Knowledge graph-based teaching resource database construction method and system

Abstract

The invention relates to a knowledge graph-based teaching resource database construction method and system, and belongs to the technical field of data resource processing, and the method comprises the following steps: S1, obtaining and preprocessing a multi-source heterogeneous teaching resource: carrying out the natural language processing of an unstructured text to recognize the association of an entity and a key concept, analyzing the user behavior log to extract hidden knowledge point association; s2, defining a multi-dimensional teaching knowledge graph mode layer: constructing a multi-dimensional graph mode layer; s3, collaborative extraction and fusion of graph entities and relationships: collaborative extraction of core teaching entities and specific relationship types from the preprocessed teaching resources to construct a knowledge graph; s4, adaptive graph updating and teaching path generation: taking the constructed knowledge graph as an underlying database, and responding to knowledge state query of a user; the method has the beneficial effects that massive and heterogeneous teaching resources and the knowledge graph are effectively fused.

Description

Technical Field

[0001] This invention belongs to the field of data resource processing technology, and specifically relates to a method and system for constructing a teaching resource database based on knowledge graphs. Background Technology

[0002] With the rapid development of educational informatization, teaching resources have experienced explosive growth, encompassing various types such as courseware, test questions, videos, and documents. However, current teaching resource databases face numerous problems in their construction and application. On the one hand, teaching resources are scattered across different platforms or systems, lacking effective connections between them, forming "information silos." This makes it difficult for teachers and students to quickly and accurately access comprehensive resources related to specific knowledge points. On the other hand, existing databases categorize teaching resources primarily based on simple attribute tags, such as subject and grade, failing to delve into the underlying knowledge logic relationships and thus hindering the development of personalized teaching and intelligent learning.

[0003] Knowledge graph technology, as a technology capable of effectively organizing, managing, and displaying knowledge, possesses powerful knowledge association and semantic expression capabilities. Applying knowledge graph technology to the construction of teaching resource databases holds promise for solving the aforementioned problems. However, currently, a mature and efficient construction method and system have not yet been established for effectively integrating massive, heterogeneous teaching resources with knowledge graphs to achieve accurate association and intelligent retrieval of teaching resources. Therefore, a knowledge graph-based method for constructing teaching resource databases is urgently needed to improve the utilization efficiency and service quality of teaching resources. Summary of the Invention

[0004] This invention provides a method and system for constructing a knowledge graph-based teaching resource database, which addresses the technical problem of effectively integrating massive, heterogeneous teaching resources with knowledge graphs. It collaboratively extracts core teaching entities and specific relationship types from preprocessed teaching resources to construct a knowledge graph, thereby achieving effective integration of massive, heterogeneous teaching resources with knowledge graphs.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] A method for constructing a knowledge graph-based teaching resource database includes the following steps:

[0007] Step S1: Acquisition and Preprocessing of Multi-Source Heterogeneous Teaching Resources: Acquire original teaching resources from at least two different types of sources, including structured course outlines, unstructured e-textbooks, and user behavior logs from online teaching platforms; preprocess the original teaching resources, including natural language processing of unstructured text to identify the association between entities and key concepts, and parsing user behavior logs to extract implicit knowledge point associations;

[0008] Step S2: Definition of Multidimensional Teaching Knowledge Graph Pattern Layer: Construct a multidimensional graph pattern layer. The graph pattern layer not only defines the core teaching entity types and their hierarchical relationships, but also defines specific relationship types used to describe the teaching logic. The specific relationship types include at least "prerequisite relationship", "increasing teaching depth relationship" and "common co-occurrence misunderstanding relationship".

[0009] Step S3: Collaborative Extraction and Fusion of Knowledge Entities and Relationships: Based on the knowledge graph schema layer, core teaching entities and specific relationship types are collaboratively extracted from preprocessed teaching resources to construct a knowledge graph. Specifically, "prerequisite relationships" and "increasing teaching depth relationships" between entities are extracted and verified by combining the analysis of structured data from the course syllabus with the semantic analysis results of the textbook content. Meanwhile, "common co-occurrence and misunderstanding relationships" are automatically discovered and generated by analyzing answer sequences and error patterns in massive user behavior logs using machine learning models.

[0010] Step S4: Adaptive Knowledge Graph Update and Teaching Path Generation: The completed knowledge graph is used as the underlying database to respond to users' knowledge status queries; based on the user's knowledge mastery, the knowledge graph is traversed in real time using graph algorithms to dynamically generate a personalized optimal learning path that spans multiple knowledge entities, and the most suitable teaching resources are recommended to the user for each node in the learning path, and the user's behavioral feedback on the new learning path is used as incremental data.

[0011] Optionally, in step S1, performing natural language processing on unstructured text to identify the associations between entities and key concepts involves extracting the logical relationships between entities / concepts in the teaching text using a relation classification model to calculate the text's association with a relation. The specific formula is as follows:

[0012] Calculate text using a pre-trained model Belongs to relational features correlation :

[0013] ;

[0014] Wherein, it indicates that the given text / entity belongs to At that time, it belongs to the relational feature. The probability, It is a relational feature Belongs to the set of relation categories ;

[0015] For text The BERT embedding vector is the text. The vector representation of the text is encoded by the BERT pre-trained model and is the model's vectorized capture of the text semantics. Relational features The corresponding weight vector is used to measure relation features. With text The degree of matching; Relational features The bias term; It is an exponential function used to convert the matching degree into a non-negative value and amplify the numerical difference, making high matching degrees stand out more.

[0016] For each relation feature With the given text / entity The correlation is calculated, specifically, The vector dot product measures the characteristics of the relationship. Weight vector and text The similarity of BERT embedding vectors, plus the bias This is to allow the model to adjust relational features more flexibly. Exponential function It is monotonically increasing; the larger the exponent, the stronger the relationship characteristics. With text The higher the matching degree;

[0017] The normalization term ensures that the sum of the probabilities of all relational features equals 1, satisfying the fundamental property of probability; specifically, This refers to traversing relation features, which belong to the set of relation categories. Each relationship in; Relational features The corresponding weight vector is used to measure the traversal relationship features. With text The degree of matching; A vector representation of the text; Features of traversal relations The bias term is used to adjust the probability offset and is also a parameter learned by the model. It is a set of relation categories All traversal relation features Summing the results;

[0018] First, calculate the matching index of a single relation to ensure the result is non-negative, amplifying the difference between high-matching and low-matching relations. Then, sum the index results of all relations to form the relation category set. All relational features The exponential results are all added together, and the numerator is a certain relationship. The exponential result, the denominator is all relational features The sum of the exponential results, by dividing the numerator by the denominator, allows the statement "the text belongs to each relation" to be transformed into a relational statement. .

