A method and system for generating a large model based on index dependency graph constraints for teaching evaluation report generation

Abstract

The application discloses a kind of based on index dependency graph constraint's big model generation control method and system for teaching evaluation report generation, belong to the field of teaching, including: constructing teaching evaluation index set, and establish the index dependency graph of the logical dependency relationship between index;In the big model inference stage, generate candidate evaluation text and extract the semantic vector of each index;According to index dependency graph, the consistency constraint calculation and conflict detection are carried out to candidate text, obtain consistency loss and conflict penalty;Based on the above loss, candidate text is scored and reordered, and the optimal text is selected;If conflict exceeds the preset threshold, trigger closed-loop rewriting mechanism to regenerate until the final teaching evaluation report that output satisfies structural consistency is obtained.The application can explicitly introduce index structure constraint in the process of teaching evaluation report generation, effectively suppress cross-index semantic conflict, improve the logical consistency, automation level and explainability of evaluation report.

Description

Technical Field

[0001] This invention belongs to the field of education, and in particular relates to a method and system for generating and controlling large models based on index dependency graph constraints for generating teaching evaluation reports. Background Technology

[0002] Currently, smart education technologies are widely used for the digital assessment of classroom teaching quality. Various teaching platforms, through behavior collection, video analysis, and learning data statistics, can obtain multi-dimensional indicators such as teacher instruction, classroom interaction, and student comprehension, and are attempting to automatically generate teaching evaluation reports using Large Language Models (LLMs). These methods can quickly generate teaching comments and improvement suggestions containing multi-indicator descriptions based on input classroom behavior indicators and student feedback data, thus improving the efficiency of teaching evaluation report generation to a certain extent.

[0003] However, existing large-model-based teaching evaluation report generation technologies still have significant shortcomings. First, there are inherent logical dependencies between teaching evaluation indicators (e.g., the clarity of teacher explanation directly affects student comprehension, and the frequency of classroom interaction constrains participation evaluation), but existing methods only input indicators as prompts into the model, lacking explicit modeling of the structural dependencies between indicators, resulting in the model's inability to understand cross-indicator logical constraints. Second, cross-indicator semantic conflicts are prone to occur in the generated text; for example, "low explanation clarity" and "high student comprehension" may appear simultaneously, and existing methods lack effective conflict quantification and detection mechanisms. Third, most systems only perform simple verification after text generation, failing to provide structured control over the generation path during the inference stage, and post-generation correction methods are insufficient to completely resolve multi-indicator contradictions. Finally, when conflicts occur, manual report modification is often required, failing to form an automated closed-loop correction process, thus limiting the intelligence and automation level of the teaching evaluation system. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a large-scale model generation control method based on index dependency graph constraints for generating teaching evaluation reports, comprising: Obtain the set of teaching evaluation indicators; Based on the set of teaching evaluation indicators, construct a teaching evaluation indicator dependency graph that represents the logical dependencies between the indicators; Candidate teaching evaluation texts are generated through a large model, and semantic vectors are obtained by extracting the semantic intensity corresponding to each teaching evaluation indicator from the candidate teaching evaluation texts based on the teaching evaluation indicator dependency graph. Based on the teaching evaluation index dependency graph and the semantic vector, constraint calculation and conflict detection are performed on the candidate teaching evaluation texts to obtain consistency constraint loss and conflict penalty. The candidate teaching evaluation texts are scored based on the consistency constraint loss and the conflict penalty, and then re-ranked according to the scoring results to select the optimal candidate teaching evaluation text. Determine whether the conflict penalty of the optimal candidate teaching evaluation text exceeds a preset threshold. If it does, trigger a closed-loop rewriting mechanism to regenerate the candidate teaching evaluation text. If it does not exceed the threshold, output the optimal candidate teaching evaluation text as the final teaching evaluation report.

