A knowledge graph-based intelligent teaching full-process intelligent auxiliary teaching system

Abstract

The present application relates to the technical field of wisdom teaching, in particular to a wisdom teaching whole-process intelligent auxiliary teaching system based on a knowledge graph. The system solves the technical problems that multi-source heterogeneous teaching data cannot be automatically fused into a unified knowledge expression structure, and that the construction of a knowledge graph and the association of resources rely on a model, resulting in uninterpretable, search-lacking, and relationship-reasoning-weak technical problems. The system includes obtaining multi-source heterogeneous teaching data, extracting knowledge point triples according to preset ontology rules, generating a knowledge graph through conflict resolution, calculating double-feature weighted confidence to construct a resource index table, and encapsulating a search and reasoning engine. The system achieves the technical effects of zero model dependence, rule result interpretability, automatic multi-source conflict resolution, and search and reasoning integration. The present application can be applied to the scenes of teacher lesson preparation, personalized learning recommendation, intelligent test paper generation, learning situation diagnosis, and automatic archiving of teaching resources in a wisdom teaching platform.

Description

Technical Field

[0001] This invention relates to the field of smart teaching technology, and in particular to a smart teaching system based on knowledge graphs that provides intelligent assistance throughout the entire teaching process. Background Technology

[0002] Data generated in smart teaching scenarios, such as textbook chapter texts, classroom teaching transcripts, student homework correction records, and quiz results, exhibit fundamental differences in format, granularity, and semantics. Current technologies cannot automatically integrate this data, scattered across the three stages of instructional planning, classroom implementation, and learning feedback, into a unified knowledge representation structure. This results in inconsistent representations of the same knowledge point across different data streams and a lack of automatic discrimination mechanisms when relational descriptions conflict. Existing educational knowledge graph construction relies on manual expert analysis or automated extraction based on machine learning models. The former is limited by expert productivity and cannot be scaled up, while the latter depends on large-scale labeled training data and the inference results are uninterpretable. Teachers and curriculum researchers cannot directly intervene in or correct the extraction logic, making it difficult for knowledge graphs to adapt to the differentiated needs of different subjects, textbook versions, and teaching styles.

[0003] When establishing associations between massive candidate teaching resources and knowledge point entities in knowledge graphs, existing technologies mainly employ two methods: First, statistical matching based solely on keyword co-occurrence frequency. While this method can cover synonyms, it fails to distinguish the precise contextual location of terms within the text, leading to the mislabeling of numerous irrelevant resources with only occasional word overlap with the knowledge points, resulting in high noise and low accuracy. Second, text similarity calculation based solely on edit distance or vector similarity. Although this method can characterize character-level or semantic-level similarity, it ignores the unique contribution of knowledge point-specific keywords, such as core subject terms and marker words in typical examples, to the association strength. When resource texts are short or contain synonyms, similarity scores are prone to drift, resulting in insufficient recall. Current technologies have not yet developed a scheme that weights and fuses edit distance similarity with keyword co-occurrence frequency for teaching resource annotation, nor have they recognized that this fusion can achieve a balance between accuracy and recall, thus resolving the inherent contradiction of balancing accuracy and comprehensiveness in resource annotation within teaching scenarios.

[0004] Furthermore, existing teaching resource retrieval systems can only perform direct matching searches based on user-input keywords or knowledge point identifiers, returning resources directly related to that knowledge point. They cannot utilize the learning sequence dependencies between knowledge points stored in the knowledge graph, i.e., predecessor / successor relationships, for reasoning and expansion. In teaching practice, students' understanding of a particular knowledge point often depends on its prerequisite knowledge, the mastery of prerequisite knowledge points, and the application of knowledge points or successor knowledge points after learning that knowledge point. Existing retrieval systems cannot automatically recall prerequisite and extension resources based on the predecessor / successor relationships between knowledge points. The search results do not conform to the cognitive patterns of teaching, leading to teachers still needing to manually supplement a large amount of related material during lesson preparation and students' self-study, thus reducing the intelligence level of the teaching support system. Summary of the Invention

[0005] To achieve the above objectives, the present invention adopts the following technical solution: One aspect of the present invention provides a knowledge graph-based intelligent teaching system for the entire teaching process, comprising: The module for acquiring heterogeneous teaching data is used to acquire multi-source heterogeneous teaching data streams in smart teaching scenarios; each multi-source heterogeneous teaching data stream contains at least two data modalities from textbook chapter text, classroom teaching recording subtitles, student homework correction records, and stage test answer results. The knowledge point triple set module is used to extract knowledge point entities and determine relation types for each multi-source heterogeneous teaching data stream according to preset subject knowledge ontology rules, thereby determining the knowledge point triple set of the multi-source heterogeneous teaching data stream. The subject knowledge ontology rules include a subject core concept dictionary, a template for superior / subordinate relations between concepts, and a table of conditions for determining predecessor / successor relations between concepts. Each triple in the knowledge point triple set describes the subject knowledge point entities appearing in the multi-source heterogeneous teaching data stream and their semantic associations with each other in a structured form of head entity, relation type, and tail entity. The structured teaching knowledge graph module is used to generate a structured teaching knowledge graph by deduplicating and merging sets of knowledge point triples corresponding to multiple heterogeneous teaching data streams from multiple sources after processing based on preset conflict resolution rules. The preset conflict resolution rules include at least the following: when there are more than two different relation types between the same head entity and the same tail entity, the relation type with the highest frequency is selected as the preferred relation; when the representation of the same entity in different data streams is inconsistent, the standard name in the subject core concept dictionary is used as the normalized entity name. The association similarity module is used to perform knowledge point annotation and association retrieval on candidate teaching resource data based on the structured teaching knowledge graph, and to determine the teaching resource annotation index table. The teaching resource annotation index table records the knowledge graph entity identifier and association confidence score associated with each candidate teaching resource. The association confidence score is calculated by weighting the edit distance similarity between the resource text and the entity description text and the keyword co-occurrence frequency.

[0006] A retrieval engine module is built to utilize the labeled index table of teaching resources to construct a knowledge graph retrieval and reasoning engine for the entire intelligent teaching process.

[0007] This invention segments and aligns standardized, multi-source, heterogeneous teaching data streams using textbook chapter titles as segmentation anchors. This ensures that each teaching unit data block carries a clear chapter affiliation identifier, resolving the structural inconsistency issue caused by different sources of the original data. It also spatially aligns teaching data scattered across textbook chapter texts, classroom teaching transcripts, student homework correction records, and stage test results according to chapter affiliation. By performing longest forward matching between the text content of the teaching unit data blocks and standard terms and equivalent aliases in the subject's core concept dictionary, prioritizing the matching of the longest terms, the unstructured teaching text is converted into a structured list of knowledge point entity names. Simultaneously, the precise location and frequency of each entity in the text are recorded, providing definite node location and frequency statistics for co-occurrence relationship determination and relationship type matching steps. Furthermore, by extracting co-occurrence sentence fragments of knowledge point entities within the context and sequentially performing pattern matching with the superordinate relation template and the predecessor-successor relation determination condition table, the text-level co-occurrence information is converted into relation type labels and direction determination results through template rules. This allows the system to extract semantic relationships from unstructured text without relying on machine learning models. This invention generates knowledge point triples by using successfully matched knowledge point entities as head and tail entities and combining them with candidate relation types determined by the corresponding matches as relation types. It then aggregates all triples from each standardized data stream to determine the knowledge point triple set, transforming each piece of knowledge association information extracted from the original teaching text into a unified graph structure unit, providing a standardized data input format for multi-data stream fusion. Furthermore, this invention uses the head and tail entity identifiers in each triple as joint grouping keys for grouping and aggregation, grouping all triples corresponding to the same pair of head and tail entities from different teaching data streams into the same group. This allows for a centralized presentation of knowledge association information extracted from the same pair of entities from different data sources, providing a complete data view based on entity pairs for relation type conflict resolution. For each candidate triple grouping result, the frequency of occurrence of each relation type within the group is statistically analyzed, and the relation type with the highest frequency is selected as the preferred relation type for that group to determine the deduplicated triples carrying the preferred relation type. This ensures that the relation types carried by the deduplicated triples comprehensively reflect the common knowledge expression of most data sources, including textbook chapter texts, classroom teaching transcripts, student homework correction records, and stage test answers, thus resolving the conflict problem of the same pair of entities carrying multiple relation types due to the heterogeneity of data sources.For each deduplicated triple carrying a preferred relation type, the head and tail entity names are compared one by one with the standard names in the subject core concept dictionary. Non-standard names are replaced with the standard names corresponding to the dictionary to determine the structured teaching knowledge graph triples with normalized entity names. This ensures that the entity names in all triples written into the structured teaching knowledge graph are unified to the standard form in the subject core concept dictionary, eliminating the problem of inconsistent representation of the same knowledge point in different data streams due to differences in textbook versions and teachers' teaching styles. This ensures that all entity nodes in the final generated structured teaching knowledge graph have a consistent naming standard. This invention obtains several entity identifiers and corresponding entity description texts in the structured teaching knowledge graph, and extracts the resource text of each candidate teaching resource into a plain text string. This plain text string removes the formatting marks and non-text characters in the original file, unifying the text content from teaching resources of different formats into the same computable data form and eliminating the interference of file format differences on the text comparison operation. For each candidate teaching resource and each entity identifier, the minimum edit distance between the resource text and the entity description text is calculated. An edit distance similarity score is then determined by normalizing the minimum edit distance and the lengths of the two texts. This score measures the precise matching degree between the two texts at the character-level operation cost, quantifying text differences into normalized values ​​to ensure comparability of similarity between texts of different lengths. A preset keyword set corresponding to the entity identifier is obtained, and word segmentation and stop word filtering are performed on the resource text. The number of times each term is hit in the keyword set after filtering is counted to determine the keyword co-occurrence frequency. This keyword co-occurrence frequency measures the semantic overlap between the resource text and the knowledge point entity at the terminology level using a word-level hit count, quantifying the coverage relationship between subject-distinguishing content words in the resource text and the preset entity keywords. By using edit distance similarity scores and keyword co-occurrence frequencies as input features and performing a weighted summation with preset weight coefficients, a correlation confidence score is obtained. This correlation confidence score simultaneously incorporates character-level precise matching information and lexical-level semantic coverage information, enabling a single value to comprehensively reflect text relevance across two different dimensions and avoiding discrimination bias caused by synonym substitutions or sparsity in short texts due to a single feature. A full list of correlation confidence scores between the same candidate resource and each entity is obtained. Each score in the full list is compared numerically with an admission threshold read from system configuration parameters. Candidate correlation pairs with scores higher than the admission threshold are selected. The resource identifier, entity identifier, and corresponding confidence score of each candidate correlation pair are extracted and written into an index table. This threshold filtering operation eliminates low-confidence pairings, ensuring that all records written to the index table are verified for confidence and valid correlation records are persistently stored in a structured format.By establishing a forward mapping relationship between resources and knowledge graph entities using resource identifiers and entity identifiers in the index table as key-value pairs, the system can directly locate all associated resource identifiers through hash mapping when receiving a query request for a specific knowledge graph entity identifier. This avoids a full table scan of the teaching resource annotation index table, reducing the time complexity of resource retrieval. Based on the forward mapping relationship and the predecessor and successor edges between entities stored in the structured teaching knowledge graph, a one-hop extended traversal is performed along the directed edges of the graph from the current entity node to the adjacent entity node to determine the set of extended associated entity identifiers for each resource beyond its directly associated entities. This set of extended associated entity identifiers enables the system to obtain resource annotation results for adjacent entities that have a learning sequence dependency relationship with the currently queried entity, expanding the retrieval scope from a single knowledge entity to its predecessor and successor knowledge entities. This ensures that the resource query results cover teaching materials that require prior preparation and extension during the teaching process. The system merges the sets of directly related entity identifiers and extended related entity identifiers contained in the forward mapping relationship to generate a unified retrieval mapping table that includes both direct and indirect relationships. This unified retrieval mapping table eliminates the data structure separation between direct and extended retrieval results, enabling the system to return both directly labeled resources and derived related resources obtained through graph edge expansion in a single query operation. Based on the unified retrieval mapping table and the graph traversal interface provided by the structured teaching knowledge graph, the key-value pair search logic and the graph edge traversal logic are encapsulated in the same software module. This allows upper-level teaching applications to call the query function without distinguishing the differences in the acquisition paths of directly and indirectly related resources, while ensuring that the key-value retrieval mode and the relational reasoning mode share the same resource association data foundation, avoiding result discrepancies between the two query modes due to inconsistent data sources. Attached Figure Description

[0008] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart is provided for the present invention of a knowledge graph-based intelligent teaching system for the entire teaching process. Figure 2 This invention provides a schematic diagram of a knowledge graph-based intelligent teaching system for the entire teaching process. Figure 3 A block diagram of the electronic device provided by the present invention.

[0009] Attached reference numerals: 1. Heterogeneous teaching data acquisition module; 2. Knowledge point triple set module; 3. Structured teaching knowledge graph module; 4. Association similarity module; 5. Search engine construction module; 6. Central processing unit / microprocessor / main control chip; 7. Storage medium; 8. Data bus; 9. Input / output bus / external bus / device bus; 10. Display; 11. Input / output device; 12. Computer-readable instructions; 13. Non-transitory computer-readable storage medium. Detailed Implementation

[0010] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0011] like Figure 1 As shown, this embodiment of the invention provides a knowledge graph-based intelligent teaching system that provides intelligent assistance throughout the entire teaching process, comprising: The heterogeneous teaching data acquisition module 1 is used to acquire multi-source heterogeneous teaching data streams in smart teaching scenarios; each multi-source heterogeneous teaching data stream contains at least two data modalities from textbook chapter text, classroom teaching recording subtitles, student homework correction records, and stage test answer results.

