Knowledge-context-subject trinity-based evolution method for constructing knowledge space

Abstract

This invention discloses a three-in-one, evolvable knowledge space construction method integrating "knowledge context and subject," relating to the fields of smart education, artificial intelligence, big data, and knowledge graph technology. First, this invention constructs a weighted knowledge dependency graph in this field and maps each knowledge concept in the discrete graph to a measurable vector space based on a "weighted geometric preservation" strategy. Next, it initializes the learner's cognitive state vector set and dynamically determines the optimal starting point and plans the learning path based on the learner's cognitive boundaries and task objectives through full-scale pre-playback. Then, it monitors the learner's interactive behavior feedback as they sequentially learn each knowledge node according to the learning path. When cognitive obstacles are detected, generative artificial intelligence is used to dynamically reconstruct learning resources. Finally, it updates the graph and vector space by comprehensively considering the characteristics of group learning behavior, achieving an adaptive mapping from "behavioral data" to "spatial structure." This invention provides a new perspective for the design of education systems.

Description

Technical Field

[0001] This invention relates to the fields of smart education, artificial intelligence, big data, and knowledge graph technology, specifically to a kind of "knowledge" Context A three-in-one, evolving knowledge space construction method for the "subject". Background Technology

[0002] With the continuous development of information technologies such as artificial intelligence, big data, and knowledge graphs, smart education is gradually shifting from standardized teaching models to personalized learning models aimed at meeting individual learning needs, promoting the formation of a new educational ecosystem characterized by human-machine collaboration. Supported by these technological conditions, intelligent learning environments that analyze and respond to learning needs are gradually becoming feasible, and the application scope of smart learning is continuously expanding.

[0003] However, as the practice of smart education continues to deepen, many problems still urgently need to be addressed. On the one hand, the knowledge systems involved in the learning content are generally fragmented, multimodal, and dynamically evolving, with complex and ever-changing internal connections between knowledge points. Learners easily find themselves struggling to effectively select and integrate information, leading to difficulties in knowledge organization and comprehension. In this situation, the main problem faced by learners is no longer insufficient information acquisition, but rather the difficulty in effectively organizing, understanding, and applying existing information.

[0004] On the other hand, in the process of deep integration of intelligent technology and educational applications, in addition to technical implementation factors, issues such as the adaptation of educational theories and application standards also arise. Existing smart education systems are mostly designed based on static content presentation or fixed knowledge structures, making it difficult to adapt to the continuous evolution of knowledge systems and the dynamic changes in the learning process, thus affecting the effectiveness of learning support and teaching.

[0005] Therefore, how to overcome the limitations of static content delivery and fixed knowledge representation while ensuring technical feasibility, and build a smart education technology solution that can adapt to dynamic changes in knowledge and support individual learning processes, remains a key issue to be addressed in existing technologies.

[0006] Currently, the field of smart education has conducted extensive research and made some progress in areas such as knowledge structure modeling, learning context modeling, and learning subject modeling. However, from a holistic system perspective, existing technologies generally suffer from fragmented modeling perspectives and insufficient collaborative mechanisms, making it difficult to support the dynamic control needs of complex learning processes.

[0007] Specifically, in knowledge-based modeling research, existing knowledge space theories and their probabilistic extension models mainly focus on the logical dependencies between knowledge points. Their modeling methods are mostly static or limited dynamic, making it difficult to depict the dynamic activation, evolution, and reorganization characteristics of knowledge generated by learning behaviors during the learning process. In context-based learning research, the modeling granularity of the learner's cognitive state is relatively coarse, making it difficult to support accurate diagnosis of the learning state and effective integration with the knowledge structure. Regarding learner modeling, existing methods are mostly based on static states or discrete time points, making it difficult to continuously reflect the dynamic evolutionary characteristics exhibited by the learner during task execution and environmental interaction.

[0008] Furthermore, existing smart education frameworks often focus on single-dimensional or two-dimensional interaction mechanisms, with knowledge structures, learning contexts, and learning subjects frequently modeled separately, lacking cross-level collaborative descriptions and coordinated control mechanisms within a unified framework. This fragmented modeling approach struggles to adapt to the real-world scenarios of continuous knowledge system evolution and dynamic changes in learning needs, thus limiting the adaptability and evolvability of smart education systems.

[0009] Therefore, existing technologies urgently need to address the following technical issues: how to achieve collaborative modeling of knowledge structure, learning context, and learning subject within a unified framework; how to characterize the dynamic evolution of knowledge state and cognitive state of learning subject during the learning process; and how to dynamically adjust learning content and learning path based on the above collaborative modeling results to enhance the personalized support capabilities and overall learning effectiveness of the smart education system. Summary of the Invention

[0010] In view of this, the present invention provides a "knowledge" Context The "subject-based" three-in-one evolutionary knowledge space construction method constructs a unified framework to dynamically and collaboratively represent the evolution of knowledge structure, changes in learning context, and cognitive development of the learning subject. Based on this, it enables the autonomous optimization and dynamic control of learning content and paths, thereby enhancing the personalized support capabilities and overall teaching effectiveness of the smart education system.