[0005] Optionally, the construction of the teaching evaluation indicator dependency graph representing the logical dependencies between indicators specifically includes: Based on the nodes, dependencies between indicators, and edge weights in the set of teaching evaluation indicators, construct the teaching evaluation indicator dependency graph; The dependency edge defines the upstream indicator node, the downstream indicator node, the dependency relationship weight, the dependency constraint type, and the dependency threshold parameter.

[0006] Optionally, the semantic intensity corresponding to each teaching evaluation indicator is extracted from the candidate teaching evaluation text to obtain a semantic vector, which specifically includes: Based on the evaluation index name, relevant paragraphs in the candidate teaching evaluation text are located, and the semantic strength is calculated through rule mapping or classification model. Multiple descriptions are then aggregated and normalized to obtain the semantic vector.

[0007] Optionally, constrained calculation and conflict detection are performed on the candidate teaching evaluation texts to obtain consistency constraint loss and conflict penalty, specifically including: Based on the monotonic constraint relationship in the indicator dependency graph, a monotonic constraint consistency constraint function is calculated to quantify the degree of violation of downstream indicators relative to upstream indicators. Based on the mutual exclusion constraints in the index dependency graph, a mutual exclusion conflict function is calculated to quantify the degree of conflict between two mutually exclusive indices. The conflict penalty is calculated based on the result of the mutual exclusion conflict function.

[0008] Optionally, scoring the candidate teaching evaluation texts and re-ranking them based on the scoring results specifically includes: The total score for each candidate teaching evaluation text is calculated based on the weighted sum of the language model generation cost, the consistency constraint loss, and the conflict penalty. The candidate teaching evaluation text with the lowest total score is selected as the optimal candidate teaching evaluation text.

[0009] Optionally, the triggering of the closed-loop rewriting mechanism to regenerate candidate teaching evaluation texts specifically includes: The large model is triggered to regenerate candidate teaching evaluation texts by one or more strategies, such as increasing the number of candidate texts, increasing the conflict penalty weight, or increasing the structural missing penalty parameter.

[0010] Optionally, obtaining the set of teaching evaluation indicators specifically includes: Based on classroom observation records, classroom behavior analysis results, learning behavior statistics, or teaching evaluation scale information, normalized indicator scores are extracted to construct the teaching evaluation indicator set.

