A teaching quality evaluation method and system based on a large model

Abstract

This invention proposes a teaching quality evaluation method and system based on a large model. The method includes: acquiring teacher explanation text, teaching behavior logs, and student feedback data to construct a triplet structured representation; constructing a bi-class node directed graph based on the triplet structured representation; performing knowledge graph state evolution based on the knowledge point subset to generate knowledge graph substructures at consecutive time steps, calculating the structural change values ​​of knowledge graph substructures at adjacent time steps, combining all structural change values ​​with a high-risk path set for regularization constraint processing, and generating a teaching behavior causal contribution matrix, wherein each row of the teaching behavior causal contribution matrix represents the structural influence of a teaching behavior on multiple knowledge points; inputting the teaching behavior causal contribution matrix, combined with the teaching behavior set and the knowledge point set, into a large language model to generate natural language suggestion content, and generating teaching suggestions based on the natural language suggestion content.

Description

Technical Field

[0001] This invention belongs to the field of large-scale models, and in particular relates to a teaching quality evaluation method and system based on large-scale models. Background Technology

[0002] Currently, teaching quality evaluation, as a crucial feedback and improvement link in the education system, has long relied on manual assessment, standardized testing, and teaching observation. While these methods have played a supervisory and feedback role to some extent, they generally suffer from strong subjectivity, coarse granularity, and poor dynamic response capabilities. Especially today, with the increasing prevalence of digital education, online classrooms, and personalized learning, traditional evaluation methods are insufficient to meet the demands for precise, real-time, and personalized analysis of the teaching process and effectiveness in new educational scenarios. In recent years, artificial intelligence technology has begun to be applied to educational assessment, such as data mining-based behavioral analysis systems and the use of simple machine learning models for performance prediction. Although some progress has been made in certain scenarios, these methods generally rely on static features and lack the ability to model the deep-seated mechanisms between teaching behavior and student responses. Therefore, they cannot accurately reveal the logical connections between teaching intentions, teaching strategies, and student learning outcomes behind teaching behaviors.

[0003] Furthermore, most existing systems assess teaching quality in an "outcome-oriented" manner, focusing on end-point data such as student test scores and homework completion, while neglecting the in-depth mining of "process information"—that is, the dynamic impact of factors such as interaction methods, knowledge presentation sequence, and classroom feedback on learning outcomes. This makes it difficult for teachers to flexibly adjust teaching strategies based on students' real-time feedback and to obtain refined, structured optimization suggestions. More importantly, these evaluation mechanisms rely almost entirely on experience-based judgment or data correlation analysis, lacking causal reasoning capabilities and unable to explain "why a certain teaching method is ineffective for a certain type of student" or "how to precisely intervene in teaching behavior to improve student performance." With the development of Large Language Model (LLM) technology, although some research has explored its applications in areas such as automatic grading and intelligent question answering, its potential in systematic modeling of teaching quality, causal judgment, and personalized suggestion generation has not yet been effectively realized.

[0004] Therefore, how to build a teaching quality evaluation system based on a large model that can understand teaching intentions, model knowledge evolution paths, and reveal the mechanistic relationship between teaching behavior and learning outcomes has become an important technical challenge in the current development of intelligent education. Summary of the Invention

[0005] The purpose of this invention is to propose a teaching quality evaluation method and system based on a large model. This invention further integrates knowledge structure expression and teaching context understanding, enabling the evaluation system to flexibly adapt to different disciplines, teaching styles and student groups, greatly improving the system's intelligence level and application value.

[0006] To achieve the above objectives, a teaching quality evaluation method based on a large model is provided in a first aspect of the present invention, the method comprising the following steps:

[0007] The teacher's explanation text, teaching behavior logs, and student feedback data are obtained to construct a triplet structured representation, which includes teaching behavior, knowledge point set, and student feedback.

[0008] A bi-class directed graph is constructed based on the triplet structured representation. The nodes of the bi-class directed graph include a set of teaching behavior nodes and a set of student feedback state nodes, as well as directed edges, representing the student states that may result from teaching behaviors. The bi-class directed graph is obtained, and the causal effect is calculated using knowledge point constraints as context filtering conditions. If the conditional causal effect of the causal path is positive and significant, and the corresponding student state belongs to the negative feedback label set, it is marked as a high-risk path, and a set of high-risk paths is output.

[0009] Based on the teaching behavior, a subset of knowledge points corresponding to the influence of the teaching behavior is determined. Based on the subset of knowledge points, the knowledge graph state evolution is performed to generate knowledge graph substructures at consecutive time steps. The structural change values ​​of the knowledge graph substructures at adjacent time steps are calculated to quantify the influence of the teaching behavior on the structural state of the knowledge points. All structural change values ​​are combined with a set of high-risk paths for regularization constraint processing to generate a causal contribution matrix of teaching behavior. Each row of the causal contribution matrix of teaching behavior represents the structural influence of a teaching behavior on multiple knowledge points.

