An emotional resonance driven course knowledge neural evolution teaching path generation method

Abstract

This invention provides a method for generating course knowledge neural evolution teaching paths driven by emotional resonance, belonging to the field of intelligent education technology. This method integrates core technologies such as large-model-driven unified representation of cross-modal semantic fields, technology-value dual-manifold implicit mapping topology discovery, emotional resonance-driven dynamic regulation of knowledge node activity, knowledge lineage inheritance-based organic growth, dual-manifold spatiotemporal coordinate system construction, semantic activation self-reconstruction of fused knowledge agents, and neuroevolutionary architecture search-driven teaching path topology evolution. This enables the emotionally perceptive intelligent generation and online adaptive reshaping of course-based educational teaching paths. Consequently, the organization of teaching content not only thoroughly explains technical principles but also naturally evokes value-based thinking in line with students' emotional responses, reducing the abruptness of value guidance and allowing for a closer integration of professional knowledge instruction with the goal of moral education.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology, specifically relating to a method for generating a course knowledge neural evolution teaching path driven by emotional resonance. Background Technology

[0002] Systematically integrating value guidance into professional courses is a core strategic deployment for higher education in the new era to fulfill its fundamental mission of cultivating talent, aiming to deeply embed value guidance into the entire process of imparting professional knowledge. However, the current value-integrated teaching system faces systemic and deep-seated technical challenges in five aspects: cross-modal resource representation, discovery of technology-value connections, dynamic evolution of knowledge graphs, precise positioning of knowledge units, and generation of personalized teaching paths, which restricts the further improvement of the effectiveness of the curriculum in cultivating talent.

[0003] In terms of cross-modal semantic unification, course resources used for value guidance are highly heterogeneous and diverse in origin. Textual lectures are conceptually clear and logically rigorous, but they struggle to convey the contemporary context and atmosphere of video case studies. Charts and data can clearly present technical structural relationships, while audio materials have a direct advantage in evoking emotional responses. However, currently, these four types of resources are usually stored and managed independently, making it difficult to manually identify the potential connections between technical knowledge points and video materials explaining their background value. Existing practices largely rely on keyword annotation or manual metadata tagging, which essentially remains at the level of surface text matching and fails to truly achieve semantic interoperability among heterogeneous modal resources. Furthermore, each time new teaching resources are added, the existing annotation system often needs to be reorganized, resulting in high maintenance costs and hindering the effective realization of the potential for collaborative teaching using multimodal resources.

[0004] In discovering the connection between technical knowledge and humanistic connotations, the current approach is to convert technical concepts and macro-level value propositions into vectors and then calculate their similarity to establish a mapping relationship between the two. However, technical concepts are more specific (such as bus arbitration timing and pipeline dependency), while value propositions are more abstract and general. The two are not at the same granularity level, and direct calculation can easily cause the macro-level description to overwhelm the technical details, or calculate a bunch of false connections that are not used in actual teaching.

[0005] Currently, determining the importance of nodes in knowledge graphs typically involves experts assigning fixed weights to each knowledge point, aiming to give important content more opportunities for recommendation in the teaching path. However, practical results show that students' acceptance of a particular value-guided point depends more on their emotional resonance with it. Experts believe that important theoretical nodes are essentially wasted in the classroom if students don't engage with them. Conversely, some cases that evoke strong resonance, even if lacking theoretical depth, often become breakthroughs in values ​​education. Static weights cannot adjust resource allocation according to changes in the students' emotional state.

[0006] Regarding knowledge graph updates, the current practice is to replace old nodes with new ones when there are new policies, new cases, or changes in the technological background, in order to keep the graph content up-to-date. The problem with this overlay update is that the historical connections carried by the old nodes are also deleted, and the evolution of the same educational concept in different periods disappears from the graph, resulting in a lack of accumulation in the time dimension of the graph.

[0007] In defining the knowledge units in a course, the conventional approach is to use the chapter order of the textbook as the coordinates for each knowledge point. The advantage of this approach is its simple structure, but a single chapter number cannot reflect the upstream and downstream relationship of a knowledge point in the technology dependency chain, nor can it reflect its role in value transmission. Teachers lack a more refined coordinate reference when deciding when to emphasize technology and when to emphasize value guidance.

[0008] In the teaching process, the emphasis on activating knowledge points currently relies mainly on teachers' experience to manually adjust the focus in different lesson types. For example, principle classes focus more on technology, while topic discussion classes connect it with humanistic connotations. This approach, which depends on individual judgment, lacks a reusable mechanism for switching between the technical and value aspects of the same knowledge point in different scenarios, and the composite function of the knowledge point itself is not fully released.

[0009] In generating personalized learning paths, most existing recommendation systems employ fixed-structure models that recommend sequences of knowledge points based on student profiles and historical records. While such models are adequate for handling purely technical knowledge, they often fall short when simultaneously considering the coherence of technical logic, the effectiveness of value infusion, and the varying levels of acceptance of value-guided content among different students.

[0010] Currently, online adjustments to teaching pathways mainly rely on teachers intervening after observing significant changes in students' states, which limits the timeliness and precision of these adjustments. Pre-planned teaching pathways cannot be fine-tuned in real time based on continuous changes in students' emotions during class, easily leading to discrepancies between the pre-planned pathways and the actual teaching situation.

[0011] The problems in each of the above-mentioned stages are intertwined, and a single fix cannot resolve the overall contradictions. Therefore, a method is needed that integrates changes in students' emotional resonance, cross-modal resource connectivity, precise technology-value correlation, continuous growth of the knowledge system, accurate positioning of knowledge points, automatic adaptation of activation focus, online evolution of teaching paths, and real-time fixes. This is precisely the core problem that this invention aims to solve. Summary of the Invention

[0012] This invention provides a method for generating course knowledge neural evolution teaching paths driven by emotional resonance. This method uses emotional resonance as its core driving force, employing an eight-in-one technical system: constructing a large model to drive unified representation of cross-modal semantic fields, discovering multi-granularity implicit mapping topologies, dynamically regulating the activity of knowledge nodes driven by emotional resonance, organically growing knowledge genealogy based on era-scene perception, constructing a technology-value dual-manifold spatiotemporal coordinate system, self-reconstructing the semantic activation of knowledge intelligent agents through teaching field perception fusion, driving the topological evolution of teaching paths through neural evolutionary architecture search, and driving online drifting and adaptive reshaping of teaching paths through emotional state fields. This system enables the emotionally perceptive intelligent generation and online adaptive reshaping of course knowledge teaching paths.

[0013] The technical solution adopted in this invention is as follows, including the following steps:

[0014] A method for generating a curriculum knowledge neuroevolutionary teaching path driven by emotional resonance includes the following steps:

