Method and device for quality of source assisted evaluation based on multi-modal large model

Abstract

This application provides a method and apparatus for auxiliary assessment of student quality based on a multimodal large model, belonging to the field of large model assessment technology. The method includes: acquiring student data of target personnel; using a multimodal large model to extract features from the student data in the following manner to obtain student quality features of the target personnel in multiple dimensions: converting structured data in the student data into corresponding student quality features according to preset mapping rules; extracting student quality features corresponding to unstructured text data in the student data using an attention mechanism based on preset prompts; and extracting student quality features corresponding to unstructured image data in the student data using a visual encoder. These multiple-dimensional student quality features are used to assist in assessing the student quality of the target personnel. This application can simultaneously process structured and unstructured data in the student data, extracting relevant features, facilitating the work of relevant personnel, and improving assessment efficiency.

Description

Technical Field

[0001] This application belongs to the field of large model evaluation technology, and more specifically, it relates to a method and device for auxiliary evaluation of student quality based on a multimodal large model. Background Technology

[0002] The current college student quality evaluation system has achieved information management to a certain extent. Most colleges and universities have established data collection platforms based on the enrollment management system, which can collect basic information such as candidates' college entrance examination scores, admission preferences, and places of origin. Based on the pre-set indicator weights, the quality of students is scored and summarized to form a structured evaluation report, which facilitates the enrollment work.

[0003] However, the current student quality process largely relies on structured data such as exam scores, and lacks automated data processing capabilities for unstructured materials submitted by students, such as recommendation letters, personal statements, descriptions of competition experiences, and portfolios. This still requires a lot of manual intervention and is inefficient.

[0004] Therefore, a method is needed to assist in the assessment of student quality that can integrate structured and unstructured data. Summary of the Invention

[0005] The purpose of this application is to provide a method and apparatus for auxiliary assessment of student quality based on a multimodal large model, which can simultaneously process structured and unstructured data in student data and extract relevant features from them, making it easier for relevant personnel to carry out their work and improving assessment efficiency.

[0006] A first aspect of this application provides a method for auxiliary assessment of student quality based on a multimodal large model, comprising: Obtain the student origin data of the target personnel; Using a multimodal large model, feature extraction is performed on the student source data in the following manner to obtain the student source quality characteristics of the target personnel in multiple dimensions, with each dimension corresponding to at least one student source quality characteristic: According to the preset mapping rules, the structured data in the student source data is converted into corresponding student source quality features; the attention mechanism of the multimodal large model is used to extract the student source quality features corresponding to the unstructured text data in the student source data based on the preset prompt words; the visual encoder of the multimodal large model is used to extract the student source quality features corresponding to the unstructured image data in the student source data. Among them, multiple dimensions of student quality characteristics are used to assist in the assessment of the student quality of target personnel; the prompt words contain target prompt words for setting target dimensions and target indicators for multimodal large models; the target indicators are the evaluation indicators corresponding to the target dimensions; the target indicators are used to construct the query vector in the attention mechanism; and unstructured text data are used to construct the key vector and value vector in the attention mechanism.

[0007] A second aspect of this application provides a student quality auxiliary assessment device based on a multimodal large model, comprising: The data acquisition module is used to acquire the student origin data of the target personnel; The feature extraction module is used to extract features from student data using a multimodal large model in the following manner, to obtain the student quality features of the target personnel in multiple dimensions, with each dimension corresponding to at least one student quality feature: According to the preset mapping rules, the structured data in the student source data is converted into corresponding student source quality features; the attention mechanism of the multimodal large model is used to extract the student source quality features corresponding to the unstructured text data in the student source data based on the preset prompt words; the visual encoder of the multimodal large model is used to extract the student source quality features corresponding to the unstructured image data in the student source data. Among them, multiple dimensions of student quality characteristics are used to assist in the assessment of the student quality of target personnel; the prompt words contain target prompt words for setting target dimensions and target indicators for multimodal large models; the target indicators are the evaluation indicators corresponding to the target dimensions; the target indicators are used to construct the query vector in the attention mechanism; and unstructured text data are used to construct the key vector and value vector in the attention mechanism.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above-described method for auxiliary assessment of student quality based on a multimodal large model.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for auxiliary assessment of student quality based on a multimodal large model.

