A knowledge graph and large language model-based adaptive learning path generation method and system

Abstract

This application relates to the field of intelligent education technology, and discloses an adaptive learning path generation method and system based on knowledge graphs and large language models. The method receives natural language input from users, combines user static profile data and historical learning records to generate a target structured intent result; constructs a dedicated knowledge subgraph based on the teaching knowledge graph and mastered boundary nodes; performs weighted topological sorting on the dedicated knowledge subgraph and supplements the semantic coherence of adjacent knowledge nodes to obtain a target learning path sequence; generates learning interaction content based on the knowledge node association text credentials associated with the current knowledge node, updates the mastery probability based on the answer results and interaction records, and reconstructs the learning path sequence accordingly. This scheme can improve the accuracy, continuity, and adaptability of learning path generation.

Description

Technical Field

[0001] This application relates to the field of intelligent education technology, and more specifically, to an adaptive learning path generation method and system based on knowledge graphs and large language models. Background Technology

[0002] Existing online learning systems typically recommend learning content to users based on course catalogs, keyword search results, or preset rules. Although some solutions introduce knowledge graphs to organize the relationships between knowledge points or introduce large language models to generate explanations, there is still a problem of discontinuity in the processing chain from learning needs identification, knowledge scope determination, path connection to learning status feedback adjustment during the learning path generation process.

[0003] On the one hand, existing solutions typically process users' learning needs expressed in natural language using keyword extraction, classification matching, or fixed-condition filtering. This makes it difficult to accurately convert learning objectives and constraints into structured results that can be directly used for knowledge retrieval and path planning, leading to a discrepancy between the subsequently determined learning scope and the user's actual needs. On the other hand, when generating learning content, existing solutions usually fail to jointly process the prerequisite dependencies, inclusion relationships, and progression relationships between historical learning records and knowledge points, making it difficult to form a unique knowledge scope that aligns with the user's current mastery.

[0004] Furthermore, even if existing solutions can provide a ranking of several knowledge points, they often remain at the level of static sorting or discrete recommendation, lacking the detection and supplementation of semantic coherence between adjacent knowledge nodes. This can easily lead to abrupt jumps in learning content and affect the continuity of the learning path. For learning interaction content generated based on large language models, if there is a lack of textual constraints corresponding to the current knowledge node, the generated content can easily deviate from the definition, relationship, scope of application, or application scenario of the current knowledge node, affecting the relevance and stability of the learning interaction content. Moreover, existing solutions usually cannot continuously update the mastery status of the current knowledge node based on the user's answers and interaction records, and further perform pre-filling, order preservation, or low-difficulty removal processing on subsequent learning paths based on the updated mastery status. This makes it difficult for the learning path to adapt to changes in the user's state during the learning process in a timely manner.

[0005] Therefore, how to provide an adaptive learning path generation method and system based on knowledge graphs and large language models to transform users' natural language learning needs into target structured intent results, and to construct exclusive knowledge subgraphs based on historical learning records and knowledge relationships to generate learning path sequences with more coherent semantic connections and dynamic reconstruction based on mastery status has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, this application provides the following technical solution: In the first aspect, this application discloses an adaptive learning path generation method based on knowledge graphs and large language models, including: Step S1: Receive user natural language input, read user static profile data and historical learning records, and parse to generate a target structured intent result including learning objectives and learning constraints; Step S2: Based on the target structured intent result, the target knowledge node is retrieved from the pre-constructed teaching knowledge graph. Based on the pre-dependency, inclusion and advancement relationships between the knowledge nodes in the teaching knowledge graph, the target knowledge node is traversed to the mastered boundary node corresponding to the historical learning record. Then, expansion and pruning are performed to obtain the exclusive knowledge subgraph. Step S3: Perform weighted topological sorting based on the exclusive knowledge subgraph to obtain the initial learning path sequence. Detect the semantic coherence of adjacent knowledge nodes in the initial learning path sequence. Generate logical transition content based on the original text fragments associated with adjacent knowledge nodes with insufficient semantic coherence, and insert them into the initial learning path sequence to obtain the target learning path sequence. Step S4: Determine the current knowledge node based on the target learning path sequence, extract the knowledge node-related text credentials associated with the current knowledge node, generate learning interaction content based on the knowledge node-related text credentials, collect the user's answer results and interaction records in response to the learning interaction content, and update the mastery probability of the current knowledge node based on the answer results and interaction records. Step S5: Reconstruct the target learning path sequence based on the mastery probability of the current knowledge node, insert the predecessor knowledge node of the current knowledge node or remove the knowledge node with lower difficulty than the current knowledge node in the same concept cluster, and output the updated learning path sequence.

[0007] Secondly, this application discloses an adaptive learning path generation system based on knowledge graphs and large language models, comprising: The intent parsing module is used to receive natural language input from users, read static user profile data and historical learning records, and parse and generate structured intent results that include learning objectives and learning constraints. The subgraph construction module is used to retrieve target knowledge nodes from the pre-built teaching knowledge graph based on the target structured intent result. Based on the pre-dependency, inclusion and advancement relationships between knowledge nodes in the teaching knowledge graph, it traverses from the target knowledge node to the mastered boundary node corresponding to the historical learning record, and then performs expansion and pruning to obtain the exclusive knowledge subgraph. The path generation module is used to perform weighted topological sorting based on the exclusive knowledge subgraph to obtain an initial learning path sequence, detect the semantic coherence of adjacent knowledge nodes in the initial learning path sequence, generate logical transition content based on the original text fragments associated with adjacent knowledge nodes with insufficient semantic coherence, and insert them into the initial learning path sequence to obtain the target learning path sequence. The interactive evaluation module is used to determine the current knowledge node based on the target learning path sequence, extract the knowledge node-related text credentials associated with the current knowledge node, generate learning interaction content based on the knowledge node-related text credentials, collect the user's answer results and interaction records in response to the learning interaction content, and update the mastery probability of the current knowledge node based on the answer results and interaction records. The path reconstruction module is used to reconstruct the target learning path sequence based on the mastery probability of the current knowledge node, insert the predecessor knowledge node of the current knowledge node or remove the knowledge node with lower difficulty within the same concept cluster, and output the updated learning path sequence.

[0008] Compared with related technologies, this application has the following advantages: This application receives natural language input from users and, combined with user static profile data and historical learning records, parses and generates a structured intent result that includes learning objectives and constraints. Then, based on the structured intent result, it retrieves target knowledge nodes in the teaching knowledge graph and performs expansion and pruning based on prerequisite dependencies, inclusion relationships, progression relationships, and already mastered boundary nodes to obtain a unique knowledge subgraph. This allows the determination of the learning scope to go beyond keyword hit results or static course catalogs, but to connect with the user's current learning objectives and existing mastery, improving the accuracy and relevance of the starting point for determining the learning path.

