An education resource semantic retrieval method based on a knowledge graph

Abstract

The application relates to the technical field of computer information processing, and provides an education resource semantic retrieval method based on a knowledge graph. The method constructs a science education knowledge graph by structurally processing course textbooks, courseware, multimedia resources and experimental videos; in the knowledge graph construction process, sequential hypothesis testing and e-process evidence accumulation mechanisms are introduced to statistically control the significance of entity recognition results, thereby improving the stability and reliability of knowledge modeling; meanwhile, a subgraph graph neural network model with relationship strength perception is constructed, a resource structure semantic representation conforming to teaching logic is obtained, and the resource structure semantic representation is fused with a text semantic representation, so that semantic retrieval and sorting of the knowledge graph are enhanced. The method can output retrieval results that are coherent in teaching logic and adapt to different teaching scenes, and improves the accuracy of education resource retrieval and the teaching auxiliary effect.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, and in particular to a semantic retrieval method for educational resources based on knowledge graphs. Background Technology

[0002] With the application of artificial intelligence, natural language processing, and deep learning technologies in the field of educational informatization, existing educational resource retrieval systems are gradually incorporating semantic vector representation, pre-trained language models, and intelligent recommendation algorithms to improve the automated retrieval and recommendation capabilities of educational resources, and to some extent, address the insufficient semantic understanding capabilities of traditional keyword matching methods. However, most existing intelligent retrieval technologies still rely primarily on textual semantic similarity, making it difficult to effectively express the objectively existing prerequisite dependencies, theoretical derivations, and structured relationships between knowledge points and experimental processes within the science knowledge system. This results in deficiencies in the retrieval results regarding the logical coherence of teaching and the completeness of knowledge evolution. Meanwhile, some educational systems that incorporate knowledge graphs typically rely on static rules or fixed confidence thresholds for entity recognition and knowledge modeling, lacking statistical control over model prediction uncertainty and sequence context. This makes them susceptible to noise or local fluctuations, affecting the stability of the knowledge graph structure and thus weakening the semantic retrieval and resource ranking effects. Furthermore, existing systems lack effective modeling of the differences in the importance of different teaching relationships during the resource ranking and result display stages, making it difficult to simultaneously address the comprehensive needs of dynamic demonstrations, experimental processes, and knowledge expansion in both classroom teaching and personalized learning scenarios. Overall, there is still room for improvement in the teaching support effect. Summary of the Invention

[0003] To address the shortcomings of existing intelligent educational resource retrieval technologies in STEM teaching scenarios, such as insufficient knowledge structure modeling capabilities, poor entity recognition stability, and difficulty in reflecting teaching logic in retrieval results, this invention proposes a knowledge graph-based semantic retrieval method for educational resources. By integrating natural language semantic modeling with structured relationship modeling of the STEM knowledge system, it achieves high-quality semantic retrieval oriented towards teaching logic. The core innovation of this invention lies in introducing a sequential hypothesis testing and e-process evidence accumulation mechanism with statistical significance guarantees during the construction of the STEM educational knowledge graph. This dynamically and statistically discriminates the confidence level of entity prediction, thereby improving the stability of entity recognition and knowledge modeling without relying on fixed thresholds or sample length assumptions. Simultaneously, considering the differences in the importance of teaching relationships such as prerequisite dependencies, theoretical derivations, and experimental correspondences, a sub-graph graph neural network model with relationship strength awareness is constructed, enabling graph embedding to highlight knowledge relationships that contribute more to teaching semantics. Based on this, textual semantic representation and graph structure embedding are integrated, along with knowledge graph-enhanced retrieval and ranking mechanisms. This transforms the retrieval results from being driven solely by semantic similarity to reflecting both knowledge structure and teaching logic, thus providing technical support for dynamic visualization-assisted teaching and personalized learning.

[0004] This invention provides a semantic retrieval method for educational resources based on knowledge graphs, applied in a teaching-assisted learning scenario. The scenario includes: a semantic retrieval server, a knowledge graph database server, and an educational resource storage server deployed in an educational cloud data center; and teacher and student terminals located in science labs, multimedia classrooms, and student study areas. The teacher and student terminals communicate with the semantic retrieval server via a campus network. The method is executed by the semantic retrieval server and includes:

[0005] Step S1: Obtain course educational resources from the educational resource storage server, and perform word segmentation, part-of-speech tagging, and named entity recognition on the title, tags, text descriptions, and subtitle information of the course educational resources to obtain structured text data;

[0006] Step S2: Knowledge Graph Construction: A Bidirectional Long Short-Term Memory Network-Conditional Random Field (BiLSTM-CRF) model is used to perform named entity recognition on structured text data. Sequential hypothesis testing and an e-process mechanism are introduced during the entity recognition process to perform sequential statistical tests on the confidence scores of the entity predictions output by the model. When the accumulated statistical evidence meets the preset significance conditions, the entity recognition result is determined, thereby automatically extracting entities from the science domain. A science education knowledge graph is formed and stored in a knowledge graph database server.