[0012] Specifically: Textbook chapter texts refer to the structured text obtained by digitizing each chapter, section, and unit of the textbook. They encompass titles, definitions, examples, and exercises, providing the skeleton framework of a knowledge graph within the system. Chapter levels directly map to knowledge level hierarchies, definitions provide a candidate set of core concepts, and the arrangement of examples provides a basis for initial determination of predecessor and successor relationships. Classroom teaching transcripts are timestamped text sequences generated by real-time transcription of teacher lectures using a speech recognition engine. As a dynamic incremental supplementary source, they can add orally introduced concepts not included in the textbook. Teachers' oral expressions of logical relationships between knowledge points serve as non-textual evidence of these relationships, while real-time error corrections are used to adjust the confidence level of these relationships. Student homework correction records are digitized data of answers, evaluation results, annotations, and error type labels after teacher corrections. As a reverse verification source, the collective error rate can reversely calibrate the confidence level of dependencies between knowledge points. Teacher annotations describing confusion with certain knowledge points provide rule input for distinguishing relationships between similar entities. The results of phase tests include individual scores, answer times, and accuracy markers for unit, midterm, or final tests. These serve as a quantitative calibration source. The strength coefficients of the antecedent and successor relationships of knowledge points are quantified and calibrated through statistical analysis of the group distribution of test scores. Cross-analysis of answer time and question complexity is used to attach cognitive load attribute labels. Two or more of the above four data modalities are forcibly selected and combined. Their overall function is to ensure the integrity of the data loop from instructional planning to classroom implementation and learning feedback, to achieve multimodal cross-validation of triple confidence, to eliminate inconsistencies in terminology caused by differences in textbook versions and teacher teaching styles, and to provide a normalized incremental update channel for the knowledge graph at different time granularities. This forms the prerequisite technical foundation for the effective execution of all processing steps, including knowledge point extraction, conflict resolution and fusion, and resource association indexing.

[0013] As a preferred approach: Module 1 for acquiring heterogeneous teaching data includes: The raw teaching data acquisition submodule is used to acquire raw multi-source heterogeneous teaching data streams in smart teaching scenarios; each raw multi-source heterogeneous teaching data stream contains at least two modalities from textbook chapter text, classroom teaching recording subtitles, student homework correction records, and stage test answer results. The teaching data stream determination submodule is used to remove noise characters and invalid records from each original multi-source heterogeneous teaching data stream according to preset data cleaning rules, and to determine a standardized multi-source heterogeneous teaching data stream. The teaching unit data block set submodule is used to segment and align the standardized multi-source heterogeneous teaching data stream according to the teaching chapter boundaries for each standardized multi-source heterogeneous teaching data stream, using the chapter title and section title in the textbook chapter structure as the segmentation anchor points, and determine the teaching unit data block set with chapter affiliation identifier. The forward matching submodule is used to perform the longest forward match between the text content of each teaching unit data block in the teaching unit data block set and the standard terms and equivalent aliases in the subject core concept dictionary. The entity name corresponding to the successfully matched dictionary entry is used as the extraction result to determine the list of knowledge point entity names for each teaching unit data block. Each extraction result in the list of knowledge point entity names is accompanied by the offset of the term's occurrence position and the frequency of occurrence in the corresponding teaching unit data block.

[0014] Specifically, the data cleaning rules include: removing headers, footers, and exercise numbers from textbook chapter texts; filtering out transcription segments with a confidence level below 85% from classroom teaching transcripts; removing blank answers and invalid annotations from student homework correction records; and filtering out abnormal submissions with an answer time of less than 500 milliseconds from stage test results. Classroom teaching transcripts are segmented after aligning with the teaching date and textbook chapter table of contents. Student homework correction records and stage test results are categorized and segmented according to the chapter number corresponding to the test questions. Invalid records refer to data collected abnormally, with corrupted formats, or with internal errors in the original multi-source heterogeneous teaching data stream. Data units that are missing and cannot be used for knowledge point identification are specifically manifested as blank lines or garbled character fragments in textbook chapter texts, unreliable transcription fragments in classroom teaching transcripts with speech recognition confidence below a preset threshold, blank records in student homework correction records with no answers, and abnormal submission records in stage test results with excessively short answer times that have no practical reference value. These invalid records are used as targets for elimination in the data cleaning process of the system. Filtering invalid records reduces noise input to downstream processing steps and improves the accuracy and efficiency of knowledge point entity extraction. Standardized multi-source heterogeneous teaching data streams refer to data sets obtained after processing the original multi-source heterogeneous teaching data streams through preset data cleaning rules. These data sets are characterized by uniform format, pure content, and consistent field structure. Specifically, textbook chapter texts are stripped of non-textual interference characters such as headers, footers, and exercise numbers, retaining only the chapter title and main text description. Classroom teaching transcripts have their timestamps and low-confidence transcription fragments removed, leaving only the teacher's valid lecture text. Student homework correction records are filtered to remove blank answers and invalid annotations, retaining only the valid question number and correction result fields. Stage test results are filtered to remove abnormal submissions, retaining only the valid question number, score, and time consumed fields. This standardized multi-source heterogeneous teaching data stream serves as the standard input format for chapter segmentation in the system, ensuring that data from different collection channels enters the processing flow with uniform text encoding, uniform field separators, and uniform record structure, avoiding logical confusion in segmentation and alignment operations due to differences in the original data format. The segmentation anchor point refers to the marker information used to determine the boundary position of the teaching chapter. Specifically, it includes the chapter title string and section title string with number structure in the textbook chapter text. In the system, the segmentation anchor point serves as the positioning mark for segmenting the standardized multi-source heterogeneous teaching data stream from a continuous long text sequence into discrete teaching unit data blocks. The system completes the segmentation operation with the starting line and ending line of the anchor point as the boundary and establishes a binding relationship between each segmented data block and the chapter number corresponding to the anchor point.The standard nomenclature is a unique, standardized name string assigned to each subject knowledge point entity in the core concept dictionary. The equivalent alias is another acceptable textual representation of the same knowledge point entity as the standard nomenclature. Specifically, it includes commonly used abbreviations of the knowledge point, synonyms used in different textbook versions, and colloquial abbreviations commonly used in classroom teaching. Together, the standard nomenclature and equivalent alias constitute the target matching dictionary for the longest forward matching operation in the system, expanding the recall scope of text matching and ensuring that the same knowledge point representation under different textbook versions and different teaching styles can be correctly identified. Longest forward matching refers to a string matching algorithm that starts from the beginning of the text content of the teaching unit data block, selects the longest continuous substring after the current pointer that completely matches the standard nomenclature or equivalent alias in the core concept dictionary as a matching result, and then moves the pointer after the matching substring to continue the matching until the end of the text. This longest forward matching is used in the system as a deterministic extraction method for identifying knowledge point entity names from unstructured teaching text, avoiding entity segmentation errors caused by prioritizing matching the longest term. An entity name refers to a continuous string of characters identified from the text content of a teaching unit data block through the longest forward matching operation, which matches the standard terminology or equivalent alias in the core concept dictionary of the subject. In the system, it serves as the raw output of the knowledge point extraction step, recorded in its actual written form in the original text and accompanied by position and frequency information before being passed to the processing stage. The extraction result refers to the structured data item generated by binding and encapsulating the standard entity name corresponding to the successfully matched dictionary entry with the matching position information. Specifically, it is a data triple consisting of the position offset, frequency, and corresponding standard entity identifier of each term in the corresponding teaching unit data block. This extraction result serves as the final output unit of the knowledge point entity extraction stage, providing deterministic entity node input for the knowledge point triple generation stage. The knowledge point entity name list refers to a collection data structure consisting of all extraction results extracted from the same teaching unit data block through the longest positive match. It is organized in the form of an array or linked list and sorted in ascending order of the offset of the occurrence position. In the system, this list serves as a complete list of knowledge point entities contained in the teaching unit data block. It is used in the steps to traverse all identified entities in the data block and perform pairwise relationship type determination processing.A teaching unit data block refers to a data segment with a clear chapter affiliation identifier obtained after dividing a standardized multi-source heterogeneous teaching data stream according to the boundary of the segmentation anchor point. Specifically, it includes all textbook text content belonging to the same chapter, classroom teaching transcripts and subtitles belonging to the same chapter, student homework correction records belonging to the same chapter, and stage test results belonging to the same chapter. This teaching unit data block serves as the smallest processing unit for knowledge point entity extraction in the system, enabling parallel execution of extraction operations at the chapter level. It also ensures that each entity name extraction result can be traced back to its specific chapter, providing contextual constraints for relation determination. The position offset refers to the character sequence number of the first character of the matched continuous character substring in the extraction result within the complete text of the teaching unit data block. This position offset is used in the system to determine the co-occurrence distance between two knowledge point entities in the text during the relation type determination step. When the difference in position offsets between two entities is less than a preset window size (e.g., 200 characters), it is determined that they have a co-occurrence relationship in the same sentence or paragraph, thus serving as one of the candidate conditions for determining a superior or predecessor / successor relationship. The frequency of occurrence refers to the total number of times the same entity name is matched by the longest positive match in all text content of the same teaching unit data block. This frequency of occurrence is used as a weighting basis for the selection of relation types in the conflict resolution step in the system. That is, when there are multiple relation type candidates between the same head entity and the same tail entity, the cumulative value of the frequency of occurrence of entities in the data source corresponding to each candidate relation type is used to calculate the confidence weight of the relation type. The relation type with higher frequency gets higher priority in conflict resolution.

[0015] In the embodiment of the heterogeneous teaching data acquisition module, raw multi-source heterogeneous teaching data streams from a smart teaching scenario are acquired. Each raw multi-source heterogeneous teaching data stream contains at least two modalities from textbook chapter text, classroom teaching recording subtitles, student homework correction records, and stage test answer results. This allows the system to simultaneously collect raw data from three different stages—teaching pre-planning, classroom implementation, and learning feedback—during the data acquisition stage, providing multi-dimensional raw materials across these stages for knowledge extraction. For each raw multi-source heterogeneous teaching data stream, noise characters and invalid records are removed according to preset data cleaning rules to determine a standardized multi-source heterogeneous teaching data stream. This data cleaning operation removes garbled characters, blank records, low-confidence transcription fragments, and abnormal submission records from the raw data, ensuring that the remaining data content is pure and formatted uniformly, thus reducing the proportion of noise input in the processing steps. For each standardized multi-source heterogeneous teaching data stream, the chapter title and section title in the textbook chapter structure are used as the segmentation anchor points. The standardized multi-source heterogeneous teaching data stream is segmented and aligned according to the teaching chapter boundaries to determine the set of teaching unit data blocks with chapter affiliation identifiers. This segmentation operation divides the continuous long text sequence into discrete unit data blocks according to the inherent hierarchical structure of the textbook, so that each teaching unit data block carries a clear chapter affiliation identifier, ensuring that the knowledge point entity extraction results can be accurately traced back to the chapter to which they belong. The text content of each teaching unit data block in the teaching unit data block set is matched with the standard terms and equivalent aliases in the subject core concept dictionary using the longest forward match. The entity name corresponding to the successfully matched dictionary entry is used as the extraction result to determine the knowledge point entity name list for each teaching unit data block. Each extraction result in this list is accompanied by the offset of the term's occurrence position and the frequency of occurrence in the corresponding teaching unit data block. This longest forward match operation converts the unstructured teaching text into a structured list of knowledge point entity names by prioritizing the matching of the longest term. At the same time, it records the precise position and occurrence number of each entity in the text, so that the co-occurrence relationship determination and relationship type matching steps have definite node position and frequency statistics.

[0016] The Knowledge Point Triple Set Module 2 is used to extract knowledge point entities and determine relation types for each multi-source heterogeneous teaching data stream according to preset subject knowledge ontology rules, thereby determining the knowledge point triple set of the multi-source heterogeneous teaching data stream. The subject knowledge ontology rules include a subject core concept dictionary, a template for superior / subordinate relations between concepts, and a table of conditions for determining predecessor / successor relations between concepts. Each triple in the knowledge point triple set describes the subject knowledge point entities appearing in the multi-source heterogeneous teaching data stream and their semantic relationships in a structured form of head entity, relation type, and tail entity.

[0017] Specifically, the pre-defined subject knowledge ontology rules are a set of predefined and solidified subject knowledge expression specifications. These include a core subject concept dictionary, templates for hierarchical and superordinate relationships between concepts, and a table of conditions for determining predecessor and successor relationships between concepts. These rules serve as a unified standard and constraint framework for knowledge extraction within the system, ensuring semantic alignment of multi-source heterogeneous data from different data modalities, textbook versions, and teaching styles under the same knowledge system. This avoids incomparable or unintegratable extraction results due to differences in terminology or diverse relationship expressions. Knowledge point entity extraction refers to the process of scanning and matching the text content in each multi-source heterogeneous teaching data stream under the guidance of the pre-defined core subject concept dictionary, identifying subject concept terms within the dictionary scope and marking them as candidate knowledge nodes. This step transforms unstructured natural language text into a structured list of knowledge nodes, a prerequisite operation for all relationship determination and graph construction. Relationship type refers to the semantic connection category between any two subject knowledge point entities in terms of teaching content or learning logic. It at least covers superior, subordinate, predecessor, and successor relationships. In the system, it assigns clear category labels to the associations between knowledge points, making the final knowledge graph not only a network of connections between knowledge point entities, but also a teaching logic relationship network with semantic category attributes that can be understood and retrieved by computers. The knowledge point triple set is a collection of multiple basic data units composed of three elements: head entity, relation type, and tail entity. Each triple is an indivisible minimum semantic knowledge unit, independently and completely describing a knowledge association. This set is stored in the system as entries in a data table or graph database and is the basic operation unit for deduplication, merging, and conflict resolution. The subject-specific core concept dictionary is a standardized terminology table compiled by subject volume. It contains the standard name of each subject knowledge point entity and its unique identifier code in the dictionary. It plays a core role in the system by normalizing entity names and eliminating ambiguity. For example, when the textbook text refers to a function's maximum or minimum value while the classroom teaching transcript uses the term "function extreme value," the system uses the standard name in the dictionary as the sole basis for unification. The hierarchical relationship template between concepts is a pre-set set of matching rules used to determine whether a hierarchical inclusion relationship exists between two knowledge point entities and to determine its directionality. Specifically, it includes several typical sentence templates and keyword combinations, such as "includes," "belongs to," and "belongs to," along with corresponding judgment priorities and direction determination rules. This template enables the system to automatically identify and determine the relationship direction of concepts in the textbook chapter hierarchy and in teacher lectures.The pre-defined criteria table for determining predecessor and successor relationships between concepts is a pre-defined set of structured criteria used to determine whether there is a learning sequence dependency relationship and its directionality between two knowledge point entities. It includes at least a list of pre-existing knowledge point conditions and a list of successor knowledge point conditions, as well as constraints on their order in textbook chapter numbers and the order of their timestamps in the teaching transcript. This criteria table provides multi-dimensional quantitative criteria for determining predecessor and successor relationships in the system, solving the problem of cross-chapter knowledge dependency identification that cannot be covered by relying solely on textbook chapter order. The head entity refers to the knowledge point entity located at the beginning of any triple structure, serving as the originating end or upstream node of the semantic association. The relationship type refers to the specific category describing the semantic connection between the head and tail entities in the triple. The tail entity refers to the knowledge point entity located at the end, serving as the receiving end or downstream node of the semantic association. Together, these three constitute a complete knowledge statement unit, ensuring that each knowledge association is described completely and unambiguously from the three dimensions of subject, predicate, and object, guaranteeing consistent understanding of the same knowledge association across different system modules or processing stages. Structured formal description refers to the processing method and final organization of a data structure that converts unstructured knowledge information contained in natural language text into a fixed format data structure consisting of head entities, relation types, and tail entities. Its function is to transform human-readable but computer-difficult-to-process teaching content text into structured data that computers can retrieve, compare, and reason about, fundamentally supporting the construction, storage, and querying of knowledge graphs. Subject knowledge point entities refer to the smallest knowledge units with independent teaching meaning and assessment value within a teaching syllabus or subject knowledge system. Specifically, they are discretized and nameable concept nodes such as functions, quadratic equations, and chemical bonds. In the system, they are not only the basic constituent nodes of the knowledge graph but also the core objects around which all resource annotation, learning analysis, and path recommendation operations revolve. Semantic associations between entities refer to directed relationships with pedagogical meaning formed between subject knowledge entities based on the logic of teaching content or the cognitive patterns of students. These relationships include hierarchical hierarchical relationships and learning sequence dependencies. In the system, isolated subject knowledge nodes are woven into a directed graph structure with pedagogical logical semantics. This enables the system to complete the retrieval of related resources, the tracing of prior knowledge, and the planning of learning paths starting from any knowledge point based on this graph structure. It is the relational foundation for the realization of the full-process auxiliary functions of smart teaching.