[0010] The causal contribution matrix of the teaching behaviors, combined with the set of teaching behaviors and the set of knowledge points, is input into a large language model to generate natural language suggestion content, and teaching suggestions are generated based on the natural language suggestion content.

[0011] Furthermore, the acquisition of teacher explanation texts, teaching behavior logs, and student feedback data to construct a triplet structured representation specifically includes:

[0012] Using a large model, embedding representations are extracted from the teacher's explanation text, teaching behavior logs, and student feedback data to generate teacher semantic embeddings, teaching behavior embeddings, and student feedback embeddings.

[0013] Structural regularization terms are introduced to process the teacher semantic embedding, teaching behavior embedding, and student feedback embedding; the structural regularization terms are used to ensure that the teacher's explanation, teaching behavior, and feedback maintain semantic consistency and to encourage the semantic direction of teaching behavior and student feedback to remain different, so as to avoid homogenization.

[0014] The set of teaching behavior labels mapped from the teaching behavior embedding is used as the teaching behavior; the set of knowledge points identified by comparing the teacher semantic embedding with the standard knowledge base is used as the knowledge point set; and the student mastery status labels inferred from the student feedback embedding are used as student feedback.

[0015] The teaching behaviors, knowledge point sets, and student feedback are represented as a triplet structured representation.

[0016] Furthermore, the calculation of causal effects using knowledge point constraints as contextual filtering conditions specifically includes:

[0017] Obtain the set of knowledge points involved in the current teaching, which is used to mark the contextual consistency between behavior and feedback;

[0018] Based on the set of knowledge points involved in the current teaching, determine the first conditional expectation of the student's state when implementing behavioral teaching behaviors;

[0019] Based on the set of knowledge points involved in the current teaching, the second conditional expectation of the student's state when no behavioral teaching behavior is implemented;

[0020] The difference between the first conditional expectation and the second conditional expectation is taken as the intervention effect under the contextual condition and thus as the causal effect.

[0021] Furthermore, the step of determining the subset of knowledge points corresponding to the influence of the teaching behavior is obtained through semantic matching of behavior and knowledge points.

[0022] The subset of knowledge points is obtained by calculating the similarity of the embedding vectors of the large model, that is, matching the behavioral semantics with the knowledge point descriptions and selecting knowledge points with similar semantics.

[0023] Furthermore, the step of performing knowledge graph state evolution based on the subset of knowledge points to generate knowledge graph substructures at consecutive time steps specifically includes:

[0024] The set of knowledge points involved in the current teaching content is called the knowledge point node set; where each knowledge point node represents a specific knowledge point.

[0025] The mastery score of the knowledge point node is determined by weighted summation of the scores of questions corresponding to the knowledge point node in the formal exam, questions involving the knowledge point in the homework, and questions related to the knowledge point in the quiz.

[0026] Based on the knowledge point dependencies in the standard teaching syllabus, add directed edges between knowledge point nodes;

[0027] Based on the set of knowledge point nodes and the set of directed edges, a continuous periodic knowledge graph substructure is constructed.

[0028] Furthermore, according to the knowledge point dependencies in the standard teaching syllabus, directed edges are added between knowledge point nodes. If the first knowledge point node is a prerequisite for the second knowledge point node, then a directed edge is added between the first and second knowledge point nodes in the graph, indicating that mastering the first knowledge point node is a prerequisite for mastering the second knowledge point node.

[0029] Furthermore, the structural change value of the knowledge graph substructure at adjacent time points is calculated as follows:

[0030]

[0031] Where, ΔS ij Indicating teaching behavior For knowledge node k j The change in mastery, i.e., the structural change value; This represents the attention weight, indicating the impact of the teaching action on knowledge node k. j The semantic relevance, with values ​​ranging from [0,1]; v t (k j () represents the student's understanding of knowledge point node k at time t. j The degree of mastery; v t-1 (k j ) represents the student's understanding of knowledge point node k at a previous time. j The degree of mastery;

[0032] If all structural change values ​​are combined with the high-risk path set and subjected to regularization constraints, a causal contribution matrix of teaching behavior is generated, which specifically includes:

[0033] If the current path is a high-risk path, the structural change value is adjusted by regularization coefficients to adjust the causal contribution matrix of the teaching behavior.

[0034] If it is not a high-risk path, then the structural change value is the causal contribution matrix of the teaching behavior.

[0035] Furthermore, the step of inputting the causal contribution matrix of the teaching behaviors, combined with the set of teaching behaviors and the set of knowledge points, into a large language model to generate natural language suggestion content, and then generating teaching suggestions based on the natural language suggestion content, specifically includes:

[0036] Based on the causal contribution matrix of the teaching behaviors, teaching behaviors with significant impact and optimization value are identified, and a feedback candidate set is constructed.

[0037] For each teaching action in the feedback candidate set, construct a structural semantic joint input vector: teaching action, sub-vector of causal influence of behavior on knowledge point, set of associated important knowledge points, and template prompt words;

[0038] Based on the structural semantic joint input vector, a large model is used to generate behavior optimization suggestions;

[0039] Based on the behavior optimization suggestions, standard suggestion items are generated, including: teaching behavior, suggestion content, knowledge point set, priority score of feedback candidate set, and execution status field;

[0040] All suggested items are compiled to form the feedback set for the current round.

