College education data fusion and intelligent evaluation method and system based on large model

Abstract

The invention belongs to the technical field of artificial intelligence and education data processing, and provides a college education data fusion and intelligent evaluation method and system based on a large model, and the method comprises the steps: constructing a college education evaluation database, carrying out the structural management and fusion processing of multi-source data related to the college education evaluation, and carrying out the intelligent evaluation of the college education data. Evaluation task adaptation is carried out on the large model, so that the large model has semantic comprehension and processing capability oriented to college education evaluation; on the basis, intention analysis is performed on an education evaluation natural language problem input by a user, a structured evaluation request is generated, screening, fusion and processing of evaluation data are completed under data and authority constraint conditions, and a corresponding college education intelligent evaluation result is generated; and meanwhile, the evaluation result data or model is updated in combination with user feedback information. According to the invention, large-model intelligent evaluation and generation output based on college education data fusion constraints are realized, and the accuracy, consistency and interpretability of college education evaluation results are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and educational data processing technology, specifically involving a method and system for the fusion and intelligent evaluation of higher education educational data based on a large model. It is used for the structured storage, access control, fusion processing, and generation of constrained evaluation results of multi-source data related to higher education teaching evaluation. Background Technology

[0002] Currently, the evaluation of higher education quality involves multiple dimensions of indicators, including teaching process, research output, student development, and management effectiveness. The data sources are diverse, complex in structure, and frequently updated. Existing higher education evaluation work typically relies on manual statistics, rule-based calculations, or expert judgment, resulting in low data utilization efficiency, long analysis cycles, and difficulties in ensuring consistency of results. With the improvement of educational informatization, universities have accumulated a large amount of structured evaluation data in their teaching management systems, research management systems, and student development platforms, providing a data foundation for education quality evaluation. However, traditional rule-based or statistical model-based evaluation methods struggle to achieve flexible querying and comprehensive processing under multi-indicator, multi-level evaluation tasks, and cannot meet the needs of administrators to quickly obtain evaluation results through natural language processing.

[0003] In recent years, large-scale models have demonstrated strong capabilities in natural language understanding and generation. Some solutions have attempted to introduce large-scale models into educational assessment question-answering tasks to achieve natural language expression of assessment results. However, existing educational assessment question-answering systems based on large-scale models typically only use the model as a general text generation tool, resulting in the following technical defects: (1) The model struggles to accurately understand the semantic boundaries and hierarchical relationships of different indicators in the higher education teaching and learning assessment indicator system; (2) There is a lack of effective linkage between the model generation process and the existing structured assessment data of higher education institutions, making it difficult to generate controlled results based on real assessment results; (3) In assessment tasks involving multiple users and multiple levels of management, existing solutions typically only implement access control during the data retrieval stage, making it difficult to prevent the model from introducing unauthorized data or inference conclusions during the generation stage, leading to problems such as unauthorized or inconsistent assessment results. Therefore, how to construct a technical solution that can collaboratively control structured assessment data, user access constraints, and the large-scale model generation process under higher education assessment tasks has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a method for data fusion and intelligent evaluation of higher education based on a large model, in order to solve the technical problems in existing higher education evaluation tasks, such as the difficulty in effectively combining structured evaluation data during the analysis and generation process, the difficulty in continuously being subject to permission constraints during the generation stage, and the tendency to produce unauthorized or inconsistent evaluation results.

[0005] By constructing a database for higher education evaluation and coordinating user permission constraints, structured evaluation data processing, and the large model generation process, the large model can generate higher education evaluation results within a limited data range and permission conditions, thereby improving the accuracy, consistency, and controllability of the evaluation results.

[0006] The objective of this invention is achieved through the following technical measures.

[0007] A method for data fusion and intelligent evaluation in higher education based on a large model includes the following steps:

[0008] (1) Construction of a Higher Education Evaluation Database. Based on a pre-set higher education evaluation index system, the basic data, evaluation index data and evaluation result data of higher education institutions are stored in a structured manner to construct a higher education evaluation database.

[0009] (2) Large-scale model evaluation task adaptation. Based on data from the field of higher education evaluation, the pre-trained large-scale model is adapted to the evaluation task, so that the large-scale model has the semantic parsing capability for higher education evaluation tasks.

[0010] (3) Natural Language Assessment Intent Analysis and Permission Determination. Semantic analysis is performed on the natural language questions on higher education assessment input by the user to identify the corresponding assessment indicators, data range and user permission information, and to generate a structured data retrieval request.

[0011] (4) Evaluation data screening and fusion processing. Based on the data retrieval request and user permission information, candidate evaluation data is obtained from the higher education evaluation database, and the candidate evaluation data is screened, associated and fused to generate structured evaluation results.