[0015] Step S1: Perform geometric semantic compression on the four modalities of text lecture notes, video cases, chart data and audio materials in the course teaching resource library, and complete equidistant alignment in the unified Riemann embedding space to obtain a unified semantic field vector set; implement an incremental alignment update mechanism for the resource units of the added course teaching resources, and perform local perturbation correction in the unified Riemann embedding space. Step S2: Use the value semantic decomposition big model to perform recursive semantic fine-grained decomposition on each coarse-grained value semantic; use the technology concept induction big model to perform value-carrying induction coarse-grained up-process operation on each fine-grained technology concept to achieve bidirectional granular alignment; design a mutual information-driven bidirectional semantic penetration and propagation mechanism between aligned semantic pairs of the same granularity, and construct a deep mapping map of technology value based on the aligned value semantic and technology concepts. Step S3: Obtain the interaction data of the learner group and extract the emotional resonance intensity vector of each knowledge node in the deep mapping graph of technology value; calculate the activity value of the knowledge node based on the emotional resonance intensity vector, so that nodes with high activity values ​​receive positive activity pulses and priority activation weights, while nodes with low activity values ​​have their activity decayed and trigger the self-repair generation of associated resource units when they fall below a threshold; and generate a neighborhood cascade propagation mechanism so that changes in node activity propagate to the neighborhood along the associated edges of the deep mapping graph of technology value, and output a dynamic activity graph; Step S4: Extract the technical concept vector, value semantic vector and era scene label of the new resource unit, and calculate its geodesic distance to the historical nodes in the dynamic active graph. Depending on whether the geodesic distance falls into the inheritance threshold or exceeds the bifurcation threshold, activate the inheritance growth operator or the bifurcation growth operator respectively, and link the new node into the dynamic active graph with the inheritance edge or the bifurcation edge to obtain the knowledge spectrum with time depth. Step S5: Construct a technical coordinate triplet on the technical manifold coordinate layer to describe the geometric position of the knowledge unit in the technical logical topology; construct a humanistic coordinate triplet on the value manifold coordinate layer to describe the emotional geometric position of the knowledge unit in the value concept manifold; design a dual-coordinate isometric projection operator on the Riemann product manifold to fuse the two types of coordinate projections into a unified spatiotemporal embedding space to obtain a spatiotemporal coordinate fingerprint set. Step S6: Construct each knowledge unit into an autonomous intelligent agent with field awareness capabilities, set up a dual-channel semantic activation competition allocation mechanism, so that the technical concept activation channel and the value semantic activation channel dynamically compete for activation dominance based on the field context vector containing course stage, learner emotional state and historical interaction trajectory; and design a local activation state negotiation and propagation protocol between intelligent agents, so that the neighboring activation state affects the routing decision of the node, and the routing parameters are continuously updated by the positive feedback signal of teaching effect, and output dynamic activation semantic representation. Step S7: Based on the spatiotemporal coordinate fingerprint of knowledge units, synthesize a dynamic three-dimensional fitness terrain composed of a technology logic coherence surface, a value penetration density surface, and a learner emotional resonance prediction surface; use the topology of the path generation network as the evolutionary search object, and use a topological mutation operator that is executed along the geodesic direction of the technology-dependent manifold and a topological recombination operator that exchanges superior gene fragments near the Pareto front to drive the population to search for the Pareto optimal teaching path in the three-dimensional fitness terrain; Step S8: The learner's discrete emotional feedback signal is interpolated into a continuous emotional state field distribution using a spatiotemporal kernel function. The spatial gradient vector of the emotional state field drives the current teaching path to continuously drift online along the emotional landscape gradient direction. A local path topology self-repair mechanism for Pareto deviation perception is set up. When the path deviates from the Pareto front by more than a threshold, the curvature of the emotional state field is used as a constraint to search for and replace local path segments. At the same time, the statistical regularity of the group's emotional state field is continuously distilled into a long-term correction term of the evolutionary population fitness function to drive the continuous calibration of the optimization objective. The specific steps of S1 are as follows: Step S101: For the text handout modality, a large text semantic encoding model is used to perform deep semantic encoding on the original text handout to extract and preserve the latent text semantic vectors that preserve the technical concept dependencies and value semantic density distribution. For video case modalities, a visual semantic coding big model is used to perform frame-by-frame semantic coding on video cases, and a cross-frame semantic state progressive anchoring mechanism is designed to preserve the dynamic changes of the context along the frame sequence, resulting in anchored semantic vectors. The anchored semantic vectors of the entire video frame sequence are weighted and pooled in the time dimension to generate the overall semantic vector of the video case. For graph data modalities, a graph neural network encoder is used to perform structured semantic encoding on the node features and edge relationships of the graph, and outputs a semantic vector that preserves the logical topological structure. For audio data modalities, an audio semantic coding big model is used to perform acoustic-semantic joint coding on the audio data. During the coding process, the semantic information of the language content and the emotional resonance intensity information carried by the acoustic features are extracted simultaneously to generate an audio semantic vector that integrates the semantic content and the emotional temperature. Step S102: Define the dimension of the unified Riemann embedding space, apply modality-specific projection layers to the semantic vectors of the four modalities respectively, and project the semantic vectors of each modality onto the unified Riemann embedding space; take the co-occurrence relationship between technical concepts and value semantics as the semantic field polarization center, construct a field polarization convergence training loss function with cross-modal synonym attraction and cross-modal heteronymy repulsion, realize the equidistant alignment of the four modalities in the unified Riemann embedding space, and output a unified semantic field vector set; Step S103: When the course teaching resource library is iteratively updated, local perturbation correction is performed in the existing equidistant embedding space, and the projection parameters of newly added resource units are optimized only. Step S2 is as follows: Step S201: Set a coarse-grained value semantic set, and perform recursive semantic fine-grained decomposition on each coarse-grained value semantic using the value semantic decomposition big model to obtain the fine-grained value elements of the recursive decomposition. The semantic decomposability score of the value elements in the current decomposition round is used as the stopping criterion for recursive decomposition. Step S202: Set a set of fine-grained technical concepts, and perform a value-carrying inductive ascending operation on each fine-grained technical concept using the large model of technical concept induction to obtain the technical principles of induction ascending. The stopping criterion of induction ascending is the value-carrying curvature score of the technical principles of the current ascending round. Step S203: After completing the bidirectional granularity adjustment, each element in the fine-grained value element set and each element in the technical concept set are at the same semantic granularity layer. The aligned technical concepts and value semantics are used as nodes. Based on the mutual information-driven bidirectional semantic penetration propagation mechanism, the bidirectional semantic penetration strength between technical concepts and value semantics is calculated. When the bidirectional semantic penetration strength is greater than the implicit mapping emergence threshold, an implicit mapping topological edge is established between technical concepts and value semantics in the deep mapping graph of technical value. The edge weight is the bidirectional semantic penetration strength. Further bidirectional semantic penetration fusion is performed. For node pairs of technical concepts and value semantics that have established implicit mapping topological edges, the corresponding nodes are fused based on the bidirectional semantic penetration strength to generate technical-value fusion nodes, and finally a deep mapping map of technical value is constructed. Step S3 is as follows: Step S301: Let the size of the learner group be U. For the u-th learner, the n-th knowledge node in the deep mapping graph of technology value... From the interaction data, we extracted emotional signals in three dimensions: emotional polarity signals of text feedback, cognitive input curve signals of response time, and attention retention distribution signals of interaction behavior; A large-scale sentiment polarity analysis model is used to perform fine-grained sentiment polarity analysis on the sentiment polarity signal of the text feedback, and the sentiment polarity score of the text feedback is obtained. A temporal convolutional network is used to extract cognitive input intensity features from the temporal cognitive input curve signal of the response; Calculate attention concentration based on the attention residence distribution signal of interactive behavior; Finally, a multilayer perceptron fusion network is used to fuse text feedback emotional polarity scores, cognitive engagement intensity features, and attentional retention concentration to obtain the emotional resonance intensity vector of the u-th learner for the n-th knowledge node; the mean value is then taken over the learner group to obtain the knowledge node... The intensity vector of group emotional resonance ; Step S302, set knowledge nodes The activity value at time t is The initial activity value is assigned a static weight by domain experts; based on the group emotional resonance intensity vector. Design the node activity excitation-inhibition dynamics update rule: in, Let be the node activity value at time t+1. The excitation gain coefficient, To suppress the attenuation coefficient, The excitation activation threshold, To suppress the activation threshold, ReLU is a linear rectified function; When a node's activity value falls below the node's self-repair trigger threshold, the node self-repair process is triggered to regenerate the associated resource units of the large model, thus regenerating the large model from node materials. For this node Regenerate related resource units with stronger potential for emotional resonance: in For the newly generated associated resource unit, For nodes Neighborhood structure information in the deep mapping map of technological value Regenerate the parameter set of the large model from the node materials; Step S303: The change in node activity value propagates to neighboring nodes along the associated topological edges of the deep mapping graph of technical value. Let the knowledge node... The set of neighboring nodes is Its neighboring nodes The active cascade is updated as follows: in For neighboring nodes The final activity value after cascading propagation, This is the cascade propagation gain coefficient. Knowledge nodes in the deep mapping graph of technological value and Edge weights between them In the deep mapping of technological value and Topological distance between them The distance attenuation coefficient for active cascade propagation is used to obtain the dynamic activity spectrum. Step S4 is as follows: Step S401, let the new resource unit set be New = { , ..., }, for the r-th new resource unit Analyzing large models using historical context Perform semantic parsing of the execution era and extract technical concept vectors Value semantic vector and era scene tags It contains information about the historical context; The semantic vector of the new resource unit semantic vectors of all existing knowledge nodes in the graph Calculate the Riemann geodesic distance distribution in a unified embedding space: in For new resource units With existing knowledge nodes in the graph The Riemann geodesic distance between them For Riemann geodesic distance function; Step S402: Based on the geodesic distance distribution, define a dual-threshold growth decision mechanism; let the inheritance threshold be... The bifurcation threshold is For new resource units The historical node with the smallest geodesic distance in the map Execution of growth judgment: when At that time, the judgment was to allow continued growth; when When, the judgment is that it is bifurcation growth; when At that time, the judgment was a newly added independent judgment; When the decision result is inheritance growth, activate the inheritance growth operator; create a new node in the graph. and by inheriting the edge Connect to the nearest predecessor node The inherited edge carries a triplet of timestamp, era / scene tag, and value transfer vector: in For the new resource unit's era timestamp; The value transfer vector is calculated using the following formula: in The value semantic vector of the new resource unit. The value semantic vector of the nearest predecessor node; Step S403, when the decision result is bifurcation growth, activate the bifurcation growth operator to trigger local topological reconstruction of the graph: in For the nearest predecessor node The subtree rooted at the root; For nodes Depth within the knowledge hierarchy tree; The bifurcation depth penalty coefficient is used; the bifurcation edge is used as the basis for the penalty coefficient. Connect to new node and the bifurcation point This forms a new branch of the knowledge system, triggers the topological reconstruction of the local region, adjusts the weights and directions of the relevant edges to adapt to the semantic geometry of the new branch, and outputs a knowledge system with time depth. Step S5 is as follows: Step S501, let the set of knowledge units formed after semantic extraction and structured organization of the course teaching resource database be . For each knowledge unit Construct technical coordinate triples at the technical manifold coordinate layer Including topological location coordinates Abstract level Riemann geodetic distance coordinates Directed path encoding of logical dependency chains ; Based on the technical logical dependencies between course knowledge units, a directed acyclic dependency graph is constructed, and the knowledge units are calculated using a topological embedding method based on graph neural networks. Topological position coordinates; The abstract level of Riemannian geodetic distance coordinates is a unit of computational knowledge. The Riemann geodesic distance vector between the set of anchor points for the abstract hierarchy of technical knowledge; Logical dependency chain directed path encoding in knowledge units The path encoder is used to generate the path from all predecessor paths in the directed acyclic dependency graph. Step S502, for each knowledge unit Constructing a humanistic coordinate triplet at the value manifold coordinate layer Including the semantic manifold coordinates of value propositions Era and scene tags Emotional resonance intensity curvature weight ; Value proposition semantic manifold coordinates For knowledge units The corresponding value semantic vector is calculated using a value semantic embedding network; Emotional resonance intensity curvature weight The following formula is used to calculate: in For knowledge units corresponding nodes The activity value at time t, This represents the average activity value of all nodes in the entire graph at the current moment. Let σ be the standard deviation, and σ be the sigmoid activation function. and These are the influence coefficients of activity value and emotional resonance intensity, respectively. Step S503: Design a dual-coordinate isometry projection operator on the Riemann product manifold. After splicing the technical coordinate triplet and the human coordinate triplet, project them together onto a unified spatiotemporal embedding space to generate a spatiotemporal coordinate fingerprint. Train the operator through isometry mapping constraints to keep the Riemann geodesic distance in the input space approximately unchanged after projection. Step S6 is as follows: Step S601, for knowledge units Corresponding fusion knowledge intelligent agent Define the technology concept activation channel and the value semantic activation channel; at each teaching moment t, use the current field context vector as the competition condition to drive the dynamic struggle for dominance between the two types of activation channels; The field context vector includes course stage labels, learner u's emotional resonance intensity vector, and a summary of historical interaction trajectories; Based on spatiotemporal coordinate fingerprints and field context vectors, the competitive scores of the concept activation channels and the value semantic activation channels are calculated by the field-aware selective routing operator, and the dynamic activation weights of the two types of activation channels are obtained by Softmax normalization. The spatiotemporal coordinate fingerprints are input into the technical concept activation channel and the value semantic activation channel, respectively. The output results are then fused using dynamic activation weights to obtain a fused knowledge intelligent agent. The activation semantic representation at time t; Step S602, Integrating Knowledge Intelligent Agent Its current active semantic representation is used as the activation message to perform local broadcasting on the knowledge spectrum, thus fusing knowledge agents. The neighborhood agent receives the activation message, uses it as context input to calculate attention weights, obtains the aggregated neighborhood activation signal based on the graph attention mechanism, integrates it into the field context vector, and re-executes the dual-channel competitive allocation in step S601 with the updated field context vector. Step S603: Obtain the knowledge unit completed by the learner. The positive feedback signal of the teaching effect after learning drives the continuous self-updating of the parameters of the field-aware selective routing operator. Step S7 is as follows: Step S701: Construct a three-dimensional adaptive landscape dynamic generator, and synthesize a dynamic three-dimensional adaptive terrain composed of a real-time technical logic coherence surface, a value penetration density surface, and a learner emotional resonance prediction surface. The construction of the technical logic coherence surface is constrained by a directed acyclic dependency graph, and the computational technical logic coherence score is obtained by using spatiotemporal coordinate fingerprints to evaluate the adjacent knowledge units in the path. Value Penetration Density Curve The construction is based on the semantic manifold coordinates of value propositions, and the density of the value proposition space covered by the path is obtained as follows: in For path The total number of knowledge units included. This is the location index of the knowledge unit in the path, with a value range from 1 to L; The kernel density estimate, taking the semantic manifold coordinates of value propositions as input, measures the knowledge unit. Local coverage density on the value concept manifold; The bandwidth parameter for kernel density estimation; For knowledge units The curvature weight of emotional resonance intensity; Learner Emotional Resonance Prediction Surface The system is constructed using learner profile vectors as input and employs an emotional resonance prediction network to predict learner u's path. The overall intensity of emotional resonance is predicted. The three-curved surface constitutes a dynamic three-dimensional adaptive terrain in the form of a Riemannian product manifold, which affects the path. The overall fitness is calculated as follows: To target candidate teaching paths With learners The three-dimensional comprehensive fitness vector; Perform a multi-objective Pareto front search in a three-dimensional fitness space; Step S702: Define a path generation network family, where each path generation network takes learner profile vectors and activation semantic representation sets as input and outputs personalized teaching paths. ; The topology mutation operator performs local topology perturbation on the network generated by the current path based on the spatiotemporal coordinate fingerprint set and the directed acyclic dependency graph; Topological recombination operators perform cross-individual gene fragment recombination in high-fitness network architectures near the Pareto front of the current population; The population evolution iteratively performs topological mutation and recombination, and drives the population to continuously converge using the gradient gravity of the three-dimensional fitness terrain: in A non-dominated sorting genetic algorithm II framework with elitist strategy is used to perform Pareto non-dominated sorting and crowding distance calculation; Generate a network population for the parent path of generation t. The family of descendant path generation networks generated by topological mutation and topological recombination operations in generation t. This is the comprehensive fitness evaluation function; Step S703: Extract the Pareto optimal path population from the evolved optimal path generation network population to generate personalized Pareto optimal teaching paths. . Step S8 is as follows: Step S801: During the path execution process, the real-time emotional feedback generated by learner u at teaching time t is a discrete signal sequence. The discrete emotional feedback signal is kernel interpolated in the path execution space to generate a continuous emotional state field distribution. The spatial gradient vector of the emotional state field is calculated and injected into the fitness function parameters of the path generation network in real time to drive the current execution path to perform continuous online drift along the emotional landscape gradient direction. Step S802: Calculate the projection distance of the current execution path relative to the Pareto front in the three-dimensional fitness space. When the projection distance is greater than the adaptive reshaping threshold, trigger the local path topology self-repair mechanism. Using the current curvature distribution of the emotional state field as a geometric constraint, perform heuristic topology repair search on the local subgraph of the knowledge system to generate a local path replacement fragment that precisely matches the curvature of the current emotional landscape. Embed the local path replacement fragment into the corresponding position of the current execution path to complete the local self-repair of the path topology. Step S803: Design a long-range knowledge distillation and accumulation mechanism for cross-learner group emotional state fields to drive the underlying optimization goal of path generation strategy to continuously calibrate towards real teaching laws. The emotional state fields of the entire learner group are weighted and superimposed, and the statistical regularity of the group's emotional state field is continuously distilled into a long-term correction term of the evolutionary population fitness function. The long-term correction term is integrated into the comprehensive fitness function, and the original comprehensive fitness function is replaced by the corrected comprehensive fitness function. The underlying optimization objective of the path generation strategy is updated once after each teaching cycle. Compared with the prior art, the present invention has the following beneficial effects:

[0016] This method, through eight sequential steps and a feedback loop, transforms cross-modal representation, latent association discovery, node activity regulation, knowledge system growth, coordinate positioning, intelligent activation, path evolution, and online reshaping into mutually reinforcing synergistic relationships, rather than operating independently.

[0017] In the representation and association discovery phase, step S1 aligns the four types of heterogeneous resources to a unified Riemannian space, providing input for step S2 to eliminate the modal gap in multi-granularity value-technology semantic penetration. Simultaneously, the implicit mapping edge weights discovered in step S2 reverse-correct the polarization direction of the semantic field in S1. These two elements form an enhanced loop of "alignment-emergence-realignment," resolving the disconnect between cross-modal and cross-granularity processing.