[0041] Furthermore, the step of identifying teaching behaviors with significant impact and optimization value based on the causal contribution matrix of the teaching behaviors and constructing a feedback candidate set specifically involves:

[0042]

[0043] in, For priority scoring, δ is the negative contribution threshold set by the system, Ω(k) j ) is knowledge point k j Importance rating It is an indicator function;

[0044] like γ is the set minimum feedback trigger value, which will then trigger the teaching action. These are considered high-priority feedback targets; ultimately forming a set of behaviors to be suggested.

[0045] In a second aspect, the present invention provides a teaching quality evaluation system based on a large model, the system comprising:

[0046] The structured unit is used to acquire teacher explanation texts, teaching behavior logs, and student feedback data to construct a triplet structured representation, which includes teaching behavior, knowledge point set, and student feedback.

[0047] A teaching risk construction set is used to construct a bi-class node directed graph based on the triple structured representation. The nodes of the bi-class node directed graph include a set of teaching behavior nodes and a set of student feedback state nodes, as well as directed edges, representing the student states that may result from teaching behaviors. The bi-class node directed graph is obtained, and the causal effect is calculated using knowledge point constraints as context filtering conditions. If the conditional causal effect of the causal path is positive and significant, and the corresponding student state belongs to the negative feedback label set, it is marked as a high-risk path, and a high-risk path set is output.

[0048] A risk evolution set is used to determine the subset of knowledge points affected by the teaching behavior, perform knowledge graph state evolution based on the subset of knowledge points, generate knowledge graph substructures at consecutive time steps, and calculate the structural change values ​​of knowledge graph substructures at adjacent time steps to quantify the impact of teaching behavior on the structural state of knowledge points. All structural change values ​​are combined with a high-risk path set for regularization constraint processing to generate a causal contribution matrix of teaching behavior, wherein each row of the causal contribution matrix of teaching behavior represents the structural impact of a teaching behavior on multiple knowledge points.

[0049] A set of teaching suggestions is generated by inputting the causal contribution matrix of the teaching behaviors, the set of teaching behaviors, and the set of knowledge points into a large language model to generate natural language suggestion content, and then using the natural language suggestion content to generate teaching suggestions.

[0050] The beneficial technical effects of the present invention are at least as follows:

[0051] To address the numerous problems existing in the current technologies, this invention provides a teaching quality evaluation method and system based on a large-scale model, focusing on solving the challenges of "missing processes, unclear causality, unreal-time feedback, and non-optimizable strategies" in current teaching evaluation. The core of this invention lies in introducing a large-scale pre-trained model with semantic understanding and generative capabilities as the cognitive core, combining teaching behavior data and student feedback information to construct a structured and dynamically evolving teaching state representation framework. The system can automatically extract key behaviors and cognitive features from the actual teaching process, depict the dynamic changes in students' knowledge acquisition, and establish a mechanistic link between teaching behaviors and learning outcomes. Furthermore, this invention achieves in-depth evaluation of the effectiveness of teaching behaviors by modeling structured causal relationships in the teaching process, and provides teachers with personalized and actionable teaching optimization suggestions by combining model generative capabilities, thereby realizing a complete closed loop of teaching quality evaluation from "observing results" to "understanding the process" and then to "generating suggestions."

[0052] Compared to existing outcome-oriented and relevance-driven evaluation systems, this invention emphasizes three key capabilities: process perception, structural modeling, and mechanism reasoning. This allows teaching evaluation to move beyond simply summarizing static indicators, enabling it to continuously perceive teaching dynamics, predict student responses, explain teaching effectiveness, and proactively provide improvement strategies. Building upon the powerful language modeling capabilities of a large-scale model, this invention further integrates knowledge structure expression and understanding of the teaching context, giving the evaluation system the ability to flexibly adapt to different disciplines, teaching styles, and student groups, greatly enhancing the system's intelligence and its application value. Attached Figure Description

[0053] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0054] Figure 1 This is a flowchart of a teaching quality evaluation method based on a large model according to the present invention. Detailed Implementation

[0055] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0056] like Figure 1 As shown in the figure, an embodiment of the present invention provides a teaching quality evaluation method based on a large model, the method comprising:

[0057] S1. Obtain teacher explanation text, teaching behavior logs, and student feedback data, and construct a triplet structured representation, wherein the triplet structured representation includes teaching behavior, knowledge point set, and student feedback.

[0058] Specifically, this step proposes a "Triple-Mapping Semantic Structure (TMSS)" model to map the original teaching process data into a structured representation S. t ={A t ,K t B t} serves as the semantic starting point input for the entire patent system.

[0059] Input data description

[0060] Teacher explains text T t Course content delivered in natural language;

[0061] Teaching Behavior Log L t Structured behavioral data, such as speech rate, interaction frequency, and multimedia usage;

[0062] Student feedback data R t Student answers, classroom responses, and incorrect answers, etc.