[0012] (5) Generation of evaluation results subject to structured data and permission constraints. The structured evaluation data, user questions, and permission constraint parameters are input into the large model generation module to generate corresponding intelligent evaluation results for higher education under the conditions of structured data and user permission constraints.

[0013] (6) Feedback-driven model update. Collect user feedback on the intelligent evaluation results of higher education, and update the higher education evaluation result data or large model based on the feedback.

[0014] Furthermore, the construction of the higher education evaluation database in step (1) includes:

[0015] (1-1) Construct a basic data table for universities to store university identification information and education evaluation related data corresponding to the universities. The basic data table for universities includes an identification field for uniquely identifying universities, an information field for representing university categories, and evaluation data fields associated with universities.

[0016] (1-2) Construct an evaluation indicator data table to store multi-level education evaluation indicator information. The evaluation indicator data table includes at least a first-level indicator table, a second-level indicator table, and a third-level indicator table. A hierarchical mapping relationship is established between the indicator tables at each level through indicator association fields, and permission fields for identifying access permissions to indicator data are set in the indicator tables at each level.

[0017] (1-3) Construct an evaluation result database to store educational evaluation result information corresponding to universities. The evaluation result database includes a university identifier field and an evaluation result index field corresponding to the university.

[0018] Furthermore, the large model evaluation task adaptation in step (2) includes:

[0019] (2-1) Load the pre-trained large model, and while keeping the main parameters of the large model unchanged, determine the parameter adjustment region for evaluating task adaptation;

[0020] (2-2) Introduce a low-rank decomposition structure based on LoRA into part of the network structure of the large model. By adjusting the corresponding parameters of the structure, the large model can obtain semantic parsing capabilities for higher education evaluation tasks.

[0021] (2-3) Construct a structured training data template that matches the higher education evaluation task. The structured training data template is used to uniformly describe evaluation instruction information, evaluation context information and corresponding target output information.

[0022] (2-4) The adaptation structure is trained based on the structured training data template, and a large model adapted to the higher education evaluation task is generated after the training is completed.

[0023] Furthermore, step (3) of natural language assessment intent parsing and permission determination includes:

[0024] (3-1) Obtain user identity-related information, which includes at least the user's unit identifier, user type, and user permission level;

[0025] (3-2) Semantically analyze the natural language questions on higher education evaluation input by users, extract information on evaluation objects, evaluation indicators and data range related to higher education evaluation, and form a structured semantic description;

[0026] (3-3) Identify the user's evaluation intent based on the structured semantic description, and map the natural language question into a structured data retrieval request containing evaluation indicators and data range information;

[0027] (3-4) Determine the corresponding permission constraint parameters based on the user identity-related information. The permission constraint parameters are used to limit the scope of accessible universities and the granularity of data access.

[0028] (3-5) Based on the structured data retrieval request and the permission constraint parameters, retrieve and return evaluation data that meets the permission constraint conditions from the higher education evaluation database, as input data for subsequent evaluation and generation processing.

[0029] Furthermore, the evaluation data screening and fusion processing in step (4) includes:

[0030] (4-1) Based on the structured data retrieval request, the candidate evaluation data is preprocessed. The preprocessing includes data validity verification and permission consistency verification to obtain a subset of evaluation data that meets the permission constraints.

[0031] (4-2) Based on the user's evaluation intent, the subset of evaluation data is filtered for relevance, and evaluation data that matches the evaluation intent is retained;

[0032] (4-3) Perform multi-dimensional correlation and fusion processing on the screened evaluation data to generate comprehensive evaluation data to characterize the status of higher education evaluation;

[0033] (4-4) When the comprehensive evaluation data involves different time dimensions or different data sources, perform cross-time dimension alignment and cross-data source topic association processing on the comprehensive evaluation data;

[0034] (4-5) Generate structured evaluation results based on the processed comprehensive evaluation data, which are used as input data for subsequent large model generation steps.

[0035] Furthermore, the generation of the evaluation results subject to structured data and permission constraints in step (5) includes:

[0036] (5-1) Obtain the structured evaluation result generated in step (4) and the permission constraint parameters corresponding to the structured evaluation result;

[0037] (5-2) Input the structured evaluation results, the user's higher education evaluation questions and the permission constraint parameters into the large model generation module to limit the generation scope of the large model, so that it generates content only based on the structured evaluation results and under the permission constraint conditions;

[0038] (5-3) Perform permission consistency verification on the evaluation results generated by the large model to determine whether the generated content conforms to the permission constraint parameters;

[0039] (5-4) If the generated content does not meet the permission constraint parameters, the generated result is adjusted to make the output intelligent evaluation result of higher education meet the permission constraint conditions and then output.

