Word knowledge point multi-modal resource generation method based on knowledge graph

Abstract

The invention discloses a word knowledge point multi-modal resource generation method based on a knowledge graph, and relates to the technical field of multi-modal resource generation, and the method comprises the steps: constructing a local knowledge point set for knowledge points contained in each word resource, building a relation matrix, calculating a shared part matrix and a direction part matrix, and forming relation complex representation; calculating a total matching value of knowledge points in combination with complex representation, extracting a total direction matching value of different target head relationships through a knowledge chain, defining a direction closing error, screening out a main teaching closed path, and combining into a unified knowledge skeleton based on the screened main teaching closed path; and performing classification mapping based on the multi-modal knowledge point resources according to a unified knowledge skeleton. According to the method, a local knowledge point set, relation matrix decomposition, knowledge point complex number representation and relation complex number representation are continuously linked, so that sharing coverage and direction offset do not stay in static co-occurrence and one-way constraint layers respectively, but enter total matching value calculation together.

Description

Technical Field

[0001] This invention relates to the field of multimodal resource generation technology, and in particular to a method for generating multimodal resources of word knowledge points based on knowledge graphs. Background Technology

[0002] With the development of intelligent technology, the generation of teaching resources around words has gradually evolved from the early dictionary-style display to a comprehensive processing method that combines knowledge graphs, recommendation algorithms and multimodal content generation. Some solutions have also begun to combine large models to directly generate multimodal learning materials based on input words.

[0003] However, the knowledge points in vocabulary teaching involve connections and correspondences in different directions. This is often the content that needs to be clarified and organized. However, existing technologies often extract single knowledge points or generate resources by sorting them by single similarity. Many existing solutions can find relevant knowledge points, but they may not know which knowledge points should be taught first, which are just supplementary explanations, and which are key points that must be reminded to students to avoid mistakes. Therefore, it is easy to have situations where the relationships seem to be connected on the surface, but the actual teaching effect is inconsistent, resulting in the drift of the teaching focus of knowledge points and difficulty in forming a stable consistency. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for generating multimodal resources of word knowledge points based on knowledge graphs, which solves the problem that existing technologies often have seemingly connected relationships but inconsistent actual teaching effects, leading to a drift in the teaching focus of knowledge points and difficulty in forming stable consistency.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for generating multimodal resources of word knowledge points based on knowledge graphs, comprising,

[0008] For each word resource, a local knowledge point set is constructed, a relation matrix is ​​established, the symmetric part of the relation matrix is ​​used as the shared part matrix, and the antisymmetric part of the relation matrix is ​​used as the directional part matrix to form a complex representation of each knowledge point. The shared coverage of each knowledge point on the relation is calculated based on the shared part matrix, and the directional bias of each knowledge point on the relation is calculated based on the directional part matrix to form a complex representation of the relation. The total matching value of the knowledge point is calculated by combining the complex representation.

[0009] The set of relation segments is filtered based on the total matching value. Relationships are defined and knowledge points associated with each other are combined into knowledge nodes. Knowledge chains are formed based on the set of relation segments. The teaching objectives are clarified and the target head relationship is determined based on the teaching relationship types corresponding to the starting point, ending point, and intermediate knowledge points of the knowledge chain. The combination consistency of the knowledge chain is calculated based on the difference between the combination matrix constructed by the knowledge chain and the relationship matrix corresponding to the target head relationship. The total directional matching value of different target head relationships is extracted through the knowledge chain. The absolute value of the difference between the total directional matching value and the total directional matching value of the target head relationship is used as the directional closure difference to filter out the main teaching closed path.

[0010] Based on the selected main teaching closed path, according to the knowledge point type to which the knowledge points in the main teaching closed path belong, four types of structured slots are constructed: core meaning, boundary meaning, establishment conditions, and extraction risk, which are combined to form a unified knowledge skeleton.

[0011] Multimodal knowledge resources are classified and mapped according to a unified knowledge skeleton to form a coverage table. Based on the coverage table, gaps are filled and deviating content is deleted.

[0012] As a preferred embodiment of the knowledge graph-based multimodal resource generation method for word knowledge points described in this invention, the step of using the symmetric part of the relation matrix as the shared part matrix and the antisymmetric part of the relation matrix as the directional part matrix includes: distinguishing and defining the knowledge points contained in each word resource, constructing a local knowledge point set, determining the association based on whether there is a learning association relationship between word knowledge points, and establishing a relation matrix.

[0013] For each relation in the relation matrix, define the shared part and the direction part, and construct the shared part matrix and the direction part matrix respectively.

[0014] As a preferred embodiment of the knowledge graph-based word knowledge point multimodal resource generation method of the present invention, wherein: the formation of the complex representation of each knowledge point includes, based on the shared part matrix, for any knowledge point, calculating the total sharing amount and constructing a shared semantic representation according to the shared proximity value accumulated between different knowledge points on all shared parts;

[0015] Based on the direction part matrix, for any knowledge point, the total direction is calculated and constructed as a direction constraint representation according to the net direction difference accumulated between different knowledge points in all directions.