[0063] Furthermore, using a large model f θ For T t ,L t ,R t Extract the embedded representation and generate:

[0064] Teacher semantic embedding e T =f θ (T t );

[0065] Embedding teaching behaviors L =f θ (L t );

[0066] Student feedback embedded in e R =f θ (R t ).

[0067] The fusion representation and the introduction of structural regularization terms (Formula (1))

[0068]

[0069] Among them, E t : The final fused structural semantic representation; W1, W2, W3: Linear weights for each type of embedding, reflecting their contribution; λ: Regularization strength control parameter; Structural regularization terms constrain the consistency and decoupling of the three types of embeddings in the semantic space.

[0070] The structure regularization term is defined as follows:

[0071]

[0072] Where α: content alignment coefficient (e.g., 1.0); β: role decoupling coefficient (e.g., 0.7);

[0073] Part 1: Maintain semantic consistency between teacher explanations, teaching behaviors, and feedback;

[0074] Part Two: Encourage semantic differences between teaching behaviors and student feedback to avoid homogenization.

[0075] Furthermore, the structured output variables are constructed as follows:

[0076] Teaching Behavior A t From e L The set of teaching behavior labels mapped (such as "low interaction", "fast pace", etc.);

[0077] Knowledge point set K t : via e T The set of knowledge points identified by comparison with a standard knowledge base;

[0078] Student Feedback B t From e R The inferred student mastery status labels (such as "high error rate" and "fluctuating attention");

[0079] Final output structure

[0080] S t ={A t ,K t B t} (3).

[0081] S2. Construct a bi-class node directed graph based on the triplet structured representation; the nodes of the bi-class node directed graph include a set of teaching behavior nodes and a set of student feedback state nodes, as well as directed edges, representing the student states that may result from teaching behaviors, thus obtaining a bi-class node directed graph, and using knowledge point constraints as context filtering conditions to calculate causal effects. If the conditional causal effect of the causal path is positive and significant, and the corresponding student state belongs to the negative feedback label set, then it is marked as a high-risk path, and the high-risk path set is output.

[0082] Specifically, this step aims to construct a causal behavior graph G. C Used for modeling structured teaching behaviors A t Student feedback status B t The causal relationship between them was investigated, and causal pathways P that may have negative impacts were identified. i This is a key step in this patented solution that connects "structured input" with "mechanism understanding," providing causal evidence for subsequent feedback on teaching strategies.

[0083] Input instructions: A t B: The set of teaching behaviors adopted by the teacher at time t (e.g., "fast speaking speed", "less blackboard writing", "low interaction"); t The set of feedback states exhibited by students after this teaching behavior (e.g., "high error rate", "attention drift", "knowledge point k") i Not yet mastered); K t The set of knowledge points involved in the current teaching is used to mark the contextual consistency between behavior and feedback.

[0084] Furthermore, construct a bi-class directed graph G.C =(V A ∪V B ,E), where: V A It is a set of teaching behavior nodes; V B It is the set of student feedback state nodes; E is a directed edge, representing "a certain teaching behavior may cause a certain student feedback state"; edge e ij ∈E represents teaching behavior This may lead to a student status Weights represent the strength of causal effects.

[0085] Specifically, this step proposes a context-aligned interventional estimator (CAIE) to extract effective causal relationships from teaching sequences; this invention introduces knowledge point constraints as context filtering conditions, applying them only to K... t Causal effects are calculated when there is consistency or high overlap to avoid misleading relevance across knowledge content. The intervention effect is calculated under contextual conditions as follows:

[0086]

[0087] Where, τ ij Teaching behavior Student status Conditional causal effect; In the context of knowledge point K t The following actions were carried out. hour, Expectations; The expected value when the behavior is not implemented in the same context.

[0088] Furthermore, define the high-risk causal path P. i Determination mechanism:

[0089] If a certain edge e ij causal effect τ ij It is positive and significant, and This belongs to the system-defined "negative feedback tag set". This approach is therefore considered to have potential risks.

[0090] The system combines all such edges into a high-risk path set P = {P i}, each path P i Shaped like The criteria for screening high-risk causal pathways are as follows:

[0091]

[0092] Among them, P i: From teaching behavior Negative state of students caused Limited to the knowledge point context K t In the middle; δ: the lower limit of the risk causal effect (e.g., δ = 0.2); The collection of negative feedback includes tags such as "decreased mastery", "increased error rate", and "decreased participation".

[0093] Final output: Causal behavior graph structure G C (Visualized teaching impact structure); High-risk path set P = {P i (For use in generating suggestions for subsequent large models).

[0094] S3. Based on the teaching behavior, determine the knowledge point subset corresponding to the influence of the teaching behavior. Based on the knowledge point subset, perform knowledge graph state evolution to generate knowledge graph substructures at consecutive time steps. Calculate the structural change values ​​of the knowledge graph substructures at adjacent time steps to quantify the influence of the teaching behavior on the structural state of the knowledge points. Combine all structural change values ​​with a high-risk path set for regularization constraint processing to generate a causal contribution matrix of the teaching behavior. Each row of the causal contribution matrix of the teaching behavior represents the structural influence of a teaching behavior on multiple knowledge points.