[0040] Furthermore, the feedback-driven model update in step (6) includes:

[0041] (6-1) Collect user feedback information on the intelligent evaluation results of higher education, the feedback information including at least result correction information, usage behavior information or evaluation mark information, and store the feedback information in a structured manner;

[0042] (6-2) Generate labeled data or training samples for evaluating task updates based on the feedback information;

[0043] (6-3) Use the labeled data or training samples to update the large model or the data of higher education evaluation results to improve the accuracy and consistency of the intelligent evaluation results of higher education.

[0044] (6-4) After the update is completed, save the updated model parameters or evaluation result data for use in the subsequent intelligent evaluation process of higher education.

[0045] This invention also provides a large-scale model-based system for the fusion and intelligent evaluation of higher education data, comprising the following modules:

[0046] The Higher Education Evaluation Database Module is used to store and manage basic data, evaluation indicator data, and evaluation result data related to higher education evaluation.

[0047] The large model evaluation task adaptation module is used to adapt the large model to evaluation tasks based on data in the field of higher education evaluation, so that the large model has the semantic parsing capability for higher education evaluation application tasks.

[0048] The Natural Language Assessment Intent Parsing and Permission Determination module is used to perform semantic parsing on the natural language questions for higher education assessment input by users, generate a structured data retrieval request containing assessment indicators and data range information, and determine the corresponding permission constraint parameters.

[0049] The evaluation data filtering and fusion module is used to obtain evaluation data from the higher education evaluation database module based on the structured data retrieval request and the permission constraint parameters, and to filter, associate and fuse the evaluation data to generate structured evaluation results.

[0050] The generation module, constrained by structured data and permissions, is used to input the structured evaluation results, the user's higher education evaluation questions, and the permission constraint parameters into the large model, and generate intelligent evaluation results for higher education under the conditions of structured data and permission constraints.

[0051] The feedback-driven model update module is used to collect user feedback on the intelligent evaluation results of higher education, and update the large model or higher education evaluation result data based on the feedback information.

[0052] The present invention provides a method and system for data fusion and intelligent evaluation of higher education based on a large model. Its beneficial effects are as follows: (1) By introducing structured higher education evaluation data into the large model generation process and imposing data range constraints on the generation stage, the risk of inconsistency between the model generation results and the actual evaluation data is reduced; (2) By introducing a unified permission constraint mechanism in the data retrieval, data fusion and generation output stages, the problem of introducing unauthorized data or reasoning conclusions into the large model during the generation process is effectively avoided; (3) By constructing structured evaluation results as model input, the collaborative control of data processing and natural language generation in the higher education evaluation process is realized, which improves the stability and interpretability of the evaluation results. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method for data fusion and intelligent evaluation in higher education according to an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the higher education evaluation database module according to an embodiment of the present invention.

[0055] Figure 3 This is a schematic diagram of the large model evaluation task adaptation process according to an embodiment of the present invention.

[0056] Figure 4 This is a schematic diagram of the natural language assessment intent parsing process according to an embodiment of the present invention.

[0057] Figure 5 This is a schematic diagram of the data filtering and fusion processing flow according to an embodiment of the present invention.

[0058] Figure 6 This is a schematic diagram of the evaluation result generation process according to an embodiment of the present invention.

[0059] Figure 7 This is a schematic diagram of the evaluation results of higher education in an embodiment of the present invention.

[0060] Figure 8 This is a schematic diagram of the feedback-driven model update process according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the specific embodiments of this invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the technical solutions of this invention and do not constitute a limitation on the scope of protection of this invention. Equivalent transformations or substitutions made by those skilled in the art to the embodiments without departing from the technical concept of this invention should all fall within the scope of protection of this invention.

[0062] like Figure 1 As shown in the figure, this invention provides a method for data fusion and intelligent evaluation of higher education based on a large model. The method includes the following steps:

[0063] (1) Constructing a Higher Education Evaluation Database. Based on intelligent evaluation indicators for higher education, establish a basic evaluation database, an evaluation indicator database, and an evaluation result database. Structure multi-source educational data (including classroom data, research data, questionnaire data, etc.) and establish semantic vector indexes for indicator descriptions. Optionally, the overall content of the constructed higher education evaluation database is as follows: Figure 2 As shown, the specific steps include the following:

[0064] (1-1) Construct a basic database of universities to store university identification information and education evaluation related data corresponding to the universities. The basic database of universities includes an identification field for uniquely identifying universities, an information field for representing university categories, and evaluation data fields associated with universities.

[0065] (1-2) Construct an evaluation indicator database to store multi-level education evaluation indicator information. The evaluation indicator data table includes at least a first-level indicator table, a second-level indicator table, and a third-level indicator table. A hierarchical mapping relationship is established between the indicator tables at each level through indicator association fields, and permission fields are set in the indicator tables at each level to identify the access permissions of indicator data.