[0009] This application further performs weighted topological sorting based on the dedicated knowledge subgraph to obtain an initial learning path sequence, and detects the semantic coherence between adjacent knowledge nodes. When the semantic coherence is insufficient, logical transition content is generated based on the original text fragments associated with adjacent knowledge nodes and inserted into the initial learning path sequence to obtain the target learning path sequence. This can supplement the connecting content between adjacent knowledge nodes, reduce the problem of abrupt jumps in learning content, and improve the continuity and orderliness of the learning path.

[0010] This application also determines the current knowledge node based on the target learning path sequence, extracts the knowledge node association text credentials associated with the current knowledge node, generates learning interaction content based on the knowledge node association text credentials, and then updates the mastery probability of the current knowledge node by combining the user's answer results and interaction records. This ensures that the learning interaction content corresponds to the definition, relationship, example and constraint of the current knowledge node, and provides clear data basis for updating the mastery status, thereby improving the relevance of the learning interaction content and the reliability of the mastery status assessment.

[0011] This application reconstructs the target learning path sequence based on the mastery probability of the current knowledge node, inserts the predecessor knowledge node of the current knowledge node or removes knowledge nodes with lower difficulty than the current knowledge node within the same concept cluster, and outputs an updated learning path sequence. This allows the subsequent learning order to be adjusted according to the user's current learning results, enabling the learning path to be dynamically updated with the learning process, thereby improving the consistency and adaptability of the learning path with the user's actual mastery status. Attached Figure Description

[0012] Figure 1 A schematic diagram of the adaptive learning path generation method based on knowledge graph and large language model provided in this application; Figure 2 This is a schematic diagram of an adaptive learning path generation system module based on knowledge graphs and large language models, which is provided for this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] Example 1 Please see Figure 1 As shown, this embodiment provides an adaptive learning path generation method based on knowledge graphs and large language models, including the following steps: Step S1, receiving user natural language input, reading user static profile data and historical learning records, and parsing to generate a target structured intent result including learning objectives and learning constraints, aims to transform the user's learning needs expressed in natural language into a target structured intent result with unified fields, definite meanings, and direct access to authoritative knowledge graph retrieval. The implementation steps include: Step 101: Receive natural language input from the user and record the user's natural language input as input text data. The input text data shall include at least one of the following: learning topic, expected learning outcome, and learning schedule.

[0015] Step 102: Read the user static profile data; the user static profile data includes at least the educational stage information, grade information, current subject information, and available learning time period information. The user static profile data is used to limit the scope and depth of subsequent learning.

[0016] Step 103: Read the historical learning record; the historical learning record includes at least the knowledge node identifier, answer result, completion status, learning time, and review record. The historical learning record is used to represent the user's past learning trajectory.

[0017] Step 104: Perform text normalization processing on the input text data. Text normalization processing includes terminology standardization, time expression standardization, and removal of irrelevant words to obtain normalized input text.

[0018] Step 105: Pre-build a large language model for intent parsing, and input the normalized input text, user static profile data and historical learning records into the large language model. Output the corresponding records of field names and field values ​​according to the preset field table to obtain the learning objectives and learning constraints.

[0019] The method for constructing a large language model for intent parsing includes: collecting textbook table of contents text, curriculum standard text, textbook body text, exercise analysis text, and supplementary question and answer text; performing sentence segmentation, deduplication, typo removal, and terminology standardization on the collected text to obtain domain training corpus; manually organizing or programmatically extracting corresponding samples of user expression content, learning target fields, and learning constraint fields based on the domain training corpus to form an intent annotation sample set; using a general language parameter file as initial parameters, continuing training with the intent annotation sample set to obtain an intent parsing parameter file; loading the intent parsing parameter file and combining it with preset prompt templates and preset field tables to form a large language model for intent parsing that outputs records corresponding to field names and field values.

[0020] The preset field table consists of a field definition table and a field value table. The field definition table includes at least the field names, meanings, and input positions for the target subject field, target knowledge scope field, learning scope constraint field, and learning depth constraint field. The field value table includes at least the set of allowed terms and standard term values ​​for each field. During parsing, the normalized input text, user static profile data, and historical learning records are concatenated into a model input sequence according to the preset prompt template. The large language model outputs records corresponding to field names and values ​​based on the preset field table, thereby obtaining the learning objectives and learning constraints.

[0021] Step 106: Perform field integrity verification on the set of required fields corresponding to the learning objectives and learning constraints and the preset field table, and call the preset terminology mapping table to convert synonyms into unified field values ​​to obtain the target structured intent result.

[0022] The method for field integrity verification includes: reading the set of required fields in the preset field table, checking each field name and field value record to see if the corresponding field exists, whether the field value is empty, and whether the field value falls into the allowed value set corresponding to the field value table; writing missing flags for missing fields, writing empty value flags for empty value fields, and writing outlier flags for field values ​​not in the allowed value set; and then, based on the user's static profile data and historical learning records, performing completion or replacement on the fields corresponding to the missing flags, empty value flags, and outlier flags to form the verified field records.

[0023] The method for constructing the pre-defined terminology mapping table includes: extracting knowledge terms, course terms, and constraint terms from textbook table of contents, curriculum standard text, textbook body text, exercise analysis text, and supplementary Q&A text; grouping multiple expressions with the same meaning or referring to the same knowledge object into the same terminology group, and assigning a uniform field value to each terminology group; and writing them into the pre-defined terminology mapping table in the form of: original expression—uniform field value—applicable field name. During mapping, the field values ​​in the validated field records are read, the corresponding original expression is retrieved from the pre-defined terminology mapping table, and the retrieved field value is replaced with the uniform field value to obtain the target structured intent result. The target structured intent result includes at least the target subject, target knowledge scope, learning scope constraints, and learning depth constraints, and serves as input for subsequent steps.

[0024] In some implementations, to illustrate the formation process of the target structured intent result, for example, the user's natural language input is: "I want to review linear equations in one variable in junior high school math this week, focusing on solving the problem of always making mistakes when transposing terms"; the user's static profile data records the current subject information as math, the grade level as junior high school, and the historical learning records as past learning situations of algebraic addition and subtraction, basic properties of equations, and solving simple equations; after performing text normalization processing on this user's natural language input, a normalized input text including "review," "this week," "junior high school math," "linear equations in one variable," and "mistakes in transposing terms" can be obtained; then... After standardizing the input text, user static profile data, and historical learning records into the large language model, it can output records corresponding to field names and field values. Among them, the target subject field is mathematics, the target knowledge scope field is linear equations in one variable, the learning scope constraint field is equation-related knowledge within the scope of junior high school mathematics, and the learning depth constraint field is error correction. Subsequently, the completeness of the fields and the consistency of terminology are verified, and expressions such as equation review and transposition exercises in the user's natural language input are uniformly mapped to standard field values ​​in the preset field table, thereby obtaining the target structured intent results that can be used for subsequent teaching knowledge graph retrieval.