[0016] Complex representations of each knowledge point are formed based on shared semantic representations and directional constraint representations.

[0017] As a preferred embodiment of the knowledge graph-based word knowledge point multimodal resource generation method of the present invention, wherein: the calculation of the total matching value of knowledge points by combining complex representation includes: calculating the shared coverage on different knowledge points based on the shared part matrix, and calculating the directional bias on different knowledge points based on the directional part matrix, which together form the complex representation of the relationship;

[0018] Based on complex representation, the complex representations of the knowledge points at the starting point and the knowledge points at the ending point of the relationship are extracted, and the shared matching value is calculated. At the same time, the directional matching value is calculated based on the complex representation of the relationship, and the sum is calculated as the total matching value of the knowledge points.

[0019] As a preferred embodiment of the knowledge graph-based multimodal resource generation method for word knowledge points described in this invention, the step of determining the target head relationship based on the teaching relationship type corresponding to the starting point, ending point and intermediate knowledge points of the knowledge chain includes: making a reliability judgment on the association relationship between knowledge points based on the total matching value; if the total matching value of the corresponding association relationship is less than 0, it is filtered out, and the remaining relationships are retained as a set of relationship segments.

[0020] Define two knowledge points related to the association relationship as the association start point and the association end point, and combine the association start point, the association end point, and the association relationship to form a knowledge node. Enumerate the knowledge nodes based on the set of relationship segments and form a knowledge chain.

[0021] Based on the structured vocabulary of the starting knowledge point, the ending knowledge point, and the structured vocabulary of the knowledge points passed through in any knowledge chain, the corresponding target head relationship is determined. Based on the target head relationship, the relationship matrix is ​​searched, the association relationships that satisfy the corresponding target head relationship are selected, and the corresponding knowledge nodes are extracted to construct the corresponding relationship matrix of the target head relationship.

[0022] As a preferred embodiment of the knowledge graph-based multimodal resource generation method for word knowledge points described in this invention, the step of calculating the combination consistency of knowledge chains based on the difference between the combination matrix constructed from the knowledge chain and the relation matrix corresponding to the target head relation includes merging the shared part matrix and the directional part matrix into a relation action matrix, and performing continuous matrix multiplication operations on the relation action matrix according to the order of relations in the knowledge chain to obtain the combination matrix, and calculating the combination consistency of each knowledge chain in combination with the corresponding relation matrix.

[0023] As a preferred embodiment of the multimodal resource generation method for word knowledge points based on knowledge graphs described in this invention, the step of using the absolute value of the difference between the total directional matching value and the total directional matching value of the target head relationship as the directional closure difference to screen out the main teaching closed path includes: extracting the total directional matching value of different associations based on the knowledge chain; and calculating the total directional matching value of different target head relationships in the corresponding relationship matrix based on the corresponding relationship matrix, and defining the directional closure difference.

[0024] The main teaching closed path is screened out based on the knowledge chain set, the consistency of each chain combination, and the directional closure difference.

[0025] As a preferred embodiment of the knowledge graph-based multimodal resource generation method for word knowledge points described in this invention, the step of constructing four types of structured slots—core meaning, boundary meaning, establishment conditions, and extraction risk—based on the knowledge point type to which the knowledge points in the main teaching closed path belong, and combining them to form a unified knowledge skeleton, includes: based on the selected main teaching closed path, querying whether the starting knowledge point belongs to the meaning table, and defining it as the core meaning slot content;

[0026] Based on the main teaching closed path of the filter, query whether there are knowledge points belonging to the discrimination table, and define them as boundary slot content;

[0027] Based on the main teaching closed path of the filter, query whether there are knowledge points belonging to the collocation table and the context table, and define them as the content of the condition slot.

[0028] Based on the main teaching closed path of the filter, query whether there are any knowledge points belonging to the error table, and define it as extracting risk slot content;

[0029] The core semantic slots, boundary semantic slots, establishment condition slots, and extracted risk slots are combined into a unified knowledge skeleton.

[0030] As a preferred embodiment of the knowledge graph-based multimodal resource generation method for word knowledge points described in this invention, the multimodal knowledge point resources are classified and mapped according to a unified knowledge skeleton to form a coverage table, including: listing the smallest set of knowledge units that must be reflected by all knowledge point resources based on the slots of the unified knowledge skeleton; and taking all the knowledge point sets of the four operations as the first knowledge unit set, the second knowledge unit set, the third knowledge unit set, and the fourth knowledge unit set, respectively.

[0031] Meanwhile, the initial drafts of multimodal resources are categorized, including text drafts, image drafts, pronunciation prompt drafts, and exercise drafts.

[0032] Map the initial draft of the resources to the four slots of the unified knowledge framework to form a coverage table.