[0066] (1-3) Construct an evaluation result database to store educational evaluation result information corresponding to universities. The evaluation result database includes a university identifier field and an evaluation result index field corresponding to the university.

[0067] (2) Assessment Task Adaptation of the Large Model. To enable the large model to understand the evaluation index system and related semantics of higher education, the pre-trained large model is adapted for the assessment task. The detailed process is as follows: Figure 3 As shown.

[0068] In one implementation, the evaluation task adaptation is achieved through efficient parameter fine-tuning, including the following process:

[0069] (2-1) Loading and freezing model parameters. Load the pre-trained large model parameters, and then freeze all parameters except for the specified modules to avoid backpropagation updates. Select the Query and Value weight matrices in the multi-head attention layer to insert fine-tuning modules.

[0070] (2-2) Inserting a low-rank adaptation structure. A low-rank adaptation structure is introduced into the fine-tuning module, and only the low-rank parameters related to the evaluation task are trained and updated, so as to achieve rapid adaptation of the model to the evaluation task of higher education teaching while keeping the main structure of the model unchanged.

[0071] A low-rank decomposition method based on LoRA (Low-Rank Adaptation) is adopted. The original linear transformation weight matrix is... Reparameterization is performed, represented as:

[0072]

[0073] in For a dimension reduction matrix, It is an upgraded matrix; The value is low rank. Only matrices A and B are trained, while the remaining parameters remain frozen; during the model's forward propagation, the output is:

[0074]

[0075] in This is an adjustable scaling factor used to balance the adaptation strength; during backpropagation, only the gradients of A and B are updated. This allows the model to be fine-tuned for the university evaluation task while keeping the main weights constant.

[0076] (2-3) Construct a structured training data template that matches the higher education evaluation task. Based on the characteristics of the higher education evaluation task, a structured training data template is constructed to standardize the input and output formats of the model. The structured training data template includes at least instruction information, context information, and target output information, where the context information describes the structured evaluation data related to higher education evaluation. Table 1 provides an example of the structured training data template.

[0077] Table 1. Examples of Structured Training Data Templates

[0078] Field Name meaning Example Instruction Instruction content Please evaluate the school's performance in the XX subject area. Context Context information The number of papers published in this discipline in the past three years is X. Target Expected output The overall quality of education in this discipline is relatively stable.

[0079] (2-4) Model Training and Parameter Saving. Based on the structured training data template, the parameters corresponding to the low-rank adaptation structure are trained and updated. After training, only the parameters related to the evaluation task adaptation are saved. During the model inference stage, the evaluation task adaptation parameters are dynamically loaded, enabling the large model to perform semantic parsing tasks related to higher education teaching evaluation.

[0080] For example, during training, the AdamW optimizer is used to update the parameters A and B of the low-rank module, and a linear warmup strategy is employed. After training, the system only saves the weights of parameters A and B, which are then used during model inference via the formula:

[0081]

[0082] When using it, the model is invoked by dynamically loading weights.

[0083] (3) Natural Language Assessment Intent Parsing and Permission Determination. The system receives user input of a natural language question related to higher education assessment, performs semantic parsing on the natural language question, and generates corresponding data retrieval requests and permission constraint parameters. The detailed process is as follows: Figure 4 As shown.

[0084] (3-1) Obtaining User Identity Information. Based on the user's login status and request context information, obtain the user's identity information, which includes at least the user's unit identifier, user type, and user permission level.

[0085] (3-2) Evaluation Semantic Parsing. Semantic parsing is performed on the natural language question 𝑄 of higher education evaluation input by the user to extract key information related to higher education evaluation, and the key information is organized into a structured semantic description.

[0086] In one implementation, the structured semantic description includes at least assessment object entity information, assessment indicator information, time range information, and assessment task type information, as shown below:

[0087]

[0088] (3-3) Evaluation Intent Recognition and Request Mapping. Based on the structured semantic description, the user's evaluation intent type is identified, and the natural language question is mapped to the corresponding structured evaluation request for subsequent evaluation data retrieval and processing.

[0089] In one implementation, the assessment intent type includes types such as indicator query, trend assessment, difference comparison, data statistics, and comprehensive assessment.

[0090] Type number meaning Example A1 Indicator Inquiry Which college has the highest teaching quality? A2 Trend Assessment What are the trends in teaching quality over the past three years? A3 Difference Comparison Is there a significant difference in teaching quality between College A and College B? A4 Data statistics How many teachers were rated as excellent this year? A5 Comprehensive assessment "Please provide an overall evaluation of the school's teaching situation."

[0091] Accepting input Next, identify the type of consciousness in the input. Sofrmax classification was used for consciousness classification.

[0092]

[0093] in This is a sentence vector representation.

[0094] After confirming the user's intent, the user's question is mapped into a structured evaluation request description based on the user's intent for subsequent data retrieval and processing.