[0025] Step S2 involves retrieving target knowledge nodes from a pre-constructed teaching knowledge graph based on the target structured intent result. Then, based on the pre-dependencies, inclusion relationships, and progression relationships between knowledge nodes in the teaching knowledge graph, the process traverses from the target knowledge nodes to the mastered boundary nodes corresponding to historical learning records. Expansion and pruning are then performed to obtain a dedicated knowledge subgraph. The purpose of this step is to extract the knowledge scope from the teaching knowledge graph that corresponds to the target structured intent result and connects with historical learning records, forming a dedicated knowledge subgraph. The implementation steps include: Step 201: Pre-construct a teaching knowledge graph. Each knowledge node in the teaching knowledge graph should at least record the knowledge node identifier, knowledge name, knowledge definition, original text fragment index, and subject information. The relation edges in the teaching knowledge graph should at least record the prerequisite dependencies, inclusion relationships, and progression relationships.

[0026] The method for constructing a teaching knowledge graph includes: extracting chapter-level relationships from the textbook's table of contents, extracting concept definition statements from the textbook's main text, extracting applicable knowledge scenarios from the exercise analysis text, and extracting the knowledge coverage from the curriculum standard text; merging text fragments representing the same knowledge object into the same knowledge node; then writing pre-dependency, inclusion, and progression relationships according to the rules of learning before learning, hierarchical relationships, and from simple to complex; and writing the original text positions supporting each knowledge node and relation edge into the original text fragment index.

[0027] Step 202: Read the target subject and target knowledge scope from the target structured intent result, perform matching on the knowledge name, knowledge definition and original text fragment in the teaching knowledge graph, retain the knowledge nodes with consistent subject and knowledge scope that fall into the target knowledge scope in the matching result, and obtain the target knowledge node set.

[0028] The matching method includes: first, filtering knowledge nodes of the same subject based on the target subject; then matching the uniform field values ​​in the target knowledge scope with the knowledge name execution name, the knowledge definition execution terminology, and the original text fragment execution fragment respectively; writing at least one matching knowledge node into the candidate results; then deleting candidate results that are inconsistent with the learning scope constraints to obtain the target knowledge node set.

[0029] Step 203: Map the knowledge node identifiers in the historical learning records to the teaching knowledge graph, count the correct answers, number of completions and number of reviews for each knowledge node, and compare them with the preset mastery judgment rules. The knowledge nodes that meet the preset mastery judgment rules are identified as mastered knowledge nodes. Then, between mastered knowledge nodes and non-mastered knowledge nodes, identify knowledge nodes that have a prerequisite dependency relationship with the target knowledge node set to obtain mastered boundary nodes.

[0030] The method for forming the pre-defined mastery judgment rules includes: pre-setting level items corresponding to correct answers, number of completions, and number of reviews, and specifying the combination results for each level item; marking knowledge nodes that achieve the pre-defined combination results as mastered knowledge nodes, and marking knowledge nodes that do not achieve the pre-defined combination results as unmastered knowledge nodes; when identifying mastered boundary nodes, checking each mastered knowledge node for a prerequisite dependency relationship with the target knowledge node set or the predecessor knowledge nodes of the target knowledge node set, and writing mastered knowledge nodes with connection relationships into the mastered boundary node set.

[0031] Step 204: Starting from the target knowledge node set, traverse forward along the preceding dependencies until the mastered boundary node is reached, and retain the knowledge nodes and relation edges passed during the traversal to obtain the dependency node set.

[0032] During traversal, the set of target knowledge nodes is written to the initial traversal queue; each time a knowledge node is read from the initial traversal queue, the predecessor knowledge node corresponding to this knowledge node is retrieved; unprocessed predecessor knowledge nodes are written to the traversal queue, and the predecessor dependency relationship between the predecessor knowledge node and the current knowledge node is recorded synchronously; when a mastered boundary node is read, the branch is stopped from being expanded forward.

[0033] Step 205: Perform expansion along the inclusion and progression relationships on the dependent node set; add knowledge nodes along the inclusion relationship to explain the composition of the target knowledge node, and add progressive knowledge nodes along the progression relationship that correspond to the target knowledge node set to obtain the expanded node set.

[0034] During expansion, first read each knowledge node in the dependent node set, retrieve the subordinate knowledge nodes that are connected to each knowledge node by inclusion, and write the subordinate knowledge nodes that are related to the target knowledge scope into the expansion node set; then retrieve the progressive knowledge nodes that are connected to each knowledge node by advancement, and write the progressive knowledge nodes that do not exceed the learning depth constraint into the expansion node set.

[0035] Step 206: Prune the extended node set according to the learning range constraint and learning depth constraint, and delete knowledge nodes and relation edges that exceed the learning range constraint, exceed the learning depth constraint, or have no connection relationship with the target knowledge node set to obtain a dedicated knowledge subgraph; the dedicated knowledge subgraph includes at least the knowledge node set, the relation edge set, and the original text fragment index set, and serves as the input for subsequent steps.

[0036] During pruning, first delete knowledge nodes that are inconsistent in subject, theme, or chapter scope according to the learning scope constraint; then delete knowledge nodes whose difficulty level exceeds the allowed level according to the learning depth constraint; then perform connectivity checks on the remaining knowledge nodes, and delete isolated knowledge nodes and corresponding relationship edges that cannot be connected to the target knowledge node set through prerequisite dependencies, inclusion relationships, or advancement relationships, to obtain a dedicated knowledge subgraph.

[0037] Furthermore, to illustrate the generation process of the dedicated knowledge subgraph, for example, if the target knowledge scope in the target structured intent result is a linear equation in one variable, and the historical learning record shows that the user has a relatively stable grasp of algebraic addition and subtraction and the basic properties of equations, the system can first obtain the set of target knowledge nodes corresponding to the linear equation in one variable after searching the teaching knowledge graph. Then, starting from this set of target knowledge nodes, it can traverse forward along the pre-dependency relationship, successively passing through knowledge nodes such as transposing terms, removing parentheses, removing denominators, and the basic properties of equations. When traversing to the basic properties of equations corresponding to the already mastered boundary node, it stops further expanding that branch. Subsequently, based on the set of dependent nodes, it adds subordinate knowledge nodes to explain the composition of the linear equation in one variable along the inclusion relationship, and adds progressive knowledge nodes that match the current learning depth constraints along the advancement relationship. After that, it deletes knowledge nodes and their relation edges that exceed the scope of junior high school mathematics or the depth of consolidation of incorrect questions, thus obtaining the dedicated knowledge subgraph. This example illustrates that the dedicated knowledge subgraph is not the entire set of knowledge in the textbook, but rather the knowledge scope that simultaneously corresponds to the target knowledge scope, the existing mastery status, and the learning depth constraints.