[0033] As a preferred embodiment of the knowledge graph-based multimodal resource generation method for word knowledge points described in this invention, the step of checking for omissions and filling gaps and deleting deviating content according to the coverage table includes: querying the corresponding coverage value of the knowledge point resource in the coverage table, determining the missing resource, filling the unrepresented slots, and deleting content that deviates from the unified knowledge skeleton, thereby obtaining text resources, image resources, pronunciation prompt resources, and practice question resources organized around the unified knowledge skeleton.

[0034] The beneficial effects of this invention are as follows: By continuously linking the local knowledge point set, relation matrix decomposition, complex representation of knowledge points, and complex representation of relations, shared coverage and directional bias no longer remain at the static co-occurrence and unidirectional constraint levels, but instead jointly enter the calculation of the total matching value. This enables the simultaneous identification of semantically similar but different teaching effects, as well as knowledge point relationships with significant surface differences but stable teaching traction, within the same computational closed loop. This significantly improves the distinguishability and sortability of local knowledge organization. Through progressive screening based on combination consistency, directional closure difference, and chain length, the screened main teaching closed path possesses comprehensive characteristics of structural integrity, target fit, and compact transmission. This provides a single path basis for subsequent unified knowledge skeleton extraction, producing a teaching closed loop determination effect that cannot be directly obtained through local relation screening. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0036] Figure 1 This is a flowchart of the multimodal resource generation method for word knowledge points based on knowledge graphs in Example 1.

[0037] Figure 2 This is a flowchart of the main teaching closed path selection process for the knowledge graph-based multimodal resource generation method for word knowledge points in Example 1.

[0038] Figure 3 This is a flowchart of the relationship filtering process for the knowledge graph-based multimodal resource generation method for word knowledge points in Example 1. Detailed Implementation

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0042] Example 1, referring to Figure 1 , Figure 2 and Figure 3 This is the first embodiment of the present invention, which provides a method for generating multimodal resources of word knowledge points based on knowledge graphs, including the following steps:

[0043] S1. For each word resource, construct a local knowledge point set and establish a relation matrix. Use the symmetric part of the relation matrix as the shared part matrix and the antisymmetric part of the relation matrix as the directional part matrix to form a complex representation of each knowledge point. Calculate the shared coverage of each knowledge point on the relation based on the shared part matrix and the directional bias of each knowledge point on the relation based on the directional part matrix to form a complex representation of the relation. Combine the complex representations to calculate the total matching value of the knowledge points.

[0044] S1.1, Define and differentiate the knowledge points contained in each word resource, construct a local knowledge point set (the total number of sets is equal to the number of local knowledge points of the target word), determine the association based on whether there is a learning relationship between the word knowledge points, and establish a relationship matrix;

[0045] For each relation in the relation matrix, define the shared part and the direction part, and construct the shared part matrix and the direction part matrix respectively, as follows:

[0046]

[0047]

[0048] in, Represents the shared portion of the matrix. The matrix represents the direction part. The relation matrix representing the association r. Indicates transpose calculation;

[0049] Specifically, for vocabulary knowledge points, there can be six types of structured vocabulary materials, including a definition table, a collocation table, a context table, a pronunciation table, a distinction table, and an error table;

[0050] Each word entry in the definition list includes the definition number, definition phrase, and part of speech for the word;

[0051] Each word entry in the collocation table includes the word's meaning number, collocation phrase, and collocation type;

[0052] Each word record in the context table includes a definition number for the word and a contextual label.

[0053] Each word entry in the pronunciation table includes the phonetic symbols for the word, the stress position, and the markers for connected speech / weak forms;

[0054] Each word entry in the discrimination table includes easily confused words, confusion types, and explanations of differences;

[0055] Each word record in the error table includes an incorrect alternative for the word and a label indicating the reason for the error.

[0056] S1.2, based on the shared part matrix, for any knowledge point, the total sharing amount is calculated and constructed as a shared semantic representation based on the accumulated shared proximity value among different knowledge points across all shared parts, as shown below:

[0057]

[0058]

[0059]

[0060]

[0061] in, Let represent the real part vector of the i-th knowledge point. R represents the total number of knowledge points, and R represents the total number of relationships. This represents the accumulated shared proximity value between knowledge points i and j across the entire shared portion. Indicates the total value close to the target value. Indicates shared semantic representation;

[0062] Based on the direction component matrix, for any knowledge point, the total direction is calculated and constructed as a direction constraint representation according to the net direction difference accumulated between different knowledge points across all directions, as shown below:

[0063]

[0064]

[0065]

[0066]

[0067] in, This represents the imaginary part vector of the i-th knowledge point. This represents the net directional difference accumulated over the entire shared portion between knowledge points i and j. Indicates the total amount in the direction. This represents the orientation constraint.

[0068] Based on the shared semantic representation and the directional constraint representation, the complex representation of each knowledge point is formed, which can be expressed as:

[0069]

[0070] in, This represents the complex number representation, and ia represents the imaginary unit;

[0071] Furthermore, regarding the determination of associations, this can be specifically manifested as follows: if there is a clear association between two knowledge points, it is recorded, and an association is established based on each type of association. Relationship matrix;

[0072] For example, if there is a "limiting" relationship between a certain meaning and a certain collocation, then it is confirmed that there is an association relationship, and the association relationship is "limiting".