[0007] Step S3: Using a pre-trained text encoding model based on the Transformer architecture, the titles, descriptions, and subtitles of the target educational resources in the course educational resources are encoded into text semantic vectors; based on the knowledge point entities, experimental entities, and chapter entities associated with the target educational resources in the science education knowledge graph, a relation strength-aware subgraph GNN model is used to perform graph embedding calculations on the relevant subgraphs to obtain graph structure embedding vectors; the text semantic vectors and graph structure embedding vectors are weighted and fused to obtain the semantic retrieval representation of educational resources and establish an index; the relation strength-aware subgraph GNN model is constructed as follows: a GNN model is established, and a relation strength-aware structured weight calculation mechanism is introduced to improve the neighborhood information propagation and aggregation process of the GNN model, thus constructing the relation strength-aware subgraph GNN model;

[0008] Step S4: When the teacher terminal and student terminal submit a search request through the search interface, the semantic search server performs word segmentation, stop word filtering and phrase recognition on the search text in the search request to obtain candidate science terms and knowledge point phrases; based on the knowledge point entity tags in the science education knowledge graph, the candidate science terms and knowledge point entities are linked to achieve the alignment of the search text and knowledge point entities, and a semantic vector representation of the search statement is generated.

[0009] Step S5 takes the knowledge point entity corresponding to the retrieval statement as the center, performs multi-hop neighborhood expansion along the prerequisite relationship, derivation relationship and experiment correspondence relationship in the science education knowledge graph, and obtains the set of extended knowledge points related to the semantics of the retrieval statement; based on the similarity between the semantic vector representation of the retrieval statement and the semantic retrieval representation of educational resources, the connection strength between extended knowledge points and curriculum educational resources in the science education knowledge graph, and the weight of curriculum educational resource type, calculates the comprehensive relevance score;

[0010] Step S6: Select educational resources from high to low according to the comprehensive relevance score, and prioritize displaying the resources marked as dynamic courseware, dynamic demonstration videos and experimental demonstration videos in the search results interface of the teacher terminal and the student terminal.

[0011] Furthermore, step S2 specifically includes the following steps:

[0012] Step S21: Perform unified encoding and concatenation on the structured fields of the structured text data to form serialized text; perform word segmentation, position alignment and label space definition on the serialized text to establish a sequence sample set for named entity recognition;

[0013] Step S22: Input each sequence sample in the sequence sample set into the BiLSTM-CRF model to obtain the label score matrix for each position in the sequence and the corresponding optimal decoded label sequence; based on the label score matrix and the decoded label sequence, parse the segments composed of consecutive labels of the same type to generate candidate entity segments, and calculate the prediction confidence of each candidate entity segment; use the candidate entity segments and their corresponding prediction confidence as the observation input for sequential statistical testing, and construct the entity candidate stream in sequence order;

[0014] Step S23: For each candidate entity fragment in the entity candidate stream, construct a corresponding sequential hypothesis test pair to statistically determine the entity validity of the candidate entity fragment. The sequential hypothesis test pair includes: 1. Null hypothesis: The candidate entity fragment does not meet the determination criteria of the target entity; 2. Alternative hypothesis: The candidate entity fragment meets the determination criteria of the target entity. Based on this, construct an evidence mapping function to map the entity prediction confidence of the model output into accumulative statistical evidence to characterize the support strength of the candidate entity fragment for the alternative hypothesis in the current observation round. When the entity prediction confidence of the candidate entity fragment is obtained in the current observation round, generate the corresponding single-step statistical evidence increment based on the evidence mapping function.

[0015] Step S24: For each candidate entity fragment in the entity candidate stream, construct the corresponding e-process statistical evidence accumulation process and initialize it; based on the single-step statistical evidence increment, update the e-process statistical evidence accumulation process according to the sequential recursion rule of satisfying the statistical constraints under the null hypothesis, and continuously obtain the e-process statistical evidence accumulation value corresponding to the candidate entity fragment.

[0016] Step S25: Set the statistical significance level and construct the sequential rejection boundary based on the significance level. For each candidate entity segment in the entity candidate stream, continuously monitor its corresponding e-process statistical evidence accumulation value, and perform dynamic stop judgment according to the first out-of-bounds rule to obtain the statistical confirmation status of the candidate entity segment and the corresponding evidence index.

[0017] Step S26: Based on the statistical confirmation status and corresponding evidence indicators of candidate entity fragments, when candidate entity fragments in the entity candidate flow conflict at the entity boundary, a conflict resolution mechanism based on sequential statistical characteristics is introduced to make a unified decision. For candidate entity fragments selected by the conflict resolution mechanism, entity confirmation and freezing processing are performed. After the conflict decision is completed, evidence suppression processing is performed on the unselected candidate entity fragments to form entity recognition results with consistent boundaries, definite types and controlled statistical significance, thereby automatically extracting entities in the science domain.