[0038] Step S3: Perform weighted topological sorting based on the exclusive knowledge subgraph to obtain the initial learning path sequence. Detect the semantic coherence of adjacent knowledge nodes in the initial learning path sequence. Generate logical transition content based on the original text fragments associated with adjacent knowledge nodes with insufficient semantic coherence, and insert them into the initial learning path sequence to obtain the target learning path sequence.

[0039] In one implementation, the purpose of performing a weighted topological sort based on the dedicated knowledge subgraph to obtain the initial learning path sequence is to transform the dedicated knowledge subgraph into an initial learning path sequence with a sequential order; the implementation steps include: Step 3011: Read the knowledge nodes and relation edges in the exclusive knowledge subgraph, use the preceding dependency relationship as the order constraint, and count the number of incoming edges for each knowledge node; during the count, perform start and end point identification for each preceding dependency relationship edge in the exclusive knowledge subgraph, and record the number of preceding dependency relationship edges pointing to a certain knowledge node as the number of incoming edges for that knowledge node.

[0040] Step 3012: A preset sorting rule table is invoked to convert the correspondence between knowledge nodes and learning objectives, the number of error records for knowledge nodes in historical learning records, and the relational distance from knowledge nodes to mastered boundary nodes into sorting sub-values. These sub-values ​​are then combined to obtain the sorting weight. The method for constructing the preset sorting rule table includes: pre-setting grading rules for the target correspondence level sub-value, the number of error records sub-value, and the relational distance sub-value; mapping different gradings to different sorting sub-values; and specifying the combination order of the sorting sub-values ​​to form the preset sorting rule table. When calculating the sorting weight, the target correspondence level sub-value is first determined based on the matching results between the target knowledge scope and the knowledge node name, knowledge definition, and original text fragment. Then, the number of error records in historical learning records is determined. Next, the relational distance sub-value is determined based on the number of relational edges traversed from the knowledge node to the mastered boundary node. Finally, the sorting sub-values ​​are combined according to the preset sorting rule table to obtain the sorting weight.

[0041] Step 3013: Select knowledge nodes with zero incoming edges to form the current candidate knowledge node set.

[0042] Step 3014: Sort the current set of candidate knowledge nodes according to the sorting weight, write the knowledge nodes with higher sorting into the initial learning path sequence, delete the preceding dependency edges of this knowledge node pointing to subsequent knowledge nodes, and then update the number of incoming edges of the affected knowledge nodes; when writing, write the knowledge node identifier, sorting position and original text fragment index into the initial learning path sequence simultaneously; after deleting the relation edge, recount the number of incoming edges of the knowledge node corresponding to the endpoint of the deleted relation edge.

[0043] Step 3015: Add the knowledge nodes whose incoming edge count is updated to zero to the current candidate knowledge node set. Repeat step 3014 until all knowledge nodes in the exclusive knowledge subgraph are written into the initial learning path sequence.

[0044] Step 3016: Output the initial learning path sequence; the initial learning path sequence includes at least the sequence position, knowledge node identifier and original text fragment index, and serves as the input for subsequent steps.

[0045] For example, to illustrate the formation process of the initial learning path sequence, assume the dedicated knowledge subgraph includes four knowledge nodes: basic properties of equations, rearranging terms, removing parentheses, and removing denominators. Rearranging terms depends on basic properties of equations, removing parentheses depends on rearranging terms, and removing denominators depends on removing parentheses. Initially, the knowledge node with zero incoming edges is the basic property of equations. After the system writes the basic property of equations into the initial learning path sequence, the number of incoming edges for rearranging terms is updated to zero. If historical learning records show that the number of errors a user makes on rearranging terms is higher than the number of errors made on removing parentheses, and the rearranging terms are more consistent with the current learning objective, then the ranking weight of the rearranging terms is higher than other current candidate knowledge nodes, and therefore it is written into the initial learning path sequence first. Then, removing parentheses and removing denominators are written sequentially, thus obtaining an initial learning path sequence that conforms to the pre-dependencies and matches the current learning objective.

[0046] In one implementation, the semantic coherence of adjacent knowledge nodes in the initial learning path sequence is detected. Logical transition content is generated based on the original text fragments associated with adjacent knowledge nodes lacking semantic coherence and inserted into the initial learning path sequence to obtain the target learning path sequence. The purpose is to supplement adjacent knowledge nodes in the initial learning path sequence with logical transition content, thereby obtaining a target learning path sequence that can be directly used for learning interaction. The implementation steps include: Step 3021: Read adjacent knowledge nodes in the initial learning path sequence in order, and extract the knowledge name, knowledge definition, relationship label and original text fragment corresponding to the adjacent knowledge node; the original text fragment is the original text reference fragment written when constructing the teaching knowledge graph.

[0047] Step 3022: Count the number of overlapping terms, the continuity of relation tags, and the overlap of subject terms between adjacent knowledge nodes, and generate a semantic coherence index based on preset coherence calculation rules. The method for constructing the preset coherence calculation rules includes: pre-setting term overlap items, relation continuation items, and subject consistency items; mapping the number of overlapping terms to the result of overlapping terms, mapping the existence of direct inheritance relationships in relation tags to the result of relation continuation items, and mapping the overlap of subject terms to the result of subject consistency items; then writing the three results into the preset coherence calculation rules according to a preset combination order; during calculation, reading the knowledge name, knowledge definition, relation tags, and original text fragments of adjacent knowledge nodes to obtain the number of overlapping terms, the result of relation continuation, and the result of subject consistency, and then outputting the semantic coherence index based on the preset coherence calculation rules.

[0048] Step 3023: Compare the semantic coherence index with the preset coherence judgment rule; when the semantic coherence index is lower than the preset coherence judgment rule, determine the corresponding adjacent knowledge nodes as knowledge node pairs to be transitioned; the preset coherence judgment rule is composed of the hierarchical results of the semantic coherence index, and write the adjacent knowledge node pairs that fall into the low coherence level into the set of knowledge node pairs to be transitioned.

[0049] Step 3024: Input the knowledge name, knowledge definition and original text fragment of the knowledge node to be transitioned into the large language model, and output the logical transition content according to the preset generation template; the preset generation template at least limits the output positions of three types of content: connecting terms, transition statements and introduction statements.

[0050] The method for constructing the preset generation template includes: pre-writing the positions of the preceding knowledge node name, the preceding knowledge node definition, the following knowledge node name, the following knowledge node definition, the original text fragment reference position, the connecting term output position, the transition statement output position, and the introduction statement output position; then specifying the output content type and output order for each position to form the preset generation template; during generation, the input content of the knowledge node pair to be transitioned is filled into the corresponding position of the preset generation template, and the large language model outputs the logical transition content according to the template position.