[0095] Specifically, this step, "Causal Modeling of Teaching Behavior on Knowledge Evolution," is the central structural step in this patent's technical roadmap, located after "Modeling the Influence of Teaching Behavior (Step 2)" and before "Generating Feedback Suggestions (Step 4)." Its function is to establish teaching behavior A. t With the evolution of knowledge structure graph G t This involves establishing a mechanistic mapping relationship to clarify how specific behaviors lead to positive or negative evolution of students' knowledge structures. The modeling results of this mechanism directly determine the "content orientation" and "basis for behavioral adjustment" of subsequent feedback suggestions. In real teaching scenarios, whether students truly "learn" a knowledge point is not determined by a single right or wrong answer, but rather by the state of that point within their knowledge network, its connections to other points, and the stability of their mastery. Therefore, this step proposes an intervention explanation mechanism based on changes in knowledge graph structure to quantify "teaching behaviors." The evolution of the state of the knowledge graph G t "Mechanism of action".

[0096] The final output of this invention is not a simple score, but a graph-structured causal weight matrix W. K It provides a score of the causal contribution of teaching behaviors to the evolution of knowledge structure, becoming the "data-driven explanation basis" for the generation of subsequent large model strategies.

[0097] Furthermore, for each teaching behavior This invention first identifies the subset of knowledge points that have a major impact through semantic matching of behavior and knowledge points.

[0098] This subset can be obtained through a large model f. θ The embedding vector similarity calculation implementation matches behavioral semantics with knowledge point descriptions, selecting semantically similar knowledge points to form a sequence.

[0099] This step does not involve embedding modeling; it only reuses the large model f defined in step one. θ And a knowledge point database.

[0100] Furthermore, construct knowledge graph substructures. and

[0101] In map G t and G t-1 Extract the corresponding from the middle. Subgraph structure;

[0102] This subgraph contains knowledge point node k. j And its edges (representing dependencies or sequence relationships), node values ​​representing mastery (e.g., v). t (k j ) represents the current state of knowledge point k. j (Level of mastery).

[0103] Among them, G t and G t-1 It is dynamically generated by the system at the end of each teaching cycle based on students' standardized learning performance data. The specific construction method is as follows:

[0104] Node construction: Based on the set K of knowledge points involved in the current teaching content t For a set of nodes; each node k j ∈K t It represents a specific knowledge point.

[0105] Node attribute assignment (mastery calculation): For each knowledge point k j The system calculates the mastery score v for this knowledge point based on the following indicators. t (k j ): ExamScore(k j HomeworkScore(k): The score for the corresponding questions on this knowledge point in formal exams such as midterms and finals; j ): Scores for questions related to this knowledge point in regular assignments; QuizScore(k) j ): The score for questions related to this knowledge point in a quiz.

[0106] The mastery level is calculated using the following weighted average formula:

[0107] v t (k j ) = 0.5 × ExamScore(k j )+0.3×HomeworkScore(k j )+0.2×QuizScore(k j (6)

[0108] Among them: all scores are normalized to the [0,1] interval; the weights (0.5,0.3,0.2) can be flexibly adjusted according to the actual teaching arrangements.

[0109] Edge relation construction: Based on the knowledge point dependencies in the standard teaching syllabus, add directed edges between knowledge points; if knowledge point k a It is knowledge point k b The prerequisite is to add directed edges (k) to the graph. a →k b ), indicating mastery of k a It is about mastering k b The premise.

[0110] Historical graph snapshot: After each teaching cycle, the system will construct the completed knowledge state graph G. t Save as a snapshot as input G for the next cycle. t-1 This ensures that subsequent analysis of the differences in how teaching behaviors affect knowledge evolution can be conducted.

[0111] Furthermore, the structural change value ΔS is calculated. ij :

[0112] Next, the present invention will be compared. and The structural differences and the impact of quantitative teaching behaviors on the structural state of knowledge points;

[0113] Differences include changes in mastery and connectivity of the graph structure. The following formula is defined for calculation:

[0114]

[0115] Where, ΔS ij : Indicates teaching behavior For knowledge node k j The degree of mastery changes; Attention weight, representing the influence of the teaching behavior on knowledge node k. j The semantic relevance, with a value range of [0,1], is given by semantic matching; v t (k j ): At time t, the student's understanding of knowledge point node kj The degree of mastery; v t-1 (k j ): Some time ago, students were working on knowledge point node k. j The degree of mastery.

[0116] Understandably, this formula is essentially a weighted model of "knowledge changes caused by behavior", focusing only on the sub-knowledge points that are actually affected by the behavior.