[0018] In the dynamic maintenance of the knowledge graph, step S3 adjusts the node activity distribution based on students' real-time emotional resonance, while step S4 uses inheritance and bifurcation operators to preserve the semantic drift trajectory of the concept of nurturing people when new resources are added. The activity distribution output by S3 provides an emotional acceptance basis for the selection of the growth position of new nodes in S4, while the historical context accumulated in S4 supplements the activity regulation of S3 with long-term historical references, making the graph both real-time sensitive and historically stable.

[0019] At the knowledge unit localization and activation level, step S5 generates a spatiotemporal coordinate fingerprint for each knowledge unit through joint projection of the technology manifold and the value manifold. Step S6 then uses this fingerprint to dynamically allocate the ratio of technology activation to value activation based on the context. The coordinate fingerprint provides structured geometric constraints for the agent's routing decisions, eliminating the reliance on global search for activation decisions and simultaneously improving response speed and adaptation accuracy.

[0020] In the closed loop of teaching path generation and execution, step S7 utilizes neuroevolutionary search to offline generate a path population with a balanced advantage in technical logic, humanistic connotation, and emotional adaptation within a three-dimensional fitness terrain, providing a Pareto-optimal baseline and repair reference for step S8. In real teaching, step S8 drives continuous path drift using the gradient of the emotional state field, triggering local self-repair when deviating from the forefront. Simultaneously, the cross-learner group emotional patterns accumulated in step S8 continuously refine the fitness function of S7 through long-range knowledge distillation, forming an adaptive closed loop of "offline evolution → online reshaping → experience feedback."

[0021] Overall, S1 to S8 form a complete chain of collaboration: "unified representation → implicit association → activity map → lineage growth → coordinate positioning → field activation → path evolution → online reshaping → feedback calibration". The output of each step is the foundation for the next, and the result of the next step, in turn, affects the parameters of the preceding steps through channels such as map weights, activity distribution, coordinate reference, and fitness correction. This ensures that any local improvement in any link can be propagated and amplified along the chain, ultimately producing a system-level superposition gain without being attenuated by upstream and downstream disconnection. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall framework of a course knowledge neural evolution teaching path generation method driven by emotional resonance provided in an embodiment of the present invention.

[0023] Figure 2 This is a detailed execution flow diagram of the large model-driven cross-modal semantic field unified representation module for heterogeneous teaching resources in an embodiment of the present invention.

[0024] Figure 3 This is a detailed execution flow diagram of the multi-granularity implicit mapping topology discovery module for technical semantics and value propositions in an embodiment of the present invention.

[0025] Figure 4 This is a detailed execution flow diagram of the emotional resonance-driven dynamic regulation module for the activity of educational knowledge nodes in this embodiment of the invention.

[0026] Figure 5 This is a detailed execution flow diagram of the era-scene-aware, knowledge-based, inheritance-style organic growth module for cultivating people through the inheritance of knowledge lineage in an embodiment of the present invention.

[0027] Figure 6 This is a detailed execution flow diagram of the module for constructing a spatiotemporal coordinate system of course knowledge on a technology-value bimanifold in an embodiment of the present invention.

[0028] Figure 7 This is a detailed execution flow diagram of the semantic activation self-reconstruction module of the teaching field perception fusion knowledge intelligent agent in this embodiment of the invention.

[0029] Figure 8 This is a detailed execution flow diagram of the teaching path topology space evolution module driven by the neuroevolutionary architecture search in this embodiment of the invention.

[0030] Figure 9 This is a detailed execution flow diagram of the online drift and adaptive reshaping module of the teaching path driven by the emotion state field in this embodiment of the invention.

[0031] Figure 10 This is a schematic diagram of the deployment architecture and data flow of the intelligent teaching adaptive evolution system for cultivating students' moral character, provided in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0033] This embodiment provides a method for generating a course knowledge neural evolution teaching path driven by emotional resonance. Based on the teaching scenario of a computer organization principles course at a university, this course covers core technical concepts such as instruction set architecture, pipeline design, memory hierarchy, bus and input / output systems, while also embedding value semantics such as independent innovation, dialectical thinking, fair resource allocation, and standardized collaborative spirit. The complete construction process of this method is described in detail, as follows: Figure 1 As shown.

[0034] S1, the specific implementation of the cross-modal semantic field unified representation module driven by the large model of heterogeneous teaching resources.

[0035] This step serves as the data foundation for the entire method. It equilaterally aligns four modalities—text lecture notes, video examples, charts, and audio materials—from the course teaching resource library within a unified Riemann embedding space, enabling resources of different modalities to be compared and correlated within the same semantic space. For example... Figure 2 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0036] S101: Dedicated semantic geometry compression channels are constructed for the four types of heterogeneous modalities, and latent semantic vector extraction is performed along the inherent information manifold structure of each modality.

[0037] For the text-based lecture note modality, its information geometry is dominated by a semantic topological network, employing a large-scale text semantic encoding model (denoted as...). Deep semantic encoding is performed on the original text lecture notes to extract latent text semantic vectors that preserve the technical concept dependencies and value semantic density distribution. Let the set of text lecture note resources be... For the i-th text lecture notes The encoding process yields the text semantic vector:

[0038]

[0039] in Let be the latent semantic vector of the i-th text lecture note, with dimension . . This is the parameter set for the large model text encoder. Indicates A text semantic encoding function with parameters.

[0040] For the video case modality, its information geometry is dominated by a temporal frame semantic manifold. Video cases not only carry the visual semantics of static images but also encapsulate the contextual tension of a historical narrative in the temporal dimension. A large-scale visual semantic encoding model (denoted as...) is adopted. ) Perform frame-by-frame semantic encoding on video cases and design a cross-frame semantic state progressive anchoring mechanism to preserve the dynamic changes in the context along the frame sequence. Let the set of video cases be V = { , , ..., }, for the j-th video segment Encoding is performed on the k-th frame. The frame-level semantic vector of the cross-frame semantic state progressive anchoring mechanism is calculated as follows:

[0041]

[0042] in Let be the anchor semantic vector of the k-th frame of the j-th video segment. This represents the original image data of the k-th frame in the j-th video segment. The parameter set for a large model of visual semantic encoding. The cross-frame anchoring coefficient (ranging from 0 to 1) controls the progressive influence of historical frame semantics on the current frame semantics. Encodes the original visual semantics of the current frame. It is the anchor semantic vector of the previous frame, and it is itself formed by the previous frame in the same way. and It is synthesized through proportional weighting. Therefore, starting from the first frame, the semantic representation of each frame inherits the semantics of the previous frame in a uniform proportion. The historical context information is mixed and passed down frame by frame along the sequence, forming a semantic chain that runs from the first frame to the last. When a new frame arrives, the narrative context is not recalculated, but rather flows naturally into the representation of the current frame along this semantic chain. Weighted pooling is performed on the anchor semantic vectors of the entire video frame sequence in the temporal dimension to generate the overall semantic vector for that video segment. .

[0043] For graph data modalities, the information geometry is dominated by a logical topological relationship graph. The graph data condenses the logical topology of the technical semantics, carrying the structured relationships between technical principles. A graph neural network encoder (denoted as...) is employed. Perform structured semantic encoding on the node features and edge relationships of the graph:

[0044]

[0045] in Let be the semantic vector of the l-th chart. It is the adjacency matrix of the graph, representing the logical dependencies between nodes within the graph. This is the initial feature matrix for the graph nodes. This is the parameter set of the graph neural network encoder. By designing a message passing mechanism to propagate and aggregate semantic information on a graph structure, a semantic vector that preserves the logical topology is output.

[0046] For audio data modalities, the information geometry is dominated by an acoustic-semantic joint manifold. Audio data not only conveys linguistic content but also combines intonation, rhythm, and emotional intensity to convey a temperature vector of emotional resonance. The large-scale audio semantic coding model (denoted as LLM_audio) performs acoustic-semantic joint coding on the audio data:

[0047]

[0048] in Let be the latent semantic vector of the p-th audio segment. This represents the original audio data of segment p. It is the parameter set of the large audio model encoder. Its encoding process simultaneously extracts the semantic information of the language content and the emotional resonance intensity information carried by the acoustic features, and generates an audio semantic vector that integrates the semantics of the content and the emotional temperature.

[0049] S102 takes the co-occurrence relationship between technical concepts and value semantics as the semantic field polarization center, designs a field polarization convergence training strategy, and drives the four types of modal semantic vectors to complete equidistant alignment in the unified Riemann embedding space.

[0050] After completing the dedicated semantic geometry compression for the four modalities, it is necessary to drive the semantic vectors of the four modalities to achieve equi-distance alignment in the unified Riemann embedding space. Let the dimension of the unified Riemann embedding space be... Modality-specific projection layers (denoted as ) are applied to the semantic vectors of the four modalities respectively. , , , The semantic vectors of each modality are projected onto a unified embedding space to obtain the aligned semantic vector of the i-th resource unit in the unified embedding space. .

[0051] Using the co-occurrence relationship between technical concepts and value semantics as the semantic field polarization center, a field polarization convergence training loss function is constructed that exhibits cross-modal synonym attraction and cross-modal heteronymy repulsion. Let the set of positive sample pairs (resource unit pairs with similar cross-modal semantics) be denoted as . The set of negative sample pairs (resource unit pairs with different semantics across modalities) is: The field polarization convergence training loss is:

[0052]

[0053] in The training loss is for field polarization convergence. To unify the Riemann geodesic distance in Riemann embedding space, and These represent the alignment semantic vectors of the i-th and j-th resource units in the unified embedding space, respectively. is the cross-modal semantic exclusion strength coefficient, and m is the exclusion interval threshold, which controls the minimum separation distance between semantically dissimilar resource unit pairs in the embedding space. and and represent the number of positive and negative sample pairs, respectively. By constructing a method that minimizes... This drives cross-modal synonymous resource units to attract and converge with each other in a unified embedding space, and cross-modal heteronymous resource units to repel and separate from each other, achieving equidistant alignment of the four types of modalities in a unified Riemann embedding space.

[0054] S103 is designed with an incremental alignment update mechanism for resource iteration update scenarios, which reduces the global reconstruction cost to local perturbation correction and completes the structural reconstruction of the unified representation of cross-modal semantic fields.

[0055] When the course teaching resource repository undergoes iterative updates (adding new resource units or revising existing ones), traditional methods require a global reconstruction of the entire embedding space, which is extremely costly. This sub-step designs an incremental alignment update mechanism to update newly added resource units. Local perturbation correction is performed within the existing equidistant embedding space, optimizing only the projection parameters of newly added resource units while fixing the embedding vectors of existing resource units:

[0056]

[0057] in Update the loss for incremental alignment. This is the set of semantic nearest neighbors of the newly added resource unit in the existing graph. The semantic similarity label between the newly added resource unit and the j-th nearest neighbor is set (1 for positive sample pairs and -1 for negative sample pairs). This is the stability regularization coefficient. The Frobenius norm difference between the old and new projection parameter matrices is used to constrain the projection parameter update magnitude to maintain the stability of the existing alignment structure. This is achieved by minimizing... This reduces the maintenance cost of the updated resource association system from global reconstruction to local perturbation correction, enabling a structural reconstruction of course value resources from isolated modal storage to deep collaborative representation across modal semantic fields. The final output is a unified set of semantic field vectors. This serves as the standardized input for step S2. The subscripts N, L, M, and P correspond to the quantity of the four resource categories: text, video, charts, and audio, respectively. The unified semantic vector representing the last resource unit in the set.

[0058] This step addresses the problems of processing four modalities independently, traditional keyword annotation only matching on the text surface, and the need for global reconstruction for resource updates. This invention, through the combination of the mid-field polarization convergence training strategy and the cross-frame semantic state progressive anchoring mechanism in step S1, completes equidistant alignment of four types of resources—text, video, charts, and audio—in a unified Riemann embedding space. This makes the cross-modal correlation between technical knowledge and humanistic connotations, which was originally difficult to discover, systematically identifiable, while reducing the cost of resource updates from global reconstruction to local perturbation correction.

[0059] The specific implementation of the S2 technology semantics and value semantics multi-granularity implicit mapping topology discovery module.