[0095] (3-4) Determination of permission constraint parameters. Based on the user identity information, determine the permission constraint parameters corresponding to the user. The permission constraint parameters include at least the range of universities that can be accessed and the data access granularity.

[0096] The permission constraint parameters are used to limit the scope of data access during the subsequent evaluation data screening and evaluation result generation process.

[0097] (3-5) Generation of evaluation data retrieval request. Based on the structured evaluation request and permission constraint parameters, an executable evaluation data retrieval request is generated and sent to the higher education evaluation database to obtain evaluation data that meets the permission constraint conditions.

[0098] In one implementation, embedding-based semantic similarity calculation is used:

[0099] First, for each database field name With input keywords Calculate similarity:

[0100]

[0101] when When the empirical threshold is 0.75, establish the mapping relationship:

[0102]

[0103] Based on the intent type and semantic mapping results, an executable query statement is automatically generated. For example, for a trend-related question (A2):

[0104]

[0105] The query statement generation logic is implemented through a template system:

[0106]

[0107] After generating the data query statement, access the university intelligent evaluation database and return the query result set. To facilitate the generation of large models, the structured data is also converted into natural language description templates. The conversion formula is as follows:

[0108]

[0109] in The formatting function represents data to natural language, which will generate natural language results such as "The teaching quality improvement rate of College A is 12.3%, College B is 11.8%, and College C is 9.6%".

[0110] (4) Evaluation data screening and fusion processing. In this embodiment, when the system retrieves candidate evaluation data from the higher education evaluation database based on the data retrieval request generated in step (3), Subsequently, data filtering and fusion processing is performed on the candidate evaluation data to generate structured evaluation results for subsequent evaluation results. The detailed process is as follows: Figure 5 As shown.

[0111] (4-1) Permission consistency verification and data preprocessing. For the candidate evaluation data... Data format validation and invalid data processing are performed, and based on the permission constraint parameters, permission consistency validation is performed on the candidate evaluation data to remove data records that do not conform to the user's permission scope.

[0112] In one implementation, the permission consistency check includes, but is not limited to:

[0113] • Remove data that is outside the scope of universities accessible to users;

[0114] • Remove highly sensitive data that users do not have access to;

[0115] • Remove data that does not belong to the current evaluation indicator system.

[0116] After the permission consistency check, an evaluation data subset that meets the permission constraints is obtained.

[0117] (4-2) Relevance Screening. Based on the evaluation object, evaluation indicators and evaluation task type in the structured evaluation request, the subset of evaluation data is screened for relevance, and evaluation data that matches the user's evaluation intent is retained.

[0118] In one implementation, the main process includes:

[0119] User input Computing vector representations ;

[0120] For each record content Generate text descriptions and compute embedding vectors ;

[0121] Calculate cosine similarity:

[0122]

[0123] Retain satisfaction Records:

[0124]

[0125] When a user asks a question, if the retrieved fields are irrelevant to the question, they will be removed during the filtering process.

[0126] (4-3) Multi-indicator correlation and fusion processing. When the evaluation data contains multiple evaluation indicator dimensions, the evaluation indicators are correlated and fused to generate comprehensive evaluation data.

[0127] In one implementation, different evaluation indicators are assigned corresponding fusion weights based on the degree of correlation between the evaluation indicators and the user's evaluation intent, so as to form a comprehensive evaluation result.

[0128] For example, the data selected and retained may include multidimensional indicators (Teacher Development Index). Paper publication index Patent conversion amount The system performs weighted fusion based on the weight vector:

[0129]

[0130] The weighting system automatically adjusts based on user semantics; the higher the similarity between the indicator and the question, the greater the weight. For example, the weight increases when the question contains the keyword "paper". value.

[0131] (4-4) Cross-time and multi-source data fusion. When the evaluation data comes from different time intervals or different data sources, time alignment and topic association processing are performed on the evaluation data to generate a unified evaluation view.

[0132] In one implementation, during time alignment, the system uses timestamps. Merge the indexes:

[0133]

[0134] In addition to time alignment, topic aggregation is also required. Topic aggregation mainly involves averaging or normalizing the weighted average of records from multiple tables for colleges and disciplines.

[0135]

[0136] in, Indicates the first The fusion result of each theme over time is used as the final evaluation result for that theme. .

[0137] (4-5) Generation of structured evaluation results. Based on the above screening and fusion processing results, structured evaluation results are generated, which are used as input data for generating evaluation results of the large model in step (5).

[0138] In one implementation, the generation of the structured evaluation results includes, but is not limited to:

[0139] • Statistical processing of indicators;

[0140] • Calculation of comprehensive score based on multiple indicators;

[0141] • Vertical trend assessment within the same university;

[0142] • Horizontal comparison within the scope of permissions allowed.