[0018] As a preferred approach: Knowledge Point Triple Set Module 2 includes: The standard teaching data acquisition submodule is used to acquire standardized multi-source heterogeneous teaching data streams; each standardized data stream has been segmented and aligned according to teaching chapters. The standardized data stream submodule is defined to perform a positive maximum string matching between the text content of each standardized multi-source heterogeneous teaching data stream and the subject core concept dictionary to determine the knowledge point entity names, their occurrence positions, and their frequencies contained in the standardized data stream. The pattern matching submodule is used to extract the co-occurring statement fragments in the context of each knowledge point entity name in each standardized multi-source heterogeneous teaching data stream, and to perform pattern matching between the co-occurring statement fragments and the superordinate relation template and the predecessor-successor relation judgment condition table in sequence to determine the candidate relation type and relation direction between each knowledge point entity. The submodule for summarizing knowledge point triples is used to generate knowledge point triples by combining the successfully matched knowledge point entities as head entities and tail entities, and the corresponding candidate relation types as relation types; it also summarizes all triples corresponding to each standardized data stream to determine the knowledge point triple set.

[0019] Specifically, the standardized multi-source heterogeneous teaching data stream refers to a set of teaching unit data blocks that, after data cleaning to remove noise characters and invalid records, are segmented and aligned according to the teaching chapter boundaries using chapter titles and section titles in the textbook chapter structure as segmentation anchors, and are accompanied by chapter affiliation identifiers. This standardized data stream serves as the input corpus for the string forward maximum matching operation in the system. Its unified text encoding, consistent field structure, and clear chapter boundaries, achieved through previous processing, provide a predictable and unambiguous processing range for entity name recognition and relation type determination, avoiding inconsistent behavior of the matching logic on different data sources due to differences in the original data format. Forward maximum string matching refers to an iterative matching algorithm that starts from the beginning character position of the text content of each teaching unit data block in a standardized multi-source heterogeneous teaching data stream. Each time, it takes the longest continuous substring after the current pointer and compares it one by one with the standard names and equivalent aliases in the subject core concept dictionary. If the longest substring matches a dictionary entry, it is recorded as a match and the pointer is moved to the character after the end position of the substring. If it does not match, the last character of the substring is shortened by one position and the comparison continues until a match is found or the character is shortened to a single character and still cannot match, at which point the character is discarded and the pointer is moved forward one position. In the system, forward maximum string matching is used as a deterministic extraction method to identify knowledge point entity names from unstructured teaching text. By prioritizing the matching of the longest term, it avoids the entity segmentation error caused by prioritizing the matching of short words. Its output entity name, its position offset and frequency of occurrence provide accurate entity location information for the extraction of co-occurring sentence fragments. Co-occurrence sentence fragments refer to a sequence of continuous character substrings containing the names of two identified knowledge point entities, centered on the offset of their respective occurrence positions within the text content of the same teaching unit data block, and extracted according to a preset window length. The preset window length is set to 200 characters by default. When the absolute value of the difference between the position offsets of the two entity names is less than this window length, all text content between them, plus 50 characters before and after each entity, is extracted as a co-occurrence sentence fragment. In the system, this co-occurrence sentence fragment is used as the object of operation for superordinate relation template matching and predecessor-successor relation determination condition table matching. It transforms the discrete information that originally only contained isolated entity positions into a matchable text unit containing natural language context information between entities, enabling the system to determine the relation type based on the actual co-occurrence context of entities in real teaching texts rather than relying on external manual annotation.The sequential matching of sentence fragments with hyperordinate relation templates refers to performing string pattern matching on each of the templates in the preset hyperordinate relation template list according to the extracted co-occurring sentence fragments. Each hyperordinate relation template contains several typical sentence structure regular expressions and keyword combinations such as contain, include, belong to, and classify as, as well as corresponding direction determination rules. When a continuous substring in a co-occurring sentence fragment successfully matches the regular expression of a certain template, the hyperordinate relation type and direction determination result corresponding to that template are output as one of the candidate relation types between the pair of knowledge point entities. This sequential matching operation transforms the natural language expression at the text level into structured relation type labels through template matching rules, enabling the system to complete the conversion from unstructured text to semantic relations without relying on machine learning models. Matching sentence fragments sequentially with the predecessor-successor relationship determination condition table refers to performing a structured comparison on each item of the same co-occurring sentence fragment according to the pre-set determination condition table. This determination condition table includes a list of prerequisite knowledge points, a list of successor knowledge points, and the order constraints between them in terms of textbook chapter numbers and the order constraints of the teaching timestamps in the teaching transcript subtitles. When the head entity and tail entity in the co-occurring sentence fragment satisfy the order constraints in terms of chapter numbers in the textbook chapter structure and simultaneously satisfy the order constraints in terms of teaching timestamps in the teaching transcript subtitles, the predecessor or successor relationship type corresponding to the determination condition table is output as a candidate relationship type. This sequential matching operation solves the problem of the learning logical sequence that cannot be determined by text co-occurrence alone by introducing independent constraints of chapter numbers and timestamps, enabling the system to distinguish between text co-occurrence relationships and cognitive dependency relationships. Candidate relation types refer to one or more possible relation category labels and their corresponding direction judgment results obtained for the same pair of knowledge point entities after sequential matching of the superior relation template and the predecessor and successor relation judgment condition table. Specifically, they include superior relations, subordinate relations, predecessor relations, successor relations, and empty relations when neither of the two matching is successful. In the system, the candidate relation type is used as the input for the conflict resolution step. After multiple candidate relation types enter the deduplication and merging process, the unique relation type that is finally written into the structured teaching knowledge graph is output under the frequency statistics and optimization logic. Relationship direction refers to the directional identifier used to clarify which entity points to which entity when determining the semantic association between any two knowledge point entities. Specifically, when the determination result is a superior relationship or predecessor relationship, the direction is from the head entity to the tail entity; when the determination result is a subordinate relationship or successor relationship, the direction is from the tail entity to the head entity. This relationship direction is reflected in the system by the positional arrangement of the head and tail entities in the triple, ensuring that each edge in the generated knowledge graph has a clear logical direction and avoiding fundamental logical reversals in retrieval and reasoning results due to missing or incorrect directions.

[0020] In the embodiment of the knowledge point triplet set module, a standardized multi-source heterogeneous teaching data stream is obtained. This standardized data stream has been segmented and aligned according to teaching chapters, enabling processing operations to be performed at a unified chapter granularity. This eliminates the structural inconsistencies caused by different sources of the original data and spatially aligns the teaching data scattered across different data modalities according to chapter affiliation. For each standardized multi-source heterogeneous teaching data stream, its text content is subjected to a forward maximum matching operation against the subject's core concept dictionary to determine the knowledge point entity names, their locations, and frequencies contained in the standardized data stream. This forward maximum matching operation converts unstructured teaching text into a structured entity node list by prioritizing the matching of the longest terms, while recording the location and frequency information of entities in the text, thus providing a definite basis for entity location in the relationship determination step. For each knowledge point entity name identified in a standardized multi-source heterogeneous teaching data stream, co-occurring sentence fragments within the context of the standardized multi-source heterogeneous teaching data stream are extracted. These co-occurring sentence fragments are then sequentially matched with a superordinate relation template and a predecessor-successor relation determination condition table to determine candidate relation types and relation directions between knowledge point entities. This pattern matching operation converts text-level co-occurrence information into relation type labels and direction determination results through template rules, enabling the system to extract semantic relationships from unstructured text without relying on machine learning models. Each successfully matched knowledge point entity is used as the head entity and tail entity, and the corresponding candidate relation type is used as the relation type. Knowledge point triples are generated by combining these triples. All triples from each standardized data stream are aggregated to determine the knowledge point triple set. This knowledge point triple set is organized in a structured form with head entities, relation types, and tail entities, transforming each knowledge relationship extracted from the original teaching text into a unified basic unit of the graph structure. This provides a standardized data input format for multi-data stream fusion and the generation of structured teaching knowledge graphs.

[0021] The Structured Teaching Knowledge Graph Module 3 is used to generate a structured teaching knowledge graph by deduplicating and merging sets of knowledge point triples corresponding to multiple heterogeneous teaching data streams from multiple sources after using preset conflict resolution rules. The preset conflict resolution rules include at least the following: when there are more than two different relation types between the same head entity and the same tail entity, the relation type with the highest frequency is selected as the preferred relation; when the representation of the same entity in different data streams is inconsistent, the standard name in the core concept dictionary of the subject is used as the normalized entity name.

[0022] Specifically, at the knowledge fusion and graph generation level, the pre-set conflict resolution rules refer to the set of judgment logic and decision conditions for handling duplicate, contradictory, or ambiguous data when merging the sets of knowledge point triples corresponding to multiple heterogeneous teaching data streams from different sources. Specifically, it includes at least the frequency selection logic for relation type conflicts and the standard nomenclature normalization logic for entity name ambiguities. This rule plays the role of adjudicator and arbitrator in the multi-source knowledge fusion process of the system, ensuring that triples from different data modalities such as textbook chapter texts, classroom teaching transcripts, student homework correction records, and stage test answers can be merged under a unified decision framework, avoiding logical contradictions in the merged knowledge graph due to the inherent heterogeneity between data sources. Deduplication and merging refers to the process of identifying and merging triples pointing to the same head entity and tail entity combination from several sets of knowledge point triples into a single record. Specifically, it is performed by grouping all triples using the head entity identifier plus the tail entity identifier as the joint primary key. Multiple triples within the same group are then resolved by a preset conflict resolution rule, and a final retained triple is output. This step merges and compresses redundant knowledge associations scattered in various data streams in the system, eliminating storage redundancy and query confusion caused by the repeated occurrence of the same knowledge point relationship in multiple source data. It is a key intermediate operation for generating a single, clean, and non-redundant structured teaching knowledge graph. A structured teaching knowledge graph is a knowledge base that organizes and stores all subject knowledge point entities and their semantic relationships in a directed graph structure consisting of head entities, relation types, and tail entities. Nodes represent subject knowledge point entities, and edges represent the relation types that are ultimately retained after conflict resolution. The direction of the edges is from the head entity to the tail entity. This knowledge graph serves as a unified knowledge representation and storage hub in the system, providing a unique and standardized source of factual evidence for all teaching resource annotation, learning path retrieval, prior knowledge tracing, and personalized recommendation operations. Its completeness and accuracy directly determine the output quality and reliability of the entire intelligent teaching process support system. The preferred relation refers to the unique relation type selected from multiple candidate relation types when two or more different relation types appear within the same data stream or across different data streams between the same head entity and the same tail entity. This is determined by a preset conflict resolution rule. The selection criterion is to count the total frequency of each relation type under the head-tail entity combination in all data streams and determine the type with the highest frequency as the preferred relation. This preferred relation is used as the unique relation type label for the pair of knowledge point entities in the structured teaching knowledge graph. This solves the selection conflict problem when multiple sources of data give different relation descriptions for the same pair of knowledge points, such as the hierarchical inclusion relationship in the textbook chapter text, the learning sequence dependency relationship in the subtitles of classroom teaching recordings, and the potential association relationship indirectly reflected in student homework correction records.Different data streams specifically refer to data sets obtained from at least two of the four data modalities within the same teaching cycle: textbook chapter text, classroom teaching transcripts, student homework correction records, and stage test results. Each set is independently extracted and processed. Within the system, these data streams represent multi-dimensional projections of the same teaching knowledge content across different teaching stages and data carriers. They provide independent evidence from multiple sources for frequency statistics in the conflict resolution process, ensuring that the final determined optimal relationship is not based on one-sided information from a single source, but rather on statistical consensus results that integrate evidence from three levels: teaching planning, classroom implementation, and learning feedback. The subject core concept dictionary is a standardized terminology table compiled and persistently stored by subject. Each subject knowledge point entity is assigned a globally unique entity identifier code, a standard name, and several allowed equivalent aliases. This dictionary serves as the sole authoritative reference for entity name normalization within the system, ensuring that regardless of the variant expressions used in the input data, the entity names ultimately written into the knowledge graph are uniformly in the standard form specified by the dictionary. A standard nomenclature refers to the unique, standardized name assigned to each subject knowledge point entity in the subject core concept dictionary. In the system, it serves as the target output for all entity name normalization operations, unifying various textual representations of the same entity—such as formal terms used in textbook chapters, colloquial abbreviations used by teachers in classroom teaching transcripts, and non-standard expressions used by students in student homework correction records—into a single string format for storage and retrieval after mapping and transformation. An entity name refers to the specific textual expression of a subject concept identified through the knowledge point entity extraction step from various multi-source heterogeneous teaching data streams. In different data streams, it may manifest as a standard nomenclature, equivalent alias, abbreviation, or colloquial variant. This entity name serves as the input for subject core concept dictionary lookup and matching operations in the system. After dictionary matching and normalization mapping, it is converted into a unified standard nomenclature, thus providing a consistent addressing basis for all graph operations using entity identifiers as keys.