[0019] Optionally, in step S1, the implicit knowledge point associations are extracted by calculating the cosine similarity between the implicit knowledge point vector and the knowledge point concept vector to determine the degree of association.

[0020] ;

[0021] in, The similarity value ranges from [−1, 1]. The closer the value is to 1, the more relevant the association; the closer the value is to -1, the less relevant the association.

[0022] For knowledge points and concepts; These are implicit knowledge points; Text embedding vectors for knowledge point concepts; The feature vectors of implicit knowledge points; and This is to transform the features of an object into a vector form;

[0023] The dot product of two vectors is calculated by multiplying the corresponding dimension components and then summing them, reflecting the degree of overlap between the vectors in the direction.

[0024] and These are the magnitudes of the text embedding vectors of knowledge points and the feature vectors of implicit knowledge points, respectively. and Used to normalize vectors and eliminate the influence of length differences;

[0025] Used to measure two vectors and The degree of correlation between vectors is measured by the cosine of the angle between the vectors.

[0026] Optionally, in step S2, the multi-dimensional map pattern layer is a sequence of teaching activities from shallow to deep, which is based on "cognitive ladder" and "activity effect feedback", and automatically selects the optimal combination sequence:

[0027] ;

[0028] in, The goal is to find the sequence that maximizes the total reward or total value of the depth-optimal sequence. , For all teaching activity sequences, i.e. Therefore, use express, The independent variable that maximizes... To select the sequence with the largest sum of cognitive ladder sum plus effect feedback weighted sum among all sequences;

[0029] For the first Teaching activities The corresponding cognitive ladder level; From arrive Feedback values ​​for the results; The weighting coefficient is used to provide feedback on the effectiveness. For the value of a single step action, the first step in the sequence One action; It is essentially a measure of the value of a single action, summation. It is to put all of the sequence The sum of the individual values ​​of each action yields the total value of a single step.

[0030] The value of the transfer between actions, It measures the action arrive The value of the transfer;

[0031] For the sequence has Each action has For each pair of adjacent actions, sum the values ​​of all adjacent transitions.

[0032] The overall goal of automatically selecting the optimal combination sequence is to find a sequence of actions. This is such that the sum of the individual values ​​of all single-step actions and the sum of the values ​​of all adjacent action transitions, multiplied by a weight. The sequence that yields the largest sum is the optimal sequence, ensuring that each action is good enough and that the connections between actions are reasonable enough.

[0033] Optionally, in step S3, for the preprocessed teaching resources, core teaching entities and specific relationship types are collaboratively extracted to construct a knowledge graph. This process of constructing the knowledge graph is a process of transforming unordered resources into an ordered structure, thereby quantifying the dynamic collaborative interaction intensity of collaborative extraction. By quantifying the dynamic synergistic effects of different extraction stages through interactive responses, the process value of collaboration is demonstrated.

[0034] ;

[0035] in, for The dynamic collaborative interaction intensity at time t represents the overall intensity of the collaborative effect of multiple entities at time t.

[0036] as the main body and Synergistic correlation factors or interaction triggering coefficients; if and exist There is always a collaborative relationship. Otherwise, it is 0; as the main body exist The state changes at any given time are fluctuations in power, activity, contribution, etc. over time;

[0037] To avoid the denominator being 0, a minimum quantity is introduced. This ensures computability;

[0038] For time decay term, It is the attenuation rate; It is the main body and The absolute value of the time difference of the interaction reflects the rule that the longer the interaction has been, the weaker its contribution to the current level of collaboration.

[0039] For double summation, it means summing over all subjects. , and Perform the traversal and calculation without repeating any steps, and then sum the results. For the sequence has One action.

[0040] Optionally, the ordered structure evolves as follows: knowledge graphs evolve dynamically. Traditional orderliness only evaluates static results, ignoring the long-term evolution of knowledge graphs through collaborative extraction. By quantifying the long-term value of collaborative extraction in continuously optimizing the ordered structure of knowledge graphs, dynamic orderliness is reflected.

[0041] ;

[0042] in, This represents the change in free energy during the evolution of an ordered structure. Let time start from the initial moment until the end time Summation is performed to reflect the cumulative effect of the entire process;

[0043] For a moment The degree of orderliness under synergy; synergy refers to the behavior of internal elements cooperating with each other to promote order.

[0044] For a moment The degree of order under independent action, that is, the degree of order when elements have no cooperation and each behaves randomly, is a reference benchmark for disorder or weak order.

[0045] It is a time-weighted factor or a discount factor. It is a rate constant or decay coefficient that controls the rate of decay of the weighted effect in the long term.

[0046] Optionally, the extraction and validation of "prerequisite relationships" and "increasing teaching depth relationships" between entities can be achieved through "structured data anchoring + semantic analysis completion":

[0047] Step a: Initially anchor relationships from the structured data of the course syllabus;

[0048] Step b: Verify the validity of the relationship by combining semantic analysis with the textbook content;

[0049] Step c: Cross-validation and relational correction.

[0050] Optionally, in step S4, the path of "perceiving user status → generating personalized solutions → providing resource support → absorbing feedback iteration" is followed.

[0051] Optional, the specific steps are as follows:

[0052] Step 1: Defining the underlying database roles of the knowledge graph;

[0053] Step II: Generation of personalized optimal learning path based on user's knowledge state;

[0054] Step 3: Recommending learning path nodes in relation to suitable teaching resources.

[0055] A knowledge graph-based teaching resource database construction system includes:

[0056] The module consists of a multi-source heterogeneous teaching resource processing module, a multi-dimensional teaching knowledge graph pattern layer definition module, a graph entity and relation collaborative processing module, and an adaptive graph application module, which are connected in sequence.

[0057] The multi-source heterogeneous teaching resource processing module is configured to obtain original teaching resources from at least two different types of sources. The original teaching resources include structured course outlines, unstructured electronic textbooks, and user behavior logs from online teaching platforms. The module also preprocesses the original teaching resources. The preprocessing operations include performing natural language processing on unstructured text to identify the association between entities and key concepts, and parsing user behavior logs to extract implicit knowledge point associations.

[0058] The multidimensional teaching knowledge graph pattern layer definition module is configured to construct a multidimensional graph pattern layer. The graph pattern layer defines both the core teaching entity types and their hierarchical relationships, as well as the specific relationship types used to describe the teaching logic. The specific relationship types include at least "prerequisite relationship", "increasing teaching depth relationship" and "common co-occurrence misunderstanding relationship".