[0073] If there is an "avoidance" relationship between a certain object of analysis and a certain risk of misuse, then it is confirmed that there is a relationship, and the relationship is "avoidance";

[0074] If a pronunciation difference and a confusing object have a "triggering reminder" relationship, then the relationship is confirmed to exist, and the relationship is "triggering reminder".

[0075] Specifically, the relation matrix can be represented as:

[0076]

[0077] in, Represents the relation matrix. and These represent the i-th and j-th knowledge points, respectively. Indicates a relationship. This indicates that there is a relationship r between the i-th and j-th knowledge points;

[0078] By combining the shared component and the directional component, the relationship retains both mutual proximity and constraint orientation, thus enabling the final matching result to have both aggregation and error-correction capabilities, achieving a teaching closure recognition effect that cannot be obtained by single similarity calculation.

[0079] S1.3, based on the shared part matrix, calculate the shared coverage on different knowledge points, and at the same time, based on the directional part matrix, calculate the directional bias on different knowledge points, which together form the complex number representation of the relation;

[0080] Based on complex representation, the complex representations of the knowledge points at the starting point and the knowledge points at the ending point of the relationship are extracted, and the shared matching value is calculated. At the same time, the directional matching value is calculated based on the complex representation of the relationship, and the sum is calculated as the total matching value of the knowledge points.

[0081] Specifically, for the construction of the complex representation of relations, the calculation of shared coverage and directional bias is performed separately. The calculation of shared coverage is expressed as follows:

[0082]

[0083]

[0084]

[0085] in, This represents the degree of shared coverage of knowledge point j with different knowledge points on relation r. Indicates the total coverage of shared areas. Indicates the real part of the relation;

[0086] The calculation for the directional offset is expressed as:

[0087]

[0088]

[0089]

[0090] in, This indicates the directional bias of knowledge point j relative to different knowledge points on relation r. Indicates the total amount in the direction. Indicates the imaginary part of the relation;

[0091] For the representation of complex relational numbers, calculated jointly by the real and imaginary parts of the relation, it is expressed as:

[0092]

[0093] in, Representing the complex representation of relations;

[0094] Furthermore, the calculation of the total matching value is expressed as follows:

[0095]

[0096]

[0097]

[0098] in, Indicates shared matching values. Indicates the direction matching value. and Let these represent the complex representations of the knowledge points at the starting point and the ending point of the relation, respectively. Indicates the total matching value;

[0099] By continuously linking local knowledge point sets, relation matrix decomposition, complex representation of knowledge points, and complex representation of relations, shared coverage and directional bias no longer remain at the levels of static co-occurrence and unidirectional constraints, but instead jointly enter the calculation of the total matching value. This enables the simultaneous identification of semantically similar but differently pedagogically useful knowledge point relationships, as well as those with significant surface differences but stable pedagogical traction, within the same computational loop. This significantly improves the discriminability and orderability of local knowledge organization. By uniformly mapping six types of heterogeneous materials—meaning, collocation, context, pronunciation, differentiation, and error—to the same complex relation space, the total matching value obtains a consistent basis for cross-material comparison. This allows subsequent rule selection and skeleton extraction to be based on a uniformly transferable relation strength, avoiding bias caused by relying solely on a single type of material.

[0100] S2, filter the set of relation segments based on the total matching value, define the association relationship and combine the associated knowledge points into knowledge nodes, and form a knowledge chain based on the set of relation segments. Clarify the teaching objectives and determine the target head relationship based on the teaching relationship type corresponding to the starting point, ending point and intermediate knowledge points of the knowledge chain. Calculate the combination consistency of the knowledge chain based on the difference between the combination matrix constructed by the knowledge chain and the relationship matrix corresponding to the target head relationship. Extract the total directional matching value of different target head relationships through the knowledge chain. Use the absolute value of the difference between the total directional matching value and the total directional matching value of the target head relationship as the directional closure difference to filter out the main teaching closed path.

[0101] S2.1, Based on the total matching value, the reliability of the relationship between knowledge points is judged. If the total matching value of the corresponding relationship is less than 0, it is filtered out and the remaining relationship is retained as a set of relationship segments.

[0102] Define two related knowledge points as the association start point and the association end point, and the association start point, association end point and association relationship are combined to form a knowledge node. Based on the set of relationship segments, enumerate the knowledge nodes and form a knowledge chain. The enumeration conditions include that the association end point of the first association relationship is equal to the association start point of the second association relationship, and the association end point of the second association relationship must be able to be closed to form a teaching objective.

[0103] Based on the structured vocabulary of the starting knowledge points, the ending knowledge points, and the structured vocabulary of the knowledge points passed through in any knowledge chain, the corresponding target head relationship is determined. Based on the target head relationship, the relationship matrix is ​​searched, the association relationships that satisfy the corresponding target head relationship are filtered, and the corresponding knowledge nodes are extracted to construct the corresponding relationship matrix of the target head relationship.