[0143] (5) Generation of Constrained Evaluation Results. In this embodiment, based on the structured evaluation results generated in step (4), higher education evaluation results are generated under permission constraints. The detailed process is as follows: Figure 6 As shown.

[0144] (5-1) Input generation. Obtain the structured evaluation results, user evaluation questions and corresponding permission constraint parameters output from step (4), and use the structured evaluation results, user evaluation questions and permission constraint parameters together as the input for generation.

[0145] (5-2) Constrained content generation. The generation input is passed to the large model generation module, which restricts the model to generate content only based on the structured evaluation results and permission constraint parameters to obtain the initial evaluation results.

[0146] (5-3) Permission Consistency Verification. Perform a permission consistency verification on the initial evaluation results to determine whether the generated content contains data references or conclusion descriptions that exceed the permission constraint parameters.

[0147] (5-4) Handling of Unauthorized Content. When generated content is detected to exceed the permission constraint parameters, the corresponding content is hidden, replaced, or downgraded to obtain and output the higher education teaching evaluation results that meet the permission constraint conditions. The evaluation results are as follows: Figure 7 As shown.

[0148] (6) Feedback-driven model update. In this embodiment, the system collects and processes feedback information from users during the use of higher education evaluation results, and updates the higher education evaluation result data or large model based on the feedback information. The detailed process is as follows: Figure 8 As shown.

[0149] (6-1) Collection of User Feedback Information. During system operation, user feedback information on the evaluation results of higher education is collected. The feedback information includes at least the user's evaluation information and error correction information on the generated results.

[0150] In one implementation, the feedback information is stored in association with the corresponding user evaluation question, the generated result, and the data source information.

[0151] (6-1-1) Classification of User Feedback Collection. User feedback is divided into three categories: explicit feedback, implicit feedback, and error correction feedback.

[0152] Explicit feedback collection: The system summarizes user feedback during usage, recording direct evaluations including likes / dislikes, "Inaccurate answer" buttons, "Insufficient explanation" prompts, and user-initiated supplementary explanations (e.g., "The number of teachers in this college is not 122, but 143."). This type of feedback is directly quantifiable, and the system saves the feedback information along with the context of the current answer and the data source as a log entry.

[0153] Implicit Feedback Collection: Since many users don't actively provide feedback, the system also uses behavioral monitoring to determine the quality of responses, including: whether the user quickly closes the conversation (indicating dissatisfaction), immediately asks a new question (indicating an incomplete answer), repeatedly modifies the same question (indicating misunderstanding), or skips the current answer to directly view data tables or other functional modules (indicating the answer is not valuable enough). The system assigns a weight coefficient to each piece of implicit feedback, as shown in the example below:

[0154] Behavior speculation Weight Immediately ask follow-up questions Inadequate response 0.6 Repeatedly modify the problem Model Misunderstanding Intent 0.9 Dwell time < 2 seconds User ignored the answer 0.3 Copy answer content Highly valuable answers 0.8

[0155] This feedback will be integrated into the quality assessment vector. .

[0156] Error correction feedback collection: Users may directly point out erroneous data or logic, including: incorrect values, reference to outdated data, unreasonable reasoning processes, and omission of a key indicator. The system provides an error correction form for this purpose. User corrections will be tagged with: correctness information, data priority, and whether it is public or personalized knowledge. This section will be used for incremental updates to the knowledge base.

[0157] (6-2) Feedback information processing and update determination. The feedback information is sorted and processed, and an update process is determined based on preset update conditions.

[0158] In one implementation, when feedback indicates that the generated results contain data errors, logical deviations, or do not meet the requirements of the higher education evaluation task, subsequent update processing is triggered. The specific processing procedure is as follows:

[0159] User feedback vector construction: Explicit feedback, implicit feedback, and error correction content are uniformly encoded into weight vectors.

[0160]

[0161] The weights of different types of feedback can be configured, with explicit feedback having the highest weight and implicit feedback having the lowest weight.

[0162] Answer Quality Score: The system generates a quality score for each answer to determine whether an update needs to be triggered. The scoring model includes: whether the answer addresses the core intent, whether the cited data is complete, whether the reasoning is logical, whether the user raises new questions, and whether the behavioral data is abnormal (such as continued user dissatisfaction). The scoring formula is shown below:

[0163]

[0164] If the score is lower than the set threshold 𝜃, the subsequent update process will begin.

[0165] (6-3) Evaluation data update. When the feedback information involves the correction or supplementation of higher education evaluation results data, the feedback information is verified, and the verified data is updated to the higher education evaluation database.

[0166] (6-3-1) Data Validation and Conflict Detection. The data provided by the user needs to be compared with the internal database. If a conflict is found, the system will perform timestamp comparison (newer data takes priority), source credibility verification (such as whether it comes from an authoritative department / user's personal opinion), and multiple data consistency checks. If the data passes the verification, it will proceed to the next step.