[0011] The present invention provides a three-in-one, evolvable knowledge space construction method based on "knowledge-context-subject," comprising:

[0012] S1, Construct a weighted knowledge dependency graph , where the set of nodes Represents core knowledge concepts within the domain; edge set Indicates at time Cognitive dependencies or prerequisite relationships between knowledge concepts; weight set This indicates the strength of cognitive support between concepts; S2, map Various knowledge concepts Mapped to a dimensional embedding vector Following the "semantic-topological dual constraint" strategy, the full knowledge space is obtained. ; S3, Constructing a set of mastered knowledge nodes based on learner identity recognition and ability assessment results. and in The text appears to be a jumbled collection of phrases and sentences, seemingly from different sources. A coherent translation isn't possible without the full context. Each node Embedded vector With unique identifier Combining into tuples Injecting into the learner's initial knowledge space middle; S4, identify the learner's learning objective knowledge nodes and starting knowledge nodes, and plan the learning path; specifically including: S41, Calculate the target semantic vector corresponding to the learner's learning needs. With the full knowledge space The similarity of all embedded vectors is used to select embedded vectors with similarity values ​​greater than a set threshold. This constitutes the set of candidate learning target embedding vectors. ; S42, Construct the global distance matrix , where matrix elements For the first The embedding vector of the candidate learning objective and the learner's first... The distance between the embedding vectors of the knowledge nodes that have been mastered; for the th The optimal dynamic anchor point of each candidate learning target embedding vector. Global distance matrix No. minimum value in row The corresponding embedded vector of the knowledge points already mastered; S43, Regarding candidate learning objectives If its Generate quick test instructions and update the embedded vectors of candidate learning objectives that pass the test to the already mastered knowledge space. In the middle, those that fail are moved into the valid candidate set. ;like Then the embedding vector of the candidate learning objective is moved into the effective candidate set. The learning objective embedding vector with the highest score is selected as the learning objective vector corresponding to that learning requirement. The corresponding optimal dynamic anchor point is the learning starting point vector. ; S44, learn the target vector and learning starting point vector Back-mapping to knowledge graph In the process, the learning target node is obtained. and learning starting point node ; Using path planning methods, in the graph The planning starts from the learning starting point node To the learning target node The path; S5, sequentially activate each node on the path described in S4; monitor the learner's progress at the current knowledge node. On-screen interactive feedback: If cognitive blockage is detected, generative artificial intelligence is used to reduce the learning difficulty coefficient. and increase the strength of scaffolding In this scenario, learning resources are dynamically reconstructed and adapted to drive learners to perform iterative interactions within the reduced cognitive context. If a learner fails to master a knowledge node under the conditions of maximum scaffolding strength or maximum number of attempts, it is marked as a failure, and the process returns to S44 to replan the path, avoiding the knowledge node and generating an alternative path pointing to the predecessor remedy node. If the learner successfully masters the knowledge node, the current knowledge node is... Add to the set of knowledge nodes already mastered In the knowledge space, the embedding vector corresponding to the knowledge node is added. middle; S6, Dependency graph based on group learning behavior characteristics Perform topology reconstruction: Specifically: Extract all flow records within the cycle from all learners' learning logs, in binary form. Use the topological key to group and aggregate records; For each group of flows Statistical analysis of the total frequency of this transaction. Calculate the throughput of this path. and the average cost of successfully passing ;based on , and Update edge set and weight set ; Based on the updated topology Reconstruct the entire knowledge space according to the S2 approach. and the knowledge space already mastered .

[0013] Preferably, in step S1, initial relational connections between concepts are constructed based on the experience of subject matter experts, the curriculum syllabus, and the structure of standard textbooks. .

[0014] Preferably, in step S2, a loss function is constructed. This is used to measure the spatial metric between node embedding vectors. By optimizing the loss function, the optimal embedding vector for each node is obtained.

[0015] Preferably, in step S2, a loss function is constructed, and an optimization algorithm is used to optimize the embedding vector of each node; wherein the loss function is:

[0016]

[0017]

[0018] in, For knowledge nodes and The strength of cognitive support between them; For knowledge nodes , The embedding vector; For knowledge nodes The corresponding initial semantic vector; These are the topology constraint weight parameters; The optimization algorithm employs gradient descent, stochastic gradient descent, momentum gradient descent, or adaptive learning rate optimization.

[0019] Preferably, in step S43, the following scoring function is used:

[0020] in, For the first Embedded vectors of candidate learning objectives The rating, For similarity calculation; The optimal cognitive span constant is set; These are the weighting coefficients.

[0021] Preferably, in S44, the path planning method adopts the A* algorithm, Dijkstra's algorithm, breadth-first search algorithm, or depth-first search algorithm.

[0022] Preferably, in S5, the learning difficulty coefficient The initial value is 1, scaffold strength The initial value is 0; Generative artificial intelligence will learn the difficulty level According to attenuation factor Scaling is applied to adjust the scaffolding strength. Increase in increments of 1 step.

[0023] Preferably, in S6, based on , and Update edge set and weight set Specifically: For edge set The existing edges in If the total interaction frequency And the pass rate And average cost If the path is determined to be an efficient cognitive path, then the edge is increased. Weighting; if the total interaction frequency However, the pass rate or average cost If the path is determined to be a cognitively hindered path, then the number of edges should be reduced. The weights; For edge set The relationship that does not exist in If the total interaction frequency And the pass rate Higher than the admission threshold Then in the edge set Create new directed edges New directed edge The weight of its pass rate ; Iterate through the updated edge weights, if a certain edge weight... Less than a tiny threshold Then remove it from the edge set. If an edge is removed from a knowledge cluster, and that edge connects two frequently accessed knowledge clusters, a "semantic gap" alert signal is generated, indicating that a gap needs to be filled. and Transitional knowledge nodes are inserted into the geometric midpoint region.

[0024] A better approach is to integrate efficient cognitive paths. The weights are updated to ; Cognitive blocking path The weights are updated to ; in, The preset gain coefficient, .