[0059] The knowledge graph entity and relation collaborative processing module is configured to extract core teaching entities and specific relation types from preprocessed teaching resources based on the knowledge graph pattern layer to construct a knowledge graph. Among them, the "prerequisite relation" and "increasing teaching depth relation" between entities are extracted and verified by combining the analysis of the structured data of the course syllabus with the semantic analysis results of the textbook content. The "common co-occurrence misunderstanding relation" is automatically discovered and generated by analyzing the answer sequences and error patterns in massive user behavior logs through machine learning models.

[0060] The adaptive knowledge graph application module is configured to use the constructed knowledge graph as the underlying database, respond to users' knowledge status queries, and dynamically generate a personalized optimal learning path that spans multiple knowledge entities by using graph algorithms to traverse the knowledge graph in real time based on the user's knowledge mastery. At the same time, it recommends the most suitable teaching resources to the user by associating each node in the learning path with the most suitable teaching resources, and uses the user's behavior feedback on the new learning path as incremental data.

[0061] The beneficial effects of this invention are:

[0062] This invention is based on a graph pattern layer, which collaboratively extracts core teaching entities and specific relationship types from preprocessed teaching resources to construct a knowledge graph. Specifically, the "prerequisite relationship" and "increasing teaching depth relationship" between entities are extracted and verified by combining the analysis of structured data of the course syllabus with the semantic analysis results of the textbook content. Meanwhile, the "common co-occurrence misunderstanding relationship" is automatically discovered and generated by analyzing the answer sequences and error patterns in massive user behavior logs through machine learning models, thus realizing the effective integration of massive, heterogeneous teaching resources with the knowledge graph. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0065] Figure 2 This is a flowchart of the process of the present invention. Detailed Implementation

[0066] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0067] Example 1;

[0068] like Figure 1 As shown, this embodiment provides a knowledge graph-based teaching resource database construction system, including:

[0069] The module consists of a multi-source heterogeneous teaching resource processing module, a multi-dimensional teaching knowledge graph pattern layer definition module, a graph entity and relation collaborative processing module, and an adaptive graph application module, which are connected in sequence.

[0070] The multi-source heterogeneous teaching resource processing module is configured to obtain original teaching resources from at least two different types of sources. The original teaching resources include structured course outlines, unstructured electronic textbooks, and user behavior logs from online teaching platforms. The module also preprocesses the original teaching resources. The preprocessing operations include performing natural language processing on unstructured text to identify the association between entities and key concepts, and parsing user behavior logs to extract implicit knowledge point associations.

[0071] The multidimensional teaching knowledge graph pattern layer definition module is configured to construct a multidimensional graph pattern layer. The graph pattern layer defines both the core teaching entity types and their hierarchical relationships, as well as the specific relationship types used to describe the teaching logic. The specific relationship types include at least "prerequisite relationship", "increasing teaching depth relationship" and "common co-occurrence misunderstanding relationship".

[0072] The knowledge graph entity and relation collaborative processing module is configured to extract core teaching entities and specific relation types from preprocessed teaching resources based on the knowledge graph pattern layer to construct a knowledge graph. Among them, the "prerequisite relation" and "increasing teaching depth relation" between entities are extracted and verified by combining the analysis of the structured data of the course syllabus with the semantic analysis results of the textbook content. The "common co-occurrence misunderstanding relation" is automatically discovered and generated by analyzing the answer sequences and error patterns in massive user behavior logs through machine learning models.

[0073] The adaptive knowledge graph application module is configured to use the constructed knowledge graph as the underlying database, respond to users' knowledge status queries, and dynamically generate a personalized optimal learning path that spans multiple knowledge entities by using graph algorithms to traverse the knowledge graph in real time based on the user's knowledge mastery. At the same time, it recommends the most suitable teaching resources to the user by associating each node in the learning path with the most suitable teaching resources, and uses the user's behavior feedback on the new learning path as incremental data.

[0074] Example 2;

[0075] like Figure 2 As shown in the figure, this embodiment provides a method for constructing a knowledge graph-based teaching resource database, including the following steps:

[0076] Step S1: Acquisition and Preprocessing of Multi-Source Heterogeneous Teaching Resources: Acquire original teaching resources from at least two different types of sources, including structured course outlines, unstructured e-textbooks, and user behavior logs from online teaching platforms; preprocess the original teaching resources, including natural language processing of unstructured text to identify the association between entities and key concepts, and parsing user behavior logs to extract implicit knowledge point associations;

[0077] Step S2: Definition of Multidimensional Teaching Knowledge Graph Pattern Layer: Construct a multidimensional graph pattern layer. The graph pattern layer not only defines the core teaching entity types and their hierarchical relationships, but also defines specific relationship types used to describe the teaching logic. The specific relationship types include at least "prerequisite relationship", "increasing teaching depth relationship" and "common co-occurrence misunderstanding relationship".

[0078] Step S3: Collaborative Extraction and Fusion of Knowledge Entities and Relationships: Based on the knowledge graph schema layer, core teaching entities and specific relationship types are collaboratively extracted from preprocessed teaching resources to construct a knowledge graph. Specifically, "prerequisite relationships" and "increasing teaching depth relationships" between entities are extracted and verified by combining the analysis of structured data from the course syllabus with the semantic analysis results of the textbook content. Meanwhile, "common co-occurrence and misunderstanding relationships" are automatically discovered and generated by analyzing answer sequences and error patterns in massive user behavior logs using machine learning models.

[0079] Step S4: Adaptive Knowledge Graph Update and Teaching Path Generation: The completed knowledge graph is used as the underlying database to respond to users' knowledge status queries; based on the user's knowledge mastery, the knowledge graph is traversed in real time using graph algorithms to dynamically generate a personalized optimal learning path that spans multiple knowledge entities, and the most suitable teaching resources are recommended to the user for each node in the learning path, and the user's behavioral feedback on the new learning path is used as incremental data.