[0117] Furthermore, after obtaining all ΔS ij Subsequently, the present invention further incorporates the high-risk path P provided in the previous step;

[0118] If behavior Located in a high-risk path (i.e. leading to) ), and knowledge point nodes Therefore, it is considered that a structural penalty factor needs to be applied to this effect to prevent subsequent systems from misjudging this behavior as a "positive contribution"; the regularization process is as follows:

[0119]

[0120] Among them, W K (i,j): The final teaching behavior For knowledge node k j The causal influence score; λ: regularity coefficient (recommended to be set to 0.2 to 0.3 times the change in mastery unit); The indicator function takes a value of 1 only if the triplet path is a risky path;

[0121] This mechanism embodies the inventive design incorporated in this step of the invention: combining behavioral-structural modeling with causal risk path regularization to make behavioral modeling more causal consistent and robust.

[0122] Ultimately, this invention outputs a behavior-knowledge point causal contribution matrix. Each line represents the structural impact of a teaching behavior on multiple knowledge points.

[0123] S4. The causal contribution matrix of the teaching behavior, combined with the set of teaching behaviors and the set of knowledge points, is input into the large language model to generate natural language suggestion content, and teaching suggestions are generated based on the natural language suggestion content.

[0124] Specifically, this step aims to utilize the causal contribution matrix W of teaching behaviors output in step three. K Based on the behavioral label A in this round of teaching t and the set of knowledge points K t Generate structured teaching suggestions This information is provided for teachers or the system to adopt, thereby completing the "assessment-reasoning-feedback" closed loop of the teaching quality evaluation system.

[0125] Furthermore, this step receives three types of structured input data from the previous stage: a set of teaching behaviors A t This represents the teaching behavior performed by the teacher in the current time period; the set of knowledge points K t This indicates all the knowledge points covered in this round of teaching;

[0126] Causal Influence Matrix of Teaching Behavior on Knowledge Points Representing each teaching behavior For knowledge point k j The intensity of positive and negative influences;

[0127] Optional input: Student's current knowledge graph state G t It is used to provide background explanation support in suggestion generation.

[0128] Furthermore, the specific steps are as follows:

[0129] Identify the target behavior that needs feedback: The system based on W K Identify teaching behaviors that have significant impact and optimization value, and construct a set of feedback candidates.

[0130] For each teaching behavior Calculate the cumulative value of its negative causal contribution:

[0131]

[0132] Where δ is the negative contribution threshold set by the system, Ω(k) j ) is knowledge point k j Importance rating (e.g., position in the course structure or difficulty level), It is an indicator function; if (γ is the set minimum feedback trigger value), then... These are considered high-priority feedback targets; ultimately forming a set of behaviors to be suggested.

[0133] Preparing the structure-semantic fusion input: for each The system constructs a structural semantic joint input vector: containing behavioral labels. The causal influence subvector W of behavior on knowledge points K (i,:); the set of important knowledge points associated with it. The system template prompt word is "Prompt," such as "Please suggest specific strategies to improve this teaching behavior in order to avoid negative impacts on key knowledge points."

[0134] Generating suggested text: Utilizing a large model f θ Generate behavior optimization suggestions:

[0135]

[0136] Among them, f θ The large language model defined in this patent is used to generate natural language suggestions; suggestion results. It is a well-structured and goal-oriented set of teaching improvement suggestions, such as "When explaining 'monotonicity of functions,' it is recommended to add more blackboard writing and real-life examples."

[0137] Establish a suggested structure representation: The structured entries are standardized for ease of display, execution, and tracking, and include the following fields: Original teaching behavior: Suggested content: The set of knowledge points involved: Priority rating: Normalization result; Execution status field: initially empty, to be filled in by the teacher or system later to determine whether it has been adopted.

[0138] All suggested items are summarized to form the feedback set for this round:

[0139]

[0140] Finally, there is a feedback and adoption mechanism: teachers receive feedback through the system interface. The suggestions in the document allow you to choose whether to adopt them; if adopted, the resulting adjustments to teaching behavior will be used as A. t+1 The components are input into the next round of teaching behavior flow; if not adopted, the system will record the feedback status and dynamically adjust the weight and priority threshold in the next suggestion generation;

[0141] This adoption mechanism forms a closed-loop teaching quality optimization process of "assessment → reasoning → feedback → implementation → reassessment".

[0142] This invention also provides a teaching quality evaluation system based on a large model, the system comprising:

[0143] The user behavior credibility modeling module is used to acquire user-side data, environmental context, and pile status information, combine them into a feature vector of the current pile location, input the feature vector into a neural network, and output the probability of the user actually appearing and starting to charge within the current scheduling cycle of the corresponding pile location and the expected real power load value, and determine the corresponding actual expected power load. The actual expected power load is the weighted product of the probability of the user actually appearing and starting to charge within the current scheduling cycle of the corresponding pile location and the expected real power load value.

[0144] The spatial thermal potential map construction module is used to embed the actual expected power load into the map structure for thermal potential diffusion analysis to obtain the thermal potential value of the corresponding node.

[0145] The resource scheduling optimization module is used to design a resource scheduling optimization function based on thermal potential diffusion analysis. The resource scheduling optimization function determines the actual power allocation value that can be obtained for each pile position according to the thermal potential value and constraints. The constraints include that the total power scheduling of the entire station must not exceed the available value of the system, the allocated power of each pile position cannot exceed its physical access limit, and the system will not allocate power to pile positions with extremely low confidence, so as to avoid resource waste caused by behavioral prediction errors.