[0023] As a preferred approach: Structured instruction knowledge graph module 3 includes: The knowledge point triple set aggregation submodule is used to obtain several knowledge point triple sets; each knowledge point triple set is generated by extracting from a single multi-source heterogeneous teaching data stream; for several knowledge point triple sets, the head entity identifier and tail entity identifier in each triple are used as joint grouping keys to perform grouping aggregation, and the candidate triples with the same head and tail entity combination are determined. The deduplication triplet submodule is used to calculate the frequency of each relation type in each candidate triplet grouping result, and take the relation type with the highest frequency as the preferred relation type for the grouping, and determine the deduplication triplet carrying the preferred relation type. The teaching knowledge graph triplet submodule is obtained, which is used to compare the head entity name and tail entity name of each triplet carrying the preferred relation type with the standard names in the subject core concept dictionary one by one; and replace the non-standard names with the standard names corresponding to the dictionary to determine the structured teaching knowledge graph triplet after entity name normalization.

[0024] Specifically, in the knowledge fusion and graph generation stage, the joint grouping key refers to the composite key value formed by combining the head entity identifier and the tail entity identifier of each triple in a set of knowledge point triples. This composite key value is used as the grouping basis by either concatenating the head entity identifier string and the tail entity identifier string in a fixed order or directly as a joint tuple of the two. In this system, it serves as a unique identifier to classify triples scattered from different data stream sources according to the same head and tail entity pairs. This ensures that even if the entity names of triples extracted from textbook chapter texts and triples extracted from classroom teaching transcripts differ, as long as the head and tail entities point to the same pair of subject knowledge point entities after normalization mapping, they can be grouped into the same group for conflict resolution. Grouping aggregation refers to a merging operation that uses a joint grouping key as the grouping basis to classify all triples in a set of knowledge point triples into different groups according to the joint key value of their head entity identifier and tail entity identifier. Specifically, it involves traversing all triples and pushing each triple into a list under the corresponding key value using the joint grouping key as the key name. The final output is a grouping mapping table with the joint grouping key as the row identifier and the list of triples with the same key as the value. In this system, grouping aggregation merges redundant or conflicting triple knowledge fragments from different teaching data streams and different teaching stages into candidate knowledge units centered on knowledge point entity pairs, providing a structured data organization foundation for frequency statistics and optimization decisions. The candidate triple grouping result refers to the list of all triples corresponding to the same joint grouping key after the grouping aggregation operation. Each triple in this list has the same head entity and tail entity, but may carry different relation types, come from different data stream sources, and have different occurrence frequency records. The candidate triple grouping result serves as the basic processing unit for conflict resolution in this system. The system performs occurrence frequency statistics on the various relation types within the grouping result and determines the preferred relation to be retained based on the statistical results. The number of triple entries contained within it is positively correlated with the number of times the knowledge point entity is simultaneously identified in the multi-source data stream. The most frequent relation type refers to the relation category that appears most frequently after counting the relation type labels carried by each triple in the same candidate triple grouping results. For example, if a triple containing the predecessor relation type appears 12 times, the successor relation type appears 3 times, and the superior relation type appears 1 time in the grouping, then the predecessor relation is the most frequent relation type. This most frequent relation type is used as the sole basis for the conflict resolution rule in this system. The principle is that the more frequently the same pair of knowledge point entities are identified as having the same relation type in multi-source heterogeneous teaching data streams, the more reliable that relation type is in a statistical sense. By selecting the most frequent relation type, the statistically optimal result is selected from multiple conflict candidate relations without manual intervention.A deduplicated triple carrying a preferred relation type refers to a new triple generated by combining the preferred relation type with the head entity and tail entity in the corresponding joint grouping key after determining the preferred relation type through conflict resolution frequency statistics. This triple no longer retains information such as the source data stream identifier, original occurrence frequency, or other candidate relation types, but only retains three fields: head entity, tail entity, and preferred relation type. In this system, it serves as a clean and conflict-free knowledge unit extracted from multi-source redundant data and enters the entity name normalization process in a concise triple form. At the same time, this deduplicated triple achieves data compression and redundancy elimination of the original multiple knowledge triple sets, ultimately reducing potentially dozens of candidate triples to a single deterministic triple record. A standard nomenclature refers to a unique, standardized name string pre-assigned to each subject knowledge point entity in the core concept dictionary. This standard nomenclature is stored in the dictionary as an entry and is accompanied by a globally unique entity identifier code. In this system, it serves as the target reference value for entity name normalization. All entity names that are not standard nomenclatures must be converted to the standard nomenclature through dictionary mapping before being written into the final structured teaching knowledge graph. This ensures that the same knowledge point in the graph is not split into multiple different nodes due to differences in the expression of different data sources. Non-standard appellations refer to entity names extracted from multi-source heterogeneous teaching data streams through knowledge point entity extraction. These names are not entirely consistent with the standard appellation strings in the core concept dictionary of the subject, but can be mapped to the same entity identifier through equivalent alias mapping entries in the dictionary. Specifically, they include abbreviated forms of standard knowledge point appellations, such as abbreviating a quadratic equation to a quadratic equation, synonymous terms used in different textbook versions, and colloquial abbreviations used by teachers in classroom teaching transcripts. In this system, non-standard appellations are treated as objects to be replaced. During the normalization operation, they are identified and replaced with standard appellations, so that non-standard appellations only exist as intermediate results in the extraction stage and do not enter the final structured teaching knowledge graph. Entity name normalization refers to using the head and tail entity name strings of each deduplicated triple as input for precise matching in the subject's core concept dictionary. If the input entity name string is completely consistent with the standard term in the dictionary, the standard term is directly retained. If the input entity name string successfully matches an equivalent alias of an entry in the dictionary, the standard term corresponding to that entry is used as the replacement value. If the input entity name string does not match either the standard term or any equivalent alias in the dictionary, an exception handling process is triggered, and the triple is marked for manual review. In this system, entity name normalization serves as the final data processing step before generating a structured teaching knowledge graph. The output standardized triples are directly written into the final structured teaching knowledge graph, ensuring that all node names in the graph are consistent, searchable, and can be referenced across modules. This is a prerequisite for the normal execution of all operations using entity identifiers as keys in the teaching resource annotation index table.

[0025] In the embodiment of the structured teaching knowledge graph module, several sets of knowledge point triples are obtained. Each set of knowledge point triples is generated by extracting from a single multi-source heterogeneous teaching data stream. The head entity identifier and tail entity identifier in each triple are used as joint grouping keys for grouping and aggregation. The grouping results of candidate triples with the same head and tail entity combination are determined. This grouping and aggregation will group all triples corresponding to the same pair of head and tail entities from different teaching data streams into the same group, so that the knowledge association information extracted from the same pair of entities from different data sources can be presented in a centralized manner, providing a complete data view based on entity pairs for resolving relation type conflicts. For each candidate triplet grouping result, the frequency of occurrence of each relation type within the group is counted, and the relation type with the highest frequency is taken as the preferred relation type corresponding to the group. The deduplicated triplet carrying the preferred relation type is determined. This frequency statistics automatically select the optimal relation type from the distribution of different relation types extracted from the same pair of head and tail entities in multi-source data streams. This ensures that the relation types carried by the deduplicated triplet comprehensively reflect the common knowledge expression of most data sources, including textbook chapter texts, classroom teaching transcripts, student homework correction records, and stage test answers. This solves the conflict problem of the same pair of entities carrying multiple relation types caused by the heterogeneity of data sources. For each deduplicated triple carrying the preferred relation type, its head entity name and tail entity name are compared one by one with the standard names in the subject core concept dictionary. Non-standard names are replaced with the standard names corresponding to the dictionary. The structured teaching knowledge graph triple with normalized entity names is determined. This name replacement operation unifies the entity names in all triples written into the structured teaching knowledge graph to the standard form in the subject core concept dictionary. This eliminates the problem of inconsistent expression of the same knowledge point in different data streams due to differences in textbook versions and teachers' teaching styles. It ensures that each entity node in the final generated structured teaching knowledge graph has a consistent naming standard.

[0026] The association similarity module 4 is used to perform knowledge point annotation and association retrieval on candidate teaching resource data based on the structured teaching knowledge graph, and to determine the teaching resource annotation index table. The teaching resource annotation index table records the knowledge graph entity identifier and association confidence score associated with each candidate teaching resource. The association confidence score is calculated by weighting the edit distance similarity between the resource text and the entity description text and the keyword co-occurrence frequency.

[0027] Specifically, the pre-defined knowledge point annotation and association retrieval refers to the process of comparing and matching the text content of candidate teaching resources with the subject knowledge point entities in the graph one by one, using the already constructed structured teaching knowledge graph as a reference benchmark. This process determines which knowledge point entity(s) in the graph each candidate teaching resource is specifically associated with and the corresponding matching strength. This step in the system plays a role in bridging the static knowledge graph with the dynamically input teaching resources across data sources. It enables a wider range of teaching resources beyond the originally isolated textbook chapter texts, classroom teaching transcripts, student homework correction records, and stage test results to be incorporated into a unified knowledge system framework for management and retrieval. This is a key transitional operation for the knowledge graph from the construction stage to the practical application stage. The teaching resource annotation index table is a collection of data records organized and persistently stored in a structured table format. Each record in the table contains at least three fields: a unique resource identifier for the candidate teaching resource, the associated knowledge graph entity identifier, and the association confidence score between the two. In the system, it serves as an intermediate mapping layer connecting the original teaching resources and the structured knowledge graph. This allows the system to quickly locate all associated candidate teaching resources when performing a query operation for any knowledge graph entity identifier, without having to re-scan the full text of all resources and calculate similarity for each query. This reduces the time complexity of resource retrieval from the level of the total number of resources multiplied by the total number of entities to a constant level operation for a single index table query. Candidate teaching resources refer to teaching material data that has not yet been labeled or incorporated into the existing knowledge graph association system. Specifically, they include, but are not limited to, question bank test texts, extracurricular reading materials, micro-lecture video subtitles, teacher lesson preparation documents, and student error notebook summaries. This type of data is used as input objects to be labeled and associated in the system. After knowledge point labeling and association retrieval processing, it is tagged with knowledge graph entity identifiers, thereby transforming it from unstructured materials outside the knowledge system into indexed teaching assets that can be directly called by the knowledge graph retrieval and reasoning engine. A knowledge graph entity identifier is a globally unique and constant coded mark assigned to each subject knowledge point entity in a structured teaching knowledge graph. Its specific forms include numeric identifiers composed of subject codes, chapter numbers, and concept sequence numbers, or string identifiers generated using universally unique identifier standards. This identifier serves as a unique identity credential for knowledge points in all storage structures, index tables, and query operations within the system. This enables knowledge point entities under the same standard name in the knowledge graph to be referenced and associated by various modules within the system using a unified key value, avoiding cross-module knowledge reference breaks or query failures due to textbook version updates or terminology adjustments.The association confidence score is a numerical measure used to quantify the reliability of the match between each candidate teaching resource and its associated knowledge graph entity identifier. Its value is a floating-point number between 0 and 1; the closer the score is to 1, the more certain the association between the resource and the entity. This score serves as the core basis for resource retrieval result ranking and filtering in the system. When performing a resource query for a specific knowledge graph entity identifier, the system only returns candidate resource records with an association confidence score higher than a preset threshold, such as 0.7, thus effectively excluding low-quality mismatch results caused by accidental keyword co-occurrence while ensuring recall. Resource text refers to the complete text content obtained after format parsing and plain text extraction of candidate teaching resources. Specifically, it includes the full text of test questions, the main text of reading materials, sentence-by-sentence sequences of video subtitles, or descriptive text paragraphs in lesson plans. This text serves as the common input material for two sub-operations in the system: edit distance similarity calculation and keyword co-occurrence statistics. Its text length and quality directly affect the accuracy of the association confidence score calculation. Before calculation, it undergoes unified word segmentation and stop word filtering preprocessing. Entity description text refers to the definitional text of knowledge points stored in the structured teaching knowledge graph and bound to the entity identifier of the knowledge graph. Specifically, it includes the standard definition of the knowledge point, the description of typical examples, and the target description in the teaching syllabus. This text serves as a fixed reference template in the system for comparing similarity with the resource text of candidate teaching resources. Before the system is deployed, it is uniformly compiled by subject teaching and research personnel based on the teaching syllabus and stored together with the knowledge graph. The standardization and completeness of its expression directly determine the distinguishability of the edit distance similarity calculation results. Edit distance similarity refers to the similarity measure between 0 and 1 obtained by calculating the minimum number of character editing operations (insertion, deletion, and replacement) between resource text and entity description text, and converting this number after length normalization. Specifically, it is calculated by dividing the Levenstein distance between two text strings by the total character length of the longer of the two strings and subtracting the ratio from 1 to obtain the similarity score. In the system, edit distance similarity reflects the text similarity between candidate resources and knowledge point entities from the degree of accurate matching at the character level. It has a high ability to distinguish and recognize test texts containing standard definitions or professional terminology, but relying solely on character edit distance cannot capture large-scale co-occurrence information at the lexical level.Keywords refer to a set of characteristic terms with subject-specific distinguishing capabilities extracted from the core concept dictionary of the subject and the entity description text of the structured teaching knowledge graph. Specifically, it includes the core nouns in the standard name of each knowledge point entity, the high-frequency subject terms in the entity description text, and a list of key subject terms preset by teaching and research personnel. This set of keywords serves as the retrieval target set for word frequency statistics of candidate teaching resources in the system. The keyword co-occurrence frequency statistics and edit distance similarity are weighted and summed by preset weight coefficients to obtain the association confidence score. Specifically, the weighting ratio is 60% for edit distance similarity and 40% for the normalized value of keyword co-occurrence frequency. This ensures that the association confidence score takes into account both the precise matching at the character level and the semantic coverage at the vocabulary level, avoiding the problem of the single edit distance indicator being insensitive to synonym substitution and the problem of the single keyword statistics being sensitive to the sparsity of short texts. This improves the accuracy and robustness of knowledge point annotation and association retrieval as a whole.