[0025] This invention provides a system employing the above method, comprising: The knowledge space geometry construction module is used to construct a weighted knowledge dependency graph and transform the weighted knowledge dependency graph into a vector space with metric properties, maintaining the topological structure of the space based on the edge weights of the graph. The subject cognitive coordinate tracking module is used to construct the learner's initial set of mastered knowledge nodes. and the corresponding initial knowledge space. And update the set based on the interaction results of the cognitive context interaction module. and space ; A geometric adaptive planning engine is used to determine the learner's learning target knowledge nodes and learning start knowledge nodes, and to plan the learning path; The cognitive context interaction module interacts with learners, monitoring their behavioral feedback as they sequentially activate nodes along the learning path planned by the geometric adaptive planning engine. When cognitive obstacles are detected, generative artificial intelligence is used to reduce the learning difficulty level. and / or increase scaffold strength In this context, learning resources are dynamically reconstructed and adapted to drive learners to perform iterative interactions in the reduced cognitive context. The collaborative evolution control module is used to drive the bidirectional, iterative evolution of the knowledge graph structure and geometric space based on the group interaction data flow.

[0026] Beneficial effects: This invention provides a "knowledge" Context The method for constructing and adaptively controlling the "subject-based" three-in-one evolvable knowledge space has the following beneficial effects: (1) The "knowledge" described in this invention Context The "subject-based" three-in-one evolutionary knowledge space construction and adaptive regulation method establishes a micro-level context-adaptive intervention mechanism, ensuring the continuity of the learning process and further improving the problem that static teaching resources are difficult to adapt to dynamic cognitive abilities in existing technologies.

[0027] (2) The "knowledge" described in this invention Context The construction and adaptive regulation method of the "subject" three-in-one evolutionary knowledge space establishes a data-driven two-way evolution mechanism for the space. The system can automatically strengthen effective connections and distance high-resistance connections based on group interaction data, and automatically detect knowledge logic gaps, so that the knowledge space has the ability to continuously iterate and self-repair with the collective wisdom.

[0028] (3) The "knowledge" described in this invention Context The construction and adaptive control method of the "subject" three-in-one evolvable knowledge space adopts incremental relaxation optimization and ID anchoring technology, which ensures the engineering robustness of large-scale systems. Under the premise of ensuring the physical stability of the coordinate system, the computational complexity of spatial reconstruction is reduced from the scale of the whole graph to the scale of a local subgraph, effectively supporting real-time response in massive concurrent scenarios.

[0029] (4) The "knowledge" described in this invention Context The "subject-based" three-in-one evolving knowledge space construction and adaptive control method realizes the coordinated control of multi-dimensional elements under a unified geometric framework, breaking the status quo of fragmented modeling of various modules in traditional education systems. This isomorphic modeling enables the system to achieve coordinated linkage from macro-path planning to micro-parameter adjustment in a closed-loop process, significantly improving the adaptability and teaching effectiveness of the smart education system. Attached Figure Description

[0030] Figure 1 This is a diagram illustrating the synergistic evolution of the "knowledge-context-subject" trinity in this invention.

[0031] Figure 2 This is the overall flowchart of the present invention.

[0032] Figure 3 This is a system architecture block diagram of the present invention.

[0033] Figure 4 This is a schematic diagram illustrating the popular embedding of knowledge space in this invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0035] This invention provides a "knowledge" Context The "subject-based" three-in-one evolving knowledge space construction method, such as Figure 1 and Figure 2 As shown, the specific steps include the following: Step 1: Construct a continuous knowledge space based on graph weights.

[0036] This step aims to construct a continuous semantic space that can represent the dynamic dependencies between knowledge, including the following sub-steps: Step 1.1: Construct a weighted knowledge dependency graph.

[0037] Define a dynamically weighted directed graph This serves as the discretized logical foundation of knowledge structure. Among them, the node set... Represents core knowledge concepts within the domain. Edge set. Indicates at time Cognitive dependencies or prerequisite relationships between knowledge concepts, initial relational connections. This refers to the topological connections of a discrete directed graph, based on knowledge space theory, which integrates the experience of subject matter experts, curriculum outlines, and the structure of standard textbooks. Weight set. For each edge Assign a positive real weight This value quantifies the concept. On the concept The strength of cognitive support. Initial weights. The initial configuration is based on a pre-defined domain standard parameter library, which is constructed by quantitatively analyzing the curriculum standard syllabus or historical teaching data.

[0038] Step 1.2: Construct the knowledge space based on the vector space mapping model with semantic-topological dual constraints.

[0039] To achieve knowledge reasoning based on geometric operations, discrete graphs are used. Each knowledge concept in Mapped to a dimensional embedding vector The vector representations of all nodes constitute the full knowledge space. In the mapping process, the system simultaneously introduces node semantic feature constraints and graph topology constraints to construct a vector space mapping model with semantic-topology dual constraints. Semantic feature constraints maintain the semantic consistency of knowledge nodes in the vector space, preventing the embedded vectors from deviating from their original semantic distribution during optimization. Topology constraints maintain the cognitive dependencies between nodes in the knowledge graph, ensuring that nodes with strong dependencies in the graph have closer geometric distances in the vector space. These semantic-topology dual constraints are implemented by constructing and optimizing a joint objective function to solve for the embedded vectors corresponding to each knowledge node.

[0040] In a preferred embodiment of the present invention, the topology constraint term is defined as:

[0041] in, Representing knowledge nodes and The dependency weights are used to adjust the strength of the distance constraint between the corresponding node embedding vectors.