[0104] Specifically, the teaching objectives include the identification of textual interpretation objectives, image discrimination objectives, pronunciation reminder objectives, and practice and assessment objectives in the teaching of knowledge points;

[0105] At the same time, the rules for changing the length of the knowledge chain are defined. If the current target word cannot be closed to form a clear teaching objective, the number of knowledge points is increased to three, and so on.

[0106] Furthermore, the target head relationship is set based on the structured vocabulary material of knowledge points contained in the knowledge chain. For example, if the chain is "meaning → collocation restriction → register boundary", then the target head relationship is determined as "textual explanation main line relationship".

[0107] If the chain is "meaning → object of analysis → risk of misuse", then the target head relationship is determined as the "error correction main line relationship";

[0108] If the chain is "pronunciation difference → object of analysis → risk of confusion", then the target head relationship is determined as "pronunciation reminder main line relationship";

[0109] S2.2, merge the shared part matrix and the direction part matrix into a relation action matrix, and perform multiplicative combination according to the knowledge chain to obtain the combination matrix. Combine the corresponding relation matrix to calculate the combination consistency degree for each knowledge chain.

[0110] Specifically, the multiplicative combination is determined by the actual number of knowledge points in the knowledge chain, and can be expressed as:

[0111]

[0112] in, Represents a combination matrix. This represents the actual number of knowledge points in the knowledge chain. Represents the relational interaction matrix;

[0113] Furthermore, combinatorial consistency can be expressed as:

[0114]

[0115] in, Indicates the degree of consistency of the combination. This represents the correspondence matrix, and P represents the knowledge chain.

[0116] It should be noted that the sum of the mean and standard deviation of the combination consistency can be used as the consistency threshold. If the combination consistency is greater than the consistency threshold, it means that the corresponding knowledge chain is superficially connected and has not truly formed the target teaching relationship. If the combination consistency is less than or equal to the consistency threshold, it means that it basically meets the teaching objectives.

[0117] By first eliminating invalid relationships based on the total matching value, then reorganizing the retainable relationships into a closable knowledge chain, and then using the structured vocabulary combination results within the chain to reverse-limit the target head relationship, the knowledge connectivity judgment no longer stops at the validity of local edges, but rises to the overall accessibility judgment oriented towards the teaching destination. This unifies the originally scattered relationship strength, chain transmission, and teaching affiliation into the same screening framework.

[0118] S2.3, extract the total directional matching value of different associations based on the knowledge chain, and calculate the total directional matching value of different target head relationships in the corresponding relationship matrix based on the corresponding relationship matrix, and define the directional closure error;

[0119] Based on the knowledge chain set, the combinatorial consistency of each chain, and the directional closure difference, the main teaching closed path is screened out. The screening rules include first selecting the knowledge chain set with the smallest combinatorial consistency; if there are multiple such sets, then selecting the knowledge chain with the smallest directional closure difference; if there are still multiple such sets, then selecting the one with the shorter knowledge chain length.

[0120] Specifically, the directional closure error can be expressed as:

[0121]

[0122]

[0123]

[0124] in, This represents the overall directional matching value for different association relationships. This represents the t-th association. This represents the overall direction matching value related to the target header. This indicates the target header relationship. Indicates directional closure error;

[0125] It should be noted that the sum of the mean and standard deviation of the directional closure error can be used as the closure error threshold. If the directional closure error is greater than the closure error threshold, it means that although the chain can be connected, it deviates from the target teaching effect in the direction. If the directional closure error is less than or equal to the closure error threshold, it means that the overall direction of the selected chain is consistent with the direction of the target teaching relationship.

[0126] By comparing the consistency between the multiplicative combination result of the knowledge chain and the correspondence matrix of the target head relationship, and then superimposing the directional closure difference for joint judgment, the chain is required not only to be connected, but also to have its overall direction of action converge in the same direction as the teaching objective. This allows for the stable elimination of interference chains that are semantically similar but have a different teaching orientation from multiple surface-reachable paths. Finally, by progressively filtering based on combination consistency, directional closure difference, and chain length, the selected main teaching closed path possesses comprehensive characteristics of structural integrity, target fit, and compact transmission. This provides a single path basis for subsequent unified knowledge skeleton extraction, producing a teaching closed loop determination effect that cannot be directly obtained by local relationship filtering.

[0127] S3. Based on the selected main teaching closed path, according to the knowledge point type to which the knowledge points in the main teaching closed path belong, construct four types of structured slots: core meaning, boundary meaning, establishment conditions, and extraction risk, and combine them to form a unified knowledge skeleton.

[0128] S3.1 Based on the main teaching closed path of the filter, query whether the starting knowledge point belongs to the sense table (if it does not belong to the sense table, backtrack to the sense node to which the knowledge point is attached), and define it as the core sense slot content.