[0011] On the other hand, the present invention also provides a large model generation and control system based on index dependency graph constraints for generating teaching evaluation reports, comprising: The data input module is used to receive classroom teaching evaluation data; The indicator modeling module is used to construct a set of teaching evaluation indicators based on the classroom teaching evaluation data. The dependency graph construction module is used to construct a teaching evaluation index dependency graph representing the logical dependencies between the indicators based on the set of teaching evaluation indicators. The large model interface module is used to call the large model to generate candidate teaching evaluation texts; The semantic alignment module is used to extract the semantic intensity corresponding to each teaching evaluation indicator from the candidate teaching evaluation text based on the teaching evaluation indicator dependency graph, and obtain a semantic vector. The constraint calculation module is used to perform constraint calculation and conflict detection on the candidate teaching evaluation text based on the teaching evaluation index dependency graph and the semantic vector, and obtain consistency constraint loss and conflict penalty. A generation control module is used to score and re-rank the candidate teaching evaluation texts based on the consistency constraint loss and the conflict penalty, so as to select the optimal candidate teaching evaluation text. The closed-loop rewriting module is used to determine whether the conflict penalty of the optimal candidate teaching evaluation text exceeds a preset threshold, and triggers a rewriting mechanism when it does, so as to control the large model interface module to regenerate the candidate teaching evaluation text.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs a dependency graph of teaching evaluation indicators, explicitly models the monotonic, interval, or mutually exclusive constraint relationships between indicators, and introduces consistency constraint loss and conflict penalty functions during the large-scale model inference stage to quantify and re-rank candidate evaluation texts. This directly suppresses cross-indicator semantic conflicts during the generation process, significantly improving the structural consistency of teaching evaluation reports. Simultaneously, this invention sets a conflict threshold and triggers a closed-loop rewriting mechanism. When a conflict exceeds the threshold, the number of candidates or penalty weights are automatically adjusted and the report is regenerated until an evaluation report that satisfies logical constraints is output, avoiding manual intervention. Experiments demonstrate that this invention effectively reduces typical conflicts such as contradictions between sub-item conclusions and comprehensive evaluations in evaluation texts, and inconsistencies between classroom behavior and teaching effectiveness. It improves the stability, interpretability, and automation level of the teaching evaluation system and can be easily deployed on smart education platforms. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the index dependency graph structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the semantic alignment module according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the constraint calculation and score construction in an embodiment of the present invention; Figure 5 This is a schematic diagram of candidate reordering and closed-loop rewriting during the inference stage of an embodiment of the present invention; Figure 6 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a large model generation control method based on indicator dependency graph constraints for generating teaching evaluation reports, including: Obtain the set of teaching evaluation indicators; Based on the set of teaching evaluation indicators, construct a teaching evaluation indicator dependency graph that represents the logical dependencies between the indicators; Candidate teaching evaluation texts are generated through a large model, and semantic vectors are obtained by extracting the semantic intensity corresponding to each teaching evaluation indicator from the candidate teaching evaluation texts based on the teaching evaluation indicator dependency graph. Based on the teaching evaluation index dependency graph and the semantic vector, constraint calculation and conflict detection are performed on the candidate teaching evaluation texts to obtain consistency constraint loss and conflict penalty. The candidate teaching evaluation texts are scored based on the consistency constraint loss and the conflict penalty, and then re-ranked according to the scoring results to select the optimal candidate teaching evaluation text. Determine whether the conflict penalty of the optimal candidate teaching evaluation text exceeds a preset threshold. If it does, trigger a closed-loop rewriting mechanism to regenerate the candidate teaching evaluation text. If it does not exceed the threshold, output the optimal candidate teaching evaluation text as the final teaching evaluation report.

[0019] The specific process is as follows: The core idea of ​​this invention is to introduce a structural constraint mechanism for teaching evaluation indicators during the generation of teaching evaluation reports. First, the classroom teaching evaluation indicators are structured and modeled, and a dependency graph between the indicators is constructed. Then, a semantic alignment module extracts the semantic strength representation of each evaluation indicator from the candidate evaluation text. Based on this, consistency constraints and conflict penalty functions are constructed according to the indicator dependencies, and the candidate evaluation texts are scored for structural consistency. During the large model inference stage, the generation path is controlled through candidate reordering or probability reweighting mechanisms, and the evaluation text is iteratively generated through a threshold-triggered closed-loop rewriting mechanism, thereby outputting a final evaluation report that meets the logical consistency requirements of teaching evaluation.

[0020] By using the above method, the present invention can explicitly introduce the structural dependencies between teaching evaluation indicators when generating teaching evaluation reports, so that the generation process is controlled by structural constraints, thereby avoiding cross-indicator semantic conflicts in the evaluation text and improving the logical consistency and reliability of automatically generated teaching evaluation reports.

[0021] (I) Control methods for generating teaching evaluation reports: This invention provides a large-scale model generation and control method based on indicator dependency graph constraints for generating teaching evaluation reports. The process is as follows: Figure 1 As shown, the main steps include the following: S1: Structured modeling of teaching evaluation indicators; In the task of generating teaching evaluation reports, the first step is to perform structured modeling of classroom teaching evaluation indicators to construct a set of teaching evaluation indicators: ; in, This represents the i-th teaching evaluation indicator.

[0022] In smart education and teaching evaluation scenarios, the aforementioned indicators may include, but are not limited to, teacher explanation clarity, student comprehension, classroom interaction frequency, classroom pacing stability, and comprehensive evaluation conclusions. These indicators typically originate from classroom observation records, classroom behavior analysis results, learning behavior statistics, or teaching evaluation scales. This step transforms the multidimensional evaluation indicators involved in the teaching evaluation task into a structured set of indicators, providing a foundation for subsequent dependency modeling and constraint calculation.