[0051] Step 3025: Perform reference verification on the logical transition content. Reference verification is used to check whether the terms and judgments in the logical transition content can be found in the corresponding original text fragments or relation tags of the knowledge nodes to be transitioned, and delete the content without basis to obtain the verified logical transition content.

[0052] During the reference verification, the logical transition content is first split into segments, and then term items and judgment items are extracted from each segment. For each term item and judgment item, supporting statements are retrieved from the corresponding original text fragments and relation tags of the knowledge node to be transitioned. When supporting statements are found, the corresponding segments are retained; when no supporting statements are found, the corresponding segments are deleted. The retained segments are combined in their original order to obtain the verified logical transition content.

[0053] Step 3026: Insert the verified logical transition content between the corresponding pairs of knowledge nodes to be transitioned in the initial learning path sequence, update the order position, and obtain the target learning path sequence. The target learning path sequence serves as the input for subsequent determination of the current knowledge node, extraction of the knowledge node associated text credentials, and updating of the mastery probability. During insertion, a transition content identifier is generated for each verified logical transition content, and the transition content identifier, the previous knowledge node identifier, the next knowledge node identifier, and the order position are written into the target learning path sequence to ensure that subsequent steps can read the knowledge nodes and logical transition content in order.

[0054] In some implementations, to illustrate the generation process of logical transition content, for example, two adjacent knowledge nodes in the initial learning path sequence are respectively the basic properties of equations and rearranging terms; after system detection, it is found that the original text fragment corresponding to the former knowledge node mainly explains the rule of performing the same processing on both sides of the equation simultaneously, while the original text fragment corresponding to the latter knowledge node mainly explains the process of rearranging the unknown term and the constant term to both sides of the equation. Since the two have little overlap in keywords in their original text fragments, and the relational markers do not directly reflect a sequential relationship, this pair of adjacent knowledge nodes is identified as the knowledge node pair to be transitioned; subsequently, this... The knowledge names, knowledge definitions, and original text fragments of the two knowledge nodes are input into the large language model, and logical transition content is generated according to the preset generation template. The generated logical transition content can be described as follows: the term transposition process can be understood as performing equal changes on both sides of the equation, and then rearranging the terms containing unknowns and constant terms so that the unknowns can be obtained later. After that, the system performs reference verification on the logical transition content, retaining only the content that can be found in the corresponding original text fragment or relation tag, and inserts the verified logical transition content between the two knowledge nodes to obtain a target learning path sequence that is more convenient for learning interaction.

[0055] Step S4: Determine the current knowledge node based on the target learning path sequence, extract the knowledge node-related text credentials associated with the current knowledge node, generate learning interaction content based on the knowledge node-related text credentials, collect the user's answer results and interaction records in response to the learning interaction content, and update the mastery probability of the current knowledge node based on the answer results and interaction records.

[0056] In one implementation, the purpose of determining the current knowledge node based on the target learning path sequence and extracting the knowledge node association text credentials associated with the current knowledge node is to identify the knowledge object currently performing the learning interaction from the target learning path sequence, and to extract the knowledge node association text credentials that support the generation of subsequent learning interaction content from the teaching knowledge graph; the implementation steps include: Step 4011: Read the sequential position, knowledge node identifier, transition content identifier and execution status marker in the target learning path sequence; the execution status marker is used to record whether the learning interaction has been completed, whether the answer has been completed and whether the mastery probability update has been completed at the corresponding sequential position.

[0057] Step 4012: Retrieve execution status markers from front to back according to their sequential positions, determine the sequential position of the incomplete learning interaction and the corresponding content type as knowledge node as the current execution position, and determine the knowledge node identifier corresponding to the current execution position as the current knowledge node identifier; the content type is used to distinguish between knowledge node content and logical transition content in the target learning path sequence.

[0058] Step 4013: Based on the current execution position, retrieve the preceding and following transition content identifiers adjacent to the current knowledge node identifier, and extract the corresponding transition content; the preceding and following transition content are used to define the connection position of the current knowledge node in the target learning path sequence.

[0059] Step 4014: Based on the current knowledge node identifier, retrieve the corresponding knowledge node in the teaching knowledge graph, and extract the knowledge name, knowledge definition, subject, prerequisite dependencies, inclusion relationships, progression relationships, and original text fragment index that are directly associated with the corresponding knowledge node.

[0060] Step 4015: Extract text fragments from the source text set associated with the teaching knowledge graph based on the original text fragment index, and generate knowledge node associated text credentials according to preset credential assembly rules; the construction method of preset credential assembly rules includes: pre-setting definition credential items, relation credential items, example credential items, and constraint credential items; definition credential items are used to record text fragments that explain the meaning of the current knowledge node, relation credential items are used to record text fragments corresponding to prerequisite dependencies, inclusion relationships, and progression relationships, example credential items are used to record text fragments that reflect the application scenario of the current knowledge node, and constraint credential items are used to record text fragments corresponding to the applicable scope or usage conditions of the current knowledge node; then specify the extraction order and writing order of each credential item to form preset credential assembly rules.

[0061] Step 4016: According to the preset voucher assembly rules, write the definition voucher item, relation voucher item, example voucher item, and constraint voucher item into the knowledge node associated text voucher, and write the current knowledge node identifier, current execution position, preceding transition content, following transition content, and knowledge node associated text voucher into the current learning task record.

[0062] In one implementation, the purpose of generating learning interaction content based on knowledge node-associated text credentials is to convert the knowledge node-associated text credentials into learning interaction content centered around the current knowledge node, so that subsequent answer results and interaction records can correspond to the current knowledge node; the implementation steps include: Step 4021: Read the current knowledge node identifier, preceding transition content, subsequent transition content, and knowledge node associated text voucher from the current learning task record, and extract the definition voucher item, relation voucher item, example voucher item, and constraint voucher item from the knowledge node associated text voucher.

[0063] Step 4022: Pre-build a learning interaction generation template. The method for building the learning interaction generation template includes: pre-setting knowledge introduction positions, knowledge explanation positions, relationship description positions, exercise question positions, answer input positions, and feedback prompt positions. The knowledge introduction position is used to write pre-transitional content, the knowledge explanation position is used to write definition voucher items and example voucher items, the relationship description position is used to write relationship voucher items, the exercise question position is used to write question content formed around the current knowledge node, the answer input position is used to limit the recording method of user input answers, and the feedback prompt position is used to limit the type of prompt statement after answering the question. Then, the output order and content type of each position are specified to form the learning interaction generation template.