[0018] Furthermore, based on the knowledge point entities, experimental entities, and chapter entities associated with the target educational resources in the science education knowledge graph, the process of using a relation strength-aware subgraph GNN model to perform graph embedding calculations on the relevant subgraphs to obtain graph structure embedding vectors includes the following steps:

[0019] Step S31: Based on the knowledge point entities, experimental entities, and chapter entities in the science education knowledge graph that are related to the target educational resources, construct a resource association subgraph for graph embedding computation;

[0020] Step S32: Through the structured weight calculation mechanism that is aware of relation strength, the relation strength weight value of each relation edge is calculated based on the semantic type identifiers corresponding to different types of relation edges in the resource association subgraph, the co-occurrence frequency of the associated entities in the teaching resources, the hierarchical distance of the associated entities in the course chapter structure, and the correspondence tightness information between the experimental entities and the knowledge point entities. The relation strength weight value is used as the structured modulation parameter in the neighborhood information aggregation process.

[0021] Step S33: On the resource association subgraph, based on the node representation update mechanism of the GNN model, multiple rounds of neighborhood information propagation and aggregation are performed on each node; during the node information aggregation process, the relation strength weight value of the relation edge is introduced to weight and modulate the information contribution from different neighboring nodes, so as to highlight the knowledge points, experiments and chapter nodes that have a higher semantic contribution to educational resources.

[0022] Step S34: After completing the resource association subgraph embedding propagation calculation for the predetermined rounds, the core knowledge point node corresponding to the target educational resource is used as the aggregation center to perform weighted aggregation processing and generate graph structure embedding vector.

[0023] Furthermore, the display methods of the search results in step S6 include:

[0024] a) On the teacher's terminal, dynamic courseware and experimental demonstration videos are displayed in a way that sorts them by the main line of knowledge points, and teachers are supported to play the dynamic derivation process and experimental process of science knowledge points in the form of a timeline.

[0025] b) On the student terminal, short-duration dynamic demonstration videos and typical experimental videos that match the difficulty of the knowledge points being searched are placed on the first screen, and a recommendation list of "prerequisite knowledge points" and "extended experiments" is provided to guide students to expand from core knowledge points to related concepts and experimental operations.

[0026] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0027] This invention introduces sequential hypothesis testing and e-process statistical evidence accumulation mechanisms into the construction of a science education knowledge graph. This enables dynamic statistical confirmation and saliency-controlled modeling of entity recognition results in the science field, improving the stability and consistency of entity extraction results in continuous text scenarios such as textbook texts, teaching instructions, and resource descriptions. It solves the problem that existing intelligent educational resource retrieval technologies rely on fixed confidence thresholds for entity confirmation, which are susceptible to model fluctuations and noise interference. This enhances the reliability and sustainable expansion capability of the science education knowledge graph structure, providing a solid data foundation for subsequent semantic retrieval and teaching assistance applications based on knowledge structures.

[0028] This invention addresses the varying importance of different types of teaching relationships in science education, such as prerequisite dependencies, theoretical derivations, and experimental correspondences. It constructs a subgraph graph neural network model that perceives relationship strength, enabling structured modeling and weighted expression of the semantic contribution of different teaching relationships. This enhances the ability of educational resource semantic representation to depict the evolutionary logic of science knowledge and the relevance of experimental teaching. It also solves the problem of existing graph embedding methods treating different teaching relationships equally, leading to distorted semantic expression and unclear teaching logic in retrieval results. Furthermore, it strengthens the rationality and interpretability of educational resource semantic retrieval results in terms of organizing the main teaching line and presenting the knowledge structure.

[0029] Building upon the above, this invention integrates text semantic representation with knowledge graph structure embedding, and combines knowledge graph-enhanced retrieval and ranking mechanisms. This enables educational resource retrieval results to comprehensively reflect semantic relevance, knowledge association strength, and resource type adaptability, thereby improving the practical value and teaching relevance of retrieval results in classroom teaching and self-directed learning scenarios. It solves the problem that existing intelligent retrieval systems cannot simultaneously meet the needs of teachers' systematic teaching presentations and students' tiered learning, and enhances the utilization efficiency and teaching assistance effect of dynamic and visualized teaching resources. As a result, it provides more accurate, coherent, and efficient intelligent support for science teaching. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a semantic retrieval method for educational resources based on knowledge graphs proposed in this invention.

[0031] Figure 2 This is a schematic diagram of the BiLSTM-CRF model proposed in Example 1; where CRF represents Conditional Random Field and LSTM represents Long Short-Term Memory Network. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] Example 1, according to Figure 1 , Figure 2 This invention provides a semantic retrieval method for educational resources based on knowledge graphs, applied in a teaching-assisted learning scenario. The scenario includes: a semantic retrieval server, a knowledge graph database server, and an educational resource storage server deployed in an educational cloud data center; and teacher and student terminals located in science labs, multimedia classrooms, and student study areas. The teacher and student terminals communicate with the semantic retrieval server via a campus network. The method is executed by the semantic retrieval server and includes:

[0034] Step S1: Structured processing of educational resources: Obtain course educational resources from the educational resource storage server. Course educational resources include: textbook chapter documents, dynamic courseware, multimedia animations, experimental demonstration videos and experimental teaching micro-lessons. Perform word segmentation, part-of-speech tagging and named entity recognition on the titles, tags, text descriptions and subtitles of the course educational resources to obtain structured text data.