[0042] The semantic feature constraint term is defined as follows:

[0043] in, For knowledge nodes The corresponding initial semantic vector.

[0044]

[0045] in, The topological constraint weight parameter is used to adjust the relative influence of topological constraints and semantic constraints during the vector space construction process.

[0046] In this preferred embodiment, the system first initializes the corresponding embedding vector for each knowledge node. The initialization method can be random initialization or initialization based on a pre-trained model. Then, an iterative optimization algorithm is used to optimize and update the joint objective function until a preset convergence condition is met, ultimately obtaining a full knowledge space that simultaneously reflects the knowledge topological dependencies and semantic structural features. .

[0047] Step 2: Initialize the learner's cognitive state vector set.

[0048] New users first need to create a personal profile through two steps: identity setting and competency assessment, to generate an initial data file. Record their identity information and ability assessment results.

[0049] This step aims to parse the structured assessment data package provided by the learner. In continuous knowledge space In this process, a precise, complete, and logically consistent "knowledge space" is constructed. This serves as the initial geometric representation of the learner's cognitive state. The space... The embedding vector of the storage node and the unique identifier (ID) of the knowledge node are used in conjunction with a real-time mapping table to adapt to the dynamic changes of the entire knowledge space.

[0050] Specifically, it includes the following sub-steps: Step 21: Attribute-driven subspace construction and baseline knowledge injection.

[0051] According to learners Submitted structured evaluation data package First, analyze its identity information and construct a cognitive attribute vector. This vector contains discrete labels such as academic stage, major, and target level. Based on It calls a pre-defined domain benchmark table to determine the set of knowledge nodes that the identity attribute has mastered by default. Subsequently, Each node Embedded vector With unique identifier Combining into tuples Inject the learner's initial set of knowledge. In this way, the baseline knowledge state initialization based on group attributes is completed.

[0052] Step 22: Potential capability level analysis and incremental coordinate mapping.

[0053] Further analysis Extract the passed items from the ability assessment results and place them in the full knowledge space. Locate the corresponding set of knowledge nodes in the middle. This completes the initial capability assessment. Next, [the process will proceed as follows]. Each node Embedded vector with identifier Combining into tuples Added as incremental information to the existing knowledge set. middle.

[0054] Step 23: Logical backtracking expansion based on topological dependencies.

[0055] To ensure the connectivity of cognitive boundaries, for the set of knowledge already mastered The corresponding set of knowledge nodes In the map Perform predecessor transitive deduction to correct boundary discontinuities caused by limited sampling of assessment items. Traverse For each node in the graph, retrieve its upstream nodes (i.e., strong logical predecessor nodes, referring to nodes that possess the prior knowledge objectively necessary to master the current node) that have a strong cognitive dependency relationship with it, and generate tuple data corresponding to these upstream nodes. Automatically append to collection Remove duplicates from the middle.

[0056] By backtracking the topological dependency logic of the graph, the initial state represented by the discretized experimental results is elevated to a cohesive and complete cognitive base in terms of dependency, realizing the leap from experimental explicit mastery to logically complete mastery of state modeling.

[0057] Step 3: Adaptive planning based on "dynamic anchor points".

[0058] This step dynamically determines the optimal starting point and plans the learning path based on the learner's cognitive boundaries and task objectives through full-scale pre-playing.

[0059] Step 31: Target selection based on vector similarity.

[0060] Analyzing each learner's learning needs Based on a text representation method consistent with the semantic features of knowledge nodes, a corresponding target semantic vector is generated. Subsequently, the target vector and the full knowledge space are calculated. All embedding vectors in The similarity can be calculated using cosine similarity, Euclidean distance, Manhattan distance, Chebyshev distance, inner product, etc. This embodiment uses cosine similarity.

[0061] Selecting similarity values ​​higher than a preset threshold Embedded vector This constitutes a preliminary set of candidate learning objectives. .

[0062] Step 32: Parallel Dynamic Anchoring Based on Matrix Operations The system does not make assumptions about a single objective, but rather about a set. Each candidate learning objective Calculate the optimal entry point for each: First, construct the bidirectional mapping matrix: First, from the set of candidate learning objectives Extract the embedding vector of each candidate learning target and stack them row by row to construct the target feature matrix. (in The number of candidate learning objectives. (For vector dimensions). Simultaneously, it considers the learner's current knowledge set. Extract the embedding vectors of all mastered knowledge nodes and construct a cognitive state matrix. (in (Number of knowledge nodes already mastered).

[0063] Next, the global distance tensor is computed in parallel: Parallel computing using broadcast mechanisms or matrix operations Each row vector in the vector and The spatial metric (such as Euclidean distance) between each row vector in the matrix is ​​used to generate a global distance matrix. Elements in the matrix Precisely characterized the first The candidate learning objective and the first Geometric proximity (i.e. cognitive span) between mastered knowledge points.

[0064] Then, row reduction and optimal anchor point locking: For the global distance matrix Perform a row-dimension minimization reduction operation. For each row of the matrix... (representing the first) (1 candidate learning objective), locate the minimum value in this row. and its corresponding column index The index of this column points to... The nodes in the data are identified as the optimal dynamic anchor points for the candidate learning objective. .

[0065] Finally, the output is a set of mapping pairs:

[0066] Step 33: Task adaptability verification based on cognitive triage strategy.