[0028] As a preferred approach: the association similarity module 4 includes: The plain text extraction submodule is used to obtain a structured teaching knowledge graph, which contains several entity identifiers and corresponding entity description texts; obtain candidate teaching resource data; and extract the resource text of each candidate teaching resource into a plain text string. The term co-occurrence frequency submodule is used to calculate the edit distance similarity between the text of the resource and the description text of the entity for each candidate teaching resource and each entity identifier; and to count the co-occurrence frequency of terms belonging to the keyword set of the entity in the resource text. The preset weighted summation submodule is used to sum the edit distance similarity and keyword co-occurrence frequency according to preset weights to calculate the association confidence score between the candidate teaching resource and the entity identifier. The correspondence writing submodule is used to write the correspondence between candidate teaching resources with association confidence scores higher than a preset threshold and entity identifiers into the index table; each record in the index table contains resource identifier, entity identifier and corresponding association confidence score.

[0029] Specifically, a plain text string refers to a continuous sequence of characters extracted from the original files of candidate teaching resources after format parsing, which contains no typesetting marks, style tags, embedded objects, or non-text characters. The specific processing methods include removing headers, footers, and image objects from PDF documents to extract the main text; stripping timestamps from video subtitle files to retain the plain text content; and extracting the question stems and option fields from question bank Excel files and concatenating them into text paragraphs. This plain text string serves as the common input basis for edit distance similarity calculation and keyword co-occurrence frequency statistics in the system, unifying the text content from teaching resources of different formats into a single calculable data form and eliminating the incommensurability of comparisons caused by file format differences. Edit distance similarity refers to the similarity measure between 0 and 1, obtained by normalizing the minimum number of character edit operations required to convert the plain text string of a candidate teaching resource to the text string describing the entity in the knowledge graph. This minimum number of character edit operations is calculated using a dynamic programming algorithm to determine the minimum total number of insertion, deletion, and replacement operations required to convert between the two strings. The normalized edit distance is obtained by dividing this total number by the total character length of the longer of the two strings. Subtracting this normalized edit distance from 1 yields the edit distance similarity. In the system, edit distance similarity reflects the textual similarity between the candidate teaching resource and the knowledge entity at the character level, and it has a high ability to distinguish and identify resources containing standard definitions or normative terminology. Co-occurrence frequency (COF) refers to the total number of times a term belonging to the pre-set keyword set of the knowledge point entity appears in the resulting sequence of terms after word segmentation and stop word filtering of the plain text string of candidate teaching resources. This pre-set keyword set consists of the standard name of the entity and its equivalent aliases in the core concept dictionary of the subject, high-discrimination subject terms selected by word frequency statistics in the entity description text, and subject key words pre-set by teaching and research personnel. COF complements edit distance similarity in the system—the former assesses similarity from the dimension of character-level precise matching, while the latter assesses relevance from the dimension of lexical-level semantic coverage. The combined use of the two can avoid the problem that the single edit distance indicator is not sensitive to synonym substitution and the problem that the single keyword statistics are too sensitive to short or sparse text. The preset weighted summation refers to reading the preset edit distance similarity weight coefficient and keyword co-occurrence frequency weight coefficient from the system configuration parameters, multiplying the edit distance similarity by its corresponding weight coefficient to obtain the first weighted value, and multiplying the keyword co-occurrence frequency by its corresponding weight coefficient after maximum frequency normalization to obtain the second weighted value. The first weighted value and the second weighted value are then added together to obtain the association confidence score. This preset weighted summation integrates two similarity measures with different dimensions and properties into a single comprehensive confidence index in the system. The default weight coefficient for edit distance similarity is set to 0.6, and the default weight coefficient for keyword co-occurrence frequency is set to 0.4.Candidate teaching resources and entity identifiers refer to the basic correspondence units processed by the system when performing knowledge point annotation and association retrieval operations. One end is the unique resource identifier of the candidate teaching resource and its corresponding plain text string, and the other end is the entity identifier of a certain knowledge graph in the structured teaching knowledge graph and its corresponding entity description text and preset keyword set. The candidate teaching resources and entity identifiers are used as input pairs for the association confidence score calculation operation in the system. The system iterates through the Cartesian product of all candidate resources and all entity identifiers to calculate the association confidence score of each pair, and writes the correspondence with scores higher than the preset threshold into the teaching resource annotation index table. Finally, each candidate teaching resource establishes a traceable semantic association with the knowledge graph through the entity identifier recorded in the index table.

[0030] In the embodiment of the association similarity module, a structured teaching knowledge graph is obtained, which contains several entity identifiers and corresponding entity description texts. Simultaneously, candidate teaching resource data is acquired, and the resource text of each candidate teaching resource is extracted into a plain text string. This plain text string removes formatting marks and non-text characters from the original file, unifying the text content from teaching resources of different formats into a single computable data form and eliminating interference from file format differences on text comparison operations. For each candidate teaching resource and each entity identifier, the edit distance similarity between the resource text and the entity description text is calculated. This edit distance similarity quantifies the textual similarity between the resource text and the knowledge point definition text from a character-level precise matching dimension. Simultaneously, the co-occurrence frequency of terms belonging to the entity's keyword set appearing in the resource text is counted. This term co-occurrence frequency quantifies the overlap between the resource text and the knowledge point terminology system from a lexical-level semantic coverage dimension. These two features characterize the association strength between candidate teaching resources and knowledge graph entities at different granularities. Edit distance similarity and keyword co-occurrence frequency are weighted and summed according to preset weights to calculate the association confidence score between the candidate teaching resource and the entity identifier. This association confidence score includes relevance information at both the character and word levels, enabling a single comprehensive indicator to reflect two different types of text similarity and avoiding discrimination bias caused by synonym substitution or text sparsity due to a single feature. The correspondence between candidate teaching resources and entity identifiers with association confidence scores higher than a preset threshold is written into an index table. Each record in this index table contains a resource identifier, an entity identifier, and the corresponding association confidence score. This threshold filtering operation eliminates low-confidence pairings, ensuring that all records written into the index table have undergone confidence verification. The index table persistently stores valid association records in a structured format, allowing queries for any entity identifier to be directly located via key-value pairs in this index table without repeatedly performing confidence calculations and threshold comparisons.

[0031] In one possible implementation: the submodule for counting term co-occurrence includes: The edit distance similarity score unit is used to obtain the resource text and entity description text corresponding to the entity identifier of the candidate teaching resources; calculate the minimum edit distance between the resource text and the entity description text; and determine the edit distance similarity score based on the minimum edit distance and the normalization of the two text lengths. The keyword co-occurrence frequency unit is determined to obtain the preset keyword set corresponding to the entity identifier; and the resource text is processed by word segmentation and stop word filtering. The number of times each word is hit in the keyword set after filtering is counted to determine the keyword co-occurrence frequency. The weighted summation unit is used to take the edit distance similarity score and keyword co-occurrence frequency as input features; and sum them according to the preset weight coefficients to obtain the association confidence score between the candidate teaching resource and the entity identifier.

[0032] Specifically, entity description text refers to the definitional description of knowledge points stored in the structured teaching knowledge graph and bound to the entity identifier of the knowledge graph. It includes the standard definition of the knowledge point, the description of typical examples, the target description in the teaching syllabus, and the contextual summary paragraph of the knowledge point in the textbook chapter. In the system, this entity description text serves as a fixed reference template for calculating the edit distance similarity with the resource text of candidate teaching resources. Its content is uniformly compiled by subject teaching and research personnel based on the teaching syllabus and textbook before the system is deployed and stored together with the knowledge graph. The standardization and length consistency of its expression directly determine the comparability of the edit distance similarity calculation results between different entities. Minimum edit distance refers to the minimum number of single-character editing operations required to convert the resource text of a candidate teaching resource into entity description text. These single-character editing operations specifically include three basic operations: inserting a character, deleting a character, and replacing a character. The minimum edit distance is calculated using the Levenstein distance dynamic programming algorithm. In this system, the minimum edit distance serves as a raw quantitative indicator to measure the degree of difference between two texts. The smaller the value, the more similar the two texts are. However, since this raw value is greatly affected by the absolute length of the two texts, it needs to be normalized before it can be used as a similarity score that can be compared across entities. The two text lengths refer to the total number of characters in the resource text of the candidate teaching resource and the total number of characters in the entity description text. These two text lengths are used in the system for length normalization of the minimum edit distance. Specifically, the normalization method involves using the larger of the two text lengths as the denominator, dividing the minimum edit distance by the larger value to obtain the normalized edit distance, and then subtracting the normalized edit distance from 1 to obtain the edit distance similarity score. This length normalization process eliminates the problem of incomparable minimum edit distance values ​​caused by the inherent length differences between entity description texts, ensuring that the similarity scores between different entities have a unified numerical range and comparability. The edit distance similarity score is a floating-point number between 0 and 1 obtained after length normalization. The closer the score is to 1, the more similar the resource text and entity description text are at the character level. This edit distance similarity score serves as the first input feature in the system for calculating the association confidence score, reflecting the degree of text similarity between candidate teaching resources and knowledge point entities from the perspective of precise character-level matching.The preset keyword set refers to a pre-constructed set composed of a number of feature terms with subject discrimination capability for each knowledge graph entity identifier in the structured teaching knowledge graph. The specific construction method of the preset keyword set is as follows: taking the standard appellation corresponding to the entity identifier and all its equivalent aliases in the subject core concept dictionary as seed words, then screening out the top 10 subject terms ranked by term frequency-inverse document frequency value from the entity description text corresponding to the entity identifier via the term frequency-inverse document frequency algorithm as expansion terms, and finally adding the relevant terms in the subject key word list preset by teaching and research personnel for each subject. The three parts are merged and deduplicated to form the preset keyword set of the entity, and this preset keyword set serves as the target matching set when performing term co-occurrence frequency statistics on the resource text of candidate teaching resources in the system. The resource text refers to the complete text content of the candidate teaching resource in the form of a plain text string, which specifically includes the full text of test question stems, the main body of reading materials, the sentence-by-sentence sequence of micro-course video subtitles or the descriptive text paragraphs of teaching plans. This resource text acts as the original input text for word segmentation processing and stop word filtering in the system, and is also one of the input texts for edit distance similarity calculation. The same piece of resource text participates in similarity calculation of two different dimensions at the same time, and its text length and quality directly affect the output results of the two sub-modules. Word segmentation processing and stop word filtering refer to a sequence of text preprocessing operations performed on the resource text in sequence. Word segmentation processing is an automated process of segmenting the continuous character sequence of the resource text into independent terms according to linguistic boundaries. Stop word filtering refers to the operation of deleting the terms belonging to the preset stop word list from the obtained term list and retaining the remaining terms after word segmentation processing. This preset stop word list at least includes high-frequency function words with no subject discrimination capability in Chinese such as "de", "le", "zai", "shi", and general question number characters in teaching scenarios such as "di", "ti", "xiaoti". The word segmentation processing and stop word filtering in the system convert the original resource text into a filtered term list containing only content terms, removing noise terms that contribute nothing to semantic discrimination, so that hit count statistics are focused on core vocabularies with subject identification capability. The hit count refers to performing equal-value matching comparison between each term in the filtered term list and each keyword in the preset keyword set, and recording the total number of successful matches. This hit count serves as the original count value of the keyword co-occurrence frequency in the system, and its value depends on the richness of subject-related content words in the resource text and the completeness of the preset keyword set. The keyword co-occurrence frequency refers to a normalized value within the range from 0 to 1 obtained by subjecting the above-mentioned hit count to maximum frequency normalization processing. The specific method of the maximum frequency normalization processing is to take the maximum hit count among all candidate teaching resources under the current entity identifier as the denominator, and divide the hit count of the current resource text by this maximum hit count to obtain the normalized keyword co-occurrence frequency. This keyword co-occurrence frequency serves as the second input feature for calculating the correlation confidence score in the system, and reflects the correlation between the candidate teaching resource and the knowledge point entity from the lexical-level semantic coverage dimension.Input features refer to the set of independent variables that are fed into the weighted summation operation as the calculation object. Specifically, they include two variables: edit distance similarity score and keyword co-occurrence frequency. In the system, these input features participate in the weighted summation calculation according to preset weight coefficients in the form of vectors. The two features characterize the association strength between candidate teaching resources and knowledge point entities from two different dimensions: precise character matching and semantic coverage of words. The two complement each other and together form the complete calculation basis for the association confidence score.

[0033] In the embodiment of the term co-occurrence frequency submodule, the resource text of the candidate teaching resources and the entity description text corresponding to the entity identifiers are obtained. The minimum edit distance between the resource text and the entity description text is calculated, and the edit distance similarity score is determined by normalization based on the minimum edit distance and the length of the two texts. This edit distance similarity score measures the degree of precise matching between the two texts at the character-level operation cost, quantifying the text differences into normalized values ​​to make the similarity of texts of different lengths comparable. A preset keyword set corresponding to the entity identifiers is obtained, and word segmentation and stop word filtering are performed on the resource text. The number of hits of each term in the keyword set after filtering is counted to determine the keyword co-occurrence frequency. This keyword co-occurrence frequency measures the degree of semantic overlap between the resource text and the knowledge point entities at the terminology level by the word-level hit count. The coverage relationship between the subject-distinguishing real words in the resource text and the preset keywords of the entities is quantified into a value. This value and the edit distance similarity score characterize the association strength between the candidate teaching resources and the knowledge point entities from two dimensions: precise character matching and word semantic coverage. Using edit distance similarity score and keyword co-occurrence frequency as input features, a weighted sum is performed according to preset weight coefficients to obtain the association confidence score between the candidate teaching resource and the entity identifier. This association confidence score simultaneously includes character-level precise matching information and lexical-level semantic coverage information, enabling a single value to comprehensively reflect the text relevance of two different dimensions. This avoids the discrimination bias caused by single features due to synonym substitution or the sparsity of short texts, and provides a comprehensive measurement basis that takes into account both accuracy and coverage for threshold screening.