[0023] S2: Construct a dependency graph of teaching evaluation indicators; To express the logical relationships between teaching evaluation indicators, this invention constructs a teaching evaluation indicator dependency graph: ; Where V represents the set of teaching evaluation indicator nodes, E represents the set of indicator-dependent edges, and W represents the set of edge weights.

[0024] For any dependent edge Its definition is: ; in, As upstream evaluation indicator nodes, For downstream evaluation indicator nodes, Indicates dependency weights. Indicates the type of dependency constraint. This indicates a dependency on a threshold parameter. The constraint type... It can include monotonic constraints, interval constraints, and mutual exclusion constraints.

[0025] For example, in classroom teaching evaluation scenarios: Clarity of teacher's explanation It will affect students' level of understanding ; Classroom interaction frequency It will affect classroom participation assessment; Multiple sub-indicators collectively influence the overall evaluation conclusion.

[0026] By constructing the aforementioned indicator dependency graph, the logical relationships between teaching evaluation indicators can be transformed into a computable structural model.

[0027] S3: Semantic alignment of teaching evaluation texts; When generating teaching evaluation reports, the large model first generates a set of candidate evaluation texts: ; in Let K represent the k-th candidate teaching evaluation text, where K represents the number of candidate texts.

[0028] To calculate the correspondence between candidate texts and teaching evaluation indicators, this invention extracts the semantic strength of the indicators from the candidate texts using a semantic alignment module, resulting in a semantic vector: ; in Indicates candidate text For indicators The semantic strength.

[0029] The semantic strength extraction process can be achieved through the following steps: 1. Locate relevant paragraphs in the candidate text based on the evaluation indicator name or structural tags; 2. Calculate the semantic strength of the indicator through rule mapping or classification models; 3. Aggregate and normalize multiple descriptions.

[0030] Meanwhile, to handle situations where candidate texts may lack descriptions of certain evaluation metrics, this invention defines a missing marker vector: ; in: ; This step transforms candidate evaluation texts into indicator semantic vectors, thereby achieving semantic alignment between the text and the evaluation indicators.

[0031] S4: Constraint calculation and conflict detection; like Figure 4 As shown, based on the teaching evaluation index dependency graph and semantic vector, this invention calculates the structural consistency constraints of candidate evaluation texts.

[0032] For monotonic constraints, define a consistency constraint function: ; in This is a proportionality coefficient. The template output is the degree of monotonic violation; a larger value indicates a more severe downstream deviation.

[0033] For interval constraints, define interval constraints as follows: ; For mutually exclusive constraints, define a conflict function: ; in This is the mutual exclusion threshold.

[0034] At the same time, conflict penalties are calculated based on mutual exclusion conflicts: ; This step allows for the quantitative detection of cross-index semantic conflicts in candidate evaluation texts.

[0035] S5: Candidate reordering during the inference phase; like Figure 5 As shown, in order to control the consistency of the evaluation text structure during the generation stage, this invention performs scoring calculations on candidate texts during the reasoning stage.

[0036] The total score of candidate texts is defined as follows: ; in: Cost of generating language models; Loss due to consistency constraints; Punishment for conflict; , These are the weight parameters.

[0037] Then, the candidate text with the lowest score is selected as the current optimal evaluation text: ; Using the above method, candidate evaluation texts with higher structural consistency can be prioritized during the reasoning stage.

[0038] S6: Closed-loop rewrite control To further improve the structural consistency of the generated results, this invention sets a conflict threshold: ; When the conflict penalty of the best candidate text exceeds the threshold, the system triggers a closed-loop rewriting mechanism to regenerate the candidate text using the following strategy: Increase the number of candidate texts K; Adjusting the weight of conflict penalties ; Adjust the missing penalty parameters.