[0035] Step S2: Knowledge Graph Construction: A Bidirectional Long Short-Term Memory Network-Conditional Random Field (BiLSTM-CRF) model is used to perform named entity recognition on structured text data. Sequential hypothesis testing and an e-process mechanism are introduced during the entity recognition process to perform sequential statistical tests on the confidence levels of the entity predictions output by the model. When the accumulated statistical evidence meets the preset significance conditions, the entity recognition result is determined, thereby automatically extracting entities from the science domain. Science domain entities include: mathematical theorems, formula names, physical concepts, chemical reactions, experimental instruments, experimental procedures, and course chapters. Based on the entity recognition results, different categories of science domain entities are mapped to multi-type nodes in the knowledge graph, and the nodes are further defined according to their relevance to textbook texts, teaching instructions, and resource descriptions. The semantic relationships described above are used to construct multi-type relation edges between entities. These relation edges include: 1. Prerequisite dependencies representing the logical order of knowledge points in the curriculum system; 2. Derivation relationships representing the theoretical deduction process between mathematical and physical concepts; 3. Experimental correspondence relationships representing the correspondence between knowledge points and experimental processes; 4. Co-occurrence relationships representing the co-occurrence of different knowledge points or experiments in the same teaching scenario; 5. Resource mapping relationships representing the pointing relationships between knowledge points, experiments, and specific teaching resources. Through the construction of the above multi-type nodes and multi-type relation edges, a science education knowledge graph with "knowledge points - experiments - courseware resources" as its core structure is formed. This science education knowledge graph is stored in a knowledge graph database server to support subsequent semantic retrieval, knowledge expansion, and educational resource recommendation.

[0036] Step S3: Construction of Semantic Representation of Educational Resources: A pre-trained text encoding model based on the Transformer architecture is used to encode the title, text description, and subtitle text of the target educational resources in the course into text semantic vectors; based on the knowledge point entities, experimental entities, and chapter entities associated with the target educational resources in the science education knowledge graph, a relation strength-aware subgraph GNN model is used to perform graph embedding calculation on the relevant subgraphs to obtain graph structure embedding vectors; the text semantic vectors and graph structure embedding vectors are weighted and fused to obtain the semantic retrieval representation of educational resources and establish an index; the construction method of the relation strength-aware subgraph GNN model is as follows: a GNN model is established, and a relation strength-aware structured weight calculation mechanism is introduced to improve the neighborhood information propagation and aggregation process of the GNN model, thus constructing the relation strength-aware subgraph GNN model;

[0037] Step S4: Semantic parsing and knowledge point alignment: When teacher terminals and student terminals submit search requests through the search interface, the semantic search server performs word segmentation, stop word filtering, and phrase recognition on the search text in the search request to obtain candidate science terms and knowledge point phrases; based on the knowledge point entity tags in the science education knowledge graph, the candidate science terms and knowledge point entities are linked to achieve the alignment of the search text and knowledge point entities, and a semantic vector representation of the search statement is generated;

[0038] Step S5: Knowledge Graph Enhanced Retrieval and Ranking: Centered on the knowledge point entity corresponding to the retrieval statement, perform multi-hop neighborhood expansion along prerequisite, derivation, and experiment correspondence relationships in the science education knowledge graph to obtain a set of extended knowledge points semantically related to the retrieval statement; based on the similarity between the semantic vector representation of the retrieval statement and the semantic retrieval representation of educational resources, the connection strength between extended knowledge points and curriculum educational resources in the science education knowledge graph, and the weight of curriculum educational resource types, calculate a comprehensive relevance score and rank the candidate educational resources;

[0039] Formula for calculating the overall relevance score:

[0040] ;

[0041] in, Indicates the first The comprehensive relevance score corresponding to each candidate educational resource; Represents the semantic vector of the search statement Representation vectors for semantic retrieval of educational resources Semantic similarity between them; Indicates the first The connection strength score of each candidate educational resource and its extended knowledge point set in the knowledge graph; Indicates the first The resource type weights corresponding to each candidate educational resource; , , Indicates the weighting coefficient;

[0042] Step S6: Generation and Push of Search Results: Select educational resources from high to low according to the comprehensive relevance score, and prioritize displaying the resources marked as dynamic courseware, dynamic demonstration videos and experimental demonstration videos in the search results interface of the teacher's terminal and the student's terminal, so as to realize the dynamic and visual auxiliary teaching display of science knowledge points such as "Pythagorean theorem", "mechanics experiment" and "circuit experiment".