[0034] In another possible implementation: a pre-defined weighted summation submodule includes: Read the weight coefficient unit to obtain the edit distance similarity score and keyword co-occurrence frequency for the same candidate resource and the same entity; read the similarity weight coefficient and frequency weight coefficient corresponding to the entity from the system configuration parameters; The second weighting unit is used to multiply the similarity score by the similarity weight coefficient to obtain the first weighting value; and to multiply the frequency by the frequency weight coefficient to obtain the second weighting value. A correlation confidence unit is determined, which is used to sum the first weighted value and the second weighted value; and the summation result is determined as the correlation confidence score between the candidate resource and the entity.

[0035] Specifically: the same candidate resource refers to a single teaching material instance with a unique resource identifier after resource text extraction. This material instance serves as the basic object entity for association calculation in the system, specifically corresponding to a test question document, a micro-lecture video subtitle file, or a reading material, etc.; the same entity refers to a single subject knowledge point node instance with a unique knowledge graph entity identifier in the structured teaching knowledge graph, corresponding to specific nameable knowledge units such as functions and quadratic equations; the same candidate resource and the same entity serve as the input pair for each weighted summation calculation operation in this step. The system performs a weighted fusion calculation on each input pair, reading the edit distance similarity score and the keyword co-occurrence frequency, to ensure that each calculation only targets the correspondence between a single resource and a single knowledge point, avoiding the feature tensor confusion problem caused by batch calculation. The edit distance similarity score refers to the normalized character-level similarity value between the resource text of the same candidate resource and the entity description text of the same entity, calculated in the previous step, with a value range between 0 and 1. The keyword co-occurrence frequency refers to the number of hits obtained by matching the resource text of the same candidate resource with a preset keyword set after word segmentation and stop word filtering, as calculated in the previous step, after normalization to the maximum frequency, with a value range between 0 and 1. The edit distance similarity score and the keyword co-occurrence frequency are read from the cache or storage structure in this step as two input feature variables for weighted summation calculation. The two are fused within the same calculation unit. This step does not involve recalculation or correction of the two, but only performs reading and operation operations. System configuration parameters refer to a set of preset numerical parameters stored in the system configuration file or database configuration table in the form of key-value pairs. This set stores the corresponding similarity weight coefficient and frequency weight coefficient for each entity identifier in the knowledge graph. Specifically, it is stored as a mapping table of entity identifier, similarity weight coefficient, and frequency weight coefficient. The frequency weight coefficient is another weight parameter that exists in pair with the edit distance similarity weight coefficient; their sum is always equal to 1. The default value is 0.6 for the similarity weight coefficient and 0.4 for the frequency weight coefficient, and this value can be configured differently for different disciplines and entity types. These two weight coefficients are used in the system to explicitly control the relative importance of two input features. By adjusting their numerical distribution, the association confidence score can be weighted according to the actual scenario requirements between the character exact matching dimension and the vocabulary semantic coverage dimension.The first weighted value is the product of the edit distance similarity score and the similarity weight coefficient; the second weighted value is the product of the keyword co-occurrence frequency and the frequency weight coefficient. The first and second weighted values ​​are calculated and temporarily stored separately as two independent weighted intermediate results in this system. They respectively carry the contribution components of the two input features after weight scaling, so that the two similarity measures of different dimensions can be added under a unified numerical scale, providing input with the same dimension for the final summation operation. The summation of the first and second weighted values ​​involves performing arithmetic addition on the two intermediate weighted results. The resulting sum is directly used as the confidence score for the association between the candidate resource and the entity. The calculation formula is: Confidence = Similarity Score × Similarity Weight + Frequency × Frequency Weight. This summation result serves as a single comprehensive confidence index in the system, used to quantify the overall association strength between the candidate teaching resource and the knowledge point entity. This comprehensive confidence score integrates the text similarity at the character level and the subject semantic coverage at the vocabulary level. Compared with a single-dimensional similarity measure, it has higher discrimination accuracy and robustness, and is the basis for comparison and screening with a preset threshold when writing into the teaching resource annotation index table.

[0036] In an embodiment of the pre-weighted summation submodule, the edit distance similarity score and keyword co-occurrence frequency of the same candidate resource are obtained. The similarity weight coefficient and frequency weight coefficient of the entity corresponding to the candidate resource are read from the system configuration parameters, allowing each of the two features of different dimensions to obtain an independent scaling factor, providing a numerical adjustment basis under the same source and scale for the weighted calculation. The edit distance similarity score and the similarity weight coefficient are multiplied to obtain a first weighted value, and the keyword co-occurrence frequency and the frequency weight coefficient are multiplied to obtain a second weighted value. These first and second weighted values ​​transform the originally different dimensions of character-level similarity and word-level co-occurrence frequency into summable intermediate values, and allow the contribution of each feature to be independently controlled by its respective weight coefficient. The first weighted value and the second weighted value are arithmetically added together, and the sum is determined as the association confidence score between the candidate resource and the entity. This association confidence score includes both the character-level textual approximation information contributed by the edit distance similarity and the lexical-level semantic coverage information contributed by the keyword co-occurrence frequency. This allows a single value to comprehensively reflect two different types of similarity measures, avoiding the deviation of confidence scores caused by the sensitivity or sparsity of any single-dimensional feature.

[0037] In another possible implementation: the correspondence is written into the submodule, including: Read the confidence threshold unit to obtain a list of full association confidence scores between the same candidate resource and each entity; read the preset confidence admission threshold from the system configuration parameters; The candidate association pair filtering unit is used to compare each score in the full association confidence score list with the admission threshold; and to filter out candidate association pairs with scores higher than the admission threshold. The confidence score extraction unit is used to extract the resource identifier, entity identifier, and corresponding confidence score from each selected candidate association pair; and write the resource identifier, entity identifier, and confidence score as a record into the index table.

[0038] Specifically, the full association confidence score list refers to a numerical sequence composed of all association confidence scores obtained by calculating one by one the entity identifiers in the structured teaching knowledge graph for the same candidate teaching resource. Each score in this numerical sequence corresponds to a different knowledge graph entity identifier. Its specific organization form is a key-value pair array or mapping table of entity identifiers and confidence scores. This full association confidence score list serves as the input data source for the threshold screening step in this system. It completely covers the association strength information between the candidate teaching resource and every knowledge point in the knowledge graph, providing the entire candidate set for the screening operation without omitting any potential association relationships. System configuration parameters refer to a set of adjustable parameters stored in system configuration files, database configuration tables, or runtime environment variables. This set of parameters must contain at least one preset confidence threshold, which is a predefined floating-point constant between 0 and 1, with a default value of 0.7. This threshold can be independently configured and dynamically adjusted for different subjects, resource types, and teaching scenarios. In this system, the system configuration parameters serve as the source of numerical standards for threshold filtering rules, making the filtering operation configurable and environmentally adaptable, and avoiding the difficulty of modification caused by hardcoding the threshold into the program logic. Candidate association pairs refer to resource-entity pairings selected from the full association confidence score list after threshold comparison, where the association confidence score is higher than a pre-set confidence threshold. Each candidate association pair contains a unique resource identifier for a candidate teaching resource and a knowledge graph entity identifier along with its corresponding confidence score. In this system, these candidate association pairs serve as the final data candidate set written into the teaching resource annotation index table. The size of this set is determined by the threshold value; a higher threshold results in fewer selected association pairs but higher average quality, while a lower threshold results in more association pairs but may introduce more low-quality matches. A resource identifier is a unique and constant coded marker assigned to each candidate teaching resource data entry. Specifically, it is a string identifier consisting of a resource type prefix, an entry timestamp, and a random sequence number. In this system, the resource identifier serves as one of the primary key fields of the index table, enabling each index table record to be uniquely traced back to its corresponding original teaching resource file or data entry. An entity identifier is a globally unique and constant coded mark assigned to each subject knowledge point entity in a structured teaching knowledge graph. Its specific forms include numeric identifiers composed of subject codes, chapter numbers, and concept sequence numbers, or string identifiers generated by universally unique identifier standards. In this system, the entity identifier serves as another primary key field of the index table, establishing a precise one-to-one correspondence between teaching resources and specific knowledge point nodes in the knowledge graph.The corresponding confidence score refers to the uniquely bound combination of the resource identifier and entity identifier, retained after threshold filtering. Its value falls within the range of the admission threshold and 1.0. This confidence score serves as a crucial component of the index table record in the system. During resource retrieval operations, it sorts multiple candidate teaching resource records under the same entity identifier from highest to lowest score, prioritizing the return of the teaching resource with the strongest association with the target knowledge point. The system also supports dynamic filtering based on minimum confidence requirements set for different application scenarios. The resource identifier, entity identifier, and corresponding confidence score together form the standard three-field structure of each record in the teaching resource annotation index table. This index table acts as an intermediate mapping layer connecting candidate teaching resources and the structured teaching knowledge graph. When performing a query operation targeting any knowledge graph entity identifier, the system can directly locate all associated candidate teaching resources through this index table, reducing the time complexity of resource retrieval from the level of multiplying the total number of resources by the total number of entities to a constant-level operation of a single index table key-value query.

[0039] In the embodiment of the correspondence writing submodule, a full list of association confidence scores between the same candidate resource and each entity in the structured teaching knowledge graph is obtained. This full list provides a complete numerical sequence of association strengths between the candidate resource and all entities in the knowledge graph, ensuring that the filtering operation has a complete candidate set and avoiding omissions due to premature truncation or sampling. Simultaneously, a preset confidence threshold is read from the system configuration parameters. This threshold, as an adjustable numerical standard, provides configuration flexibility for the filtering rules, preventing the judgment boundary from being fixed in the program logic. Each score in the full association confidence score list is compared with the threshold. The Boolean judgment result output by the comparison operation is directly used to filter out candidate association pairs with scores higher than the threshold. This set of candidate association pairs is smaller in number than the full list, eliminating low-confidence pairings below the threshold. This ensures that the records written to the index table only retain reliable associations verified by the threshold, reducing the proportion of low-quality matching entries in the index table from the data source. For each candidate association pair selected, the resource identifier, entity identifier, and corresponding confidence score are extracted from the association pair. The resource identifier, entity identifier, and confidence score are written as a record into the index table. The index table is stored in the form of a structured table, so that each valid association is persisted in the format of a ternary field. The query operation for any entity identifier can use the index table as the direct data source for key-value location without re-executing the full confidence calculation and threshold comparison process. The computational overhead of the query response stage is transferred to the index building stage.

[0040] Module 5 is designed to build a retrieval engine, which utilizes teaching resources to annotate the index table and constructs a knowledge graph retrieval and reasoning engine for the entire intelligent teaching process.

[0041] Specifically, the teaching resource annotation index table is a persistent collection of data records in a structured table format. Each record in this table contains at least three fields: a unique resource identifier for the candidate teaching resource, the associated knowledge graph entity identifier, and the association confidence score between the two. In the system, it serves as an intermediate mapping layer connecting the original candidate teaching resources and the structured teaching knowledge graph. This allows the system to directly locate all associated candidate teaching resources when performing a query operation for any knowledge graph entity identifier, without having to recalculate the full text of all resources and the distance similarity and keyword co-occurrence frequency for each query. This reduces the time complexity of resource retrieval from the level of the total number of resources multiplied by the total number of entities to the constant level of a single index table key-value query. The index table is incrementally updated during system operation as new candidate teaching resources are input, and maintains consistency with the current structured teaching knowledge graph through full reconstruction or difference merging when the system restarts or the graph version is upgraded. The intelligent teaching full-process knowledge graph retrieval refers to an automated process that uses a teaching resource annotation index table as the main data access medium and a structured teaching knowledge graph as the logical reasoning basis. Responding to externally inputted query requests in the form of knowledge graph entity identifiers or natural language descriptions, it retrieves all candidate teaching resource records with a correlation confidence score higher than a preset threshold (e.g., 0.7) from the teaching resource annotation index table and outputs the retrieval results after sorting them from highest to lowest score. This retrieval operation in the system transforms the node and edge structure of the knowledge graph into a list of specific teaching materials that can be directly used in teaching scenarios. This allows teachers to instantly obtain all indexed teaching assets related to that knowledge point, such as test questions, lesson plans, reading materials, and video subtitles, after inputting a chapter name or a concept / term. Furthermore, this retrieval operation supports compound query conditions, meaning the system allows the simultaneous specification of multiple knowledge graph entity identifiers for union or intersection retrieval, meeting the practical needs of cross-knowledge point comprehensive review or chapter / unit-based lesson preparation in teaching.The intelligent teaching full-process knowledge graph retrieval and reasoning engine is a software system module that integrates the rapid retrieval capability of the teaching resource annotation index table with the semantic reasoning capability of the structured teaching knowledge graph into the same operating environment and provides a unified external call interface. It undertakes a functional upgrade from passive querying to active reasoning within the system. Specifically, when the externally input query request contains only the entity identifier of a specific subject knowledge point, the engine first directly retrieves all candidate teaching resources associated with that entity identifier from the teaching resource annotation index table. Simultaneously, based on the predecessor and successor relationships stored in the structured teaching knowledge graph, it performs a one-hop or multi-hop traversal along the directed edge direction. The traversed predecessor and successor knowledge point entity identifiers are used as extended query keys to perform another retrieval from the teaching resource annotation index table. The direct query results and extended query results are merged, deduplicated, and output together. The upper limit of the number of hops for this multi-hop reasoning is set by the system configuration parameters. The default value is one hop, i.e., only including direct predecessors and direct successors. When the system configuration parameters are adjusted to two hops... After completing a one-hop expansion, the engine continues to expand outward along the graph edges to second-level neighbor nodes and performs corresponding retrievals. The engine also supports a reverse reasoning query mode, that is, when the external input query request contains a unique resource identifier of a candidate teaching resource, the engine retrieves all knowledge graph entity identifiers associated with the resource and their corresponding confidence scores from the teaching resource annotation index table, thereby supporting the reverse query requirement of tracing the knowledge point system to which any teaching material belongs. In the overall system architecture, the engine serves as the sole entry and exit point for all knowledge graph-related operations for end users and upper-level teaching applications. It completely shields the internal implementation details such as the underlying teaching resource annotation index table storage structure, the structured teaching knowledge graph database, and the preset conflict resolution rules from external callers, enabling specific teaching application functions such as personalized learning path recommendation, intelligent test paper generation, learning diagnosis, and lesson preparation assistance to be independently developed and iterated based on a unified retrieval and reasoning interface, without having to repeatedly implement the knowledge graph traversal logic and resource index query logic.