[0064] Step 4023: Fill the current knowledge node identifier, preceding transition content, subsequent transition content, and knowledge node associated text credentials into the corresponding positions of the learning interaction generation template, and input them into the large language model to obtain candidate learning interaction content; the candidate learning interaction content includes at least knowledge introduction content, knowledge explanation content, relationship explanation content, exercise content, answer input items, and feedback prompt items.

[0065] Step 4024: Perform a credential consistency check on the candidate learning interaction content. The credential consistency check method includes: splitting the candidate learning interaction content into segments and extracting term items and judgment items for each segment; for each term item and judgment item, retrieving the corresponding supporting fragment in the knowledge node associated text credential; retaining the corresponding segment when a supporting fragment is retrieved, and deleting or regenerating the corresponding segment when no supporting fragment is retrieved, to obtain the verified learning interaction content.

[0066] Step 4025: Associate the verified learning interaction content with the current knowledge node identifier, the current execution position, and the answer input item identifier to obtain the learning interaction content record.

[0067] Furthermore, to illustrate the generation process of learning interaction content, for example, the current knowledge node is solving a linear equation in one variable by removing denominators; the system extracts definition vouchers based on the original text fragment index to explain the meaning of removing denominators, relation vouchers to explain the connection between removing denominators and the basic properties of equations and transposition, example vouchers to explain the application of removing denominators in specific equation simplification, and constraint vouchers to explain that the processing on both sides of the equation should be consistent when removing denominators; after the system fills the preceding transition content, the above vouchers, and the current knowledge node identifier into the learning interaction generation template, it can obtain candidate learning interaction content, including knowledge introduction content, knowledge explanation content, relation explanation content, and practice questions; then, based on the text vouchers associated with the knowledge node, it performs voucher consistency verification on the candidate learning interaction content, deletes sentences lacking supporting evidence, obtains the verified learning interaction content, and writes it into the learning interaction content record for users to answer later.

[0068] In one implementation, the purpose of collecting user responses and interaction records related to learning content, and updating the mastery probability of the current knowledge node based on these responses and records, is to convert the user's learning behavior towards the current knowledge node into a mastery probability that can represent the current mastery status of the knowledge node. The implementation steps include: Step 4031: Read the current knowledge node identifier, current execution position, exercise content, answer input items, and feedback prompts from the learning interaction content record, and output the learning interaction content to the user.

[0069] Step 4032: Collect user responses and interaction records related to the learning interaction content; the response results should include at least the answer content, submission time, and answer completion status; the interaction records should include at least the duration of time spent on the knowledge explanation content, the duration of time spent answering practice questions, the number of modifications, the number of times the prompts were viewed, and the number of retries.

[0070] Step 4033: Perform result determination on the answer; the method of result determination includes: comparing the answer content with the preset reference answer or preset scoring rules to obtain the correct result, partially correct result or incorrect result; and then writing the answer completion status and submission time into the answer result record.

[0071] Step 4034: Perform behavior quantification processing on the interaction record; the behavior quantification processing method includes: generating dwell time item results based on dwell time, generating response time item results based on response time, generating modification item results based on number of modifications, generating prompt item results based on number of times prompts are viewed, generating retry item results based on number of retries, and writing each result into the interaction quantification record.

[0072] Step 4035: Pre-construct mastery probability update rules; the construction method of mastery probability update rules includes: pre-setting the hierarchical results of answer result items, pause items, answer duration items, modification items, prompt items and retry items; then specifying the incremental direction and decrement direction corresponding to each hierarchical result, and forming the mastery probability update order for the same knowledge node.

[0073] Step 4036: Based on the answer result record and interaction quantification record, update the mastery probability of the current knowledge node according to the mastery probability update rule; during the update, first read the mastery probability of the current knowledge node in the previous round in the historical learning record; then write the current round increase / decrease result according to the mastery probability update rule; then combine the previous round mastery probability with the current round increase / decrease result to obtain the mastery probability update result of the current knowledge node.

[0074] Step 4037: Write the current knowledge node identifier, current execution position, answer result record, interaction quantification record, and mastery probability update result into the mastery status record.

[0075] As another example, to illustrate the mastery probability update process, suppose a user completes the exercises for the current knowledge node, and the answer is judged to be partially correct. The interaction log shows that the user spent a relatively long time on the knowledge explanation, viewed prompts frequently, but retried relatively few times. The system first generates the corresponding answer result item based on the answer result, and then generates the dwell time, prompts, and retry results in the interaction quantification log based on the dwell time, number of prompts viewed, and number of retryes, respectively. Subsequently, the system reads the previous round's mastery probability of the current knowledge node in the historical learning record and, according to the mastery probability update rules, performs combined processing on each sub-item result of the current round to obtain the updated mastery probability result of the current knowledge node. If the current round result indicates that the user has not yet reached a stable mastery state of the knowledge node, this updated mastery probability result is used for pre-processing or sequence maintenance in subsequent trigger path reconstruction.

[0076] Step S5: Reconstruct the target learning path sequence based on the mastery probability of the current knowledge node, insert the predecessor knowledge node of the current knowledge node or remove the knowledge node with lower difficulty than the current knowledge node in the same concept cluster, and output the updated learning path sequence.

[0077] In one implementation, the purpose of reconstructing the target learning path sequence based on the mastery probability of the current knowledge node is to adjust the subsequent learning order according to the mastery status of the current knowledge node, so that the target learning path sequence is consistent with the learning result corresponding to the current knowledge node; the implementation steps include: Step 5011: Read the current knowledge node identifier, current execution position, and mastery probability update result from the mastery status record, and read the content of subsequent knowledge nodes located after the current execution position in the target learning path sequence.

[0078] Step 5012: Pre-construct path reconstruction rules; the method for constructing path reconstruction rules includes: pre-setting the mastery probability classification results of low mastery level, medium mastery level and high mastery level; specifying pre-filling processing for low mastery level, specifying order-preserving processing for medium mastery level, and specifying low-difficulty removal processing for high mastery level; and specifying the node retrieval order, insertion position and deletion range corresponding to each processing to form path reconstruction rules.

[0079] Step 5013: Compare the obtained probability update results with the path reconstruction rules to obtain path reconstruction instructions; the path reconstruction instructions include at least one of the following: pre-fill instructions, order preservation instructions, and low-difficulty removal instructions.

[0080] Step 5014: Retrieve reconstruction objects in the teaching knowledge graph and the target learning path sequence according to the path reconstruction instructions; the reconstruction object corresponding to the pre-fill instruction is the pre-knowledge node of the current knowledge node; the reconstruction object corresponding to the low difficulty removal instruction is the low difficulty knowledge node in the target learning path sequence that belongs to the same concept cluster as the current knowledge node and whose sequential position is after the current execution position; the reconstruction object corresponding to the order preservation instruction is empty.