[0043] Example 2, this example is based on Example 1. In this example, step S2 specifically includes the following steps:

[0044] Step S21: Perform unified encoding and concatenation on the structured fields of the structured text data to form serialized text; perform word segmentation, position alignment and label space definition on the serialized text to establish a sequence sample set for named entity recognition, wherein the label space includes entity labels corresponding to mathematical theorems, formula names, physical concepts, chemical reactions, experimental instruments, experimental steps and course chapters;

[0045] Step S22: Input each sequence sample in the sequence sample set into the BiLSTM-CRF model to obtain the label score matrix for each position in the sequence and the corresponding optimal decoded label sequence; based on the label score matrix and the decoded label sequence, parse the segments composed of consecutive labels of the same type to generate candidate entity segments, and calculate the prediction confidence of each candidate entity segment; wherein, the prediction confidence includes: 1. boundary confidence representing the reliability of the start and end positions of the candidate entity segment; 2. intra-segment label consistency confidence representing the degree of label consistency within the candidate entity segment; 3. path probability confidence representing the credibility of the overall decoding path corresponding to the candidate entity segment; use the candidate entity segments and their corresponding prediction confidence as the observation input for sequential statistical testing, construct the entity candidate stream in sequence order, and use it as the test object for subsequent sequential hypothesis testing and e-process mechanism;

[0046] Step S23: For each candidate entity fragment in the entity candidate stream, construct a corresponding sequential hypothesis test pair to statistically determine the entity validity of the candidate entity fragment. The sequential hypothesis test pair includes: 1. Null hypothesis: The candidate entity fragment does not meet the determination criteria of the target entity; 2. Alternative hypothesis: The candidate entity fragment meets the determination criteria of the target entity. Based on this, construct an evidence mapping function to map the entity prediction confidence of the model output into accumulative statistical evidence to characterize the support strength of the candidate entity fragment for the alternative hypothesis in the current observation round. When the entity prediction confidence of the candidate entity fragment is obtained in the current observation round, generate the corresponding single-step statistical evidence increment based on the evidence mapping function. Use the single-step statistical evidence increment as input evidence in the sequential hypothesis testing process for subsequent cumulative statistical determination.

[0047] The formula used by the evidence mapping function to generate the corresponding single-step statistical evidence increment is as follows:

[0048] ;

[0049] in, Indicates the time step index (observation round) of the sequential test. Indicates the first The incremental statistical evidence generated in each round measures the strength of support for the alternative hypothesis from the confidence observations of the current round. Indicates the first The entity prediction confidence is calculated for the candidate entity fragment. Represents the evidence mapping function; This indicates that, under the null hypothesis, the mapping function... The normalized upper bound;

[0050] Step S24: For each candidate entity segment in the entity candidate stream, construct and initialize the corresponding e-process statistical evidence accumulation process; based on the single-step statistical evidence increment, update the e-process statistical evidence accumulation process according to the sequential recursion rule that satisfies the statistical constraints under the null hypothesis, and continuously obtain the e-process statistical evidence accumulation value corresponding to the candidate entity segment; thereby modeling the entity prediction result of the neural network as a sequential evidence evolution process with strict statistical guarantees, so as to realize the continuous accumulation and dynamic evolution of supporting evidence for candidate entity segments;

[0051] The sequential recursive rule for statistical constraints under the null hypothesis specifically includes: in the t-th observation round, the incremental single-step statistical evidence generated in the t-th round is recursively updated with the cumulative e-process statistical evidence value in the (t-1)-th round, resulting in:

[0052] ;

[0053] in, Indicates the first Cumulative e-process statistical evidence value at each observation round. Indicates the first The cumulative evidence value across rounds is used as a benchmark for comparison;

[0054] The e-process statistical evidence accumulation process satisfies the supermartingale constraint under the condition that the null hypothesis is true. Specifically, the conditional expectation of the statistical evidence accumulation process under the condition that the null hypothesis is true does not increase with the increase of the number of observation rounds. This ensures that the statistical evidence is controlled at the preset significance level in any observation round, and achieves effective statistical significance control at any time without relying on a fixed sample length or static threshold.

[0055] The e-process statistical evidence accumulation process satisfies the supermartingar constraint under the null hypothesis:

[0056] ;

[0057] in, Indicates the null hypothesis, Indicates as of the date The historical σ-algebra up to the next round; Indicates that in the known historical information Under the premise, and in Upon establishment, the next step is to accumulate evidence. The expected value of the condition;

[0058] Through the above-mentioned e-process evidence accumulation and significance control mechanism, the confidence of entity prediction output by the model is subjected to sequential statistical test, avoiding premature judgment, random fluctuation amplification or bias accumulation caused by entity confirmation based on a single confidence threshold, and improving the stability and reliability of entity recognition results in the continuous observation process.

[0059] Step S25: Set the statistical significance level and construct the sequential rejection boundary based on the significance level. For each candidate entity segment in the entity candidate stream, continuously monitor its corresponding e-process statistical evidence accumulation value, and perform dynamic stop judgment according to the first out-of-bounds rule to obtain the statistical confirmation status of the candidate entity segment and the corresponding evidence index.