[0167] (6-3-2) Structured Data Storage. Through data extraction and structuring algorithms, the error correction content or newly added data is converted into a searchable JSON format and then written to the database. For newly added or modified data, the system automatically generates a data change log, the scope of impact (which answer logic will be affected), and source tracing tags.

[0168] (6-4) Large model update. When feedback indicates that the generated results continue to perform poorly under a specific higher education evaluation task, a lightweight update is performed on the large model to make the model better adapt to the higher education evaluation task.

[0169] In one implementation, the large model update only adjusts model parameters or configurations relevant to higher education evaluation tasks.

[0170] (6-4-1) Set trigger conditions. The system can trigger an update when any condition is met, including: a large number of negative feedbacks are continuously received for a certain type of question, the same knowledge point is constantly corrected by users, the model performs poorly for a certain type of task (such as "trend assessment"), and new knowledge base content affects the answer results multiple times.

[0171] (6-4-2) Model Update. The system automatically constructs training samples, including: the original user question. Model incorrect answers Correct answer after correction and related data fragments Then, the model is updated through LoRA, AdapterTuning, Prompt-Tuning, and domain vocabulary. The updated model parameters and configurations will be recorded and written back to the system, including: new version markers, adaptation parameters for specific tasks, and correspondence with the knowledge base, thereby realizing the iterative evolution of the model version.

[0172] After the model is updated, the system continuously monitors user feedback for the next stage, forming a closed-loop process of "prediction → usage → feedback → improvement". Ultimately, the system becomes more accurate, more personalized, and more adept at understanding educational assessment tasks with each use, gradually evolving into a high-precision, logically robust, and continuously iterating educational model.

[0173] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0174] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for data fusion and intelligent evaluation of higher education based on a large model, characterized in that... The method includes the following steps: (1) Construction of a higher education evaluation database; Based on the pre-set higher education evaluation index system, the basic data, evaluation index data and evaluation result data of higher education institutions are stored in a structured manner to construct a higher education evaluation database; (2) Large model evaluation task adaptation; based on data in the field of higher education evaluation, the pre-trained large model is adapted to the evaluation task, so that the large model has the semantic parsing ability for higher education evaluation tasks; (3) Natural language assessment intent analysis and permission determination: Semantic analysis of the natural language questions on higher education assessment input by users, identification of corresponding assessment indicators, data range and user permission information, and generation of structured data retrieval requests; (4) Evaluation data screening and fusion processing; Based on the data retrieval request and user permission information, candidate evaluation data is obtained from the higher education evaluation database, and the candidate evaluation data is screened, correlated and integrated to generate structured evaluation results. (5) Generation of evaluation results subject to structured data and permission constraints; The structured evaluation data, user questions and permission constraint parameters are input into the large model generation module to generate the corresponding intelligent evaluation results of higher education under the conditions of structured data and user permission constraints; (6) Feedback-driven model update: Collect user feedback information on the intelligent evaluation results of higher education, and update the higher education evaluation result data or large model based on the feedback information.

2. The method for data fusion and intelligent evaluation of higher education based on a large model as described in claim 1, characterized in that... Step (1) involves constructing a higher education evaluation database, which includes: (1-1) Construct a basic data table for universities to store university identification information and education evaluation related data corresponding to the universities. The basic data table for universities includes an identification field for uniquely identifying universities, an information field for representing university categories, and evaluation data fields associated with universities. (1-2) Construct an evaluation indicator data table to store multi-level education evaluation indicator information. The evaluation indicator data table includes at least a first-level indicator table, a second-level indicator table, and a third-level indicator table. A hierarchical mapping relationship is established between the indicator tables at each level through indicator association fields, and permission fields for identifying access permissions to indicator data are set in the indicator tables at each level. (1-3) Construct an evaluation result database to store educational evaluation result information corresponding to universities. The evaluation result database includes a university identifier field and an evaluation result index field corresponding to the university.

3. The method for data fusion and intelligent evaluation of higher education based on a large model as described in claim 1, characterized in that... Step (2) of the large model evaluation task adaptation includes: (2-1) Load the pre-trained large model, and while keeping the main parameters of the large model unchanged, determine the parameter adjustment region for evaluating task adaptation; (2-2) Introduce a low-rank decomposition structure based on LoRA into part of the network structure of the large model. By adjusting the corresponding parameters of the structure, the large model can obtain semantic parsing capabilities for higher education evaluation tasks. (2-3) Construct a structured training data template that matches the higher education evaluation task. The structured training data template is used to uniformly describe evaluation instruction information, evaluation context information and corresponding target output information. (2-4) The adaptation structure is trained based on the structured training data template, and a large model adapted to the higher education evaluation task is generated after the training is completed.