[0060] This step follows the unified semantic field vector set E output from step S1. It addresses the granularity gap across multiple semantic abstraction levels between the value semantic abstraction layer (such as the systematic cultivation of independent innovation spirit and the deep construction of dialectical thinking) and the technical semantic primitive layer (such as the timing constraints of bus arbitration logic, dependency resolution of pipeline hazards, and the principle of locality in storage layers). It constructs a deep mapping map of technical value with topological completeness and semantic interpretability. For example... Figure 3 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0061] S201, construct a recursive fine-grained decomposer for value semantics, which decomposes the coarse-grained value semantics of the value semantic abstraction layer into fine-grained value semantic units that can establish direct fiber mappings with technical semantic primitives.

[0062] Based on the semantic type labels of each vector in the unified semantic field vector set E, a subset of nodes belonging to the value semantic abstraction layer is extracted, and the coarse-grained value semantics in this subset are denoted as... The value semantic decomposition big model (denoted as) For each coarse-grained value semantic Perform recursive semantic fine-grained decomposition. The operation of the r-th round of decomposition is defined as follows:

[0063]

[0064] in This is the first sub-proposition generated in the r-th round of decomposition for the k-th coarse-grained value semantic. Let be the number of subpropositions generated in the r-th round of decomposition. To decompose the parameter set of a large model. Initial input. = This refers to the original macroscopic value semantics.

[0065] To avoid over-decomposition leading to the fragmentation of value semantics, semantic decomposability scoring is used as the stopping criterion for recursive decomposition. Let the j-th subproposition in the r-th round of decomposition be... The semantic decomposability score is:

[0066]

[0067] in The value ranges from [0,1]. A larger value indicates that the subproposition still has a semantic space that can be further decomposed at the current granularity level. This is the parameter set for the large-scale semantic decomposability scoring model. When the semantic decomposability score is less than a preset decomposition stopping threshold, further decomposition of the sub-proposition is stopped, and it is marked as a fine-grained value element. And incorporate it into a set of fine-grained value semantic units. .

[0068] S202 constructs a technology concept value-carrying inductive ascender, which elevates the fine-grained technology concepts at the technology semantic primitive layer along the inductive path to a technology principle manifold with value-carrying curvature, achieving bidirectional granular alignment.

[0069] Similarly, a subset of nodes belonging to the technological concept primitive layer is extracted from the unified semantic field vector set E, and the fine-grained technological concepts in this subset are denoted as... The large model is summarized using technical concepts (denoted as...). For each fine-grained technology concept Execute the value-carrying inductive uplift operation; the definition of the s-th round of inductive uplift operation is:

[0070]

[0071] in This is the abstract technical principle after the q-th fine-grained technical concept is inductively derived in the s-th round. This involves summarizing the parameter set of a large model to represent the technical concepts. Initial input. = This is the original concept of fine-grained technology.

[0072] The inductive criterion for halting ascent is a value-bearing curvature score based on technical principles:

[0073]

[0074] in The value range is [0, 1]. The larger the value, the higher the value carrying capacity of the technical principle in terms of value concept manifold. This is the parameter set for the large-scale model of value-bearing curvature scoring. When the value-bearing curvature score is greater than the preset value-bearing curvature threshold, it is determined that the technical principle has reached the manifold of technical principles with value-bearing curvature, and the inductive improvement stops. Marked as a technical principle manifold node And incorporated into the set of technical principles .

[0075] After completing the two-way granularity adjustment, the set of fine-grained value elements Each element in the set of technical principles Each element in the model is at the same semantic granularity level, laying the foundation for granular alignment for subsequent bidirectional semantic penetration and propagation driven by mutual information.

[0076] S203 designs a mutual information-driven bidirectional semantic penetration and propagation mechanism between aligned semantic pairs of the same granularity, which naturally emerges deep implicit mapping topologies in the semantic field and constructs a deep mapping map of technical value.

[0077] Let the value semantic vector after alignment at the same granularity be... The technical concept vector is This corresponds to the node representation in the unified semantic field vector set E of step S1, after granular alignment is completed through value semantic decomposition in S201 and technical concept induction in S202. A mutual information-driven bidirectional semantic penetration and propagation mechanism replaces the traditional unidirectional attention calculation, defining the bidirectional semantic penetration strength between technical concepts and value semantics:

[0078]

[0079] In practical calculations, a neural network mutual information estimator (denoted as ) is used. Differentiable approximation of mutual information:

[0080]

[0081] Representation of technology concept vectors With value semantic vector Mutual information between variables measures the amount of information shared between two variables. Indicates in and Find the expectation under the joint distribution. To use a neural network estimator The estimated value is obtained by differentiable approximation of the mutual information. It is a concatenation of technical concept vectors and value semantic vectors. This is the parameter set of the neural network mutual information estimator. When... ( When the latent mapping emergent threshold is reached, in the deep mapping map of technical value... and Establish an implicit mapping topological edge between them, with an edge weight of . .

[0082] Further perform bidirectional semantic penetration fusion to generate a technology-value fusion node representation. (This involves processing the technology concept vector.) It receives semantic vectors with high mutual information value. Semantic penetration:

[0083]

[0084] in This indicates that the estimated mutual information with the j-th technical principle node exceeds the threshold. A set of value element indexes. This is a fusion representation of the technical concept vector after receiving the permeation of value semantics. A semantic permeation projection matrix from value to technology. This is the normalization coefficient, equal to the sum of all mutual information estimates exceeding the threshold. Symmetrically, value element nodes also receive semantic permeation from high-mutual-information principles, generating... .

[0085] Through bidirectional semantic permeation and propagation, deep implicit mapping topologies such as the value resonance fibers between storage hierarchy and resource allocation fairness, and the semantic manifold homeomorphism between instruction set design philosophy and standardized collaborative spirit naturally emerge in the semantic field, ultimately constructing a deep mapping map of technological value. ,in This is a set of all technical concept nodes and value semantic nodes. The vector representation of each node has been updated to a version fused through bidirectional semantic penetration. and , Output the set of all implicit mapping topological edges that exceed the threshold to step S3.

[0086] This step addresses the problem that direct vector fusion between coarse-grained value semantics and fine-grained technical concepts, due to their vastly different semantic granularities, easily leads to macro-level semantics overshadowing micro-level features, while forced granularity matching generates false associations. This invention achieves bidirectional granularity alignment between the value proposition recursive decomposer and the technical concept inductive ascender in step S2, combined with a bidirectional semantic penetration and propagation mechanism driven by mutual information. This allows deep implicit cross-granularity associations such as "storage hierarchy and resource allocation fairness" and "instruction set philosophy and standardized collaboration" to naturally emerge in the semantic field, breaking through the bottleneck of true associations being undiscovered due to granularity misalignment.

[0087] The specific implementation of the S3 Emotional Resonance-Driven Dynamic Regulation Module for Educational Knowledge Node Activity.

[0088] This step builds upon the deep mapping map of technical value constructed in step S2. To address the fundamental flaw in this learning graph—that the technical principle nodes and value element nodes represent teaching importance with static weights and fail to perceive the dynamic distribution of learners' emotional states—the driving force for node activity is reconstructed from static expert weighting to dynamic emotional field-driven mechanisms. For example... Figure 4 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0089] S301 constructs a multi-dimensional emotional response deep parser to extract the emotional resonance intensity vector for each knowledge node.

[0090] Let the size of the learner group be U. For the u-th learner and the n-th knowledge node... The interaction data can be used to extract emotional signals in three dimensions: the emotional polarity signal of text feedback. (i.e., open-ended text comments written by learners after completing the learning of this knowledge node), and the cognitive input curve signal of the response sequence. Distribution signal of attention dwell time in interaction behavior .

[0091] The emotional polarity signal of the text feedback is analyzed using a large emotional polarity analysis model (denoted as...). Perform fine-grained sentiment polarity analysis:

[0092]

[0093] in [-1, 1] represents the emotional polarity score of the text feedback of the u-th learner to the n-th knowledge node. The closer the value is to 1, the stronger the emotional resonance. The closer the value is to -1, the stronger the emotional resistance.

[0094] Regarding the cognitive input curve signal of response timing, let the learner's response timing at the nth knowledge node be a time series. Temporal convolutional networks (denoted as ) are used. Extracting cognitive input intensity features:

[0095]

[0096] in This is the parameter set for a temporal convolutional network.

[0097] For the attention dwell distribution signal of interactive behavior, statistically analyze the dwell time distribution of learners on each sub-content block of the nth knowledge node. Calculate attention retention concentration:

[0098]

[0099] Note: k represents a knowledge node. The index of the k-th sub-content block. Here, the negative value of the information entropy is used; the smaller the entropy, the more focused the attention. The larger the value, the higher the concentration of cognitive input.

[0100] By combining emotional signals from three dimensions, a multilayer perceptron fusion network (denoted as...) is used. Generate the emotional resonance intensity vector of the u-th learner to the n-th knowledge node:

[0101]

[0102] The average value of all learners is used to obtain the knowledge nodes. The intensity vector of group emotional resonance :

[0103]

[0104] Where U represents the size of the learner group.

[0105] S302 designs an emotional resonance-driven node activity excitation-inhibition dynamics update mechanism to achieve dynamic regulation of knowledge node activity.

[0106] Set knowledge nodes The activity value at time t is The initial activity value is assigned a static weight by domain experts. Based on the intensity vector of group emotional resonance Design the node activity excitation-inhibition dynamics update rule:

[0107]

[0108] in Let be the node activity value at time t+1. This represents the excitation gain coefficient. To suppress the attenuation coefficient. The activation threshold for excitement is determined by the intensity of group emotional resonance. Exceed The time node receives a positive active pulse. To suppress the activation threshold, when the intensity of group emotional resonance remains below a certain level... The inhibition signal received at the time node causes activity decay. ReLU is a linear rectified function, ensuring the unidirectionality of excitation and inhibition signals.

[0109] When knowledge nodes The activity continues to decline to a low activity state. , When the node self-healing trigger threshold is reached, the node self-healing process of regenerating associated resource units in the large model is triggered.

[0110] ]

[0111] in This is the node self-repair trigger flag; a value of 1 triggers the self-repair process, and a value of 0 does not trigger it. `I[·]` is an indicator function; it takes a value of 1 when the condition within the parentheses is true, and 0 otherwise. After triggering self-repair, the large model (denoted as `I[·]`) is regenerated from the node assets. ) Targeting low-activity nodes Regenerate related resource units with stronger potential for emotional resonance:

[0112]

[0113] in Regenerate large models from node materials to create low-activity nodes. The newly generated associated resource unit. For knowledge nodes The contextual information includes its neighborhood structure information in the deep mapping map of technological value. This is the parameter set for a self-healing large model. Newly generated associated resource units. This will replace the original low-activity associated resource unit of the node, as the node. The updated teaching resource units are used to recommend and present subsequent teaching paths in order to enhance the potential for emotional resonance in the teaching process of low-activity nodes.

[0114] S303, design a neighborhood cascade propagation mechanism for active changes to form a knowledge active propagation wave perceived in the teaching field.

[0115] Knowledge Nodes Activity changes Deep mapping of technological value The associated topological edges propagate to neighboring nodes. Let the knowledge node... The set of neighboring nodes is Neighboring nodes The active cascade is updated as follows:

[0116]

[0117] in Neighboring nodes The final activity value after cascading propagation. This is the cascade propagation gain coefficient. Nodes in the graph and The implicit mapping topological edge weights between them (i.e., the mutual information estimates calculated in step S2) ). For the atlas and The topological distance between them (expressed as the shortest path length). This is the distance attenuation coefficient for active cascade propagation, controlling the rate attenuation of the active propagation wave as the topological distance increases. Through a neighborhood cascade propagation mechanism, the knowledge active propagation wave diffuses into the neighborhood structure of the graph, forming a knowledge active propagation wave perceived in the teaching field. This achieves continuous dynamic adaptive reshaping of the overall active distribution of the graph according to the learner's emotional state field, outputting a dynamic active graph. Proceed to step S4. Where A(t) is the set of all node activity values ​​in the time t graph, which is continuously and dynamically updated with the emotional state field.

[0118] This step addresses the problems of traditional knowledge graphs, such as assigning fixed weights to nodes based on expert experience, failing to perceive changes in students' emotional states, and the long-term rigid allocation of teaching resources between high-activity and low-activity nodes. This invention, through the triple synergy of a multi-dimensional emotional response parser, an excitation-inhibition dynamics update mechanism, and neighborhood cascading propagation in step S3, enables the graph activity distribution to be continuously and dynamically adjusted according to the learner's emotional state field, thus opening up the graph-level response channel for improving the effectiveness of teaching and communication through emotional resonance.

[0119] The specific implementation of the S4 era scene-aware education knowledge genealogy inheritance organic growth module.