[0067] This step receives the set of mapping pairs output from step 32. Based on geometric proximity The candidate learning objectives are divided into three cognitive intervals, and the optimal learning path endpoint at the current time is determined using a multi-objective utility function. Specifically, this includes the following sub-steps: Step 331, Traverse Each mapping element in Geometric proximity Compared to the preset cognitive development interval (based on the Vygotsky zone of proximal development theory) threshold Perform comparison and differentiate processing: (1) Judgment and handling of over-proficiency zone: like If the knowledge span corresponding to a candidate learning objective is determined to be less than the minimum cognitive threshold, it is considered repetitive or low-return learning. This candidate learning objective is marked as a "node to be determined," and a quick verification instruction (such as a lightweight test) is generated. The system performs a two-way operation based on the verification result: if the user passes the verification, it is determined to have effectively mastered the skill, and the node is directly updated to the learning space at low cost. In this process, the knowledge boundary is rapidly expanded; if a user fails the verification, it is determined that there is a "weak foundation," and the node and its corresponding dynamic anchor point are moved into the valid candidate set. middle.

[0068] (2) Panic Zone Determination and Circuit Breaker like If the geometric distance to the candidate learning objective exceeds the learner's current maximum cognitive load boundary, forcing learning would easily lead to frustration. Therefore, the candidate learning objective is marked as an "unreachable task," and a protective circuit breaker is implemented, removing the learner directly from the candidate learning objective set. Remove from the list.

[0069] (3) The zone of proximal development shall be retained: If satisfied The candidate learning objective is determined to be within the learner's "zone of proximal development," i.e., the optimal match between challenge and skill. The candidate learning objective and its corresponding dynamic anchor point are then moved into the valid candidate set. Then, the scoring process begins.

[0070] Step 332, for the valid candidate set For all tasks, construct a comprehensive utility scoring function. To balance "user intent preference" and "cognitive adaptation accuracy":

[0071] in For similarity calculation, For the effective candidate set Effective candidate learning objectives in User demand vector; The current spatial metric and the set optimal cognitive span constant. The absolute value of the deviation between them. These are adjustable weighting coefficients, corresponding to the weights of "learning willingness" and "cognitive damping," respectively.

[0072] Step 333: Perform the maximization reduction operation and select... The highest value is the effective candidate learning target vector. As the final learning target vector Its corresponding dynamic anchor point Embed vector for the starting node Output tuple instruction To the path planning module (step 34). Among them, It is not the user's current static position, but the optimal entry point for this specific learning objective, which is dynamically calculated by step 32.

[0073] Step 34: Path planning based on geometry heuristics.

[0074] Query starting vector and learning target vector In reliance on knowledge graphs The corresponding knowledge nodes and Plan the learning path that connects the two. The specific process is as follows: Path search: in knowledge dependency graphs Above, with Starting from, To reach the endpoint, run the A* graph search algorithm.

[0075] Heuristic function design: To improve search efficiency, the entire knowledge space is used. Geometric distance between intermediate nodes Directly used as a heuristic function in the A* algorithm This design leverages the strong correlation between geometric distance and cognitive relevance to effectively guide the search direction and accelerate algorithm convergence.

[0076] Path Output: The algorithm outputs a path consisting of multiple knowledge nodes. The resulting learning sequence serves as a structured learning path recommended to learners.

[0077] This invention employs a hybrid strategy of "manifold anchoring and graph pathfinding." Addressing the inevitable topological distortions and loss of logical directionality (i.e., semantic drift) that occur when embedding knowledge graphs into low-dimensional vectors, this invention utilizes only the vector space for semantic anchoring and heuristic guidance of start and end points. The specific path generation process is then traced back to the original weighted dependency graph for execution. While ensuring search efficiency, this invention strictly leverages the edge connection constraints of the graph to guarantee strong logical predecessor dependencies between knowledge points, completely eliminating the risk of "cognitive logic inversion" that may arise from pure vector retrieval.

[0078] Step 4: Interactive execution based on contextual parameters.

[0079] This step executes a sequence of tasks, runs a contextual state machine to handle micro-interaction anomalies, and records edge-based traversal costs. Specifically, it includes the following sub-steps: Step 41: Sequence loading and topology predecessor locking.

[0080] The sequence output in step 34 Load into the execution queue and activate nodes in sequence. Activate the current node At that time, read the last node that has completed execution from the queue. This node is identified as a topological predecessor node. A temporary logical flow association is established in memory. At the same time, load the initial scenario configuration vector. ,in, Difficulty level For scaffold strength. Scaffold strength It is a preset dynamic adjustment variable used to quantify the current node in the task execution process. The degree of guidance, assistance, or constraint provided.

[0081] Step 42: Adaptive adjustment of context parameters and resource reconfiguration based on interactive feedback.

[0082] This step monitors the learner's interactive behavior at the current knowledge node in real time. When cognitive obstacles are detected (such as multiple consecutive incorrect answers, operation timeouts, or requests for help), the task is not directly judged as a failure. Instead, an "in-situ dimensionality reduction adjustment loop" is triggered, which dynamically reconstructs learning resources using generative artificial intelligence (AIGC).

[0083] Real-time data collection for this (the first) Feedback data from each interaction (including but not limited to answer accuracy, response latency, and operation trajectory). If the feedback data shows "Pass," mark the status as "Successful" and proceed to step 43. If the data shows "Failed / Cognitive Impeded," first read the retry counter of the current node. .