[0146] The control command execution and feedback module is used to send the actual available power allocation value of each charging pile to the actual charging pile control equipment, drive it to execute power output behavior according to the system strategy within the current scheduling cycle, and monitor the actual control execution status of each charging pile in real time, and make dynamic adjustments when necessary.

[0147] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0148] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0149] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0150] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A teaching quality evaluation method based on a large model, characterized in that, The method includes the following steps: The teacher's explanation text, teaching behavior logs, and student feedback data are obtained to construct a triplet structured representation, which includes teaching behavior, knowledge point set, and student feedback. A bi-class directed graph is constructed based on the triplet structured representation. The nodes of the bi-class directed graph include a set of teaching behavior nodes and a set of student feedback state nodes, as well as directed edges, representing the student states that may result from teaching behaviors. The bi-class directed graph is obtained, and the causal effect is calculated using knowledge point constraints as context filtering conditions. If the conditional causal effect of the causal path is positive and significant, and the corresponding student state belongs to the negative feedback label set, it is marked as a high-risk path, and the high-risk path set is output. Based on the teaching behavior, a subset of knowledge points corresponding to the influence of the teaching behavior is determined. Based on the subset of knowledge points, the knowledge graph state evolution is performed to generate knowledge graph substructures at consecutive time steps. The structural change values ​​of the knowledge graph substructures at adjacent time steps are calculated to quantify the influence of the teaching behavior on the structural state of the knowledge points. All structural change values ​​are combined with a set of high-risk paths for regularization constraint processing to generate a causal contribution matrix of teaching behavior. Each row of the causal contribution matrix of teaching behavior represents the structural influence of a teaching behavior on multiple knowledge points. The causal contribution matrix of the teaching behavior, combined with the set of teaching behaviors and the set of knowledge points, is input into the large language model to generate natural language suggestion content, and teaching suggestions are generated based on the natural language suggestion content. The calculation of the structural change value of the knowledge graph substructure at adjacent time points is as follows: ； in, Indicating teaching behavior Knowledge Points The change in mastery, i.e., the structural change value; This represents the attention weight, indicating the impact of the teaching behavior on knowledge points. semantic relevance, range of values ; Indicates time Students' understanding of knowledge points The degree of mastery; This indicates that students' understanding of the knowledge points at a certain point in time was... The degree of mastery; If all structural change values ​​are combined with the high-risk path set and subjected to regularization constraints, a causal contribution matrix of teaching behavior is generated, which specifically includes: If the current path is a high-risk path, the structural change value is adjusted by regularization coefficients to adjust the causal contribution matrix of the teaching behavior. If it is not a high-risk path, then the structural change value is the causal contribution matrix of the teaching behavior.

2. The teaching quality evaluation method based on a large model according to claim 1, characterized in that, The acquisition of teacher explanation texts, teaching behavior logs, and student feedback data, and the construction of a triplet structured representation, specifically includes: Using a large model, embedding representations are extracted from the teacher's explanation text, teaching behavior logs, and student feedback data to generate teacher semantic embeddings, teaching behavior embeddings, and student feedback embeddings. Structural regularization terms are introduced to process the teacher semantic embedding, teaching behavior embedding, and student feedback embedding; the structural regularization terms are used to ensure that the teacher's explanation, teaching behavior, and feedback maintain semantic consistency and to encourage the semantic direction of teaching behavior and student feedback to remain different, so as to avoid homogenization. The set of teaching behavior labels mapped from the teaching behavior embedding is used as the teaching behavior; the set of knowledge points identified by comparing the teacher semantic embedding with the standard knowledge base is used as the knowledge point set; and the student mastery status labels inferred from the student feedback embedding are used as student feedback. The teaching behaviors, knowledge point sets, and student feedback are represented as a triplet structured representation.

3. The teaching quality evaluation method based on a large model according to claim 1, characterized in that, The calculation of causal effects using knowledge point constraints as contextual filtering conditions specifically includes: Obtain the set of knowledge points involved in the current teaching, which is used to mark the contextual consistency between behavior and feedback; Based on the set of knowledge points involved in the current teaching, determine the first conditional expectation of the student's state when implementing behavioral teaching behaviors; Based on the set of knowledge points involved in the current teaching, the second conditional expectation of the student's state when no behavioral teaching behavior is implemented; The difference between the first conditional expectation and the second conditional expectation is taken as the intervention effect under the contextual condition and thus as the causal effect.

4. The teaching quality evaluation method based on a large model according to claim 1, characterized in that, The step of determining the subset of knowledge points corresponding to the influence of the teaching behavior is obtained through semantic matching of behavior and knowledge points. The subset of knowledge points is obtained by calculating the similarity of the embedding vectors of the large model, that is, matching the behavioral semantics with the knowledge point descriptions and selecting knowledge points with similar semantics.