[0060] The dynamic stop determination of the first out-of-bounds rule is as follows: when the cumulative value of statistical evidence corresponding to a candidate entity fragment exceeds the sequential rejection boundary in a certain observation round, it is determined that the cumulative statistical evidence of the candidate entity fragment has met the preset significance condition, and then the candidate entity fragment is confirmed as the target entity. The entity boundary and entity type are confirmed and frozen, and the subsequent evidence update process of the candidate entity fragment is terminated.

[0061] If the cumulative statistical evidence value corresponding to a candidate entity fragment does not reach the sequential rejection boundary within the preset maximum number of observation rounds, the null hypothesis is maintained and the candidate entity fragment is marked as an unconfirmed fragment to suppress the risk of false confirmation caused by noisy samples or instantaneous high confidence fluctuations.

[0062] Sequential rejection boundary formula:

[0063] ;

[0064] in, Indicates the sequential rejection boundary. Indicates the pre-set statistical significance level;

[0065] The formula for dynamically stopping the execution of the first out-of-bounds rule (core sequential mechanism) is as follows:

[0066] ;

[0067] in, Indicates the moment of the first boundary crossing. Indicates the index of the infimum / minimum satisfaction;

[0068] Step S26: Based on the statistical confirmation status and corresponding evidence indicators of candidate entity fragments, when candidate entity fragments in the entity candidate stream conflict at the entity boundary, a conflict resolution mechanism based on sequential statistical characteristics is introduced for unified adjudication. For candidate entity fragments selected by the conflict resolution mechanism, entity confirmation and freezing processing are performed. After completing the conflict adjudication, evidence suppression processing is performed on the unselected candidate entity fragments. Evidence suppression processing includes: 1. resetting its statistical evidence accumulation process; 2. applying a decay weight to its subsequent single-step statistical evidence increment; thereby avoiding duplicate confirmation or conflict propagation. Through the above parallel sequential testing and conflict resolution mechanism, entity identification results with consistent boundaries, definite types, and controlled statistical significance are formed, thereby automatically extracting entities in the science domain.

[0069] The conflict resolution mechanism includes the following judgment rules: 1. Priority rule for crossing the boundary: Among candidate entity segments that meet the significance condition, the candidate entity segment that first crosses the sequential rejection boundary and has a smaller corresponding stopping time is selected as the entity result with earlier stable statistical evidence; 2. Priority rule for evidence strength: When multiple candidate entity segments have the same or similar crossing times, their cumulative e-process statistical evidence values ​​at the stopping time are further compared, and the candidate entity segment with higher statistical evidence strength is selected as the final confirmation result.

[0070] In the conventional technical field, step S2 involves using a bidirectional long short-term memory network-conditional random field (BiLSTM-CRF) model to perform named entity recognition on structured text data, determining the entity recognition results, and thus automatically extracting entities from the science and technology domain; specifically, this includes the following steps:

[0071] Step C1: Sequence Sample Construction: Concatenate the structured fields of the structured text data to form a serialized text; perform word segmentation, position alignment and label space definition on the serialized text to construct a sequence sample set for named entity recognition, where the label space includes entity labels corresponding to mathematical theorems, formula names, physical concepts, chemical reactions, experimental instruments, experimental steps and course chapters;

[0072] Step C2: Sequence Feature Encoding and Label Prediction: Input each sequence sample in the sequence sample set into the BiLSTM network to encode the sequence context features bidirectionally and obtain the hidden state representation at each position; based on the hidden state representation, jointly decode the sequence labels through the Conditional Random Field (CRF) layer to obtain the label score for each position in the sequence and the corresponding optimal label sequence;

[0073] Step C3: Entity Fragment Parsing: Based on the optimal label sequence, parse the text fragments corresponding to consecutive labels of the same type, extract the corresponding candidate entity fragments, and determine the entity boundaries and entity types of the candidate entity fragments;

[0074] Step C4: Entity Result Confirmation: Output the parsed candidate entity fragments as the final entity recognition result, which serves as the entity extraction result in the science domain.

[0075] Example 3, based on Example 2, describes a process for obtaining graph structure embedding vectors by using a relation strength-aware subgraph GNN model to perform graph embedding calculations on relevant subgraphs based on knowledge point entities, experimental entities, and chapter entities associated with the target educational resources in the science education knowledge graph. The specific steps include:

[0076] Step S31: Based on the knowledge point entities, experimental entities, and chapter entities in the science education knowledge graph that are related to the target educational resources, construct a resource association subgraph for graph embedding computation; wherein, the resource association subgraph includes multiple types of nodes and multiple types of relation edges, the multiple types of nodes include knowledge point nodes, experimental nodes, and chapter nodes, and the multiple types of relation edges include prerequisite dependencies, derivation relations, experimental correspondence relations, and resource mapping relations;

[0077] Step S32: Through the structured weight calculation mechanism that is aware of relation strength, the relation strength weight value of each relation edge is calculated based on the semantic type identifiers corresponding to different types of relation edges in the resource association subgraph, the co-occurrence frequency of the associated entities in the teaching resources, the hierarchical distance of the associated entities in the course chapter structure, and the correspondence tightness information between the experimental entities and the knowledge point entities. The relation strength weight value is used as the structured modulation parameter in the neighborhood information aggregation process.