[0039] The system repeats steps S3 to S5 until a teaching evaluation report that meets the constraints is generated. The final output is a teaching evaluation report text with a consistent structure. .

[0040] like Figure 6 As shown, on the other hand, this embodiment also provides a large model generation and control system based on index dependency graph constraints for generating teaching evaluation reports, including: The data input module is used to receive classroom teaching evaluation data; The indicator modeling module is used to construct a set of teaching evaluation indicators based on the classroom teaching evaluation data. The dependency graph construction module is used to construct a teaching evaluation index dependency graph representing the logical dependencies between the indicators based on the set of teaching evaluation indicators. The large model interface module is used to call the large model to generate candidate teaching evaluation texts; The semantic alignment module is used to extract the semantic intensity corresponding to each teaching evaluation indicator from the candidate teaching evaluation text based on the teaching evaluation indicator dependency graph, and obtain a semantic vector. The constraint calculation module is used to perform constraint calculation and conflict detection on the candidate teaching evaluation text based on the teaching evaluation index dependency graph and the semantic vector, and obtain consistency constraint loss and conflict penalty. A generation control module is used to score and re-rank the candidate teaching evaluation texts based on the consistency constraint loss and the conflict penalty, so as to select the optimal candidate teaching evaluation text. The closed-loop rewriting module is used to determine whether the conflict penalty of the optimal candidate teaching evaluation text exceeds a preset threshold, and triggers a rewriting mechanism when it does, so as to control the large model interface module to regenerate the candidate teaching evaluation text.

[0041] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0042] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0043] Example 2 This embodiment provides a large model generation control method based on indicator dependency graph constraints for generating teaching evaluation reports, including: In a smart education and teaching evaluation system, the classroom teaching process typically generates various types of teaching behavior data, such as classroom behavior records, classroom observation data, learning behavior statistics, and teaching evaluation scales. The system first receives classroom teaching evaluation data as input data X.

[0044] Based on the requirements of teaching evaluation tasks, a structured model of classroom teaching evaluation indicators is constructed to create a set of teaching evaluation indicators: ; in: Teachers explain the clarity index. Student comprehension level indicators Classroom interaction frequency index Classroom rhythm stability index Classroom comprehensive evaluation indicators.

[0045] The above indicators can be derived from information such as classroom observation records, classroom video analysis results, learning behavior statistics, and teaching evaluation scale scores.

[0046] The original teaching data is processed using an indicator extraction function to obtain an indicator state vector: ; in This represents the normalized score of the i-th teaching evaluation indicator, with a value range of [0,1].

[0047] For example, in the evaluation of a certain classroom lesson, the system obtains the following indicator scores: ; Among them: the teacher's explanation clarity score is 0.70, the student's comprehension score is 0.50, the classroom interaction frequency score is 0.60, the classroom rhythm stability score is 0.55, and the overall evaluation index is initially set to 0, which will be used to obtain it later through semantic mapping of the evaluation text.

[0048] This step transforms the multidimensional indicator information in the teaching evaluation task into a structured set of indicators, providing a foundation for subsequent dependency modeling and generation control.

[0049] like Figure 2 As shown, in order to express the logical relationship between teaching evaluation indicators, this invention constructs a teaching evaluation indicator dependency graph: G = (V, E, W); Where: V represents the set of teaching evaluation index nodes, E represents the set of index dependency edges, and W represents the set of dependency weights; For any dependent edge, it is represented as: ; in: Indicates the upstream indicator node; Indicates downstream indicator nodes; Indicates dependency weights; In teaching evaluation scenarios, the following indicator dependency relationship can be established: clarity of teacher explanation It will affect students' level of understanding Classroom interaction frequency It will affect students' level of understanding Classroom rhythm stability It will affect classroom interaction. Multiple sub-evaluation indicators jointly influence the comprehensive evaluation conclusion. ; The aforementioned dependencies can be determined using teaching evaluation rules or historical teaching evaluation data.