4. The method for data fusion and intelligent evaluation of higher education based on a large model as described in claim 1, characterized in that... Step (3) of natural language assessment intent parsing and permission determination includes: (3-1) Obtain user identity-related information, which includes at least the user's affiliated unit identifier, user type, and user permission level; (3-2) Semantically analyze the natural language questions on higher education evaluation input by users, extract information on evaluation objects, evaluation indicators and data range related to higher education evaluation, and form a structured semantic description; (3-3) Identify the user's evaluation intent based on the structured semantic description, and map the natural language question into a structured data retrieval request containing evaluation indicators and data range information; (3-4) Determine the corresponding permission constraint parameters based on the user identity-related information. The permission constraint parameters are used to limit the scope of accessible universities and the granularity of data access. (3-5) Based on the structured data retrieval request and the permission constraint parameters, retrieve and return evaluation data that meets the permission constraint conditions from the higher education evaluation database, as input data for subsequent evaluation and generation processing.

5. The method for data fusion and intelligent evaluation of higher education based on a large model as described in claim 1, characterized in that... Step (4) of the evaluation data screening and fusion process includes: (4-1) Based on the structured data retrieval request, the candidate evaluation data is preprocessed, including data validity verification and permission consistency verification, so as to obtain a subset of evaluation data that meets the permission constraints. (4-2) Based on the user's evaluation intent, the subset of evaluation data is filtered for relevance, and evaluation data that matches the evaluation intent is retained; (4-3) Perform multi-dimensional correlation and fusion processing on the screened evaluation data to generate comprehensive evaluation data to characterize the status of higher education evaluation; (4-4) When the comprehensive evaluation data involves different time dimensions or different data sources, perform cross-time dimension alignment and cross-data source topic association processing on the comprehensive evaluation data; (4-5) Generate structured evaluation results based on the processed comprehensive evaluation data, which are used as input data for subsequent large model generation steps.

6. The method for data fusion and intelligent evaluation of higher education based on a large model as described in claim 1, characterized in that... Step (5) involves generating the evaluation results subject to structured data and permission constraints, including: (5-1) Obtain the structured evaluation result generated in step (4) and the permission constraint parameters corresponding to the structured evaluation result; (5-2) Input the structured evaluation results, the user's higher education evaluation questions and the permission constraint parameters into the large model generation module to limit the generation scope of the large model, so that it generates content only based on the structured evaluation results and under the permission constraint conditions; (5-3) Perform permission consistency verification on the evaluation results generated by the large model to determine whether the generated content conforms to the permission constraint parameters; (5-4) If the generated content does not meet the permission constraint parameters, the generated result is adjusted to make the output intelligent evaluation result of higher education meet the permission constraint conditions and then output.

7. The method for data fusion and intelligent evaluation of higher education based on a large model as described in claim 1, characterized in that... The feedback-driven model update in step (6) includes: (6-1) Collect user feedback information on the intelligent evaluation results of higher education, the feedback information including at least result correction information, usage behavior information or evaluation mark information, and store the feedback information in a structured manner; (6-2) Generate labeled data or training samples for evaluating task updates based on the feedback information; (6-3) Use the labeled data or training samples to update the large model or the data of higher education evaluation results to improve the accuracy and consistency of the intelligent evaluation results of higher education. (6-4) After the update is completed, save the updated model parameters or evaluation result data for use in the subsequent intelligent evaluation process of higher education.

8. A data fusion and intelligent evaluation system for higher education based on a large model, characterized in that, The method for implementing the large-model-based higher education data fusion and intelligent evaluation method according to any one of claims 1 to 7 includes: The Higher Education Evaluation Database Module is used to store and manage basic data, evaluation indicator data, and evaluation result data related to higher education evaluation. The large model evaluation task adaptation module is used to adapt the large model to evaluation tasks based on data in the field of higher education evaluation, so that the large model has the semantic parsing capability for higher education evaluation application tasks. The Natural Language Assessment Intent Parsing and Permission Determination module is used to perform semantic parsing on the natural language questions for higher education assessment input by users, generate a structured data retrieval request containing assessment indicators and data range information, and determine the corresponding permission constraint parameters. The evaluation data filtering and fusion module is used to obtain evaluation data from the higher education evaluation database module based on the structured data retrieval request and the permission constraint parameters, and to filter, associate and fuse the evaluation data to generate structured evaluation results. The generation module, constrained by structured data and permissions, is used to input the structured evaluation results, the user's higher education evaluation questions, and the permission constraint parameters into the large model, and generate intelligent evaluation results for higher education under the conditions of structured data and permission constraints. The feedback-driven model update module is used to collect user feedback on the intelligent evaluation results of higher education, and update the large model or higher education evaluation result data based on the feedback information.

Images