[0084] like and The condition is determined to be temporary cognitive difficulty, indicating a reparable cognitive blockage. A nonlinear adaptive correction loop for contextual parameters is then triggered to calculate a target parameter set that significantly reduces cognitive load. Based on this, adaptive learning resources are dynamically reconstructed, driving learners to perform iterative interactions within the dimensionality-reduced cognitive context. Specifically, this includes: (1) Downgrade the difficulty level: reduce the difficulty level According to attenuation factor Scaling .parameter It represents the entropy value of information or the depth of logical nesting.

[0085] (2) Upgrading the scaffolding level: increasing the scaffolding strength parameters Execute incremental steps This parameter determines the level of system intervention and assistance (e.g., No prompt. Highlight key points. To explain by analogy, ... For the highest level of guidance, such as step-by-step demonstrations or direct examples.

[0086] (3) AIGC resource reconstruction based on parameter constraints: Activate the generative artificial intelligence engine, firstly, the corrected parameters are... The mapping is done into structured prompt engineering instructions, and then the current knowledge point is extracted. The core semantic vector or original text is used as the "content base" for generation. Finally, the difficulty level is... and scaffolding strength level These constraints are converted into language style constraints and logical structure constraints, respectively. The "content base" is then rewritten in real time to generate a new version of interactive resources adapted to the current low cognitive load, and pushed to learners for the next step. This is the first attempt.

[0087] like (Preset circuit breaker threshold) or (Highest intensity guidance) to determine the learner's understanding of the knowledge point. If a deep cognitive impairment exists, the process will not be retried. Instead, a "cognitive blocking signal" will be generated directly, marking the final state of this interaction as "failure," and step 43 will be executed.

[0088] Step 43: Record the associated costs of crossing edges.

[0089] After the interaction ends, calculate the scalar cost of this interaction. ,in These are weighting coefficients. (In the learner's learning log) Construct and write a flow attribute record, with the data structure defined as:

[0090] Where u represents the learner; Generate timestamps for events, used for system timing synchronization and causal analysis. This is a binary final state (1 = success, 0 = failure), used to distinguish between "high-cost passage" and "path blockage" in subsequent steps. The scalar cost of this interaction represents the learner's journey from knowledge nodes. Crossing to knowledge nodes The system resources and cognitive load consumed.

[0091] Step 5: Data update and anomaly handling of the cognitive state set.

[0092] This step is based on the interaction log. Status flags in Execute the set in memory Write operations or generate system interrupt instructions.

[0093] Step 51: Logical routing based on status flags.

[0094] The system retrieves data from the logs. Read the current interactive knowledge node The latest records, analyzed within. Fields directly execute binary logic splitting: Forward path ( ): This indicates that the learner has successfully crossed the node, generating an update enable signal and triggering step 52.

[0095] Negative path ( ): This indicates that the learner still failed to master the node under the conditions of maximum scaffolding strength or maximum number of attempts, generating a cognitive blocking signal and triggering step 53.

[0096] Step 52: Union operation of coordinate sets.

[0097] Will Each node Embedded vector with identifier Combining into tuples Added as incremental information to the existing knowledge set. middle; In response to the node generated in step 51 The data update enable signal is used to update the set stored in memory. Perform a set union operation once:

[0098] Once completed, return to step 41 until the final learning task is finished.

[0099] Step 53: Trigger replanning based on execution exceptions.

[0100] In response to the cognitive blocking signal generated in step 51, the exception handling procedure is executed: (1) Queue Suspend: Freeze the current execution queue and terminate the activation of subsequent nodes.

[0101] (2) Cognitive blockade marker: maintaining set Keep it unchanged, and set the nodes This is marked as a cognitive blockage state in the path search algorithm.

[0102] (3) Replanning Request: Send a global replanning instruction to step 34. During the replanning process, the system maintains the knowledge dependency graph. Under the premise of keeping the global edge weights unchanged, a path cost adjustment mechanism is introduced for the current learner: for knowledge nodes that are judged as cognitive obstacles during the learning process... During the path search phase, additional cost penalties are imposed on the dependent edges associated with the node, which significantly increases the cumulative cost of the path through the node. This is then automatically suppressed during the A* search process. The algorithm prioritizes generating alternative learning paths that bypass the node and point to its strong logical predecessor remedy node.

[0103] Step 6: Bidirectional evolution of the knowledge space based on group cost.

[0104] This step performs batch statistical analysis on the log database, and analyzes the dependency graph based on the characteristics of group learning behavior. Perform topological reconstruction and drive incremental geometric evolution of the knowledge space to achieve an adaptive mapping from "behavioral data" to "spatial structure." Specifically, this includes the following sub-steps: Step 61: Aggregation and feature extraction of the circulating data.

[0105] From learners' learning logs Extract all flow records (including status) within the cycle. All records marked as "success" and "failure"). In pairs. Using topological keys, records are grouped and aggregated. For each group of flow relationships, a multidimensional statistical feature vector is calculated. :

[0106] in, This represents the total frequency of the flow (i.e., the total number of attempts), used to measure the statistical significance of the path. Indicates path pass rate ( This characterizes the logical accessibility and cognitive feasibility of the path. The average cost of successful samples is calculated only. In the records The arithmetic mean of the path represents the range of difficulty under the premise of logical coherence.

[0107] Step 62: Bidirectional modulation of topology connection weights.

[0108] Step 61: Traverse all aggregated flow relationship groups and determine their positions in the graph. The existence state in, combined with "pass rate" "and average cost" The two-factor characteristic of "clustering reward" or "alienation penalty" is used to perform clustering reward or alienation penalty operations.