[0120] This step follows the dynamic activity spectrum output from step S3. To address the need for updating the knowledge graph due to the continuous evolution of the external knowledge context, the operational semantics of graph updating are reconstructed from overwriting and replacement to inheritance and growth. This transforms the value in the course from a static accumulation of items into a dynamic knowledge system with temporal depth and historical connections. For example... Figure 5 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0121] S401 constructs a new resource unit era scene semantic parser, extracts the technical concept vector and value semantic vector of the new course resource unit, and calculates its geodesic distance distribution with the historical nodes of the map.

[0122] The newly added resource units during the dynamic iteration of the course teaching resource database include, but are not limited to, newly compiled teaching videos, text examples, charts and data, and audio materials. Let the set of new resource units be New={ ,..., }, for the r-th new resource unit Analyzing large models using era-specific scenarios (denoted as...) Perform semantic parsing of the era and scene to extract technical concept vectors. With value semantic vector :

[0123]

[0124] in The era-specific tags for new resource units include information on historical background, period markers, and technological paradigms. This is the parameter set of a large model for semantic parsing of era-specific scenarios. and These are the technical concept vector and value semantic vector of the resource unit in the unified semantic field, respectively. Both, along with the technical principle node and value semantic unit node in step S2, belong to the embedding space defined by the unified semantic field vector set E. This provides a unified geometric metric basis for subsequent geodesic distance calculations, allowing the semantic proximity between the new resource unit and historical nodes in the map to be directly calculated in the same space. The semantic vector of the new resource unit... semantic vectors of all existing nodes in the graph Calculate the Riemann geodesic distance distribution in a unified embedding space:

[0125]

[0126] in For new resource units With existing nodes in the graph The Riemann geodesic distance between them. To unify the Riemann geodesic distance function in the Riemann embedding space.

[0127] S402, a dual-threshold growth decision mechanism based on geodesic distance is designed, which activates the corresponding growth operator according to the inheritance threshold and the bifurcation threshold.

[0128] Based on geodesic distance distribution, a dual-threshold growth decision mechanism is defined. Let the inheritance threshold be... The bifurcation threshold is For new curriculum teaching resources The closest historical node to the geodesic distance in the atlas (denoted as the nearest predecessor node) Execute the growth decision:

[0129] when At that time, the judgment was to allow continued growth. At that time, the judgment was that it was bifurcation growth. At that time, the judgment was a newly added independent one.

[0130] When the decision result is inheritance growth, activate the inheritance growth operator. Create a new node in the graph. and by inheriting the edge Connect to the nearest predecessor node The inherited edge carries a triplet of timestamp, era / scene tag, and value transfer vector:

[0131]

[0132] in This serves as the era timestamp for the new resource unit. The era scene label is the one parsed in step S401. The value migration vector is calculated using the following formula: The difference between the value semantic vector of the new resource unit and the value semantic vector of the nearest predecessor node is used to calculate the value migration vector.

[0133]

[0134] in The value semantic vector of the new curriculum teaching resources This is the value semantic vector of the nearest predecessor node. The difference vector... The direction represents the drift of value semantics over time, and the modulus quantifies the drift amplitude, thus accurately depicting the semantic evolution trajectory of the same value semantics in different eras and contexts. The design of value transfer vector triplets carried by inheritance edges fully preserves the evolutionary trajectory of the semantic drift of the same educational concept across eras, enabling the historical inheritance relationship of knowledge to be structurally preserved at the graph level.

[0135] S403 enables the coordinated driving of both inheritance and bifurcation growth operators, resulting in an organic tree-like evolutionary pattern for the graph. When the decision result is bifurcation growth, the bifurcation growth operator is activated, triggering local topological reconstruction of the graph.

[0136]

[0137] in For the nearest predecessor node The subtree with the root. For nodes Depth in the knowledge hierarchy tree. The branching depth penalty coefficient tends to encourage new branches to form at appropriate levels in the knowledge tree rather than at excessively deep levels where branches become too fine. This is based on the branching edge... Connect to new node and the bifurcation point This process creates new branches within the knowledge system, triggering topological reconstruction of the local region and adjusting the weights and directions of relevant edges to adapt to the semantic geometry of the new branch. Driven by the synergistic growth operators of inheritance and branching, the graph exhibits an organic tree-like evolutionary form, enabling the structured presentation and reuse of cross-generational value inheritance in teaching, and outputting a knowledge system with temporal depth. Proceed to step S5.

[0138] This step addresses the problem that traditional graph updates replace old nodes with new ones, resulting in the erasure of the semantic evolution trajectory of the concept of education across eras and the break in the inheritance relationship of knowledge history. The present invention uses a dual-threshold growth decision mechanism based on geodesic distance in step S4 to collaboratively drive the inheritance growth operator and the branching growth operator, so that the graph takes the form of a tree-like organic evolution, preserves the semantic drift trajectory of the same concept of education in different eras, and realizes the structured accumulation of the temporal depth of knowledge system and the value inheritance context.

[0139] The specific implementation of the module for constructing a spatiotemporal coordinate system of course knowledge on the S5 technology-value bimanifold.

[0140] This step follows the knowledge hierarchy with time depth output from step S4. To address the fundamental geometric differences between the hierarchical structure of technical logic and the distribution of emotional resonance based on value concepts, which traditional linear chapter number coordinates cannot capture, this paper constructs two independent geometric coordinate spaces: a technical manifold and a value manifold. A dual-coordinate joint embedding operator is then designed to fuse the two types of coordinate projections into a unified spatiotemporal embedding space. For example... Figure 6 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0141] S501 constructs technical coordinate triples in the technical manifold coordinate layer to accurately describe the geometric position of knowledge units in the technical logical topology.

[0142] Let the set of course knowledge units formed after semantic extraction and structured organization of the course teaching resource database in step S1 be . For each knowledge unit Construct technical coordinate triples at the technical manifold coordinate layer Including topological location coordinates Abstract level Riemann geodetic distance coordinates Directed path encoding of logical dependency chains .

[0143] Among them, topological position coordinates A topological embedding method based on graph neural networks is adopted, which is based on the principle module dependency graph (directed acyclic dependency graph, denoted as ). ) computational knowledge unit Topological location representation:

[0144]

[0145] in For topological embedding graph neural networks, there is a directed acyclic dependency graph. This step is based on the technical logical dependencies between course knowledge units. Perform directed message passing on this graph, Topological locations in the dependency graph are encoded as low-dimensional vectors. This is the set of parameters for the topology-embedded network.

[0146] Abstract level Riemannian geodetic distance coordinates Computational knowledge unit The Riemann geodesic distance vector between the set of anchor points for the abstract hierarchy of technical knowledge. For knowledge units The technical semantic components corresponding to the unified semantic field vector set E in step S1. Anchor point set. This is a pre-defined set of anchor points based on the technical abstraction levels of the course knowledge system. Each anchor point corresponds to a representative technical unit at each level of abstraction, from basic principles to advanced concepts. Let K be the technical concept vector of the k-th anchor point in the unified semantic field, where K is the number of anchor points. The specific calculation is as follows:

[0147]

[0148] Directed path encoding of logical dependency chains Based on knowledge units In a directed acyclic dependency graph All predecessor paths are taken as input, and a directed path code is generated using a path encoder (denoted as PathEncoder):

[0149]

[0150] in In a directed acyclic dependency graph, The set of all predecessor paths leading to the destination. This is the parameter set for the path encoder.

[0151] S502 constructs a humanistic coordinate triplet at the value manifold coordinate layer to accurately describe the emotional geometric position of knowledge units in the value concept manifold.

[0152] For each knowledge unit Constructing a humanistic coordinate triplet at the value manifold coordinate layer Including the semantic manifold coordinates of value propositions Historical context, time scale tags Emotional resonance intensity curvature weight .

[0153] Value proposition semantic manifold coordinates For knowledge units The coordinates of the corresponding value semantics on the value concept semantic manifold are calculated using a value semantic embedding network (denoted as ValEmbed):

[0154] = ValE

[0155] in For knowledge units The corresponding value semantic vector, It is the parameter set of the value semantic embedding network, and its output is the coordinate vector on the value concept semantic manifold, which preserves the manifold intrinsic distance structure between value semantics.

[0156] Historical context, time scale tags Inherited from the era-specific labels assigned during the organic growth of the knowledge system in step S4, the era-specific labels of original knowledge units that have not undergone organic growth and updates are manually marked with initial values ​​by the curriculum design experts.

[0157] Emotional resonance intensity curvature weight The group emotional resonance intensity vector calculated in step S3 Based on the modulus, and combined with the node activity values ​​after neighborhood activity cascade propagation, the curvature weight is calculated:

[0158]

[0159] in For knowledge units corresponding nodes The activity value at time t. This represents the average activity value of all nodes in the entire graph at the current moment. σ is the standard deviation. σ is the sigmoid activation function that maps the curvature weights to the (0,1) interval. and The optimal values ​​are determined by combining the influence coefficients of activity value and emotional resonance intensity, respectively, through hyperparameter search.

[0160] S503 designs a dual-coordinate isometry projection operator on a Riemann product manifold to generate a spatiotemporal coordinate fingerprint carrying complete technical and value semantics.

[0161] Set up a technical coordinate triplet Humanistic coordinates triplet The concatenated vector is Design a bi-coordinate iso-projection operator (denoted as IsoProj) on the Riemann product manifold to... Jointly projected onto a unified spatiotemporal embedding space, a spatiotemporal coordinate fingerprint is generated:

[0162]

[0163] in For knowledge units Spatiotemporal coordinate fingerprint. The weight matrix is ​​an isometric projection matrix, which is trained by isometric mapping constraints (i.e., keeping the Riemann geodesic distance in the input space approximately unchanged after projection). Here is the bias vector. tanh is the hyperbolic tangent activation function, which maps the coordinate fingerprint to... space, Let be the dimension of the spatiotemporal coordinate fingerprint. The isometric projection constraint is implemented using the following isometric loss function:

[0164]

[0165] in for and Geodesic distance on the Riemann product manifold. By minimizing This ensures that the isometric projection operator maintains the geometric relationships between knowledge units in the spatiotemporal coordinate fingerprint space, guaranteeing that the teaching function and value effectiveness of any knowledge unit can be uniquely and accurately located by the spatiotemporal coordinate fingerprint. The output is a set of spatiotemporal coordinate fingerprints.

[0166] This step addresses the problem that traditional linear chapter number coordinates cannot simultaneously characterize the hierarchy of technical logic dependencies and the distribution of value and emotion, and the ambiguity in the teaching function positioning of knowledge units. In step S5, this invention uses a joint projection of technical coordinate triples and humanistic coordinate triples onto a Riemannian product manifold to generate a spatiotemporal coordinate fingerprint for each knowledge unit, carrying complete technical and value semantics. This allows the teaching function and value role of any knowledge unit to be uniquely and accurately located, providing a precise coordinate navigation foundation for subsequent field perception activation.

[0167] The specific implementation of the S6 teaching field perception fusion knowledge intelligent agent semantic activation self-reconstruction module.

[0168] This step follows the spatiotemporal coordinate fingerprint set generated in step S5. Addressing the fundamental deficiency of fixed activation weights for knowledge units across different teaching contexts and the long-term inability to fully unleash their teaching potential, this step constructs each knowledge unit as an autonomous intelligent agent with contextual awareness. The weights of the technical activation channel and the value activation channel are adaptively allocated using the teaching context vector as a conditional input. For example... Figure 7 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0169] S601 designs a dual-channel semantic activation competition allocation mechanism to realize the dynamic competition for dominance between the technical semantic activation channel and the value proposition activation channel.

[0170] Knowledge Units Corresponding fusion knowledge intelligent agent Define the technical concept to activate the channel With value semantic activation channel At each teaching moment t, the current field context vector is used. Competition conditions drive the dynamic struggle for dominance between the two types of activation channels.

[0171] Field context vector It consists of three parts: course stage labels (Using one-hot encoding to represent the technical principle introduction stage, the value topic in-depth stage, or the comprehensive practice stage), a snapshot of the learner's emotional state. (Take the real-time output of the multidimensional emotion response deep parser in step S3) Summary of the interaction trajectory between learner u at the current moment (i.e., the emotional resonance intensity vector of learner u) and its historical interaction trajectory. (Represented by the summary vector of the K most recent interactions):

[0172]

[0173] The competition scores for the technology concept activation channel and the value semantic activation channel are calculated by the field-aware selective routing operator (denoted as Router):

[0174]

[0175] in The competitive score for activating channels for technological concepts. The competitive score for the value semantic activation channel. For field-aware selective routing operators, the spatiotemporal coordinate fingerprint of knowledge units is used. With field context vector The concatenation of is the input. Internally, the system learns the activation efficiency patterns of the two types of channels under different field scenarios in historical data to determine the competitive scores of the current input and output channels. This is the parameter set for the routing operator. The dynamic activation weight allocation for the two types of activation channels is obtained through Softmax normalization:

[0176]

[0177]

[0178] Knowledge Unit The activation semantics at time t are represented as:

[0179]

[0180] in This is an activation representation of the technology concept dominated by technology manifold coordinates. It is a value semantic activation representation dominated by value manifold coordinates. To activate the semantic representation of the integrated knowledge units, the activation ratio of the technical concept side and the value semantic side is dynamically adjusted according to the teaching context.