[0042] In another possible implementation: construct a retrieval engine module 5, including: A submodule for establishing a positive mapping relationship is used to obtain the teaching resource annotation index table and the structured teaching knowledge graph; using the resource identifier and entity identifier in the index table as key-value pairs, a positive mapping relationship is established from resources to entities in the knowledge graph. The extended associated entity identifier set submodule is used to determine the extended associated entity identifier set of each resource by performing a one-hop extended traversal along the graph edges based on the forward mapping relationship and the predecessor and successor edges between entities stored in the knowledge graph. The module for generating a retrieval mapping table is used to merge the forward mapping relationship with the extended associated entity identifier set to generate a unified retrieval mapping table containing direct and indirect associations. Based on the unified retrieval mapping table and the graph traversal interface of the knowledge graph, a query engine that supports both key-value retrieval and relation reasoning is encapsulated.

[0043] Specifically, the teaching resource annotation index table and the structured teaching knowledge graph refer to two completed and persistently stored data structures. The teaching resource annotation index table focuses on the correspondence between resource identifiers and entity identifiers, while the structured teaching knowledge graph focuses on a graph structure where nodes represent subject knowledge point entities and directed edges represent predecessor-successor and hierarchical relationships between entities. These two serve as the two fundamental data sources required to build the query engine in this system. The former provides direct association information from resources to knowledge points, while the latter provides logical relationship information between knowledge points. They complement each other, providing complete data support for establishing and expanding the forward mapping relationship. A key-value pair is a data mapping unit consisting of the entity identifier of each record in the teaching resource annotation index table as the key and the corresponding resource identifier as the value. This key-value pair is used in this system to establish a forward mapping relationship from resources to entities in the knowledge graph. Its specific implementation is a hash mapping table with entity identifiers as keys, where each key corresponds to a set of one or more resource identifiers. Predecessor-successor edges between entities refer to directed edges connecting two subject knowledge point entity nodes in a structured teaching knowledge graph. The direction of these directed edges points from the predecessor knowledge point entity to the successor knowledge point entity, representing a dependency relationship between them based on the order of learning. These edges serve as the path for graph traversal operations in the system, allowing the system to move from the current entity node to an adjacent entity node along the direction of the directed edge to expand knowledge associations. Forward mapping relationships refer to the mapping from knowledge graph entity identifiers to the set of associated candidate teaching resource identifiers. Specifically, this is implemented as a hash mapping table with entity identifiers as keys and a list of resource identifiers as values. These forward mapping relationships serve as the basis for direct resource retrieval operations in the system. When a query request specifies a particular entity identifier, the system directly returns all candidate teaching resources directly associated with that entity through these forward mapping relationships. Merging the forward mapping relationship and the extended associated entity identifier set refers to performing a union operation on the direct resource retrieval results and the results obtained through graph extension traversal. This results in each entity identifier in the final generated retrieval mapping table being associated not only with the resource identifiers directly labeled through the index table, but also with the indirectly associated resource identifiers obtained through predecessor and successor edge extension traversal. The merging of this forward mapping relationship and the extended associated entity identifier set generates a unified retrieval mapping table that includes both direct and indirect associations.Direct association refers to the direct correspondence between candidate teaching resources and knowledge point entities that are explicitly recorded in the teaching resource labeling index table and written after being calculated and filtered by association confidence scores. Indirect association refers to the fact that the candidate teaching resource itself is not directly labeled to a certain knowledge graph entity, but there is a predecessor and successor relationship between the knowledge graph entity directly associated with the candidate teaching resource and the target entity currently being queried. Thus, the derived association is formed through graph edge traversal. Direct and indirect associations together expand the coverage of single resource retrieval in this system, enabling the system to return not only resources directly related to the knowledge point when performing knowledge point queries, but also resources related to the predecessor or successor knowledge points of the knowledge point, meeting the resource acquisition needs for knowledge pre-laying and extension in teaching scenarios. The graph traversal interface refers to the standard application programming interface (API) provided by the structured teaching knowledge graph storage engine for accessing nodes along graph edges. This interface provides the function of retrieving a list of adjacent nodes starting from a specified starting node and proceeding along directed edges. When encapsulating the query engine, the system performs a one-hop extended traversal operation based on this interface. This graph traversal interface serves as the basic entry point for indirect association extension operations in the system, enabling the system to complete graph relationship reasoning without directly manipulating the underlying storage structure of the graph database. Key-value retrieval and relational reasoning refer to the two different query modes provided by the query engine. The key-value retrieval mode directly returns a list of corresponding resource identifiers based on a forward mapping relationship using entity identifiers as keys. The relational reasoning mode extends along the predecessor and successor edges between entities using the graph traversal interface and returns a list of all related entities and their corresponding resource identifiers after the extension. Key-value retrieval and relational reasoning together constitute the core functions of the intelligent teaching full-process knowledge graph retrieval and reasoning engine in this system. Key-value retrieval is used to meet the resource search needs for precise knowledge point location, while relational reasoning is used to meet complex teaching needs requiring knowledge logic deduction, such as comprehensive review across knowledge points, tracing prior knowledge, and planning learning paths. A query engine is a software system module that integrates the aforementioned key-value retrieval capabilities and relational reasoning capabilities into the same operating environment and provides a unified external calling interface. In this system, the query engine serves as the sole entry and exit point for all knowledge graph-related operations for end users and upper-level teaching applications. It completely shields the internal implementation details of the underlying teaching resource annotation index table storage structure, structured teaching knowledge graph database, and graph traversal interface from external callers. This allows specific teaching application functions such as personalized learning path recommendation, intelligent test paper generation, learning diagnosis, and lesson preparation assistance to be independently developed and iterated based on the unified retrieval and reasoning interface, without having to repeatedly implement the knowledge graph traversal logic and resource index query logic.

[0044] In the embodiment of constructing the retrieval engine module, based on the teaching resource annotation index table and the structured teaching knowledge graph, a positive mapping relationship is established from resources to knowledge graph entities using key-value pairs composed of resource identifiers and entity identifiers in the index table. This positive mapping relationship enables the system to directly locate all associated resource identifiers through hash mapping when it receives a query request for a specific knowledge graph entity identifier, avoiding a full table scan of the teaching resource annotation index table and reducing the time complexity of resource retrieval. At the same time, this positive mapping relationship serves as the basic node index for graph traversal expansion operations, ensuring that the starting node of the expansion operation has a definite resource association anchor point. Based on the aforementioned positive mapping relationship and the predecessor and successor edges between entities stored in the structured teaching knowledge graph, a one-hop expansion traversal is performed from the current entity node to the adjacent entity node along the direction of the directed edges of the graph. This determines the set of extended related entity identifiers for each resource, beyond the directly related entities. This set of extended related entity identifiers enables the system to obtain resource annotation results for adjacent entities that have a learning sequence dependency relationship with the currently queried entity, expanding the retrieval scope from a single knowledge entity to its predecessor and successor knowledge entities. This ensures that the resource query results cover teaching materials that require prior preparation and extension during the teaching process. The set of directly related entity identifiers and the set of extended related entity identifiers contained in the positive mapping relationship are merged to generate a unified retrieval mapping table that simultaneously contains direct and indirect associations. This unified retrieval mapping table eliminates the data structure separation between direct retrieval results and extended retrieval results, allowing the system to simultaneously return directly annotated resources and derived related resources obtained through graph edge expansion in a single query operation, avoiding the redundant overhead caused by multiple independent queries. Based on the unified retrieval mapping table and the graph traversal interface provided by the structured teaching knowledge graph, the key-value pair search logic and the graph edge traversal logic are encapsulated in the same software module. This software module provides a unified query entry point and hides the underlying hash mapping storage structure and the graph directed edge storage structure from external callers. This allows upper-level teaching applications to call the query function without distinguishing the differences in the acquisition paths of directly related resources and indirectly related resources. At the same time, it ensures that the key-value retrieval mode and the relational reasoning mode share the same resource association data foundation, avoiding discrepancies in results between the two query modes due to inconsistent data sources.

[0045] In the above-mentioned embodiment of the intelligent auxiliary teaching system for the entire process of intelligent teaching based on knowledge graphs, multi-source heterogeneous teaching data streams in intelligent teaching scenarios are acquired. Each multi-source heterogeneous teaching data stream contains at least two or more data modalities from textbook chapter texts, classroom teaching recordings, student homework correction records, and stage test answer results. This allows the system to simultaneously collect raw data from three different stages—teaching pre-planning, classroom implementation, and learning feedback—during the data acquisition stage. This provides multi-dimensional raw materials across stages for knowledge extraction and avoids knowledge blind spots caused by incomplete information from a single data source. For each multi-source heterogeneous teaching data stream, according to the preset subject knowledge ontology rules, which include a dictionary of core subject concepts, a template for superior-subordinate relationships between concepts, and a table of conditions for determining predecessor-successor relationships between concepts, knowledge point entity extraction and relation type determination are performed on the multi-source heterogeneous teaching data stream. This determines the set of knowledge point triples for the multi-source heterogeneous teaching data stream. Each triple describes the subject knowledge point entities appearing in the data stream and their semantic relationships in a structured form, consisting of a head entity, a relation type, and a tail entity. The subject knowledge ontology rules provide a unified knowledge extraction constraint framework for multi-source heterogeneous data from different textbook versions and different teachers' teaching styles, enabling the entity extraction and relation determination results to be aligned under the same knowledge system. At the same time, unstructured natural language text is converted into structured graph data basic units. By deduplicating and merging sets of knowledge point triples corresponding to multiple heterogeneous teaching data streams from multiple sources, a structured teaching knowledge graph is generated. The pre-set conflict resolution rules include prioritizing the relationship type with the highest frequency when there are two or more different relationship types between the same head entity and the same tail entity, and using the standard name in the subject core concept dictionary as the normalized entity name when the expression of the same entity in different data streams is inconsistent. This conflict resolution rule enables multiple candidate triples corresponding to the same pair of head and tail entities from multiple sources to be automatically merged into a single deterministic triple through a frequency statistics mechanism, and ensures that all entity names are unified in the dictionary standard form. This solves the problems of relationship conflict and terminology ambiguity caused by the heterogeneity of multi-source data, and generates a structured teaching knowledge graph with consistent node names and unique relationship types. Based on the structured teaching knowledge graph, knowledge point annotation and association retrieval are performed on the candidate teaching resource data to determine the teaching resource annotation index table. This index table records the knowledge graph entity identifier and association confidence score associated with each candidate teaching resource. The association confidence score is calculated by weighting the edit distance similarity between the resource text and the entity description text and the keyword co-occurrence frequency. This confidence score also includes relevance information from two dimensions: character-level exact matching and lexical-level semantic coverage. This quantifies the association strength between teaching resources and knowledge graph entities into a single comprehensive indicator, and the effective association records are persistently stored in the index table in the form of structured ternary fields.A smart teaching knowledge graph retrieval and reasoning engine is constructed using a teaching resource annotation index table. This engine supports key-value location of directly related resources based on the positive mapping relationship of entity identifiers in the index table, reducing the time complexity of query operations from the level of the total number of resources multiplied by the total number of entities to the constant level. At the same time, based on the predecessor and successor edges between entities stored in the structured teaching knowledge graph, it supports extended traversal along the graph direction, returning indirect related resources corresponding to predecessor and successor knowledge entities in addition to directly related resources. This ensures that the query results cover teaching materials that need to be laid out and extended in the teaching process. The key-value retrieval mode and relational reasoning mode are encapsulated in a unified external calling interface, so that upper-level teaching applications can complete comprehensive resource queries across knowledge points without distinguishing the differences in the acquisition paths of direct and indirect associations.

[0046] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0047] To verify the superiority of the association annotation method based on the weighted fusion of edit distance similarity and keyword co-occurrence frequency in this invention over existing single-index methods, a comparative experiment was designed in this embodiment.

[0048] Experimental setup: A teaching resource database related to the high school mathematics function chapter was selected, containing 2000 candidate resources including test question texts, lesson plan excerpts, and micro-lesson subtitles. The structured teaching knowledge graph contains 15 knowledge point entities under the chapter, such as the definition, domain, range, monotonicity, and parity of functions. Each entity is accompanied by a standard description text and a preset set of keywords, which were developed by subject teaching and research personnel based on the teaching syllabus.

[0049] Control group design: Control group 1 uses only keyword co-occurrence frequency: only keyword co-occurrence frequency is used as the association confidence, and edit distance similarity is not calculated. The threshold screening method is the same as that of this invention.

[0050] Control group 2 only uses edit distance similarity: only edit distance similarity is used as the association confidence score, keyword co-occurrence frequency is not counted, and the threshold screening method is the same as that of this invention.

[0051] The present invention provides the following solution: As described in claim 1, the association confidence score is obtained by weighted summation of edit distance similarity (weight 0.6) and keyword co-occurrence frequency (weight 0.4), with a threshold of 0.7.

[0052] Evaluation indicators: Precision = Correctly labeled resources - Number of entity pairs / Total labeled resources - Number of entity pairs; Recall = Correctly labeled resources - Number of entity pairs / Resources that should be labeled - Total number of entities.