[0078] Formula for relation strength weighting:

[0079] ;

[0080] in, Represents nodes in the resource association subgraph With nodes The relationship strength weight value of the edges between them The semantic type identifier representing the relation edge. Represents the node With nodes Hierarchical spacing within the course chapter structure; Represents a node With nodes Co-occurrence frequency in teaching resources; A mapping function representing the strength of a relationship;

[0081] Step S33: On the resource association subgraph, based on the node representation update mechanism of the GNN model, multiple rounds of neighborhood information propagation and aggregation are performed on each node; during the node information aggregation process, the relation strength weight value of the relation edge is introduced to weight and modulate the information contribution from different neighboring nodes, so as to highlight the knowledge points, experiments and chapter nodes that have a higher semantic contribution to educational resources.

[0082] Step S34: After completing the resource association subgraph embedding propagation calculation for a predetermined number of rounds, the core knowledge point node corresponding to the target educational resource is used as the aggregation center to perform weighted aggregation processing and generate a graph structure embedding vector to represent the structural semantic features of the target educational resource.

[0083] Example 4, based on Example 3, in which the display method of the search results in step S6 includes:

[0084] a) On the teacher's terminal, dynamic courseware and experimental demonstration videos are displayed in a way that sorts them by the main line of knowledge points, and teachers are supported to play the dynamic derivation process and experimental process of science knowledge points in the form of a timeline.

[0085] b) On the student terminal, short-duration dynamic demonstration videos and typical experimental videos that match the difficulty of the knowledge points being searched are placed on the first screen, and a recommendation list of "prerequisite knowledge points" and "extended experiments" is provided to guide students to expand from core knowledge points to related concepts and experimental operations.

[0086] Example 5, this example is based on Example 4, in this example,

[0087] Step S5: Knowledge Graph Enhanced Retrieval and Ranking: Centered on the knowledge point entity corresponding to the retrieval statement, perform multi-hop neighborhood expansion along prerequisite relations, derivation relations, and experiment correspondence relations in the science education knowledge graph to obtain a set of extended knowledge points semantically related to the retrieval statement; calculate the comprehensive relevance score based on the similarity between the semantic vector representation of the retrieval statement and the semantic retrieval representation of educational resources, the connection strength between extended knowledge points and curriculum educational resources in the science education knowledge graph, and the weight of curriculum educational resource types.

[0088] In this embodiment, a student terminal inputs the search query "Pythagorean theorem, experimental demonstration" in a mathematical learning scenario;

[0089] The overall relevance score is calculated, as shown in Table 1:

[0090] Table 1

[0091]

[0092] Step S6: Generation and Push of Search Results: Select educational resources from high to low according to the comprehensive relevance score, and prioritize displaying the resources marked as dynamic courseware, dynamic demonstration videos and experimental demonstration videos in the search results interface of the teacher's terminal and the student's terminal, so as to realize the dynamic and visual auxiliary teaching display of science knowledge points such as "Pythagorean theorem", "mechanics experiment" and "circuit experiment".

[0093] (a) Results screening and display strategy:

[0094] The semantic retrieval server selects the top 3 educational resources from highest to lowest based on their comprehensive relevance score and executes the following display strategy:

[0095] 1. Prioritize displaying dynamic resources:

[0096] Experimental demonstration video;

[0097] Dynamic courseware;

[0098] Dynamic demonstration video;

[0099] 2. Static resources are used as a supplementary display:

[0100] Textbook chapter documents;

[0101] Textual explanation materials.

[0102] (II) Terminal Display Example:

[0103] The system generates the following display structure in the search results interface of student terminals and teacher terminals:

[0104] First: Pythagorean theorem jigsaw puzzle experiment demonstration video (R2);

[0105] Second item: Dynamic courseware on the geometric proof of the Pythagorean theorem (R1);

[0106] Third: Micro-lecture video on the derivation of the Pythagorean theorem formula (R3).