[0080] Example 3;

[0081] Based on Example 2, in step S1, natural language processing is performed on unstructured text to identify the association between entities and key concepts. The logical relationships between entities / concepts in the teaching text need to be extracted through a relation classification model to calculate the text's relation. The specific formula is as follows:

[0082] Let the text segment to be analyzed be... (e.g., calculating the side length of a right triangle), the set of relation categories is... Calculate text using a pre-trained model (i.e., the BERT model). Belongs to relational features correlation :

[0083] ;

[0084] Wherein, it indicates that the given text / entity belongs to At that time, it belongs to the relational feature. The probability (e.g., text belonging to a certain category in text classification, or the relationship between two entities in a knowledge graph). The probability of ( ). It is a relational feature Belongs to the set of relation categories (Relationship category set) For all possible categories / relationship types, such as: belonging to the same category);

[0085] For text The BERT embedding vector is the text. The vector representation of the text is encoded by the BERT pre-trained model and is the model's vectorized capture of the text semantics. Relational features The corresponding weight vector (the parameters the model learns for each relation) is used to measure relation features. With text The degree of matching; Relational features The bias term (adjusting the probability offset, which is also a parameter learned by the model); It is an exponential function used to convert the "match degree" into a non-negative value and amplify the numerical difference, making high match degree more prominent;

[0086] For each relation feature With the given text / entity The correlation is calculated, specifically, The vector dot product measures the characteristics of the relationship. Weight vector and text The similarity of BERT embedding vectors (the larger the dot product, the higher the semantic matching), plus the bias. This is to allow the model to adjust relational features more flexibly. Exponential function It is monotonically increasing; the larger the exponent, the stronger the relationship characteristics. With text The higher the matching degree;

[0087] The normalization term ensures that the sum of the probabilities of all relational features equals 1, satisfying the fundamental property of probability; specifically, This refers to traversing relation features, which belong to the set of relation categories. Each relationship in; Relational features The corresponding weight vector is used to measure the traversal relationship features. With text The degree of matching; It is a vector representation of text (or entity) (encoded by a BERT pre-trained model, which is a vectorization of text semantics). Features of traversal relations The bias term is used to adjust the probability offset and is also a parameter learned by the model. It is a set of relation categories All traversal relation features The results are summed.

[0088] First, calculate the matching exponentiation of a single relation to ensure the result is non-negative (because the output of the exponential function is always greater than 0), amplifying the difference between high-matching and low-matching relations (the exponential function is monotonically increasing; the larger the exponent, the higher the matching degree). Then, sum the exponentiation results of all relations, thus summing the relation categories. All relational features The exponential results are all added together. The numerator is a certain relationship. The exponential result, the denominator is all relational features The sum of the exponential results, by dividing the numerator by the denominator, allows the statement "the text belongs to each relation" to be transformed into a relational statement. (Satisfies "non-negative" and "sum of 1").

[0089] The entire process from original teaching texts to structured entities / concepts can be summarized by the following formula:

[0090] Original text Word segmentation results Text vector Entity recognition Key concept relationships Semantic relationships cover the entire process of NLP for teaching texts, from preprocessing to feature generation to semantic extraction. This enables the structuring of teaching resources (e.g., knowledge point base construction and intelligent retrieval).

[0091] The implicit knowledge point associations are extracted by calculating the cosine similarity between the implicit knowledge point vector and the knowledge point concept vector to determine the degree of association.

[0092] ;

[0093] in, The similarity value ranges from [−1, 1]. The closer the value is to 1, the more relevant the association; the closer the value is to -1, the less relevant the association.

[0094] For knowledge points and concepts; These are implicit knowledge points; The text embedding vector for knowledge point concepts (obtained by encoding the concept name and definition using the BERT model); The feature vectors of implicit knowledge points; and This is to transform the features of an object into a vector form, such as: word frequency vector of text, attribute vector of an item;

[0095] The dot product (inner product) of two vectors is calculated by multiplying the corresponding dimension components and then summing them, reflecting the degree of overlap between the vectors in the direction.

[0096] and These are the magnitudes of the text embedding vectors of knowledge points and the feature vectors of implicit knowledge points, respectively. and Used to normalize vectors and eliminate the influence of length differences;

[0097] Used to measure two vectors and The degree of correlation (similarity) between vectors is measured by cosine similarity, which uses the cosine of the angle between the vectors.

[0098] If the two vectors have the same direction (the included angle is 0°). ,at this time ,express and High similarity; if the two vectors are in completely opposite directions (with an angle of 180°). ,at this time ,express and The heights are different; if the two vectors are perpendicular (with an angle of 90°). ,at this time ,express and No similarity.

[0099] In fields such as text similarity analysis (e.g., the topic similarity between two articles, calculated through word frequency vectors), recommendation systems (e.g., similarity matching of item attribute vectors), and image feature matching, the core principle is to measure the degree of association through "directional similarity" rather than "numerical magnitude".

[0100] Example 4;

[0101] Based on Example 2, in step S2, the knowledge graph is divided into a schema layer and a data layer:

[0102] The data layer consists of specific "examples of teaching knowledge" (such as "linear equation in one variable", "Newton's first law", "subject judgment of 'ba' clauses"), as well as specific relationships between examples (such as "'linear equation in one variable' is a prerequisite for 'linear equation in two variables'").

[0103] The pattern layer is a higher-dimensional "abstract framework" that does not focus on specific knowledge points, but rather predefines "which types of things can become the core entities in teaching", "how these entities are layered", and "which relationships can be used to describe the teaching logic", which is equivalent to setting a "unified standard" for filling in specific teaching content later.

[0104] The multi-dimensional approach emphasizes that the design of the model layer should cover multiple key dimensions of the teaching process (such as knowledge progression dimension, cognitive difficulty dimension, and error diagnosis dimension), rather than focusing on a single "knowledge point classification".

[0105] The core tasks of the graph pattern layer are "defining the core teaching entity types and their hierarchical relationships" and "defining the specific relationship types that describe the teaching logic," which together support the knowledge organization of the teaching scenario.

[0106] The core teaching entity type refers to the most critical and fundamental knowledge carrier category in the teaching scenario, rather than specific knowledge points; the hierarchical relationship is the inclusion / subordination / abstract-concrete relationship between these types, ensuring the structure of the knowledge system;

[0107] The specific relational types in teaching logic are such that entity types form the framework, while specific relational types give the framework the vitality of teaching. These relations must align with the core logic of teaching (e.g., knowledge progression, difficulty escalation, and error association), rather than the generic "association" or "containment." They must include at least three types of relations, as detailed below:

[0108] Prerequisite-relation is used to solve the logic of "what to learn first and what to learn next". It describes the dependency relationship between two knowledge points (or knowledge modules) that "A must be mastered before B can be learned". It is the core basis for designing the teaching sequence (corresponding to the "sequential order of knowledge"). The relationship is unidirectional (A→B, A is a prerequisite for B, and B depends on A). It avoids "teaching ahead of time" (e.g., teaching "linear equations in one variable" directly without teaching "addition and subtraction of algebras", which students cannot understand transposition). It helps to develop personalized learning paths (e.g., if a student cannot learn "linear equations in two variables", we can check whether they have mastered the prerequisite knowledge "linear equations in one variable").