5. The teaching quality evaluation method based on a large model according to claim 4, characterized in that, The step of performing knowledge graph state evolution based on the subset of knowledge points to generate knowledge graph substructures at consecutive time steps specifically includes: The set of knowledge points involved in the current teaching content is called the knowledge point node set; where each knowledge point node represents a specific knowledge point. The mastery score of the knowledge point node is determined by weighted summation of the scores of questions corresponding to the knowledge point node in the formal exam, questions involving the knowledge point in the homework, and questions related to the knowledge point in the quiz. Based on the knowledge point dependencies in the standard teaching syllabus, add directed edges between knowledge point nodes; Based on the set of knowledge point nodes and the set of directed edges, a continuous periodic knowledge graph substructure is constructed.

6. The teaching quality evaluation method based on a large model according to claim 5, characterized in that, According to the knowledge point dependencies in the standard teaching syllabus, directed edges are added between knowledge point nodes. If the first knowledge point node is a prerequisite for the second knowledge point node, then a directed edge is added between the first and second knowledge point nodes in the graph, indicating that mastering the first knowledge point node is a prerequisite for mastering the second knowledge point node.

7. The teaching quality evaluation method based on a large model according to claim 1, characterized in that, The step of inputting the causal contribution matrix of the teaching behaviors, combined with the set of teaching behaviors and the set of knowledge points, into a large language model to generate natural language suggestion content, and then generating teaching suggestions based on the natural language suggestion content, specifically includes: Based on the causal contribution matrix of the teaching behaviors, teaching behaviors with significant impact and optimization value are identified, and a feedback candidate set is constructed. For each teaching action in the feedback candidate set, construct a structural semantic joint input vector: teaching action, sub-vector of causal influence of behavior on knowledge point, set of associated important knowledge points, and template prompt words; Based on the structural semantic joint input vector, a large model is used to generate behavior optimization suggestions; Based on the behavior optimization suggestions, standard suggestion items are generated, including: teaching behavior, suggestion content, knowledge point set, priority score of feedback candidate set, and execution status field; All suggested items are compiled to form the feedback set for the current round.

8. The teaching quality evaluation method based on a large model according to claim 7, characterized in that, The step of identifying teaching behaviors with significant impact and optimization value based on the causal contribution matrix of the teaching behaviors and constructing a feedback candidate set specifically involves: ； in, Score based on priority. It is the negative contribution threshold set by the system. It is a knowledge point Importance rating It is an indicator function; For teaching behavior Knowledge Points The causal effect score; like , If it is the set minimum feedback trigger value, then the teaching action will be... These are considered high-priority feedback targets; ultimately forming a set of behaviors to be suggested. .

9. A teaching quality evaluation system based on a large model, characterized in that, The system includes: The structured unit is used to acquire teacher explanation texts, teaching behavior logs, and student feedback data to construct a triplet structured representation, which includes teaching behavior, knowledge point set, and student feedback. A teaching risk construction set is used to construct a bi-class node directed graph based on the triple structured representation. The nodes of the bi-class node directed graph include a set of teaching behavior nodes and a set of student feedback state nodes, as well as directed edges, representing the student states that may result from teaching behaviors. The bi-class node directed graph is obtained, and the causal effect is calculated using knowledge point constraints as context filtering conditions. If the conditional causal effect of the causal path is positive and significant, and the corresponding student state belongs to the negative feedback label set, it is marked as a high-risk path, and a high-risk path set is output. A risk evolution set is used to determine the subset of knowledge points affected by the teaching behavior, perform knowledge graph state evolution based on the subset of knowledge points, generate knowledge graph substructures at consecutive time steps, and calculate the structural change values ​​of knowledge graph substructures at adjacent time steps to quantify the impact of teaching behavior on the structural state of knowledge points. All structural change values ​​are combined with a high-risk path set for regularization constraint processing to generate a causal contribution matrix of teaching behavior, wherein each row of the causal contribution matrix of teaching behavior represents the structural impact of a teaching behavior on multiple knowledge points. A set of teaching suggestions is generated by inputting the causal contribution matrix of the teaching behavior, the set of teaching behaviors, and the set of knowledge points into a large language model to generate natural language suggestion content, and then generating teaching suggestions based on the natural language suggestion content. The calculation of the structural change value of the knowledge graph substructure at adjacent time points is as follows: ； in, Indicating teaching behavior Knowledge Points The change in mastery, i.e., the structural change value; This represents the attention weight, indicating the impact of the teaching behavior on knowledge points. semantic relevance, range of values ; Indicates time Students' understanding of knowledge points The degree of mastery; This indicates that students' understanding of the knowledge points at a certain point in time was... The degree of mastery; If all structural change values ​​are combined with the high-risk path set and subjected to regularization constraints, a causal contribution matrix of teaching behavior is generated, which specifically includes: If the current path is a high-risk path, the structural change value is adjusted by regularization coefficients to adjust the causal contribution matrix of the teaching behavior. If it is not a high-risk path, then the structural change value is the causal contribution matrix of the teaching behavior.

Images