By using the resource crawling engine configured in the education cloud platform, based on predefined resource types, subject tags and update cycle strategies, the system actively scans and discovers newly released or updated educational resource datasets within the education cloud platform. For the original content of educational resources in different modalities and formats, the corresponding standardized adapter is called for unified access, and the unstructured or semi-structured original data is initially converted into an intermediate representation. The intermediate representation of the data is cleaned, deduplicated, and formatted, and core metadata related to the resources is automatically extracted or supplemented. The dataset is scanned using pre-defined filtering rules and a content security model to identify and filter out content containing sensitive information, copyright disputes, or content that does not comply with educational standards.

3. The method of claim 1, wherein, The pre-trained resource extraction model is a lightweight large language model fine-tuned in the education domain. Specifically, the pre-trained resource extraction model is a model built on the Transformer architecture and obtained by domain-adaptive pre-training or supervised fine-tuning on educational professional corpora. It is then deployed locally on the service nodes or edge devices of the education cloud platform.

4. The method of claim 3, wherein, The process of performing deep semantic understanding and information extraction on resource content using a resource extraction model, and simultaneously outputting structured recognition results, specifically includes: By designing specific prompt word engineering, the lightweight large language model is guided to generate structured output containing the resource summary, the global semantics of the resource, the types of resource elements, and the keywords of the resource; The resource keywords and their semantic relationships are represented in the form of a semantic network or graph structure.

5. The method of claim 4, wherein, The resource extraction model incorporates an educational terminology knowledge graph for reinforcement learning during the training phase, specifically: Establish a terminology knowledge graph that includes core concepts in the field of education and their relationships; During the training or inference process of the resource extraction model, when a word in the input text is identified as matching a standard node in the terminology knowledge graph, the corresponding semantic representation within the resource extraction model is weighted and strengthened or the nodes are linked.

6. The method of claim 3, wherein, During the training of the resource extraction model, a strong regularization strategy for text data was adopted, including: The training data was augmented using synonym replacement, sentence back-translation, and random deletion techniques based on an education-related terminology database. During model training, Dropout, weight decay, and layer normalization techniques are applied. And an early stopping method is used to terminate training based on validation set performance monitoring.

7. The method of claim 1, wherein, The specific steps of generating and exporting a generic resource template that matches the resource content include: Using the denoised resource semantic representation as input, the resource extraction model or a dedicated template generation module is called again to generate template creation instructions that describe the target resource framework and filling requirements. Based on the template creation instructions, an initial template framework is matched or dynamically created from the template library; The initial template framework is iteratively adjusted and verified by importing standardized sample data, and a general resource template that meets the requirements is output.

8. An education resource sharing system based on an education cloud platform for implementing the method of any one of claims 1 to 6, characterized in that, include: Resource dataset acquisition module: Acquires the service cloud platform resource dataset to be processed from the education cloud platform; Deep semantic understanding and extraction module: It has a built-in pre-trained resource extraction model for deep analysis of resource datasets and outputting structured recognition results; Semantic denoising and representation module: Cleans and refines the structured recognition results to obtain a pure resource semantic representation; Resource template generation module: Based on the resource semantic representation, it generates and exports a concrete general resource template.

9. The system of claim 8, wherein, The deep semantic understanding and extraction module includes a localized model deployment unit and a structured output parsing unit. The localized model deployment unit deploys the pre-trained resource extraction model on a private server or edge computing device. The structured output parsing unit parses and verifies the structured data output by the model and converts it into a unified semantic representation format.

10. The system according to claim 8, characterized in that, The resource template generation module includes an instruction generation unit, a template adaptation and adjustment unit, and a template export unit. The instruction generation unit generates executable template construction instructions based on semantic representation. The template adaptation and adjustment unit matches or initializes the template according to the instructions and performs multiple rounds of optimization using standard data. The template export unit outputs the final determined general resource template to the cloud platform shared resource library in a standardized format.

Images