[0109] The Depth-Increment Relation is used to address the logic of "how deep to learn, how many steps to take," describing the progressive relationship of the same knowledge point (or the same topic) in terms of "cognitive difficulty, application scope, and assessment requirements," from shallow to deep. This corresponds to the gradual improvement of students' cognitive abilities (distinct from the "prerequisite relationship" which involves "different knowledge in sequence," this relationship represents "layering the depth of the same knowledge"). The relationship is unidirectional (shallow layer A → deep layer B), avoiding a "one-size-fits-all" approach to teaching (e.g., teaching "shallow application" to students with weak foundations first, and "deep analysis" to high-achieving students). It also helps in designing differentiated assignments and periodic tests (e.g., unit tests focus on "intermediate understanding," and final tests focus on "deep application").

[0110] Co-Occurrence-Misconception-Relation is used to solve the logic of "which errors often occur together". It describes the "highly co-occurring" relationship between two or more "common misunderstandings". That is, if a student has misunderstanding A, he is also likely to have misunderstanding B. The essence is the "correlation of erroneous cognition" (originating from the inherent connection between knowledge points or the common cognitive biases of students). The relationship is bidirectional or multidirectional (A↔B, A and B co-occur).

[0111] The multi-dimensional map model layer consists of a sequence of teaching activities from shallow to deep, based on "cognitive ladder" and "activity effect feedback," automatically selecting the optimal combination sequence:

[0112] ;

[0113] in, The total reward or total value of the depth-optimal sequence is represented by the sequence that maximizes the total reward. , For all teaching activity sequences, i.e. Therefore, use express, The independent variable that maximizes... To select the sequence with the largest sum of cognitive ladder sum plus effect feedback weighted sum among all sequences;

[0114] For the first Teaching activities The corresponding cognitive ladder level, i.e. Levels are 1-6, where 1 = memory, 2 = comprehension, ..., 6 = creation;

[0115] From arrive Feedback values ​​for the effect The score is calculated from 0 to 10 based on student participation and accuracy. The higher the value, the smoother the transition;

[0116] This is a weighting coefficient for performance feedback, with a value of 0-1, used to balance depth and actual effect; it controls the transfer of value between actions and their importance in the total reward. For example, if more emphasis is placed on the coherence of the sequence, then... Increase the intensity; if more emphasis is placed on the quality of the individual step movements themselves, then... Turn it down.

[0117] For the value of a single step action, the first step in the sequence An action (or behavior, event, depending on the specific scenario, such as: moving, grasping, words or phrases in natural language).

[0118] It measures the value of an individual action itself, such as the action's "benefits," "reasonableness," and "semantic fluency," defined by the specific task. In reinforcement learning, it's the immediate reward; in NLP, it's the language model probability of words, summed together. It is to put all of the sequence The sum of the individual values ​​of each action yields the total value of a single step.

[0119] The value of the transfer between actions, It measures the action arrive The value of transfer, such as: the coherence between actions, logical rationality, and probability transferability (i.e., the smoothness of action connection, and the grammatical coherence between phrases in NLP).

[0120] For the sequence has Each action has For each pair of adjacent actions, sum the values ​​of all adjacent transitions.

[0121] The overall goal of automatically selecting the optimal combination sequence is to find a sequence of actions. This is such that the sum of the individual values ​​of all single-step actions and the sum of the values ​​of all adjacent action transitions, multiplied by a weight. The sequence that yields the largest sum is the optimal sequence, ensuring that each action is good enough and that the connections between actions are reasonable enough.

[0122] Example 5;

[0123] Based on Example 2, in step S3, the core teaching entities and specific relationship types are collaboratively extracted from the preprocessed teaching resources to construct a knowledge graph. This process of constructing the knowledge graph is a transformation from disordered resources to an ordered structure, thereby quantifying the dynamic collaborative interaction intensity of the collaborative extraction. By quantifying the dynamic synergistic effects of different extraction stages (such as entity recognition stage and relation classification stage) through interactive responses, the value of the synergistic process is demonstrated.

[0124] ;

[0125] in, for The dynamic collaborative interaction intensity at time t represents the overall intensity of the collaborative effect of multiple entities at time t.

[0126] as the main body and Synergistic correlation factors or interaction triggering coefficients; if and exist There is always a collaborative relationship (such as cooperation and information transmission). Otherwise, it is 0;

[0127] as the main body exist The state changes at any given time are fluctuations in power, activity, contribution, etc. over time;

[0128] To avoid the denominator being 0, a minimum quantity is introduced. This ensures computability;

[0129] For time decay term, It is the attenuation rate ( >0, the larger the value, the faster the effect of the time difference decays. It is the main body and The absolute value of the time difference between interactions reflects the principle that the longer the interaction takes place, the weaker its contribution to the current level of collaboration.

[0130] For double summation, it means summing over all subjects. ( and Perform the calculations (without repetition) on each iteration, and then sum the results. For the sequence has One action;

[0131] Calculate the dynamic collaborative contribution of each pair of agents, and then obtain the overall collaborative strength by summing them up. ,use Measurement Subject Relative activity (due to addition) ,avoid (Time denominator is undefined); use To determine whether these two entities currently have a collaborative relationship, use The impact of time difference is attenuated, reflecting that near-time interactions have a greater impact on current collaboration than far-time interactions. Summing up the collaborative contributions of all subject pairs yields the result... The total intensity of dynamic collaborative interaction at any given moment;

[0132] It represents the amount of information propagated by a node and is used to measure the dynamic strength of information collaboration and interaction between nodes in the network. It optimizes information flow and quantifies the degree of dynamic collaboration among multiple subjects from three dimensions: "individual state change", "interaction association" and "time decay". It quantifies the overall strength of real-time dynamic collaboration among multiple subjects, taking into account which subjects are collaborating, the contribution of each subject, and whether the interaction time is recent enough. Finally, it obtains the system-level collaboration metric through double summation.

[0133] Furthermore, the evolution of ordered structures means that knowledge graphs are dynamically evolving (e.g., new data supplements triples and corrects errors). Traditional orderliness only evaluates static results, ignoring the long-term evolution of knowledge graphs through collaborative extraction (e.g., highly consistent triples extracted through collaborative extraction can reduce conflicts and accelerate fusion when new data is added). By quantifying the long-term value of collaborative extraction in continuously optimizing the ordered structure of knowledge graphs, dynamic orderliness is demonstrated.

[0134] ;

[0135] in, It represents the change in free energy during the evolution of an ordered structure. Free energy is the core quantity in thermodynamics that measures available energy, and structural evolution is usually accompanied by energy changes.

[0136] Let time start from the initial moment until the end time Summation is performed to reflect the cumulative effect of the entire process; For a moment The degree of orderliness under synergy; synergy refers to the behavior of internal elements cooperating with each other to promote order.