[0129] Based on the main teaching closed path of the filter, query whether there are knowledge points belonging to the discrimination table. (If the knowledge points of the discrimination table do not appear directly in the main teaching closed path, but the knowledge points of the pronunciation table or error table appear, then find the discrimination object most directly connected to it along the knowledge chain) and define it as the content of the boundary semantic slot.

[0130] Based on the main teaching closed path of the filter, query whether there are knowledge points belonging to the collocation table and the context table, and define them as the content of the condition slot.

[0131] Based on the main teaching closed path of the filter, query whether there are any knowledge points belonging to the error table, and define it as extracting risk slot content;

[0132] The core meaning slots, boundary meaning slots, establishment condition slots, and extracted risk slots are combined into a unified knowledge skeleton.

[0133] By reorganizing the meanings, distinctions, conditions, and risk information in the main teaching closed path into a unified knowledge skeleton according to fixed slots, the teaching main line obtained from the previous path selection is stably precipitated into the same semantic reference that can be reused across modalities, thereby transforming the teaching constraints that were originally attached to the link sequence into a skeleton structure that can be directly checked.

[0134] S4. Based on the multimodal knowledge point resources, the knowledge point resources are classified and mapped according to the unified knowledge skeleton to form a coverage table. Based on the coverage table, gaps are filled and deviations are deleted.

[0135] S4.1, based on the slots of the unified knowledge skeleton, lists the smallest set of knowledge units that must be reflected by all knowledge point resources, and takes the set of all knowledge points of the four operations as the first set of knowledge units, the second set of knowledge units, the third set of knowledge units, and the fourth set of knowledge units, respectively.

[0136] Meanwhile, the initial drafts of multimodal resources are categorized, including text drafts, image drafts, pronunciation prompt drafts, and exercise drafts.

[0137] Map the initial draft of the resources to the four slots of the unified knowledge framework and form a coverage table;

[0138] Specifically, the mapping between the initial draft of the resource and the four slots of the unified knowledge framework may include:

[0139] Text resources must contain: core meaning slots and conditional meaning slots;

[0140] Image resources must contain: core slot content and boundary slot content;

[0141] Pronunciation cues must contain: boundary definition slot content and extraction risk slot content;

[0142] Exercises must include: boundary definition slot content and extraction risk slot content;

[0143] It should be further explained that the constructed coverage table includes row 1: text, row 2: image, row 3: pronunciation prompts, and row 4: practice questions. The four columns correspond to the four slots of the unified knowledge skeleton.

[0144] S4.2, query the corresponding coverage value of the knowledge point resources in the coverage table according to the coverage table, determine the missing item resources, fill in the slots that are not reflected, and delete the content that deviates from the unified knowledge skeleton, so as to obtain text resources, image resources, pronunciation prompt resources and exercise resources organized around the unified knowledge skeleton respectively.

[0145] Specifically, taking the teaching of the English word "run" as an example, the system first extracts various knowledge points related to "run" from the knowledge graph and categorizes them. For example, its core meaning of "running" (verb) is marked as the core meaning; the easily confused word "jog" and its differences (speech speed and intensity) are marked as distinguishing points; common correct usage scenarios, such as collocations with "run a marathon" and the context of "sports", are marked as conditions for validity; and common student errors, such as misusing "run" to express "operating a machine" (using "operate"), are marked as extraction risks. Next, the system, based on a pre-set teaching logic, combines these scattered knowledge points—namely, the core meaning, distinguishing points, conditions for validity, and common errors—into a structured teaching framework as a unified knowledge skeleton.

[0146] Further based on the knowledge framework, supporting multimedia teaching materials were generated, including text explanations, illustrations, pronunciation tips, and practice questions. After generating the initial draft, each type of material was checked to ensure it covered all four categories of information required by the framework. For example, the check revealed that the text explanation mentioned the meaning of "running," the ability to use "running a marathon," and pointed out the incorrect usage of "operating a machine," but did not explain the difference between it and "jog." The illustrations showed the action of running and the comparison between "run" and "jog," but did not reflect the specific scenario of "marathon," nor did they point out the error of "operating a machine." The pronunciation tips included a comparison of the pronunciations of "run" and "jog," as well as the pronunciation of the incorrect example sentence "run a machine," but lacked a standard example sentence pronunciation for the core meaning of "running." The practice questions tested the difference between "run" and "jog" and the collocation of "running a marathon," but did not focus on the core meaning of "running" itself, and also omitted the analysis of the error of "operating a machine."

[0147] To address these deficiencies, for the missing section on "difference between 'run' and 'jog'" in the text explanation, we retrieved the explanatory text from the pre-set knowledge base explaining that "run emphasizes fast running, while jog refers to slow jogging for fitness" and inserted it into the text. For the missing "marathon" scene and the "operating machine" error message in the pictures, we retrieved and added a photo of a marathon race and an icon with the word "machine" crossed out in red from the image library. For pronunciation tips, we added a standard example sentence demonstrating the usage of "run" in the context of "running," such as the pronunciation of "He can run very fast." For practice questions, we retrieved or automatically generated new questions from the question bank, such as a question to fill in the blank with the correct form of "run" and a question to correct the error "I run this machine.", and added them to the workbook.