[0050] By constructing an indicator dependency graph, the logical relationships between teaching evaluation indicators can be transformed into a computable structured model.

[0051] In the process of generating teaching evaluation reports, the large model first generates a set of candidate teaching evaluation texts based on the input teaching evaluation indicators: ; in: Let K represent the k-th candidate teaching evaluation report text; K represents the number of candidate texts. In this embodiment, the system sets the number of candidate texts K = 5.

[0052] To ensure consistency in the evaluation report structure, the system requires the generated teaching evaluation report to include the following structure: evaluation of teacher explanation clarity; evaluation of student comprehension; evaluation of classroom interaction; evaluation of classroom pacing control; and evaluation of overall teaching effectiveness. Each evaluation section includes the following: evaluation conclusion; evaluation basis; improvement suggestions; like Figure 3 As shown, the semantic intensity corresponding to each evaluation index is then extracted from the candidate evaluation text using the semantic alignment module, resulting in a semantic vector: ; in: Indicates candidate text The semantic strength corresponding to the i-th evaluation index.

[0053] For example, for the candidate evaluation text Y^{(1)}, the semantic vector is obtained: ; At this point, the student's comprehension level was rated at 0.30, while the explanation clarity was rated at 0.85, indicating a possible inconsistency in the structure.

[0054] Based on the index dependency graph and semantic vector, this invention calculates the consistency constraints of candidate texts.

[0055] For monotonic dependencies, define a consistency constraint function: ; in: This is the proportionality coefficient.

[0056] For example, for candidate text : when = 0.85, = 0.30, and = 0.6; but: ; This result indicates that students' comprehension assessment was lower than expected for the clarity of the explanation.

[0057] Then, the consistency loss is calculated: ; Simultaneously define conflict penalties: ; The above method can be used to quantify cross-index semantic conflicts in candidate evaluation texts.

[0058] In order to control the consistency of the evaluation report structure during the generation stage, this invention scores the candidate evaluation texts during the reasoning stage.

[0059] The total score of candidate texts is defined as follows: ; in: Cost of generating language models; Loss due to consistency constraints; Punishment for conflict; , These are weight parameters; The system scores and ranks all candidate texts, and selects the candidate text with the lowest score as the current optimal teaching evaluation report. ; To further ensure the structural consistency of teaching evaluation reports, this invention sets a conflict threshold: When the optimal candidate text satisfies: ; Then output the final teaching evaluation report: ; like: ; This triggers a closed-loop rewriting mechanism, which regenerates candidate text using the following strategies: Increase the number of candidate texts K; increase the conflict penalty weight. Add a structural missing penalty parameter; The system re-executes the inference phase to generate candidate evaluation texts and performs semantic alignment to the candidate reordering in the inference phase until a teaching evaluation report text that meets the structural constraints is generated.

[0060] Through the above process, the present invention can achieve cross-indicator structural consistency control in the process of generating teaching evaluation reports, thereby outputting logically consistent and structurally complete teaching evaluation reports.

[0061] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating and controlling a large model based on index dependency graph constraints for teaching evaluation report generation, characterized in that, include: Obtain the set of teaching evaluation indicators; Based on the set of teaching evaluation indicators, construct a teaching evaluation indicator dependency graph that represents the logical dependencies between the indicators; Candidate teaching evaluation texts are generated through a large model, and semantic vectors are obtained by extracting the semantic intensity corresponding to each teaching evaluation indicator from the candidate teaching evaluation texts based on the teaching evaluation indicator dependency graph. Based on the teaching evaluation index dependency graph and the semantic vector, constraint calculation and conflict detection are performed on the candidate teaching evaluation texts to obtain consistency constraint loss and conflict penalty. The candidate teaching evaluation texts are scored based on the consistency constraint loss and the conflict penalty, and then re-ranked according to the scoring results to select the optimal candidate teaching evaluation text. Determine whether the conflict penalty of the optimal candidate teaching evaluation text exceeds a preset threshold. If it does, trigger a closed-loop rewriting mechanism to regenerate the candidate teaching evaluation text. If it does not exceed the threshold, output the optimal candidate teaching evaluation text as the final teaching evaluation report.