[0137] For a moment The degree of order under independent action, that is, the degree of order when elements have no cooperation and each behaves randomly, is a reference benchmark for disorder or weak order.

[0138] It is a time-weighted factor or a discount factor. It is a rate constant (or decay coefficient) that controls the rate of decay of the weighted effect in the long term; The larger the scale, the more emphasis is placed on the near term, and the faster the long-term contribution diminishes. Indicates from time until the end time The remaining time, the longer the remaining time is The smaller the value, the closer it is to the initial moment. The larger, The smaller the value, the less the contribution of early actions to the final free energy is discounted; conversely, The closer , The smaller, The closer the value is to 1, the more attention is paid to the contribution of late-stage behavior.

[0139] First calculate each time step The difference between cooperative order and independent order. This reflects the difference between collaboration and independence. The degree of orderliness of contributions made at any given moment; then multiplied by the time weight. This reflects the differences in the weight of contributions at different times in the final free energy; that is, the early contributions are weakened by time decay, while the later contributions are more prominent, across the entire time range. Summing these values ​​yields the total change in free energy resulting from the synergistic effect throughout the entire evolutionary process.

[0140] From a dynamic evolutionary perspective, it breaks through the limitations of static evaluation, linking collaborative optimization with the long-term viability of the knowledge graph. For example, high-quality triples extracted collaboratively can reduce conflicts in subsequent data fusion and lower maintenance costs; the time discount term aligns with practical application scenarios (e.g., knowledge graphs focus more on the orderliness of recent data, as it directly affects current decisions; the value range is...). , The larger the value, the stronger the benefit of collaborative extraction to the long-term orderly evolution of the knowledge graph, and the more persistent the optimization effect. (Ordered structure evolution) Evaluation of the long-term value optimization of collaborative extraction.

[0141] The extraction and verification of the "prerequisite relationship" and "increasing teaching depth relationship" between entities are as follows: a certain teaching entity (e.g., the knowledge point "linear equation in one variable") requires another entity (e.g., the knowledge point "operations with rational numbers") as a learning foundation in order to be effectively mastered. This extraction and verification is achieved through "structured data anchoring + semantic analysis completion".

[0142] Step a: Initially anchor relationships from the structured data of the course syllabus;

[0143] The course syllabus, as the core document of the teaching plan, contains structured information such as a clear sequence of teaching modules and prerequisite requirements. By parsing the tag fields of "prerequisite knowledge" and "preparatory modules" in the syllabus, preliminary prerequisite relationship pairs can be directly extracted.

[0144] Step b: Verify the validity of the relationship by combining semantic analysis with the textbook content;

[0145] As the core carrier of teaching, textbooks directly reflect the prerequisite dependencies between entities through their content arrangement and logical connections. Natural language processing technology is used to perform semantic analysis on textbook texts, verifying prerequisite relationships from three aspects: concept definitions, example analyses, and exercise connections.

[0146] Step c: Cross-validation and relational correction;

[0147] The prerequisite relationships extracted from the course syllabus are cross-referenced with the relationships verified by semantic analysis of the textbook. Relationships that match are directly confirmed, while those that differ are verified a second time.

[0148] "Common co-occurrence misconceptions" can easily lead to confusion between two or more teaching entities (such as "sufficient condition" and "necessary condition"), resulting in errors occurring simultaneously. This is addressed through in-depth analysis of massive user behavior logs using machine learning models, with the following specific process:

[0149] User behavior logs contain massive amounts of unstructured and semi-structured data, including learners' answer records (question ID, answer time, answer correctness, and answer duration), knowledge point association records (knowledge points tested by the questions), and error feedback records (incorrect question annotations and error reason descriptions). Data cleaning and feature extraction are required first.

[0150] The automatic discovery of user behavior logs based on answer sequences and error patterns involves the following steps: First, invalid logs (such as records of erroneous operations with answer times less than 3 seconds or duplicate submissions) are removed, while valid answer sequences (user answer IDs and results arranged chronologically) are retained. Second, a "question-knowledge point" mapping matrix is ​​constructed, and the core knowledge points tested for each question are determined through teaching expert annotation and text matching technology (e.g., the question "Determine the condition type of 'If x>2, then x>1'" corresponds to the knowledge points "sufficient condition" and "necessary condition"). Finally, key features are extracted: the total number of incorrect questions is extracted from the answer sequences. The metrics include "Frequency" (the number of times the same user answers two questions corresponding to two knowledge points incorrectly within the same answering time period), "Error Interval Time" (the time difference between the same user answering two questions corresponding to two knowledge points incorrectly; if the interval is less than 10 minutes, it is considered short-term co-occurrence), and "Error Pattern Similarity" (analyzing the error descriptions in the user's incorrect question feedback through natural language processing, such as "confusing the definitions of sufficient and necessary conditions" or "reversing the direction of the judgment of sufficient conditions," and calculating the semantic similarity of the error descriptions). At the same time, user attribute features (such as grade and learning progress) are extracted to provide multi-dimensional feature data for subsequent model training.

[0151] Example 6;

[0152] Based on Example 2, in step S4, following the path of "perceiving user status → generating personalized solutions → providing resource support → absorbing feedback iterations", the specific steps are as follows:

[0153] Step 1: Defining the underlying database roles of the knowledge graph;

[0154] The completed knowledge graph (including core teaching entities, "prerequisite relationships" and "misunderstanding relationships") is the data foundation, and this information is the core basis for subsequent personalized services.

[0155] Meanwhile, the graph supports real-time query capabilities: when the system receives a user's knowledge status query request (such as: a student completing a post-class test, or a teacher querying a student's weak points), it can quickly match the knowledge entities that the user has mastered / not mastered from the graph, providing a user profile for subsequent analysis.

[0156] Step II: Generation of personalized optimal learning path based on user's knowledge state;

[0157] The goal is to design the most suitable learning sequence for users with different knowledge levels, avoiding a one-size-fits-all teaching approach. Key points are as follows:

[0158] Input: User's knowledge mastery status. The system first needs to obtain the user's knowledge status data, which commonly comes from:

[0159] Explicit data: test scores (e.g., 100% accuracy for knowledge point 1, 30% accuracy for knowledge point 2), homework completion status, course progress, etc.

[0160] Implicit data: learning time (e.g., spending 3 times more time on knowledge point 2 than on knowledge point 1), number of times hesitating when answering questions (e.g., repeatedly revising answers when doing questions on knowledge point 2).

[0161] Core Technology: Real-time traversal system using graph algorithms (such as shortest path algorithms and weighted path planning algorithms) combined with relationship rules in knowledge graphs to dynamically generate learning paths. Example:

[0162] Assuming the user has not yet mastered knowledge point 2, the graph clearly states that "the prerequisite for knowledge point 2 is knowledge point 1" and "a common misconception about knowledge point 2 is that it is confused with knowledge point 3".