[0181] S602, design a local activation state negotiation and propagation protocol among intelligent agents to form a knowledge activation collaborative propagation wave for the perception of the teaching field.

[0182] Knowledge Intelligent Agent Its currently active semantic representation As a message in the knowledge spectrum Local broadcast is performed. Neighborhood intelligent agent. ( ∈ ,in For knowledge nodes In the knowledge spectrum The set of neighboring nodes on the network receives data from... activation semantic representation It also incorporates the neighborhood activation signal into the context input of its own routing decision:

[0183]

[0184] in For time t intelligent agent For neighborhood intelligent agents The attention weights are calculated using a graph attention mechanism:

[0185]

[0186] Where a is the attention vector. This is the attention transformation matrix. For neighborhood set Perform normalization. LeakyReLU is a linear rectified activation function with leakage. This is the message transformation matrix, used to perform a linear transformation on the activation messages sent by the neighboring agents.

[0187] Aggregated neighborhood activation signals Incorporate the context vector to update the agent Field context vector:

[0188]

[0189] With the updated The dual-channel competitive allocation in step S601 is re-executed to form a knowledge activation collaborative propagation wave of teaching field perception, so that the activation state of the neighboring agent has a collaborative influence on the routing decision of the current agent.

[0190] S603 uses positive feedback signals of teaching effectiveness as a driving force to achieve continuous self-updating of routing decision parameters.

[0191] intelligent agent Routing decision parameters It continuously self-updates, driven by positive feedback signals of teaching effectiveness. Let the positive feedback signal of teaching effectiveness be... Includes the learner completing a knowledge unit at time t. Post-learning assessment scores Enhanced by emotional resonance ,in Obtained from in-class assessment records of the teaching system. The difference between the emotional resonance vectors extracted before and after learning by the multi-dimensional emotional response deep parser in step S3 is obtained by calculating the difference:

[0192]

[0193] in and These are the assessment gain weight and the emotional resonance gain weight, respectively. and Identify the course phase function. This is achieved by constructing a minimized... This allows routing decision parameters to continuously converge towards channel allocation strategies that have historically activated better performance, enabling the intelligent knowledge ecosystem characteristics of dynamic evolution of course content within the teaching context, semantic self-reconstruction, and precise value embedding. The output is a set of activation semantic representations. Proceed to step S7.

[0194] This step addresses the problem that fixed activation weights cannot respond to the dynamic emphasis on the technical and value sides in different teaching scenarios such as technical introduction, in-depth topic study, and comprehensive practice, and that the teaching potential of knowledge units is not fully released due to the solidification of activation modes. This invention, through the combination of the dual-channel semantic activation competition allocation mechanism and the inter-agent neighborhood activation negotiation and propagation protocol in step S6, enables the course content to automatically adjust the activation ratio of technical and value semantics in different teaching situations, thus breaking through the suppression of knowledge teaching potential by activation solidification.

[0195] The specific implementation of the S7 neuroevolutionary architecture search-driven instructional path topology space evolution module.

[0196] This step follows the set of activation semantic representations output in step S6. To address the triple computational challenges faced in generating optimal personalized learning paths—heterogeneous constraint manifold exponential expansion, multi-objective Pareto front search dilemma, and insufficient cross-learner fast topological adaptability—this paper treats the topology of the path generation network itself as a search object that dynamically evolves with the emotional landscape. For example... Figure 8 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0197] S701 constructs a dynamic 3D adaptive landscape generator, which uses real-time synthesis technology to create a dynamic 3D adaptive terrain composed of a logical coherence surface, a value penetration density surface, and a learner emotional resonance prediction surface.

[0198] The three-dimensional adaptive landscape is based on the spatiotemporal coordinate fingerprint of knowledge units, which jointly drives the real-time synthesis of three surfaces.

[0199] Technical Logic Coherence Surface surface The construction of (the technical dependencies used to measure the rationality and logical fluency of the sequence of knowledge points in the teaching path) is based on a directed acyclic dependency graph. For constraints, the path Adjacent knowledge units The score reflects the logical coherence of the computational technique (where L is the total number of knowledge units included in the instructional path). (representing the last knowledge unit of the path)

[0200]

[0201] in For knowledge units Technical manifold coordinate components in spatiotemporal coordinate fingerprinting. The temperature coefficient for technical consistency. [This is] a directed dependency constraint indicator function that ensures that adjacent knowledge units satisfy the dependency constraints of the technical logic. This function is a pre-path validity check rule independent of the scoring formula mentioned above, used to ensure that adjacent knowledge units satisfy the dependency constraints of the technical logic. Knowledge point pairs conform to... The function takes the value 1 when there is a dependency, otherwise it takes the value 0.

[0202] Value Penetration Density Surface surface The construction is based on value manifold coordinates, and the calculation path Density of the space of covered value propositions:

[0203]

[0204] in For path The total number of knowledge units included. This is the location index of the knowledge unit in the path, with a value range from 1 to L. The kernel density estimate, taken as input value manifold coordinates, measures the knowledge unit. Local coverage density on the value concept manifold. This is the bandwidth parameter for kernel density estimation. The curvature weight of the emotional resonance intensity is calculated in step S5.

[0205] Learner Emotional Resonance Prediction Surface The construction of learner profile vectors Using historical emotional resonance records and cognitive style features as input, an emotional resonance prediction network (denoted as EmoPredict) is employed to predict learner u's path. The overall intensity of emotional resonance predicted above:

[0206]

[0207] in For path The l-th knowledge unit in the corresponding teaching time The activation semantic representation sequence, This is the parameter set for the emotional resonance prediction network, which includes the weight matrix and bias vector of each layer of the network.

[0208] The three-curved surface constitutes a dynamic three-dimensional adaptive terrain in the form of a Riemannian product manifold, which affects the path. The overall fitness is calculated as follows:

[0209]

[0210] To target specific candidate teaching paths With specific learners The three-dimensional comprehensive fitness vector is used to quantify the overall teaching effectiveness of the path for the learner. Let u be any candidate teaching path consisting of a sequence of knowledge units, and u be a unique identifier for a single learner. The three components of the vector correspond to three independent optimization objectives: For path The overall technical logic coherence score is determined solely by the sequential order of knowledge units within the path itself, and does not vary with individual learner differences. For path The density score of the covered value proposition space reflects the extent to which the path covers the pre-set value guidance objectives of the course, and is also only related to the knowledge unit composition of the path. For learner u in the path The overall intensity score for predicting emotional resonance depends on both the content structure of the path and the learner's individual emotional preferences and cognitive style characteristics. Using a three-dimensional vector format instead of a single scalar weighting preserves the independent evaluation information of all three dimensions, providing a complete decision-making basis for subsequent multi-objective optimization.

[0211] By replacing scalar weighted summation with Pareto dominance relation, a multi-objective Pareto front search is performed in the three-dimensional fitness space, avoiding the fundamental defect that single-objective optimization strategies cannot approximate the true Pareto optimal surface.

[0212] S702 designs a semantically aware dual-operator evolution mechanism to drive the path generation network population to continuously converge toward the Pareto optimal solution in three-dimensional fitness terrain.

[0213] Let the family of path generation networks be { Each path generates a network. Learner profile vectors and activation semantic representation set Input, output personalized teaching path . A profile vector for a single learner u. Let T be the complete set of dynamic activation semantic representations of all knowledge units.

[0214] The topological mutation operator TopMutate generates a network along the geodesic direction of the technology-dependent constrained manifold for the current path. Perform local topology perturbation:

[0215]

[0216] This formula defines the input-output relationship of the topological mutation operator. Generate a network for the original path to be mutated. The technical logic of the course is represented by a directed acyclic dependency graph. For the complete set of spatiotemporal coordinate fingerprints of all knowledge units, It is the set of hyperparameters for the topological mutation operator, which controls the magnitude, range, and probability of topological perturbations.

[0217] The core constraint of the topological mutation operator is: the mutated network Generated arbitrary paths The technical logic dependency constraints must be met, that is... All adjacent knowledge unit pairs are in a directed acyclic dependency graph. There are legitimate dependency edges in the structure, ensuring that the continuity of technical logic is not disrupted by topological disturbances.

[0218] The topological recombination operator TopCrossover performs cross-individual gene segment recombination in high-fitness network architectures near the Pareto front of the current population.

[0219]

[0220] in and These are two highly fit parent networks selected from the current Pareto fronts. ParetoFront(Pop) is the set of Pareto fronts for the current population. This is the set of hyperparameters for topological recombination operators, controlling the length of gene fragment exchange and the recombination probability. Topological recombination operator identification. and Complementary superior gene fragments in emotional adaptation patterns are used to perform cross-individual recombination at the gene fragment level, integrating and disseminating superior genes of different learners' emotional adaptation patterns at the population level.

[0221] The population evolution iteratively performs topological mutation and recombination, and drives the population to continuously converge using the gradient gravity of the three-dimensional fitness terrain:

[0222]

[0223] in The Pareto non-dominated sorting and crowding distance calculation are performed using the Non-dominated Sorting Genetic Algorithm II (NSGA-II) framework with an elitist strategy. Generate a network population for the parent path of generation t. The family of descendant path generation networks generated by topological mutation and topological recombination operations in generation t. This is the three-dimensional comprehensive fitness evaluation function defined above. The selection operator selects a next-generation population of size S from the union of the current population and the offspring population, ensuring that the population continuously converges towards the dominant solution in the three-dimensional equilibrium of technical logic rigor, humanistic connotation penetration, and emotional adaptation flexibility.

[0224] S703 extracts the Pareto optimal path population from the evolved optimal path generation network population.

[0225] After the prescribed algebra After evolutionary iterations, the optimal set of path generation networks is extracted from the Pareto front of the path generation network family Pop. For a specific learner u, its learner profile vector is obtained from ParetoOpt. To guide the selection of the path with the highest fitness, a network is generated. Generate personalized Pareto optimal teaching paths :

[0226]

[0227] Output Pareto optimal path population The corresponding three-dimensional fitness landscape state is then processed to step S8 to support the online dynamic reshaping driven by the emotional state field.

[0228] This step addresses the triple dilemma of the legal path space expanding exponentially with the number of knowledge points, the multi-objective Pareto search problem arising from the joint optimization of technical coherence and value penetration, and the insufficient generalization ability of fixed structure strategies for students with heterogeneous emotions. This invention, through the synergy of the three-dimensional fitness landscape dynamic generator and the semantic perception dual-operator evolution mechanism in step S7, uses the path generation network topology itself as the evolutionary search object to generate an optimal path generation network population that is balanced and dominant in three dimensions: rigorous technical logic, permeability of humanistic connotation, and flexibility of emotional adaptation.

[0229] The specific implementation of the S8 Emotional State Field-Driven Online Drift and Adaptive Reshaping Module for Teaching Paths.

[0230] This step follows up on the Pareto optimal path population output in step S7. And the three-dimensional fitness landscape state, addressing the fundamental flaw of how the dynamic evolution of learners' emotional states in real teaching practice continuously perturbs the curvature distribution of the fitness landscape, causing the offline optimal path to deviate from the Pareto front, the path adjustment mechanism is reconstructed from discrete triggering to continuous online drift driven by the emotional field gradient, ultimately forming an eight-in-one intelligent teaching adaptive evolution system for curriculum development. For example... Figure 9 As shown, the specific implementation process of this step is divided into the following three sub-steps.

[0231] S801 constructs a continuous estimator of the emotional state field for real-time emotional feedback of learners, interpolating the discrete emotional feedback signal into a continuous emotional state field distribution.

[0232] During the path execution process, the real-time emotional feedback generated by learner u at instruction time t is a discrete signal sequence. Where u is the unique identifier of a single learner, T is the total number of emotional feedback signals generated by learner u during the execution of the current path, and the feedback signals at each time point are extracted in real time by the multidimensional emotional response deep parser in step S3.