[0107] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A semantic retrieval method for educational resources based on knowledge graphs, characterized in that, The method includes: Step S1: Obtain course educational resources, and perform word segmentation, part-of-speech tagging, and named entity recognition on the titles, tags, text descriptions, and subtitles of the course educational resources to obtain structured text data; Step S2: Use the BiLSTM-CRF model to perform named entity recognition on structured text data, and introduce sequential hypothesis testing and e-process mechanism in the entity recognition process to perform sequential statistical testing and determine the entity recognition results; thereby automatically extracting entities in the science field to form a science education knowledge graph. Step S3: Use a text encoding model to encode the title, text description, and subtitle text of the target educational resources in the course educational resources into text semantic vectors; based on the knowledge point entities, experiment entities, and chapter entities associated with the target educational resources in the science education knowledge graph, use the relation strength-aware subgraph GNN model to perform graph embedding calculation to obtain graph structure embedding vectors; weight and fuse the text semantic vectors and graph structure embedding vectors to obtain the semantic retrieval representation of educational resources; Step S4: Teacher terminals and student terminals submit search requests through the search interface, generating semantic vector representations of the search statements; Step S5: Calculate the comprehensive relevance score based on the semantic vector representation of the search statement and the semantic retrieval representation of educational resources; Step S6: Select educational resources from high to low according to the comprehensive relevance score, and prioritize displaying resources marked as dynamic courseware, dynamic demonstration videos and experimental demonstration videos in the search results interface of the teacher terminal and student terminal. The construction method of the relation strength-aware subgraph GNN model is as follows: establish a GNN model, introduce a relation strength-aware structured weight calculation mechanism to improve the neighborhood information propagation and aggregation process of the GNN model, and construct the relation strength-aware subgraph GNN model. The process of using a relation strength-aware subgraph GNN model to perform graph embedding computation and obtain graph structure embedding vectors includes the following steps: Step S31: Based on the knowledge point entities, experimental entities, and chapter entities in the science education knowledge graph that are related to the target educational resources, construct a resource association subgraph; Step S32: Through the structured weight calculation mechanism that is aware of relation strength, the relation strength weight value of each relation edge is calculated based on the semantic type identifiers corresponding to different types of relation edges in the resource association subgraph, the co-occurrence frequency of the associated entities in the teaching resources, the hierarchical distance of the associated entities in the course chapter structure, and the correspondence tightness information between the experimental entities and the knowledge point entities. The relation strength weight value is used as the structured modulation parameter in the neighborhood information aggregation process. Step S33: On the resource association subgraph, based on the node representation update mechanism of the GNN model, perform multiple rounds of neighborhood information propagation and aggregation for each node; during the node information aggregation process, the relation strength weight value of each relation edge is introduced for weighted modulation; Step S34: After completing the resource association subgraph embedding propagation calculation, generate the graph structure embedding vector.

2. The semantic retrieval method for educational resources based on knowledge graphs according to claim 1, characterized in that: Step S2 specifically includes the following steps: Step S21: Perform unified encoding and concatenation on the structured fields of the structured text data to establish a sequence sample set; Step S22: Input each sequence sample in the sequence sample set into the BiLSTM-CRF model to obtain the label score matrix and decode the label sequence, and parse the fragments composed of consecutive labels of the same type to generate candidate entity fragments and construct the entity candidate stream; Step S23: For each candidate entity fragment in the entity candidate stream, construct a sequential hypothesis test pair to statistically determine the entity validity of the candidate entity fragment; based on this, construct an evidence mapping function; and generate a single-step statistical evidence increment based on the evidence mapping function. Step S24: For each candidate entity fragment in the entity candidate stream, construct and initialize the e-process statistical evidence accumulation process; based on the single-step statistical evidence increment, update the e-process statistical evidence accumulation process according to the sequential recursive rule of satisfying the statistical constraints under the null hypothesis, and continuously obtain the e-process statistical evidence accumulation value corresponding to the candidate entity fragment. Step S25: Set the statistical significance level and construct the sequential rejection boundary based on the significance level. For each candidate entity segment in the entity candidate stream, continuously monitor the cumulative value of e-process statistical evidence and execute dynamic stop judgment according to the first out-of-bounds rule to obtain the statistical confirmation status of the candidate entity segment and the corresponding evidence index. Step S26: Based on the statistical confirmation status and corresponding evidence indicators of candidate entity fragments, when candidate entity fragments in the entity candidate flow conflict at the entity boundary, a conflict resolution mechanism based on sequential statistical characteristics is introduced for unified adjudication. For candidate entity fragments selected by the conflict resolution mechanism, entity confirmation and freezing processing are performed. After the conflict adjudication is completed, evidence suppression processing is performed on the unselected candidate entity fragments. Through the above sequential hypothesis testing and conflict resolution mechanism, entity recognition results are formed, and entities in the science domain are automatically extracted.

3. The semantic retrieval method for educational resources based on knowledge graphs according to claim 2, characterized in that: The sequential hypothesis testing pairs include:

1. Null hypothesis: The candidate entity fragment does not meet the determination criteria of the target entity; 2. Alternative hypothesis: The candidate entity fragment meets the determination criteria of the target entity.

4. The semantic retrieval method for educational resources based on knowledge graphs according to claim 2, characterized in that: The e-process statistical evidence accumulation process satisfies the supermartingale constraint under the null hypothesis, specifically, the conditional expectation of the statistical evidence accumulation process under the null hypothesis does not increase with the number of observation rounds.

5. The semantic retrieval method for educational resources based on knowledge graphs according to claim 2, characterized in that: The conflict resolution mechanism includes the following judgment rules:

1. Priority rule for the moment of boundary crossing; 2. Priority rule for the strength of evidence.

Images