[0109] Scenario 1: Weight Gain of Cognitive Accommodation Path For edge set The existing edges in An edge is considered an efficient cognitive path if and only if the statistical features corresponding to that edge simultaneously satisfy the following three conditions: (1.1) Total interaction frequency (Significance threshold) (1.2) Pass rate (High-order threshold) (1.3) Average cost (Low threshold) For efficient cognitive paths, perform weight gain and update the weights. ,in The preset gain coefficient ( ).

[0110] Scenario 2: Weight decay of cognitive blocking pathways For edge set The existing edges in When the edge satisfies the significance condition A path is considered a cognitive impairment pathway if it meets at least one of the following negative performance criteria: (2.1) Pass rate Below the low threshold

[0111] (2.2) Average cost Above the high threshold

[0112] For cognitive blocking paths, perform weight decay and update the weights. .

[0113] Scenario 3: Instantiation of Implicit Relations For edge set The relationship that does not exist in A node pair is considered a newly emerging valid learning path when its statistical characteristics simultaneously meet the following two conditions: (3.1) Total frequency Greater than the significance threshold

[0114] (3.2) Pass rate Higher than the admission threshold

[0115] For new effective learning paths, in the edge set Create new directed edges The weight of the edge is initialized based on its statistical characteristics (such as the initial pass rate). Its technical effect is that it can dynamically discover and solidify "cognitive shortcuts" revealed by group learning behavior that were not defined in the initial knowledge design, so that the graph structure can continuously approach the real and efficient knowledge topology.

[0116] Finally, iterate through the updated edge weights, and if a certain edge weight... Less than a tiny threshold If the cognitive connection of an edge is determined to be invalid or too difficult to reach due to excessive resistance, it is removed from the edge set. Physically remove the edge. If the removed edge connects two frequently accessed knowledge clusters, a "semantic gap" alert signal is generated, indicating that a gap needs to be filled. and Transitional knowledge nodes are inserted into the geometric midpoint region.

[0117] Step 63: Incremental spatial update and structural anomaly detection.

[0118] Based on the updated topology The optimization algorithm is executed in the manner described in step 1.2 to complete the local incremental update and generate a new generation of manifold space. .

[0119] In response to changes in the spatial coordinate system ( This triggers a synchronization calibration mechanism to ensure that all active users have access to the master set. Maintain consistency with the new spatial coordinate system. Specific implementation methods may include: re-querying the new coordinates based on the knowledge node ID for a full refresh, or using an affine transformation matrix to map the old coordinates to the new space, so as to ensure that subsequent path planning (S200) is based on a unified metric.

[0120] This invention also provides a system for constructing an evolvable knowledge space using the above method, the system framework of which is as follows: Figure 3 As shown, it includes: The knowledge space geometry construction module is used to construct a weighted knowledge dependency graph and transform the weighted knowledge dependency graph into a vector space with metric properties, maintaining the topological structure of the space based on the edge weights of the graph. The subject cognitive coordinate tracking module is used to construct the learner's initial set of mastered knowledge nodes. and the corresponding initial knowledge space. And update the set based on the interaction results of the cognitive context interaction module. and space ; A geometric adaptive planning engine is used to determine the learner's learning target knowledge nodes and learning start knowledge nodes, and to plan the learning path; The cognitive context interaction module interacts with learners, monitoring their behavioral feedback as they sequentially activate nodes along the learning path planned by the geometric adaptive planning engine. When cognitive obstacles are detected, generative artificial intelligence is used to reduce the learning difficulty level. and / or increase scaffold strength In this context, learning resources are dynamically reconstructed and adapted to drive learners to perform iterative interactions in the reduced cognitive context. The collaborative evolution control module is used to drive the bidirectional, iterative evolution of the knowledge graph structure and geometric space based on the group interaction data flow.

[0121] The present invention also provides a computer system comprising one or more processors, a memory, and computer program instructions stored in the memory, wherein when the instructions are executed by the processor, individual adaptation and group evolution are realized.