[0053] The experimental results are shown in the table below: Control group 1: Precision 68.3%, Recall 81.2%, F1-score 0.742; Control group 2: Precision 79.5%, Recall 62.7%, F1-score 0.701; The solution of this invention achieves a precision of 87.6%, a recall of 85.4%, and an F1-score of 85.4%. Results Analysis: Although control group 1 had a high recall rate (81.2%), its accuracy was only 68.3% due to numerous accidental word overlaps leading to mislabeling. Control group 2, while having a higher accuracy rate (79.5%), had a significantly lower recall rate (62.7%) due to its sensitivity to synonyms and short texts caused by edit distance. The proposed solution, by weighting and fusing the two factors, improved accuracy to 87.6%, recall to 85.4%, and achieved an F1-score of 0.865, representing improvements of 16.6% and 23.4% compared to control groups 1 and 2, respectively. These results demonstrate that the dual-factor weighting is not a simple superposition of two single indicators, but rather a synergistic effect achieved through the complementarity of character-level precise matching and word-level semantic coverage, thus solving the technical challenge of balancing accuracy and recall in teaching resource annotation. When faced with resource labeling problems, those skilled in the art usually only choose one mature solution, either keyword matching or text similarity, and lack the motivation to merge the two with specific weights. They also do not expect that fusion can significantly improve both accuracy and recall. Therefore, this invention has outstanding substantive features and significant progress.

[0054] In certain application scenarios of this application: teacher lesson preparation and instructional design scenarios, student personalized learning path recommendation scenarios, intelligent test paper generation and assessment scenarios, learning situation diagnosis and weak point location scenarios, intelligent classroom teaching assistance scenarios, automatic annotation and archiving of teaching resources scenarios, and cross-knowledge point comprehensive review and exam preparation scenarios.

[0055] In the aforementioned application scenarios, during teacher lesson preparation and instructional design, the system inputs a chapter name or knowledge point name. The system then uses a retrieval and reasoning engine to quickly locate and return all teaching resources directly related to that knowledge point, as well as those indirectly related through predecessor-successor relationships, from the teaching resource annotation index table. These resources include lesson plan templates, test questions, micro-lecture videos, and supplementary reading materials, which teachers can refer to during lesson preparation, reducing the time spent searching and filtering teaching materials one by one. In the scenario of personalized learning path recommendation for students, the system, based on the predecessor-successor dependencies between knowledge point entities in a structured teaching knowledge graph and combining the confidence scores of the associations between each resource and knowledge point in the teaching resource annotation index table, generates an ordered learning path from currently mastered knowledge points to the target knowledge point. It then sequentially pushes corresponding teaching resources according to each prerequisite knowledge point in the path, supporting students' progressive self-directed learning based on the cognitive dependency order of knowledge points. In the intelligent test paper generation and assessment scenario, the system retrieves candidate test question resources corresponding to each entity from the teaching resource annotation index table based on all knowledge point entities within a specified range in the structured teaching knowledge graph. It then filters test questions according to preset constraints on the number of questions, difficulty distribution, and knowledge point coverage, and generates unit test papers or mid-term and final exam papers. This transforms the test paper generation process from manual question-by-question selection to automated matching based on the knowledge graph's association index. In the learning diagnosis and weak point location scenario, the system imports student homework correction records and stage test answers. It maps students' answered questions to corresponding knowledge graph entity identifiers through the teaching resource annotation index table. It calculates the error rate of students under each entity identifier, identifies entities with error rates exceeding a preset threshold as weak knowledge points, and traces all prerequisite knowledge points based on the predecessor and successor edges of the structured teaching knowledge graph. This generates a complete diagnostic report containing the weak point and its prerequisite knowledge points, allowing teachers to provide targeted remedial instruction or students to engage in targeted tutoring. In the intelligent classroom teaching assistance scenario, the system receives real-time subtitles from the recorded classroom lectures. For the currently being taught text segment, it extracts and determines the relevant knowledge entities through knowledge point entity extraction and relationship analysis. It then instantly retrieves related examples, variation exercises, and common error messages from the teaching resource annotation index table, and pushes the search results to the teacher's terminal in card format, assisting the teacher in instantly accessing targeted training materials during the lesson. In the automatic teaching resource annotation and archiving scenario, the system continuously receives new candidate teaching resource data. For each new resource, it automatically performs resource text extraction, edit distance similarity calculation, keyword co-occurrence frequency statistics, and association confidence score calculation. Associations with confidence scores higher than a threshold are automatically written into the teaching resource annotation index table, enabling new resources to be included in the knowledge graph retrieval system without manual annotation, achieving automated archiving and semantic indexing of teaching assets.In scenarios involving comprehensive review and exam preparation across multiple knowledge points, the system accepts composite query requests composed of multiple entity identifiers for comprehensive exams covering multiple knowledge points. It locates all specified entities and all their predecessors and successors in a one-hop extension from the structured teaching knowledge graph, recalls all relevant candidate resources based on the teaching resource annotation index table, clusters and groups the resources according to the coverage of knowledge points, and generates a comprehensive review material package covering all queried knowledge points and their dependencies.

[0056] like Figure 2 As shown, Figure 2 This paper demonstrates the core data processing logic of a knowledge graph-based intelligent teaching system, which goes through five stages from top to bottom: data acquisition, knowledge extraction, graph construction, resource annotation, and retrieval engine encapsulation. The process begins with the acquisition of multi-source heterogeneous teaching data streams, which contain at least two modalities from textbook chapter text, classroom teaching transcripts, student homework correction records, and quiz results. Then, according to pre-defined subject knowledge ontology rules, knowledge point entity extraction and relation type determination are performed, transforming the unstructured teaching text into a set of knowledge point triples. Next, the triple sets from multiple data streams are grouped, aggregated, and conflict resolved. Grouping is performed using the head and tail entities as joint grouping keys. The frequency of each relation type within each group is counted, and the relation type with the highest frequency is selected as the preferred relation. Entity names are normalized to standard terms in the subject's core concept dictionary, outputting a structured teaching knowledge graph. After the knowledge graph is constructed, the process enters the resource annotation stage. For each candidate teaching resource, the resource text is extracted, and the edit distance similarity between it and the description text of each entity is calculated. The co-occurrence frequency of terms belonging to the entity's preset keyword set in the resource text is also counted. These two values ​​are then weighted and summed according to preset weights to obtain the association confidence score. Finally, a confidence threshold is used as the criterion for filtering. Associations exceeding the threshold are written into the teaching resource annotation index table. Based on the index table, a positive mapping relationship is established using resource identifiers and entity identifiers. A one-hop expansion traversal is performed along the predecessor and successor edges in the knowledge graph. Direct and indirect associations are merged and encapsulated to obtain a retrieval engine that simultaneously supports key-value retrieval and relational reasoning.

[0057] Figure 3 A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown.

[0058] The electronic device may include a central processing unit / microprocessor / main control chip 6; a storage medium 7 is coupled to the central processing unit / microprocessor / main control chip 6 and stores computer-executable instructions therein for performing the steps of various methods of embodiments of the present invention when executed by the processor.

[0059] The central processing unit / microprocessor / main control chip 6 may include, but is not limited to, one or more processors or microprocessors.

[0060] Storage medium 7 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CD-ROM, DVD-ROM, Blu-ray disc, etc.).

[0061] In addition, the electronic device may include (but is not limited to) a data bus 8, an input / output bus / external bus / device bus 9, a display 10, and input / output devices 11 (e.g., keyboard, mouse, speaker, etc.).

[0062] The central processing unit / microprocessor / main control chip 7 can communicate with external devices (10, 11, etc.) via wired or wireless networks (not shown) through the input / output bus / external bus / device bus 9.

[0063] The storage medium 7 may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when run by the central processing unit / microprocessor / main control chip 6.

[0064] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0065] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0066] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge graph-based intelligent teaching system for the entire teaching process, characterized in that, Include: The module for acquiring heterogeneous teaching data is used to acquire multi-source heterogeneous teaching data streams in smart teaching scenarios. The knowledge point triple set module is used to extract knowledge point entities and determine relation types for each multi-source heterogeneous teaching data stream according to the preset subject knowledge ontology rules, and determine the knowledge point triple set of the multi-source heterogeneous teaching data stream. The structured teaching knowledge graph module is used to generate a structured teaching knowledge graph by deduplicating and merging a set of knowledge point triples corresponding to multiple heterogeneous teaching data streams from multiple sources after deduplication and merging based on preset conflict resolution rules. The structured teaching knowledge graph stores the predecessor and successor relationships between knowledge point entities, and the predecessor and successor relationships represent the learning sequence dependence between subject knowledge points. The association similarity module is used to perform the following operations on each candidate teaching resource based on the entity identifiers and their corresponding entity description texts in the structured teaching knowledge graph: the edit distance similarity between the resource text of the candidate teaching resource and each entity description text, and the co-occurrence frequency of terms belonging to the entity's preset keyword set in the resource text are weighted and summed according to preset weights to calculate the association confidence score between the candidate teaching resource and each entity. The correspondence between candidate teaching resources with association confidence scores higher than a preset threshold and entity identifiers is written into the teaching resource annotation index table; A retrieval engine module is constructed to obtain a teaching resource annotation index table and a structured teaching knowledge graph. A positive mapping relationship between resources and entities in the knowledge graph is established using resource identifiers and entity identifiers in the index table. Based on the forward mapping relationship and the predecessor-successor relationship stored in the structured teaching knowledge graph, a one-hop extended traversal is performed along the directed edge direction to determine the extended associated entity identifier set of each resource; the forward mapping relationship and the extended associated entity identifier set are merged to generate a unified retrieval mapping table containing direct and indirect associations; based on the unified retrieval mapping table and the graph traversal interface of the structured teaching knowledge graph, a query engine that supports both key-value retrieval and relation reasoning is encapsulated.

2. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 1, characterized in that, In the module for acquiring heterogeneous teaching data, each multi-source heterogeneous teaching data stream contains at least two data modalities from textbook chapter text, classroom teaching recording subtitles, student homework correction records, and stage test answer results.

3. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 1, characterized in that, In the knowledge point triple set module, the subject knowledge ontology rules include a dictionary of core subject concepts, a template for superior / subordinate relationships between concepts, and a table of conditions for determining predecessor / successor relationships between concepts. Each triple in the knowledge point triple set describes the subject knowledge point entities that appear in the multi-source heterogeneous teaching data stream and their semantic relationships with each other in a structured form of head entity, relation type, and tail entity.

4. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 1, characterized in that, In the structured teaching knowledge graph module, the pre-set conflict resolution rules include at least the following: when there are two or more different relation types between the same head entity and the same tail entity, the relation type with the highest frequency is selected as the preferred relation; when the representation of the same entity in different data streams is inconsistent, the standard name in the core concept dictionary of the subject is used as the normalized entity name.

5. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 1, characterized in that, In the association similarity module, the preset weights are: edit distance similarity weight coefficient is 0.6, and keyword co-occurrence frequency weight coefficient is 0.

4.

6. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 1, characterized in that, The association similarity module includes: The plain text extraction submodule is used to obtain a structured teaching knowledge graph, which contains several entity identifiers and corresponding entity description texts; obtain candidate teaching resource data; and extract the resource text of each candidate teaching resource into a plain text string. The term co-occurrence frequency submodule is used to calculate the edit distance similarity between the resource text and the entity description text for each candidate teaching resource and each entity identifier; and to count the co-occurrence frequency of terms belonging to the entity keyword set in the resource text. The preset weighted summation submodule is used to sum the edit distance similarity and keyword co-occurrence frequency according to preset weights to calculate the association confidence score between candidate teaching resources and entity identifiers. The correspondence writing submodule is used to write the correspondence between candidate teaching resources with association confidence scores higher than a preset threshold and entity identifiers into the index table; each record in the index table contains resource identifier, entity identifier and corresponding association confidence score.

7. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 6, characterized in that, The submodule for counting term co-occurrence includes: The edit distance similarity score unit is used to obtain the resource text and entity description text corresponding to the entity identifier of the candidate teaching resources; calculate the minimum edit distance between the resource text and the entity description text; and determine the edit distance similarity score based on the minimum edit distance and the normalization of the two text lengths. The keyword co-occurrence frequency unit is determined to obtain the preset keyword set corresponding to the entity identifier; and the resource text is processed by word segmentation and stop word filtering. The number of times each word is hit in the keyword set after filtering is counted to determine the keyword co-occurrence frequency. The weighted summation unit is used to take the edit distance similarity score and keyword co-occurrence frequency as input features; and sums them according to preset weight coefficients to obtain the association confidence score between candidate teaching resources and entity identifiers.

8. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 6, characterized in that, The pre-defined weighted summation submodule includes: Read the weight coefficient unit to obtain the edit distance similarity score and keyword co-occurrence frequency for the same candidate resource and the same entity; read the similarity weight coefficient and frequency weight coefficient corresponding to the entity from the system configuration parameters; The second weighting unit is used to multiply the similarity score by the similarity weight coefficient to obtain the first weighting value; and to multiply the frequency by the frequency weight coefficient to obtain the second weighting value. A correlation confidence unit is determined to sum the first weighted value and the second weighted value; and the summation result is determined as the correlation confidence score between the candidate resource and the entity.

9. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 6, characterized in that, The correspondence relationship is written into the submodule, including: Read the confidence threshold unit to obtain a list of full association confidence scores between the same candidate resource and each entity; read the preset confidence admission threshold from the system configuration parameters; The candidate association pair filtering unit is used to compare each score in the full association confidence score list with the admission threshold; and to filter out candidate association pairs with scores higher than the admission threshold. The confidence score extraction unit is used to extract the resource identifier, entity identifier, and corresponding confidence score from each selected candidate association pair; and write the resource identifier, entity identifier, and confidence score as a record into the index table.

10. The intelligent auxiliary teaching system based on knowledge graphs for the entire intelligent teaching process as described in claim 1, characterized in that, The module for building the search engine includes: A submodule for establishing a positive mapping relationship is used to obtain the teaching resource annotation index table and the structured teaching knowledge graph; using the resource identifier and entity identifier in the index table as key-value pairs, a positive mapping relationship is established from resources to entities in the knowledge graph. The extended associated entity identifier set submodule is used to perform a one-hop extended traversal along the directed edge direction of the graph based on the forward mapping relationship and the predecessor and successor edges between entities stored in the knowledge graph, to determine the extended associated entity identifier set of each resource. The module for generating a retrieval mapping table is used to merge the forward mapping relationship with the extended associated entity identifier set to generate a unified retrieval mapping table containing direct and indirect associations. Based on the unified retrieval mapping table and the graph traversal interface of the knowledge graph, a query engine that supports both key-value retrieval and relation reasoning is encapsulated.

Images