[0163] The graph algorithm first determines whether the user has mastered knowledge point 1; if not, the starting point of the path is set to knowledge point 1; if the user has mastered it, it directly points to knowledge point 2, and at the same time, it adds a step in the path to distinguish between knowledge point 2 and knowledge point 3 (to avoid misunderstanding).

[0164] The final personalized learning path might be: learn knowledge point 1 → differentiate between knowledge point 2 and knowledge point 3 → learn knowledge point 2 → practice typical examples of knowledge point 2 (rather than directly having users learn knowledge point 2, avoiding learning difficulties due to insufficient prior knowledge or misunderstandings). The "optimal" criterion for judging the path is not "shortest," but rather "most suitable for user efficiency and effectiveness."

[0165] Step 3: Recommending the association between learning path nodes and suitable teaching resources;

[0166] After generating the learning path, the system needs to match the "most suitable teaching resources" for each "node" (i.e., each knowledge point to be learned) in the path to ensure that users can efficiently master the knowledge of that node. The core logic is the "precise binding of resources and knowledge entities":

[0167] Personalized adaptation: It not only matches resources corresponding to knowledge points, but also adjusts them based on user characteristics. For example:

[0168] For visual learners, we recommend animated video explanations for knowledge point 2;

[0169] For users with weak foundations, we recommend introductory-level examples (rather than difficult ones) for knowledge point 2.

[0170] For users who may easily confuse knowledge points 2 and 3, we additionally recommend a comparative analysis document for knowledge points 2 and 3, which uses the "co-occurrence misunderstanding relationship" in the graph for comparison and analysis.

[0171] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope described in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a teaching resource database based on knowledge graphs, characterized in that, Includes the following steps: Step S1: Acquisition and Preprocessing of Multi-Source Heterogeneous Teaching Resources: Acquire original teaching resources from at least two different types of sources, including structured course outlines, unstructured e-textbooks, and user behavior logs from online teaching platforms; preprocess the original teaching resources, including natural language processing of unstructured text to identify the association between entities and key concepts, and parsing user behavior logs to extract implicit knowledge point associations; Step S2: Definition of Multidimensional Teaching Knowledge Graph Pattern Layer: Construct a multidimensional graph pattern layer. The graph pattern layer not only defines the core teaching entity types and their hierarchical relationships, but also defines specific relationship types used to describe the teaching logic. The specific relationship types include at least "prerequisite relationship", "increasing teaching depth relationship" and "common co-occurrence misunderstanding relationship". Step S3: Collaborative Extraction and Fusion of Knowledge Entities and Relationships: Based on the knowledge graph schema layer, core teaching entities and specific relationship types are collaboratively extracted from preprocessed teaching resources to construct a knowledge graph. Specifically, "prerequisite relationships" and "increasing teaching depth relationships" between entities are extracted and verified by combining the analysis of structured data from the course syllabus with the semantic analysis results of the textbook content. Meanwhile, "common co-occurrence and misunderstanding relationships" are automatically discovered and generated by analyzing answer sequences and error patterns in massive user behavior logs using machine learning models. Step S4: Adaptive Knowledge Graph Update and Teaching Path Generation: The completed knowledge graph is used as the underlying database to respond to users' knowledge status queries; based on the user's knowledge mastery, the knowledge graph is traversed in real time using graph algorithms to dynamically generate a personalized optimal learning path that spans multiple knowledge entities, and the most suitable teaching resources are recommended to the user for each node in the learning path, and the user's behavioral feedback on the new learning path is used as incremental data.

2. The method for constructing a knowledge graph-based teaching resource database according to claim 1, characterized in that, In step S1, performing natural language processing on the unstructured text to identify the associations between entities and key concepts involves extracting the logical relationships between entities / concepts in the teaching text using a relation classification model to calculate the text's association with a relation. The specific formula is as follows: Calculate text using a pre-trained model Belongs to relational features correlation : ； Wherein, it indicates that the given text / entity belongs to At that time, it belongs to the relational feature. The probability, It is a relational feature Belongs to the set of relation categories ; For text The BERT embedding vector is the text. The vector representation of the text is encoded by the BERT pre-trained model and is the model's vectorized capture of the text semantics. Relational features The corresponding weight vector is used to measure relation features. With text The degree of matching; Relational features The bias term; It is an exponential function used to convert the matching degree into a non-negative value and amplify the numerical difference, making high matching degrees stand out more. For each relation feature With the given text / entity The correlation is calculated, specifically, The vector dot product measures the characteristics of the relationship. Weight vector and text The similarity of BERT embedding vectors, plus the bias This is to allow the model to adjust relational features more flexibly. Exponential function It is monotonically increasing; the larger the exponent, the stronger the relationship characteristics. With text The higher the matching degree; The normalization term ensures that the sum of the probabilities of all relational features equals 1, satisfying the fundamental property of probability; specifically, This refers to traversing relation features, which belong to the set of relation categories. Each relationship in; Relational features The corresponding weight vector is used to measure the traversal relationship features. With text The degree of matching; A vector representation of the text; Features of traversal relations The bias term is used to adjust the probability offset and is also a parameter learned by the model. It is a set of relation categories All traversal relation features Summing the results; First, calculate the matching index of a single relation to ensure the result is non-negative, amplifying the difference between high-matching and low-matching relations. Then, sum the index results of all relations to form the relation category set. All relational features The exponential results are all added together, and the numerator is a certain relationship. The exponential result, the denominator is all relational features The sum of the exponential results, by dividing the numerator by the denominator, allows "text belonging to various relations" to be transformed into relevance. .

3. The method for constructing a knowledge graph-based teaching resource database according to claim 1, characterized in that, In step S1, the extraction of implicit knowledge point associations involves determining the degree of association by calculating the cosine similarity between the implicit knowledge point vector and the knowledge point concept vector. ； in, The similarity value ranges from [−1, 1]. The closer the value is to 1, the more relevant the association; the closer the value is to -1, the less relevant the association. For knowledge points and concepts; These are implicit knowledge points; Text embedding vectors for knowledge point concepts; The feature vectors of implicit knowledge points; and This is to transform the features of an object into a vector form; The dot product of two vectors is calculated by multiplying the corresponding dimension components and then summing them, reflecting the degree of overlap between the vectors in the direction. and These are the magnitudes of the text embedding vectors of knowledge points and the feature vectors of implicit knowledge points, respectively. and Used to normalize vectors and eliminate the influence of length differences; Used to measure two vectors and The degree of correlation between vectors is measured by the cosine of the angle between the vectors.