2. The method according to claim 1, characterized in that, The construction of the teaching evaluation indicator dependency graph, which represents the logical dependencies between indicators, specifically includes: Based on the nodes, dependencies between indicators, and edge weights in the set of teaching evaluation indicators, construct the teaching evaluation indicator dependency graph; The dependency edge defines the upstream indicator node, the downstream indicator node, the dependency relationship weight, the dependency constraint type, and the dependency threshold parameter.

3. The method according to claim 1, characterized in that, The semantic intensity corresponding to each teaching evaluation indicator is extracted from the candidate teaching evaluation texts to obtain the semantic vector, which specifically includes: Based on the evaluation index name, the relevant paragraphs in the candidate teaching evaluation text are located, and the semantic strength is calculated through rule mapping or classification model. Multiple descriptions are then aggregated and normalized to obtain the semantic vector.

4. The method according to claim 1, characterized in that, The constraint calculation and conflict detection of the candidate teaching evaluation texts to obtain the consistency constraint loss and conflict penalty specifically include: Based on the monotonic constraint relationship in the indicator dependency graph, a monotonic constraint consistency constraint function is calculated to quantify the degree of violation of downstream indicators relative to upstream indicators. Based on the mutual exclusion constraint relationship in the index dependency graph, a mutual exclusion conflict function is calculated to quantify the degree of conflict between two mutually exclusive indices. The conflict penalty is calculated based on the result of the mutual exclusion conflict function.

5. The method according to claim 1, characterized in that, Scoring the candidate teaching evaluation texts and re-ranking them based on the scoring results specifically includes: The total score for each candidate teaching evaluation text is calculated based on the weighted sum of the language model generation cost, the consistency constraint loss, and the conflict penalty. The candidate teaching evaluation text with the lowest total score is selected as the optimal candidate teaching evaluation text.

6. The method according to claim 1, characterized in that, The triggering closed-loop rewriting mechanism to regenerate candidate teaching evaluation texts specifically includes: The large model is triggered to regenerate candidate teaching evaluation texts by one or more strategies, such as increasing the number of candidate texts, increasing the conflict penalty weight, or increasing the structural missing penalty parameter.

7. The method according to claim 1, characterized in that, The specific set of teaching evaluation indicators includes: Based on classroom observation records, classroom behavior analysis results, learning behavior statistics, or teaching evaluation scale information, normalized indicator scores are extracted to construct the teaching evaluation indicator set.

8. A large-scale model generation and control system based on index dependency graph constraints for generating teaching evaluation reports, characterized in that, include: The data input module is used to receive classroom teaching evaluation data; The indicator modeling module is used to construct a set of teaching evaluation indicators based on the classroom teaching evaluation data. The dependency graph construction module is used to construct a teaching evaluation index dependency graph representing the logical dependencies between the indicators based on the set of teaching evaluation indicators. The large model interface module is used to call the large model to generate candidate teaching evaluation texts; The semantic alignment module is used to extract the semantic intensity corresponding to each teaching evaluation indicator from the candidate teaching evaluation text based on the teaching evaluation indicator dependency graph, and obtain a semantic vector. The constraint calculation module is used to perform constraint calculation and conflict detection on the candidate teaching evaluation text based on the teaching evaluation index dependency graph and the semantic vector, and obtain consistency constraint loss and conflict penalty. A generation control module is used to score and re-rank the candidate teaching evaluation texts based on the consistency constraint loss and the conflict penalty, so as to select the optimal candidate teaching evaluation text. The closed-loop rewriting module is used to determine whether the conflict penalty of the optimal candidate teaching evaluation text exceeds a preset threshold, and triggers a rewriting mechanism when it does, so as to control the large model interface module to regenerate the candidate teaching evaluation text.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-7.

Images