[0233] The discrete emotional feedback signal is kernel interpolated in the path execution space (with the knowledge unit sequence index as the coordinate axis) to generate a continuous emotional state field distribution. (x, t):

[0234]

[0235] Where x is a continuous coordinate in the path execution space (parameterized by path progress). i is the index of the discrete feedback signal, ranging from 1 to T. T is the total number of discrete emotional feedback signals generated by learner u during the execution of the current teaching path. The coordinates represent the path execution progress when the i-th emotional feedback occurs. The kernel bandwidth in the spatial dimension controls the spatial impact range of emotional feedback. The kernel bandwidth in the spatial dimension controls the spatial impact range of emotional feedback. The kernel bandwidth in the time dimension controls the time decay rate of historical emotional feedback. For the spacetime kernel function, an anisotropic Gaussian kernel is used:

[0236]

[0237] in The position offset in the path execution space is equal to the current coordinate x and the path coordinates when the i-th emotion feedback occurs. The difference. t is the time offset, which is equal to the time t between the current time and the time when the i-th emotional feedback occurs. The difference. In this formula Pi is a mathematical constant.

[0238] Calculate the spatial gradient vector of the emotional state field It uses the fitness function parameters of the real-time injected path generation network to drive the current execution path to perform continuous online drift along the emotional landscape gradient direction:

[0239]

[0240] in To determine the optimal teaching path for learner u at time t. t represents the time step for online path updates. For the process The new optimal teaching path after the time step t is updated. To control the magnitude of a single path adjustment during online drifting, and to prevent overly drastic adjustments from disrupting the teaching logic. The coordinates of the current path execution progress in the path execution space. The gradient vector of the emotional field at the current location points in the direction of the greatest emotional resonance gain, driving the path to continuously drift to the high resonance area in the emotional landscape.

[0241] S802 designs a local path topology self-repair mechanism for Pareto deviation perception, generating a path that aligns with the current emotional trajectory. The projected distance relative to the Pareto front in the three-dimensional fitness space:

[0242]

[0243] in For any reference optimal path on the Pareto front, when ( When the adaptive reshaping threshold is reached, a local path topology self-repair mechanism is triggered:

[0244]

[0245] After the reshaping is triggered, the current curvature distribution of the emotional state field is used as the geometric constraint. (Curvature distribution of the emotional state field) The Laplace operator of the emotion state field is obtained by taking the second-order partial derivative of the emotion state field with respect to the path spatial coordinate x.

[0246]

[0247] The curvature distribution reflects the local rate of change in the emotional state field: positive curvature indicates an increasing trend in learner emotional resonance in that region, while negative curvature indicates a decreasing trend. Using this as a geometric constraint, in the knowledge unit neighborhood graph (i.e., the knowledge spectrum)... Perform a heuristic topology repair search on a local subgraph to generate local path replacement fragments that precisely match the curvature of the current mood landscape. :

[0248]

[0249] TopRepairSearch is a heuristic topology repair search algorithm that uses the current emotional landscape curvature distribution as heuristic information, while satisfying the technology dependency constraint (based on a directed acyclic dependency graph). Under the premise of ensuring (the path is properly positioned), search for local path replacement fragments so that the 3D fitness vector of the replaced path falls back into the Pareto front neighborhood. Replace the local path fragments... Embed the current execution path At the corresponding location, complete the local self-repair of the path topology:

[0250]

[0251] in This is the index of the segment to be replaced in the current path. Replace performs a path segment replacement operation.

[0252] S803 designs a long-range knowledge distillation and accumulation mechanism across learner groups' emotional state fields, driving the underlying optimization goals of path generation strategies to continuously calibrate towards real teaching patterns.

[0253] For the entire learner group { , , ..., The emotional state field By performing weighted superposition, the statistical regularities of the group's emotional state field are continuously distilled into long-term correction terms of the evolutionary population fitness function:

[0254]

[0255] in The group weight for the i-th learner is obtained by normalizing it using their historical learning performance scores:

[0256]

[0257] Average value of group emotional state Guided by statistical regularities, a long-range knowledge distillation accumulation loss function is constructed to distill the patterns of group emotions into a long-term correction term Δ of the evolutionary population fitness function. :

[0258]

[0259] in To incorporate the correction fitness after incorporating long-term correction terms. To construct a fitness correction parameter oriented towards the group's emotional state field, a minimum is required. This allows the fitness function parameters of the evolutionary population to be continuously calibrated towards the true laws of teaching.

[0260] With the corrected fitness function By replacing the original fitness function Fitness in step S7, the underlying optimization objective of the driving path generation strategy is updated and corrected once after each teaching cycle. This allows the Pareto front of the evolutionary population to continuously drift towards a better solution as real teaching patterns accumulate. Ultimately, this forms a complete system of intelligent teaching adaptation and evolution that integrates cross-modal semantic field representation, latent association topological emergence, node activity and emotional regulation, organic growth of knowledge system, precise spatiotemporal dual-coordinate positioning, knowledge intelligent agent field self-reconstruction, neural evolution path topological evolution, and online reshaping of the emotional field-driven path. This achieves a systematic leap in the effectiveness of curriculum-based education.

[0261] This step addresses the problem that value identification mutations, cognitive fatigue accumulation, and the continuous perturbation of fitness landscape by the emergence of group emotional resonance cause the offline optimal path to gradually deviate from the Pareto front during online execution. This invention achieves online dynamic maintenance of the Pareto optimality of the teaching path by using a three-in-one synergy of the emotional state field continuous estimator, the Pareto deviation perception local topology self-repair mechanism, and the long-range knowledge distillation accumulation across learner groups in step S8. This changes the path adjustment from discrete triggering to continuous online drift driven by the emotional field gradient.

[0262] The functional modules constructed in steps S1 to S8 above are interconnected in actual deployment through two stages: offline preprocessing and online inference. Together, they constitute the complete technical framework of the intelligent adaptive evolution system for curriculum-based moral education described in this embodiment. To intuitively demonstrate the data flow connections and overall deployment logic between the modules, Figure 10 A schematic diagram of the system deployment architecture and data flow in this embodiment is provided. See also... Figure 10The system is divided into two main streams: offline data stream and real-time data stream. In the offline phase, after course resources are uploaded, they sequentially undergo S1 cross-modal semantic unified encoding, S2 technology-value implicit mapping discovery, S3 emotional resonance-driven offline node activity initialization, and S4 knowledge lineage inheritance-based organic growth to construct a static knowledge graph foundation with temporal depth. In the online phase, S5 dual-manifold spatiotemporal coordinate positioning and S6 field-aware agent semantic activation self-reconstruction receive real-time learning context input, providing dynamic semantic input conditions for S7 neural evolution path search. Finally, the S8 emotion state field-driven path online drift and local self-repair module injects learners' real-time emotional feedback as a continuous state field into the path optimization process, forming a closed-loop real-time data stream of "emotion measurement—path dynamic adjustment—terminal presentation." Simultaneously, the cross-learner group emotional patterns accumulated in the online phase are fed back to the offline knowledge base through a long-range distillation mechanism, driving the periodic self-evolution and updating of knowledge node activity distribution and fitness functions.

[0263] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for generating a curriculum knowledge neural evolution teaching path driven by emotional resonance, characterized in that, Includes the following steps: Step S1: Perform geometric semantic compression on the four modalities of text lecture notes, video cases, chart data and audio materials in the course teaching resource library, and complete equidistant alignment in the unified Riemann embedding space to obtain a unified semantic field vector set; implement an incremental alignment update mechanism for the resource units of the added course teaching resources, and perform local perturbation correction in the unified Riemann embedding space. Step S2: Use the value semantic decomposition big model to perform recursive semantic fine-grained decomposition on each coarse-grained value semantic; use the technology concept induction big model to perform value-carrying induction coarse-grained up-process operation on each fine-grained technology concept to achieve bidirectional granular alignment; design a mutual information-driven bidirectional semantic penetration and propagation mechanism between aligned semantic pairs of the same granularity, and construct a deep mapping map of technology value based on the aligned value semantic and technology concepts. Step S3: Obtain the interaction data of the learner group and extract the emotional resonance intensity vector of each knowledge node in the deep mapping graph of technology value; The activity value of knowledge nodes is calculated based on the emotional resonance intensity vector, so that nodes with high activity values ​​receive positive activity pulses and priority activation weights, while nodes with low activity values ​​have their activity decayed and trigger the self-repair generation of associated resource units when they fall below a threshold; and a neighborhood cascade propagation mechanism is generated so that changes in node activity propagate to the neighborhood along the associated edges of the deep mapping graph of technical value, and outputs a dynamic activity graph. Step S4: Extract the technical concept vector, value semantic vector and era scene label of the new resource unit, and calculate its geodesic distance to the historical nodes in the dynamic active graph. Depending on whether the geodesic distance falls into the inheritance threshold or exceeds the bifurcation threshold, activate the inheritance growth operator or the bifurcation growth operator respectively, and link the new node into the dynamic active graph with the inheritance edge or the bifurcation edge to obtain the knowledge spectrum with time depth. Step S5: Construct a technical coordinate triplet on the technical manifold coordinate layer to describe the geometric position of the knowledge unit in the technical logical topology; construct a humanistic coordinate triplet on the value manifold coordinate layer to describe the emotional geometric position of the knowledge unit in the value concept manifold; design a dual-coordinate isometric projection operator on the Riemann product manifold to fuse the two types of coordinate projections into a unified spatiotemporal embedding space to obtain a spatiotemporal coordinate fingerprint set. Step S6: Construct each knowledge unit into an autonomous intelligent agent with field awareness capabilities, set up a dual-channel semantic activation competition allocation mechanism, so that the technical concept activation channel and the value semantic activation channel dynamically compete for activation dominance based on the field context vector containing course stage, learner emotional state and historical interaction trajectory; and design a local activation state negotiation and propagation protocol between intelligent agents, so that the neighboring activation state affects the routing decision of the node, and the routing parameters are continuously updated by the positive feedback signal of teaching effect, and output dynamic activation semantic representation. Step S7: Based on the spatiotemporal coordinate fingerprint of knowledge units, synthesize a dynamic three-dimensional fitness terrain composed of a technology logic coherence surface, a value penetration density surface, and a learner emotional resonance prediction surface; use the topology of the path generation network as the evolutionary search object, and use a topological mutation operator that is executed along the geodesic direction of the technology-dependent manifold and a topological recombination operator that exchanges superior gene fragments near the Pareto front to drive the population to search for the Pareto optimal teaching path in the three-dimensional fitness terrain; Step S8: The learner's discrete emotional feedback signal is interpolated into a continuous emotional state field distribution using a spatiotemporal kernel function. The spatial gradient vector of the emotional state field drives the current teaching path to continuously drift online along the emotional landscape gradient direction. A local path topology self-repair mechanism for Pareto deviation perception is set up. When the path deviates from the Pareto front by more than a threshold, the curvature of the emotional state field is used as a constraint to search for and replace local path segments. At the same time, the statistical regularity of the group's emotional state field is continuously distilled into a long-term correction term of the evolutionary population fitness function to drive the continuous calibration of the optimization objective.

2. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 1, characterized in that, The specific steps of S1 are as follows: Step S101: For the text handout modality, a large text semantic encoding model is used to perform deep semantic encoding on the original text handout to extract and preserve the latent text semantic vectors that preserve the technical concept dependencies and value semantic density distribution. For video case modalities, a visual semantic coding big model is used to perform frame-by-frame semantic coding on video cases, and a cross-frame semantic state progressive anchoring mechanism is designed to preserve the dynamic changes of the context along the frame sequence, resulting in anchored semantic vectors. The anchored semantic vectors of the entire video frame sequence are weighted and pooled in the time dimension to generate the overall semantic vector of the video case. For graph data modalities, a graph neural network encoder is used to perform structured semantic encoding on the node features and edge relationships of the graph, and outputs a semantic vector that preserves the logical topological structure. For audio data modalities, an audio semantic coding big model is used to perform acoustic-semantic joint coding on the audio data. During the coding process, the semantic information of the language content and the emotional resonance intensity information carried by the acoustic features are extracted simultaneously to generate an audio semantic vector that integrates the semantic content and the emotional temperature. Step S102: Set the dimension of the unified Riemann embedding space, apply a modality-specific projection layer to the semantic vectors of the four modalities respectively, and project the semantic vectors of each modality onto the unified Riemann embedding space. Taking the co-occurrence relationship between technical concepts and value semantics as the semantic field polarization center, a field polarization convergence training loss function with cross-modal synonym attraction and cross-modal heteronymy repulsion is constructed to achieve equidistant alignment of four types of modalities in a unified Riemann embedding space and output a unified semantic field vector set. Step S103: When the course teaching resource library is iteratively updated, local perturbation correction is performed in the existing equidistant embedding space, and the projection parameters of newly added resource units are optimized only.

3. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 2, characterized in that, Step S2 is as follows: Step S201: Set a coarse-grained value semantic set, and perform recursive semantic fine-grained decomposition on each coarse-grained value semantic using the value semantic decomposition big model to obtain the fine-grained value elements of the recursive decomposition. The semantic decomposability score of the value elements in the current decomposition round is used as the stopping criterion for recursive decomposition. Step S202: Set a set of fine-grained technical concepts, and perform a value-carrying inductive ascending operation on each fine-grained technical concept using the large model of technical concept induction to obtain the technical principles of induction ascending. The stopping criterion of induction ascending is the value-carrying curvature score of the technical principles of the current ascending round. Step S203: After completing the bidirectional granularity adjustment, each element in the fine-grained value element set and each element in the technical concept set are at the same semantic granularity layer. The aligned technical concepts and value semantics are used as nodes. Based on the mutual information-driven bidirectional semantic penetration propagation mechanism, the bidirectional semantic penetration strength between technical concepts and value semantics is calculated. When the bidirectional semantic penetration strength is greater than the implicit mapping emergence threshold, an implicit mapping topological edge is established between technical concepts and value semantics in the deep mapping graph of technical value. The edge weight is the bidirectional semantic penetration strength. Further bidirectional semantic penetration fusion is performed. For node pairs of technical concepts and value semantics that have established implicit mapping topological edges, the corresponding nodes are fused based on the bidirectional semantic penetration strength to generate technical-value fusion nodes, and finally a deep mapping map of technical value is constructed.

4. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 3, characterized in that, Step S3 is as follows: Step S301: Let the size of the learner group be U. For the u-th learner, the n-th knowledge node in the deep mapping graph of technology value... From the interaction data, we extracted emotional signals in three dimensions: emotional polarity signals of text feedback, cognitive input curve signals of response time, and attention retention distribution signals of interaction behavior; A large-scale sentiment polarity analysis model is used to perform fine-grained sentiment polarity analysis on the sentiment polarity signal of the text feedback, and the sentiment polarity score of the text feedback is obtained. A temporal convolutional network is used to extract cognitive input intensity features from the temporal cognitive input curve signal of the response; Calculate attention concentration based on the attention residence distribution signal of interactive behavior; Finally, a multilayer perceptron fusion network is used to fuse text feedback emotional polarity scores, cognitive engagement intensity features, and attentional retention concentration to obtain the emotional resonance intensity vector of the u-th learner for the n-th knowledge node; the mean value is then taken over the learner group to obtain the knowledge node... The intensity vector of group emotional resonance ; Step S302, set knowledge nodes The activity value at time t is The initial activity value is assigned a static weight by domain experts; Based on the intensity vector of group emotional resonance Design the node activity excitation-inhibition dynamics update rule: in, Let be the node activity value at time t+1. The excitation gain coefficient, To suppress the attenuation coefficient, The excitation activation threshold, To suppress the activation threshold, ReLU is a linear rectified function; When a node's activity value falls below the node's self-repair trigger threshold, the node self-repair process is triggered to regenerate the associated resource units of the large model, thus regenerating the large model from node materials. For this node Regenerate related resource units with stronger potential for emotional resonance: in For the newly generated associated resource unit, For nodes Neighborhood structure information in the deep mapping map of technological value Regenerate the parameter set of the large model from the node materials; Step S303: The change in node activity value propagates to neighboring nodes along the associated topological edges of the deep mapping graph of technical value. Let the knowledge node... The set of neighboring nodes is Its neighboring nodes The active cascade is updated as follows: in Neighboring nodes The final activity value after cascading propagation, This is the cascade propagation gain coefficient. Knowledge nodes in the deep mapping graph of technological value and Edge weights between them In the deep mapping of technological value and Topological distance between them The distance attenuation coefficient for active cascade propagation is used to obtain the dynamic activity spectrum.

5. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 4, characterized in that, Step S4 is as follows: Step S401, let the new resource unit set be New = { , ..., }, for the r-th new resource unit Analyzing large models using historical context Perform semantic parsing of the execution era and extract technical concept vectors Value semantic vector and era scene tags It contains information about the historical context; The semantic vector of the new resource unit semantic vectors of all existing knowledge nodes in the graph Calculate the Riemann geodesic distance distribution in a unified embedding space: in For new resource units With existing knowledge nodes in the graph The Riemann geodesic distance between them For Riemann geodesic distance function; Step S402: Based on the geodesic distance distribution, define a dual-threshold growth decision mechanism; let the inheritance threshold be... The bifurcation threshold is For new resource units The historical node with the smallest geodesic distance in the map Execution of growth judgment: when At that time, the judgment was to allow continued growth; when When, the judgment is that it is bifurcation growth; when At that time, the judgment was a newly added independent judgment; When the decision result is inheritance growth, activate the inheritance growth operator; create a new node in the graph. and by inheriting the edge Connect to the nearest predecessor node The inherited edge carries a triplet of timestamp, era / scene tag, and value transfer vector: in For the new resource unit's era timestamp; The value transfer vector is calculated using the following formula: in The value semantic vector of the new resource unit. The value semantic vector of the nearest predecessor node; Step S403, when the decision result is bifurcation growth, activate the bifurcation growth operator to trigger local topological reconstruction of the graph: in For the nearest predecessor node The subtree rooted at the root; For nodes Depth within the knowledge hierarchy tree; The bifurcation depth penalty coefficient is used; the bifurcation edge is used as the basis for the penalty coefficient. Connect to new node and the bifurcation point This forms a new branch of the knowledge system, triggers the topological reconstruction of the local region, adjusts the weights and directions of the relevant edges to adapt to the semantic geometry of the new branch, and outputs a knowledge system with time depth.

6. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 5, characterized in that, Step S5 is as follows: Step S501, let the set of knowledge units formed after semantic extraction and structured organization of the course teaching resource database be . For each knowledge unit Construct technical coordinate triples at the technical manifold coordinate layer Including topological location coordinates Abstract level Riemann geodetic distance coordinates Directed path encoding of logical dependency chains ; Based on the technical logical dependencies between course knowledge units, a directed acyclic dependency graph is constructed, and the knowledge units are calculated using a topological embedding method based on graph neural networks. Topological position coordinates; The abstract level of Riemannian geodetic distance coordinates is a unit of computational knowledge. The Riemann geodesic distance vector between the set of anchor points for the abstract hierarchy of technical knowledge; Logical dependency chain directed path encoding in knowledge units The path encoder is used to generate the path from all predecessor paths in the directed acyclic dependency graph. Step S502, for each knowledge unit Constructing a humanistic coordinate triplet at the value manifold coordinate layer Including the semantic manifold coordinates of value propositions Era and scene tags Emotional resonance intensity curvature weight ; Value proposition semantic manifold coordinates For knowledge units The corresponding value semantic vector is calculated using a value semantic embedding network; Emotional resonance intensity curvature weight The following formula is used to calculate: in For knowledge units corresponding nodes The activity value at time t, This represents the average activity value of all nodes in the entire graph at the current moment. Let σ be the standard deviation, and σ be the sigmoid activation function. and These are the influence coefficients of activity value and emotional resonance intensity, respectively. Step S503: Design a dual-coordinate isometry projection operator on the Riemann product manifold. After splicing the technical coordinate triplet and the human coordinate triplet, project them together onto a unified spatiotemporal embedding space to generate a spatiotemporal coordinate fingerprint. Train the operator through isometry mapping constraints to keep the Riemann geodesic distance in the input space approximately unchanged after projection.

7. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 6, characterized in that, Step S6 is as follows: Step S601, for knowledge units Corresponding fusion knowledge intelligent agent Define the activation channels for technical concepts and the activation channels for value semantics; At each teaching moment t, the current field context vector serves as the competitive condition, driving the dynamic struggle for dominance between the two types of activation channels; The field context vector includes course stage labels, learner u's emotional resonance intensity vector, and a summary of historical interaction trajectories; Based on spatiotemporal coordinate fingerprints and field context vectors, the competitive scores of the concept activation channels and the value semantic activation channels are calculated by the field-aware selective routing operator, and the dynamic activation weights of the two types of activation channels are obtained by Softmax normalization. The spatiotemporal coordinate fingerprints are input into the technical concept activation channel and the value semantic activation channel, respectively. The output results are then fused using dynamic activation weights to obtain a fused knowledge intelligent agent. The activation semantic representation at time t; Step S602, Integrating Knowledge Intelligent Agent Its current active semantic representation is used as the activation message to perform local broadcasting on the knowledge spectrum, thus fusing knowledge agents. The neighborhood agent receives the activation message, uses it as context input to calculate attention weights, obtains the aggregated neighborhood activation signal based on the graph attention mechanism, integrates it into the field context vector, and re-executes the dual-channel competitive allocation in step S601 with the updated field context vector. Step S603: Obtain the knowledge unit completed by the learner. The positive feedback signal of the teaching effect after learning drives the continuous self-updating of the parameters of the field-aware selective routing operator.

8. The method for generating a course knowledge neural evolution teaching path driven by emotional resonance according to claim 7, characterized in that, Step S7 is as follows: Step S701: Construct a three-dimensional adaptive landscape dynamic generator, and synthesize a dynamic three-dimensional adaptive terrain composed of a real-time technical logic coherence surface, a value penetration density surface, and a learner emotional resonance prediction surface. The construction of the technical logic coherence surface is constrained by a directed acyclic dependency graph, and the computational technical logic coherence score is obtained by using spatiotemporal coordinate fingerprints to evaluate the adjacent knowledge units in the path. Value Penetration Density Curve The construction is based on the semantic manifold coordinates of value propositions, and the density of the value proposition space covered by the path is obtained as follows: in For path The total number of knowledge units included. This is the location index of the knowledge unit in the path, with a value range from 1 to L; The kernel density estimate, taking the semantic manifold coordinates of value propositions as input, measures the knowledge unit. Local coverage density on the value concept manifold; The bandwidth parameter for kernel density estimation; For knowledge units The curvature weight of emotional resonance intensity; Learner Emotional Resonance Prediction Surface The system is constructed using learner profile vectors as input and employs an emotional resonance prediction network to predict learner u's path. The overall intensity of emotional resonance is predicted. The three-curved surface constitutes a dynamic three-dimensional adaptive terrain in the form of a Riemannian product manifold, which affects the path. The overall fitness is calculated as follows: To target candidate teaching paths With learners The three-dimensional comprehensive fitness vector; Perform a multi-objective Pareto front search in a three-dimensional fitness space; Step S702: Define a path generation network family, where each path generation network takes learner profile vectors and activation semantic representation sets as input and outputs personalized teaching paths. ; The topology mutation operator performs local topology perturbation on the network generated by the current path based on the spatiotemporal coordinate fingerprint set and the directed acyclic dependency graph; Topological recombination operators perform cross-individual gene fragment recombination in high-fitness network architectures near the Pareto front of the current population; The population evolution iteratively performs topological mutation and recombination, and drives the population to continuously converge using the gradient gravity of the three-dimensional fitness terrain: in A non-dominated sorting genetic algorithm II framework with elitist strategy is used to perform Pareto non-dominated sorting and crowding distance calculation; Generate a network population for the parent path of generation t. The family of descendant path generation networks generated by topological mutation and topological recombination operations in generation t. This is the comprehensive fitness evaluation function; Step S703: Extract the Pareto optimal path population from the evolved optimal path generation network population to generate personalized Pareto optimal teaching paths. .

9. The method for generating a course knowledge neuroevolution teaching path driven by emotional resonance according to claim 8, characterized in that, Step S8 is as follows: Step S801: During the path execution process, the real-time emotional feedback generated by learner u at teaching time t is a discrete signal sequence. The discrete emotional feedback signal is kernel interpolated in the path execution space to generate a continuous emotional state field distribution. The spatial gradient vector of the emotional state field is calculated and injected into the fitness function parameters of the path generation network in real time to drive the current execution path to perform continuous online drift along the emotional landscape gradient direction. Step S802: Calculate the projection distance of the current execution path relative to the Pareto front in the three-dimensional fitness space. When the projection distance is greater than the adaptive reshaping threshold, trigger the local path topology self-repair mechanism. Using the current curvature distribution of the emotional state field as a geometric constraint, perform heuristic topology repair search on the local subgraph of the knowledge system to generate a local path replacement fragment that precisely matches the curvature of the current emotional landscape. Embed the local path replacement fragment into the corresponding position of the current execution path to complete the local self-repair of the path topology. Step S803: Design a long-range knowledge distillation and accumulation mechanism for cross-learner group emotional state fields to drive the underlying optimization goal of path generation strategy to continuously calibrate towards real teaching laws. The emotional state fields of the entire learner group are weighted and superimposed, and the statistical regularity of the group's emotional state field is continuously distilled into a long-term correction term of the evolutionary population fitness function. The long-term correction term is integrated into the comprehensive fitness function, and the original comprehensive fitness function is replaced by the corrected comprehensive fitness function. The underlying optimization objective of the path generation strategy is updated once after each teaching cycle.

Images