[0081] Step 5015: Write the path reconstruction instruction, reconstruction object identifier, and current execution position into the path reconstruction record.

[0082] In one implementation, inserting a predecessor knowledge node of the current knowledge node or removing a knowledge node within the same concept cluster that is less difficult than the current knowledge node, and outputting an updated learning path sequence, aims to perform specific adjustments on the target learning path sequence according to the path reconstruction record to obtain the updated learning path sequence; the implementation steps include: Step 5021: Read the path reconstruction instructions, reconstruction object identifiers and current execution positions from the path reconstruction records, and read all sequential position records from the target learning path sequence.

[0083] Step 5022: When the path reconstruction instruction is a pre-implementation instruction, the pre-knowledge node is retrieved based on the pre-dependency relationship of the current knowledge node in the teaching knowledge graph, and the pre-knowledge node is checked in the target learning path sequence to see if it already exists before the current execution position. If it does not exist, the pre-knowledge node is inserted before the current execution position. If it already exists, the original order position remains unchanged. When inserting, the original text fragment index associated with the pre-knowledge node is retrieved synchronously, and a new order position is generated for the inserted content.

[0084] Step 5023: When the path reconstruction instruction is a low-difficulty removal instruction, the same concept cluster identifier is retrieved based on the current knowledge node identifier, and knowledge nodes within the same concept cluster that are in the order position after the current execution position are selected in the target learning path sequence; then, knowledge nodes with a difficulty level lower than the current knowledge node are identified as knowledge nodes to be removed, and the order position record corresponding to the knowledge nodes to be removed is deleted from the target learning path sequence.

[0085] Step 5024: When the path reconstruction instruction is a sequence-preserving instruction, the sequence position records in the target learning path sequence are kept unchanged.

[0086] Step 5025: Re-sort the order of the target learning path sequence after the insertion or removal processing, and synchronously update the correspondence between the knowledge node identifier, the transition content identifier, and the execution status marker to obtain the updated learning path sequence.

[0087] Step 5026: Output the updated learning path sequence and write the updated learning path sequence back to the path version record corresponding to the historical learning record, for subsequent determination of the next current knowledge node.

[0088] As another example, if the current knowledge node involves solving a linear equation with denominators and the mastery probability update corresponds to a low mastery level, the system generates a pre-emptive insertion instruction based on the path reconstruction rules and searches for the pre-emptive knowledge node in the teaching knowledge graph. If the retrieved pre-emptive knowledge node involves removing parentheses and the parentheses do not yet exist before the current execution position, the parentheses are inserted before the current execution position, and the order of the target learning path sequence is rearranged. Conversely, if the mastery probability update of the current knowledge node corresponds to a high mastery level, and there are still knowledge nodes in the target learning path sequence after the current execution position that belong to the same concept cluster as the current knowledge node and have a lower difficulty level, the system generates a low-difficulty removal instruction and deletes the corresponding knowledge node from the target learning path sequence. In this way, the updated learning path sequence can be consistent with the user's actual mastery status of the current knowledge node.

[0089] Example 2 See Figure 2As shown, this embodiment provides an adaptive learning path generation system based on knowledge graphs and large language models. Since this system uses an adaptive learning path generation method based on knowledge graphs and large language models from Embodiment 1, it has the same effect, and will not be described again here. The system includes: The intent parsing module is used to receive natural language input from users, read static user profile data and historical learning records, and parse and generate structured intent results that include learning objectives and learning constraints. The subgraph construction module is used to retrieve target knowledge nodes from the pre-built teaching knowledge graph based on the target structured intent result. Based on the pre-dependency, inclusion and advancement relationships between knowledge nodes in the teaching knowledge graph, it traverses from the target knowledge node to the mastered boundary node corresponding to the historical learning record, and then performs expansion and pruning to obtain the exclusive knowledge subgraph. The path generation module is used to perform weighted topological sorting based on the exclusive knowledge subgraph to obtain an initial learning path sequence, detect the semantic coherence of adjacent knowledge nodes in the initial learning path sequence, generate logical transition content based on the original text fragments associated with adjacent knowledge nodes with insufficient semantic coherence, and insert them into the initial learning path sequence to obtain the target learning path sequence. The interactive evaluation module is used to determine the current knowledge node based on the target learning path sequence, extract the knowledge node-related text credentials associated with the current knowledge node, generate learning interaction content based on the knowledge node-related text credentials, collect the user's answer results and interaction records in response to the learning interaction content, and update the mastery probability of the current knowledge node based on the answer results and interaction records. The path reconstruction module is used to reconstruct the target learning path sequence based on the mastery probability of the current knowledge node, insert the predecessor knowledge node of the current knowledge node or remove the knowledge node with lower difficulty within the same concept cluster, and output the updated learning path sequence.

[0090] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. An adaptive learning path generation method based on knowledge graphs and large language models, characterized in that, include: Step S1: Receive user natural language input, read user static profile data and historical learning records, and parse to generate a target structured intent result including learning objectives and learning constraints; Step S2: Based on the target structured intent result, the target knowledge node is retrieved from the pre-constructed teaching knowledge graph. Based on the pre-dependency, inclusion and advancement relationships between the knowledge nodes in the teaching knowledge graph, the target knowledge node is traversed to the mastered boundary node corresponding to the historical learning record. Then, expansion and pruning are performed to obtain the exclusive knowledge subgraph. Step S3: Perform weighted topological sorting based on the exclusive knowledge subgraph to obtain the initial learning path sequence. Detect the semantic coherence of adjacent knowledge nodes in the initial learning path sequence. Generate logical transition content based on the original text fragments associated with adjacent knowledge nodes with insufficient semantic coherence, and insert them into the initial learning path sequence to obtain the target learning path sequence. Step S4: Determine the current knowledge node based on the target learning path sequence, extract the knowledge node-related text credentials associated with the current knowledge node, generate learning interaction content based on the knowledge node-related text credentials, collect the user's answer results and interaction records in response to the learning interaction content, and update the mastery probability of the current knowledge node based on the answer results and interaction records. Step S5: Reconstruct the target learning path sequence based on the mastery probability of the current knowledge node, insert the predecessor knowledge node of the current knowledge node or remove the knowledge node with lower difficulty than the current knowledge node in the same concept cluster, and output the updated learning path sequence.

2. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 1, characterized in that, Methods for parsing and generating structured intent results that include learning objectives and learning constraints include: The system standardizes the input text by unifying terminology, time expressions, and removing irrelevant words from the user's natural language input. It then uses the standardized input text, user static profile data, and historical learning records to analyze the input intent using a large language model. The target subject and target knowledge scope are extracted as learning objectives, and learning scope constraints and learning depth constraints are extracted as learning constraints. Based on the required field set and allowed value set, the system performs field integrity checks and calls a pre-defined terminology mapping table to convert synonyms into unified field values, resulting in a structured intent result that includes learning objectives and learning constraints.

3. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 1, characterized in that, Methods for obtaining target knowledge nodes include: In the teaching knowledge graph, knowledge node identifiers, knowledge names, knowledge definitions, original text fragment indexes, and subject information are pre-recorded, along with prerequisite dependencies, inclusion relationships, and progression relationships. The target subject and target knowledge scope are read from the target structured intent result, and matching is performed on the knowledge name, knowledge definition, and original text fragment respectively. Knowledge nodes with consistent subject and knowledge scope falling within the target knowledge scope are retained to obtain the target knowledge node set.

4. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 2, characterized in that, Methods for obtaining a dedicated knowledge subgraph include: The knowledge node identifiers in the historical learning records are mapped to the teaching knowledge graph. Based on the correctness of answers, the number of times completed and reviewed, and in combination with the preset mastery judgment rules, the mastered knowledge nodes are determined. Then, the mastered boundary nodes that are connected to the target knowledge node set through pre-dependent relationships are identified. Starting from the target knowledge node set, the system traverses forward along the pre-dependent relationships to the mastered boundary nodes, and performs expansion along the inclusion and advancement relationships. Finally, pruning is performed based on the learning scope constraints and learning depth constraints to obtain the exclusive knowledge subgraph.

5. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 1, characterized in that, Methods for obtaining the initial learning path sequence include: Using the pre-dependencies in the dedicated knowledge subgraph as order constraints, the number of incoming edges for each knowledge node is counted. A preset sorting rule table is called to convert the correspondence between the knowledge node and the learning objective, the number of error records of the knowledge node in the historical learning record, and the relational distance from the knowledge node to the mastered boundary node into sorting sub-values. The sorting sub-values ​​are combined to obtain the sorting weight. Knowledge nodes with zero incoming edges are selected to form a candidate knowledge node set, and sorted according to the sorting weight to obtain the initial learning path sequence.

6. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 5, characterized in that, Methods for obtaining the target learning path sequence include: The system reads adjacent knowledge nodes in the initial learning path sequence in sequence, extracts knowledge names, knowledge definitions, relation tags, and original text fragments; generates semantic coherence indicators based on the number of overlapping terms, the continuity of relation tags, and the overlap of subject terms, combined with preset coherence calculation rules; for adjacent knowledge nodes with insufficient semantic coherence, it generates logical transition content based on preset generation templates, and inserts it into the initial learning path sequence after citation verification to obtain the target learning path sequence.

7. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 1, characterized in that, Methods for extracting the associated text credentials of the current knowledge node include: The system reads the sequential position, content type, and execution status marker in the target learning path sequence, and identifies the knowledge node corresponding to the sequential position of the incomplete learning interaction and the content type as a knowledge node as the current knowledge node. Then, based on the current knowledge node, it extracts the knowledge name, knowledge definition, prerequisite dependencies, inclusion relationships, progression relationships, and original text fragment index in the teaching knowledge graph. According to the preset voucher assembly rules, it writes the text fragment used to explain the meaning of the current knowledge node into the definition voucher item, the text fragment used to represent the prerequisite dependencies, inclusion relationships, and progression relationships into the relationship voucher item, the text fragment used to represent the application scenario of the current knowledge node into the example voucher item, and the text fragment used to represent the applicable scope or usage conditions of the current knowledge node into the constraint voucher item, thus obtaining the knowledge node associated text voucher.

8. The adaptive learning path generation method based on knowledge graphs and large language models according to claim 7, characterized in that, Methods for updating the mastery probability of the current knowledge node include: Based on the definition voucher, relation voucher, example voucher, and constraint voucher items in the knowledge node associated text voucher, and based on the preceding and following transitional content of the current knowledge node in the target learning path sequence, candidate learning interaction content is generated according to the learning interaction generation template; after performing voucher consistency verification on the candidate learning interaction content, the learning interaction content is output; the user's answer results and interaction records generated in response to the learning interaction content are collected, and according to the preset mastery probability update rule, the quantitative results corresponding to the answer results, dwell time, answer duration, number of modifications, number of times of viewing prompts, and number of retries are combined to update the mastery probability of the current knowledge node.

9. The adaptive learning path generation method based on knowledge graph and large language model according to claim 1, characterized in that, Methods for reconstructing the target learning path sequence include: The mastery probability of the current knowledge node is read and compared with the preset path reconstruction rules to obtain the preceding insertion instruction, the order preservation instruction, or the low-difficulty removal instruction. When the preceding insertion instruction is in effect, the preceding knowledge node is retrieved based on the preceding dependency relationship of the current knowledge node in the teaching knowledge graph, and it is checked whether the preceding knowledge node is located before the current knowledge node in the target learning path sequence. If it is not located before the current knowledge node, insertion is performed. When the low-difficulty removal instruction is in effect, knowledge nodes in the same concept cluster in the target learning path sequence that are lower in difficulty than the current knowledge node are deleted. Then the order positions are rearranged to obtain the updated learning path sequence.

10. An adaptive learning path generation system based on knowledge graphs and large language models, used to implement the adaptive learning path generation method based on knowledge graphs and large language models as described in any one of claims 1-9, characterized in that, The system includes: The intent parsing module is used to receive user natural language input, read user static profile data and historical learning records, and parse and generate a structured intent result that includes learning objectives and learning constraints. The subgraph construction module is used to retrieve target knowledge nodes from the pre-built teaching knowledge graph based on the target structured intent result. Based on the pre-dependency, inclusion and advancement relationships between knowledge nodes in the teaching knowledge graph, it traverses from the target knowledge node to the mastered boundary node corresponding to the historical learning record, and then performs expansion and pruning to obtain the exclusive knowledge subgraph. The path generation module is used to perform weighted topological sorting based on the exclusive knowledge subgraph to obtain an initial learning path sequence, detect the semantic coherence of adjacent knowledge nodes in the initial learning path sequence, generate logical transition content based on the original text fragments associated with adjacent knowledge nodes with insufficient semantic coherence, and insert them into the initial learning path sequence to obtain the target learning path sequence. The interactive evaluation module is used to determine the current knowledge node based on the target learning path sequence, extract the knowledge node-related text credentials associated with the current knowledge node, generate learning interaction content based on the knowledge node-related text credentials, collect the user's answer results and interaction records in response to the learning interaction content, and update the mastery probability of the current knowledge node based on the answer results and interaction records. The path reconstruction module is used to reconstruct the target learning path sequence based on the mastery probability of the current knowledge node, insert the predecessor knowledge node of the current knowledge node or remove the knowledge node with lower difficulty within the same concept cluster, and output the updated learning path sequence.

Images