4. The method for constructing a knowledge graph-based teaching resource database according to claim 1, characterized in that, In step S2, the multi-dimensional map pattern layer is a sequence of teaching activities from shallow to deep, which is based on "cognitive ladder" and "activity effect feedback", automatically selecting the optimal combination sequence: ； in, The goal is to find the sequence that maximizes the total reward or total value of the depth-optimal sequence. , For all teaching activity sequences, i.e. Therefore, use express, The independent variable that maximizes... To select the sequence with the largest sum of cognitive ladder sum plus effect feedback weighted sum among all sequences; For the first Teaching activities The corresponding cognitive ladder level; From arrive Feedback values ​​for the results; The weighting coefficient is used to provide feedback on the effectiveness. For the value of a single step action, the first step in the sequence One action; It is essentially a measure of the value of a single action, summation. It is to put all of the sequence The sum of the individual values ​​of each action yields the total value of a single step. The value of the transfer between actions, It measures the action arrive The value of the transfer; For the sequence has Each action has For each pair of adjacent actions, sum the values ​​of all adjacent transitions. The overall goal of automatically selecting the optimal combination sequence is to find a sequence of actions. This is such that the sum of the individual values ​​of all single-step actions and the sum of the values ​​of all adjacent action transitions, multiplied by a weight. The sequence that yields the largest sum is the optimal sequence, ensuring that each action is good enough and that the connections between actions are reasonable enough.

5. The method for constructing a knowledge graph-based teaching resource database according to claim 1, characterized in that, In step S3, the core teaching entities and specific relationship types are collaboratively extracted from the preprocessed teaching resources to construct a knowledge graph. This process of constructing the knowledge graph is a transformation from unordered resources to an ordered structure, thereby quantifying the dynamic collaborative interaction intensity of the collaborative extraction. By quantifying the dynamic synergistic effects of different extraction stages through interactive responses, the process value of collaboration is demonstrated. ； in, for The dynamic collaborative interaction intensity at time t represents the overall intensity of the collaborative effect of multiple entities at time t. as the main body and Synergistic correlation factors or interaction triggering coefficients; if and exist There is always a collaborative relationship. Otherwise, it is 0; as the main body exist The state changes at any given time are fluctuations in power, activity, contribution, etc. over time; To avoid the denominator being 0, a minimum quantity is introduced. This ensures computability; For time decay term, It is the attenuation rate; It is the main body and The absolute value of the time difference of the interaction reflects the rule that the longer the interaction has been, the weaker its contribution to the current level of collaboration. For double summation, it means summing over all subjects. , and Perform the traversal and calculation without repeating any steps, and then sum the results. For the sequence has One action.

6. The method for constructing a knowledge graph-based teaching resource database according to claim 5, characterized in that, The evolution of ordered structures is a dynamic process in knowledge graphs. Traditional orderliness assessments only evaluate static results, neglecting the impact of collaborative extraction on the long-term evolution of knowledge graphs. By quantifying the long-term value of collaborative extraction in continuously optimizing the ordered structure of knowledge graphs, dynamic orderliness is demonstrated. ； in, This represents the change in free energy during the evolution of an ordered structure. Let time start from the initial moment until the end time Summation is performed to reflect the cumulative effect of the entire process; For a moment The degree of orderliness under synergy; synergy refers to the behavior of internal elements cooperating with each other to promote order. For a moment The degree of order under independent action, that is, the degree of order when elements have no cooperation and each behaves randomly, is a reference benchmark for disorder or weak order. It is a time-weighted factor or a discount factor. It is a rate constant or decay coefficient that controls the rate of decay of the weighted effect in the long term.

7. The method for constructing a knowledge graph-based teaching resource database according to claim 5, characterized in that, The extraction and verification of the "prerequisite relationship" and "increasing teaching depth relationship" between the entities are achieved through "structured data anchoring + semantic analysis completion". Step a: Initially anchor relationships from the structured data of the course syllabus; Step b: Verify the validity of the relationship by combining semantic analysis with the textbook content; Step c: Cross-validation and relational correction.

8. The method for constructing a knowledge graph-based teaching resource database according to claim 1, characterized in that, In step S4, the path of "perceiving user status → generating personalized solutions → providing resource support → absorbing feedback iteration" is followed.

9. A method for constructing a knowledge graph-based teaching resource database according to claim 8, characterized in that, The specific steps are as follows: Step 1: Defining the underlying database roles of the knowledge graph; Step II: Generation of personalized optimal learning path based on user's knowledge state; Step 3: Recommending learning path nodes in relation to suitable teaching resources.

10. A knowledge graph-based teaching resource database construction system, used to execute the knowledge graph-based teaching resource database construction method according to any one of claims 1-9, characterized in that, include: The module consists of a multi-source heterogeneous teaching resource processing module, a multi-dimensional teaching knowledge graph pattern layer definition module, a graph entity and relation collaborative processing module, and an adaptive graph application module, which are connected in sequence. The multi-source heterogeneous teaching resource processing module is configured to obtain original teaching resources from at least two different types of sources. The original teaching resources include structured course outlines, unstructured electronic textbooks, and user behavior logs from online teaching platforms. The module also preprocesses the original teaching resources. The preprocessing operations include performing natural language processing on the unstructured text to identify the association between entities and key concepts, and parsing the user behavior logs to extract implicit knowledge point associations. The multidimensional teaching knowledge graph pattern layer definition module is configured to construct a multidimensional graph pattern layer. The graph pattern layer defines both the core teaching entity types and their hierarchical relationships, and specific relationship types used to describe teaching logic. The specific relationship types include at least "prerequisite relationships", "increasing teaching depth relationships", and "common co-occurrence misunderstanding relationships". The graph entity and relation collaborative processing module is configured to collaboratively extract core teaching entities and specific relation types from preprocessed teaching resources based on the graph pattern layer to construct a knowledge graph. The "prerequisite relationship" and "increasing teaching depth relationship" between entities are extracted and verified by combining the analysis of structured data of the course syllabus with the semantic analysis results of the textbook content. The "common co-occurrence misunderstanding relationship" is automatically discovered and generated by analyzing the answer sequences and error patterns in massive user behavior logs through machine learning models. The adaptive knowledge graph application module is configured to use the constructed knowledge graph as the underlying database, respond to the user's knowledge status query, and, based on the user's knowledge mastery, use graph algorithms to traverse the knowledge graph in real time to dynamically generate a personalized optimal learning path that spans multiple knowledge entities. At the same time, it recommends the most suitable teaching resources associated with each node in the learning path to the user, and uses the user's behavior feedback on the new learning path as incremental data.

Images