[0122] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A kind of "knowledge" Context The "subject-based" three-in-one evolving knowledge space construction method is characterized by: include: S1, Construct a weighted knowledge dependency graph , where the set of nodes Represents core knowledge concepts within the domain; edge set Indicates at time Cognitive dependencies or prerequisite relationships between knowledge concepts; weight set This indicates the strength of cognitive support between concepts; S2, map Various knowledge concepts Mapped to a dimensional embedding vector Following the "semantic-topological dual constraint" strategy, the full knowledge space is obtained. ; S3, Constructing a set of mastered knowledge nodes based on learner identity recognition and ability assessment results. and in The text appears to be a jumbled collection of phrases and sentences, seemingly from different sources. A coherent translation isn't possible without the full context. Each node Embedded vector With unique identifier Combining into tuples Injecting into the learner's initial knowledge space middle; S4, identify the learner's learning objective knowledge nodes and starting knowledge nodes, and plan the learning path; specifically including: S41, Calculate the target semantic vector corresponding to the learner's learning needs. With the full knowledge space The similarity of all embedded vectors is used to select embedded vectors with similarity values ​​greater than a set threshold. This constitutes the set of candidate learning target embedding vectors. ; S42, Construct the global distance matrix , where matrix elements For the first The embedding vector of the candidate learning objective and the learner's first... The distance between the embedding vectors of the knowledge nodes that have been mastered; for the th The optimal dynamic anchor point of each candidate learning target embedding vector. Global distance matrix No. minimum value in row The corresponding embedded vector of the knowledge points already mastered; S43, Regarding candidate learning objectives If its Generate quick test instructions and update the embedded vectors of candidate learning objectives that pass the test to the already mastered knowledge space. In the middle, those that fail are moved into the valid candidate set. ;like Then the embedding vector of the candidate learning objective is moved into the effective candidate set. The learning objective embedding vector with the highest score is selected as the learning objective vector corresponding to that learning requirement. The corresponding optimal dynamic anchor point is the learning starting point vector. ; S44, learn the target vector and learning starting point vector Back-mapping to knowledge graph In the process, the learning target node is obtained. and learning starting point node ; Using path planning methods, in the graph The planning starts from the learning starting point node To the learning target node The path; S5, sequentially activate each node on the path described in S4; monitor the learner's progress at the current knowledge node. On-screen interactive feedback: If cognitive blockage is detected, generative artificial intelligence is used to reduce the learning difficulty coefficient. and increase the strength of scaffolding In this scenario, learning resources are dynamically reconstructed and adapted to drive learners to perform iterative interactions within the reduced cognitive context. If a learner fails to master a knowledge node under the conditions of maximum scaffolding strength or maximum number of attempts, it is marked as a failure, and the process returns to S44 to replan the path, avoiding the knowledge node and generating an alternative path pointing to the predecessor remedy node. If the learner successfully masters the knowledge node, the current knowledge node is... Add to the set of knowledge nodes already mastered In the process, the embedding vector corresponding to the knowledge node is added to the already mastered knowledge space. middle; S6, Dependency graph based on group learning behavior characteristics Perform topology reconstruction: Specifically: Extract all flow records within the cycle from all learners' learning logs, in binary form. Use the topological key to group and aggregate records; For each group of flows Statistical analysis of the total frequency of this transaction. Calculate the throughput of this path. and the average cost of successfully passing ;based on , and Update edge set and weight set ; Based on the updated topology Reconstruct the entire knowledge space according to the S2 approach. and the knowledge space already mastered .

2. The method as described in claim 1, characterized in that, In step S1, initial relational connections between concepts are constructed based on the experience of subject matter experts, the curriculum syllabus, and the structure of standard textbooks. .

3. The method as described in claim 1, characterized in that, In S2, the loss function is constructed. This is used to measure the spatial metric between node embedding vectors. By optimizing the loss function, the optimal embedding vector for each node is obtained.

4. The method as described in claim 1, characterized in that, In step S2, a loss function is constructed, and an optimization algorithm is used to optimize the embedding vector of each node; the loss function is: in, For knowledge nodes and The strength of cognitive support between them; For knowledge nodes , The embedding vector; For knowledge nodes The corresponding initial semantic vector; These are the topology constraint weight parameters; The optimization algorithm employs gradient descent, stochastic gradient descent, momentum gradient descent, or adaptive learning rate optimization.

5. The method as described in claim 1, characterized in that, In S43, the following scoring function is used: in, For the first Embedded vectors of candidate learning objectives The rating, For similarity calculation; The optimal cognitive span constant is set; These are the weighting coefficients.

6. The method as described in claim 1, characterized in that, In S44, the path planning method adopts the A* algorithm, Dijkstra's algorithm, breadth-first search algorithm, or depth-first search algorithm.

7. The method as described in claim 1, characterized in that, In S5, the learning difficulty coefficient The initial value is 1, scaffold strength The initial value is 0; Generative artificial intelligence will learn the difficulty level According to attenuation factor Scaling is applied to adjust the scaffolding strength. Increase in increments of 1 step.

8. The method as described in claim 1, characterized in that, In S6, based on , and Update edge set and weight set Specifically: For edge set The existing edges in If the total interaction frequency And the pass rate And average cost If the path is determined to be an efficient cognitive path, then the edge is increased. Weighting; if the total interaction frequency However, the pass rate or average cost If the path is determined to be a cognitively hindered path, then the number of edges should be reduced. The weights; For edge set The relationship that does not exist in If the total interaction frequency And the pass rate Higher than the admission threshold Then in the edge set Create new directed edges New directed edge The weight of its pass rate ; Iterate through the updated edge weights, if a certain edge weight... Less than a tiny threshold Then remove it from the edge set. If an edge is removed from a knowledge cluster, and that edge connects two frequently accessed knowledge clusters, a "semantic gap" alert signal is generated, indicating that a gap needs to be filled. and Transitional knowledge nodes are inserted into the geometric midpoint region.

9. The method as described in claim 8, characterized in that, Efficient cognitive path The weights are updated to ; Cognitive blocking path The weights are updated to ; in, The preset gain coefficient, .

10. A system employing the method as described in any one of claims 1 to 9, characterized in that, include: The knowledge space geometry construction module is used to construct a weighted knowledge dependency graph and transform the weighted knowledge dependency graph into a vector space with metric properties, maintaining the topological structure of the space based on the edge weights of the graph. The subject cognitive coordinate tracking module is used to construct the learner's initial set of mastered knowledge nodes. and the corresponding initial knowledge space. And update the set based on the interaction results of the cognitive context interaction module. and space ; A geometric adaptive planning engine is used to determine the learner's learning target knowledge nodes and learning start knowledge nodes, and to plan the learning path; The cognitive context interaction module interacts with learners, monitoring their interactive behavior as they sequentially activate nodes along the learning path planned by the geometric adaptive planning engine. When cognitive obstacles are detected, generative artificial intelligence is used to reduce the learning difficulty level. and / or increase scaffold strength In this context, learning resources are dynamically reconstructed and adapted to drive learners to perform iterative interactions in the reduced cognitive context. The collaborative evolution control module is used to drive the bidirectional, iterative evolution of the knowledge graph structure and geometric space based on the group interaction data flow.

Images