[0148] While supplementing the missing content, it can also review and delete redundant information that is irrelevant to the current knowledge framework. It compares every sentence, every picture, and every question in the material with the four information categories in the framework. If some content cannot be classified into any of the core meaning, distinguishing points, conditions for establishment, or common errors, it will be judged as off-topic. For example, a text introducing "run" as "noun, a period of operation" (such as "a successful run of the play") is correct in itself, but it has nothing to do with the teaching framework of this lesson, which revolves around the meaning of the verb "run". This text can be removed from the material of this lesson.

[0149] Through this series of "check-supplement-clean" operations, the text, images, pronunciation, and practice materials generated for the word "run" will strictly revolve around the core concept of "running" and fully include all the key information on distinguishing it from "jog," its use in "marathon," and avoiding misuse as "operation." This will form a set of multimodal teaching resources that are goal-oriented, complementary in content, and rigorous and consistent.

[0150] By classifying and mapping text, images, pronunciation prompts, and exercises according to a unified framework, and combining this with a coverage table to perform gap filling and deviation deletion, different modalities are no longer organized independently around local highlights, but rather collaboratively converge around the same core meaning, boundary meaning, conditions for establishment, and risk points, forming a resource combination with clear content division and mutual reinforcement. Through the continuous cooperation of framework extraction and coverage correction, the results of the previous rule selection can be stably inherited by the subsequent resource generation, ultimately obtaining an overall teaching output that combines main line consistency, modal complementarity, and error constraint capabilities.

[0151] This embodiment also provides a computer device applicable to the knowledge graph-based multimodal resource generation method for word knowledge points, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the knowledge graph-based multimodal resource generation method for word knowledge points as proposed in the above embodiment.

[0152] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0153] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for generating multimodal resources of word knowledge points based on knowledge graphs as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0154] In summary, this invention, by continuously linking local knowledge point sets, relation matrix decomposition, complex representation of knowledge points, and complex representation of relations, enables shared coverage and directional bias to move beyond static co-occurrence and unidirectional constraint levels and instead jointly enter the calculation of the total matching value. This allows for the simultaneous identification of semantically similar but differently pedagogically useful, and superficially different but stably pedagogically supportive, knowledge point relationships within the same computational loop. This significantly improves the discriminability and orderability of local knowledge organization. Through progressive filtering based on combinatorial consistency, directional closure difference, and chain length, the selected main pedagogical closed path possesses comprehensive characteristics of structural integrity, target alignment, and compact transmission. This provides a single-path foundation for subsequent unified knowledge skeleton extraction, producing a pedagogical closed-loop determination effect that cannot be directly obtained through local relation filtering.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for generating multimodal resources of word knowledge points based on knowledge graphs, characterized in that: include, For each word resource, a local knowledge point set is constructed, a relation matrix is ​​established, the symmetric part of the relation matrix is ​​used as the shared part matrix, and the antisymmetric part of the relation matrix is ​​used as the directional part matrix to form a complex representation of each knowledge point. The shared coverage of each knowledge point on the relation is calculated based on the shared part matrix, and the directional bias of each knowledge point on the relation is calculated based on the directional part matrix to form a complex representation of the relation. The total matching value of the knowledge point is calculated by combining the complex representation. The set of relation segments is filtered based on the total matching value. Relationships are defined and knowledge points associated with each other are combined into knowledge nodes. Knowledge chains are formed based on the set of relation segments. The teaching objectives are clarified and the target head relationship is determined based on the teaching relationship types corresponding to the starting point, ending point, and intermediate knowledge points of the knowledge chain. The combination consistency of the knowledge chain is calculated based on the difference between the combination matrix constructed by the knowledge chain and the relationship matrix corresponding to the target head relationship. The total directional matching value of different target head relationships is extracted through the knowledge chain. The absolute value of the difference between the total directional matching value and the total directional matching value of the target head relationship is used as the directional closure difference to filter out the main teaching closed path. Based on the selected main teaching closed path, according to the knowledge point type to which the knowledge points in the main teaching closed path belong, four types of structured slots are constructed: core meaning, boundary meaning, establishment conditions, and extraction risk, which are combined to form a unified knowledge skeleton. Multimodal knowledge resources are classified and mapped according to a unified knowledge skeleton to form a coverage table. Based on the coverage table, gaps are filled and deviating content is deleted.

2. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 1, characterized in that: The method of using the symmetric part of the relation matrix as the shared part matrix and the antisymmetric part of the relation matrix as the directional part matrix includes: distinguishing and defining the knowledge points contained in each word resource, constructing a local knowledge point set, determining the association based on whether there is a learning association between word knowledge points, and establishing a relation matrix. For each relation in the relation matrix, define the shared part and the direction part, and construct the shared part matrix and the direction part matrix respectively.

3. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 2, characterized in that: The formation of the complex representation of each knowledge point includes, based on the shared part matrix, for any knowledge point, calculating the total shared amount and constructing a shared semantic representation according to the shared proximity value accumulated between different knowledge points on all shared parts; Based on the direction part matrix, for any knowledge point, the total direction is calculated and constructed as a direction constraint representation according to the net direction difference accumulated between different knowledge points in all directions. Complex representations of each knowledge point are formed based on shared semantic representations and directional constraint representations.

4. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 3, characterized in that: The calculation of the total matching value of knowledge points by combining complex representations includes calculating the shared coverage of different knowledge points based on the shared part matrix, and calculating the directional bias of different knowledge points based on the directional part matrix, which together form the complex representation of the relationship. Based on complex representation, the complex representations of the knowledge points at the starting point and the knowledge points at the ending point of the relationship are extracted, and the shared matching value is calculated. At the same time, the directional matching value is calculated based on the complex representation of the relationship, and the sum is calculated as the total matching value of the knowledge points.

5. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 4, characterized in that: The target head relationship is determined based on the teaching relationship types corresponding to the starting point, ending point, and intermediate knowledge points of the knowledge chain. This includes determining the reliability of the relationships between knowledge points based on the total matching value; if the total matching value of a corresponding relationship is less than 0, it is filtered out, and the remaining relationships are retained as a set of relationship segments. Define two knowledge points related to the association relationship as the association start point and the association end point, and combine the association start point, the association end point, and the association relationship to form a knowledge node. Enumerate the knowledge nodes based on the set of relationship segments and form a knowledge chain. Based on the structured vocabulary of the starting knowledge point, the ending knowledge point, and the structured vocabulary of the knowledge points passed through in any knowledge chain, the corresponding target head relationship is determined. Based on the target head relationship, the relationship matrix is ​​searched, the association relationships that satisfy the corresponding target head relationship are selected, and the corresponding knowledge nodes are extracted to construct the corresponding relationship matrix of the target head relationship.

6. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 5, characterized in that: The step of calculating the combination consistency of the knowledge chain based on the difference between the combined matrix constructed according to the knowledge chain and the relation matrix corresponding to the target head relation includes merging the shared part matrix and the directional part matrix into a relation action matrix, and performing continuous matrix multiplication operations on the relation action matrix according to the order of relations in the knowledge chain to obtain the combined matrix, and calculating the combination consistency of each knowledge chain in combination with the corresponding relation matrix.

7. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 6, characterized in that: The step of using the absolute value of the difference between the total direction matching value and the total direction matching value of the target head relationship as the direction closure difference to screen out the main teaching closed path includes: extracting the total direction matching value of different associations based on the knowledge chain; calculating the total direction matching value of different target head relationships in the corresponding relationship matrix based on the corresponding relationship matrix; and defining the direction closure difference. The main teaching closed path is screened out based on the knowledge chain set, the consistency of each chain combination, and the directional closure difference.

8. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 7, characterized in that: Based on the knowledge point type to which the knowledge points in the main teaching closed path belong, four types of structured slots are constructed: core meaning, boundary meaning, conditions for establishment, and extraction risk. These slots are combined to form a unified knowledge framework. This includes, based on the main teaching closed path of the filter, querying whether the starting knowledge point belongs to the sense table, and defining it as the core sense slot content; Based on the main teaching closed path of the filter, query whether there are knowledge points belonging to the discrimination table, and define them as boundary slot content; Based on the main teaching closed path of the filter, query whether there are knowledge points belonging to the collocation table and the context table, and define them as the content of the condition slot. Based on the main teaching closed path of the filter, query whether there are any knowledge points belonging to the error table, and define it as extracting risk slot content; The core semantic slots, boundary semantic slots, establishment condition slots, and extracted risk slots are combined into a unified knowledge skeleton.

9. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 8, characterized in that: The multimodal knowledge point resources are classified and mapped according to a unified knowledge skeleton to form a coverage table, including: based on the slots of the unified knowledge skeleton, listing the smallest set of knowledge units that must be reflected by all knowledge point resources, and taking all the knowledge point sets of the four operations as the first knowledge unit set, the second knowledge unit set, the third knowledge unit set, and the fourth knowledge unit set, respectively. Meanwhile, the initial drafts of multimodal resources are categorized, including text drafts, image drafts, pronunciation prompt drafts, and exercise drafts. Map the initial draft of the resources to the four slots of the unified knowledge framework to form a coverage table.

10. The method for generating multimodal resources of word knowledge points based on knowledge graphs as described in claim 9, characterized in that: The process of filling in gaps and deleting deviating content based on the coverage table includes querying the corresponding coverage value of knowledge point resources in the coverage table, identifying missing resources, filling in the unrepresented slots, and deleting content that deviates from the unified knowledge framework, thereby obtaining text resources, image resources, pronunciation prompt resources, and practice question resources organized around the unified knowledge framework.

Images