[0010] The beneficial effects of the student quality auxiliary assessment method and device based on a multimodal large model provided in this application are as follows: This application's embodiments employ preset mapping rules to transform structured data, converting standardized data such as grades, preferences, and place of origin into calculable student quality characteristics. For unstructured text data, this embodiment uses prompts containing target dimensions and indicator settings to transform evaluation indicators into query vectors with an attention mechanism. Key-value vectors are constructed from the text to achieve targeted extraction of text semantics, replacing manual reading of recommendation letters, personal statements, competition experience descriptions, and other materials, thus solving the problem of unstructured text being difficult to process automatically. This embodiment also extracts features from unstructured image data through a visual encoder, enabling automated evaluation of visual materials such as portfolios and scanned certificates. This embodiment can simultaneously process both structured and unstructured data in student data and extract relevant features, facilitating work for relevant personnel and improving evaluation efficiency. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a method for auxiliary assessment of student quality based on a multimodal large model, provided in an embodiment of this application; Figure 2 A flowchart illustrating another method for auxiliary assessment of student quality based on a multimodal large model provided in an embodiment of this application; Figure 3 A structural block diagram of a student quality auxiliary assessment device based on a multimodal large model provided in an embodiment of this application; Figure 4 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0015] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a method for auxiliary assessment of student quality based on a multimodal large model, provided in an embodiment of this application. The method can be executed by an electronic device and may include: S101-S102.

[0016] S101: Obtain the student source data of the target personnel.

[0017] In this embodiment, the target personnel can be applicants. Student source data refers to the materials submitted by the target personnel during the admission application process and the materials collected by the admission management system, used to characterize their student source qualifications. This data can contain structured data, unstructured text data, and unstructured image data. In the current scenario, common structured data can include total Gaokao score, individual subject scores, Gaokao ranking, postgraduate entrance examination total score, province of origin, and order of application. Unstructured text data can be documents submitted by applicants without a fixed format and free semantic expression, such as personal statements, letters of recommendation, and descriptions of competition participation experience. Unstructured image data can include scanned copies of award certificates, photos of qualification certificates, and photos of competition entries.

[0018] In this embodiment, the acquired raw student data can also be preprocessed. For example, structured data can be processed by filling missing values ​​and detecting outliers, unstructured text data such as recommendation letters and personal statements can be preprocessed by sentence segmentation, noise reduction and stop word removal, and unstructured image data such as portfolio images and scanned certificates can be processed by image size normalization (uniformly scaled to 224×224 pixels) and color channel standardization (mean and variance normalization).

[0019] S102: Using a multimodal large model, feature extraction is performed on the student source data in the following ways to obtain the student source quality features of the target personnel in multiple dimensions: the structured data in the student source data is converted into corresponding student source quality features according to the preset mapping rules; the attention mechanism of the multimodal large model is used to extract the student source quality features corresponding to the unstructured text data in the student source data based on preset prompt words; the visual encoder of the multimodal large model is used to extract the student source quality features corresponding to the unstructured image data in the student source data.

[0020] In this embodiment, student quality characteristics refer to standardized numerical values ​​or labeled features extracted from student data that can quantify the qualifications of target personnel. Each dimension corresponds to at least one student quality characteristic. Multiple dimensions may include: academic foundation, student structure, application status, comprehensive quality, and development potential. Among them, the student quality characteristics corresponding to the academic foundation dimension may include the total score of the college entrance examination, the level of the undergraduate institution, the ranking of professional grades, and the level of awards in subject competitions, etc. The student quality characteristics corresponding to the student structure data dimension may include the province of origin, the level of the graduating high school, the admission batch, and the household registration attribute, etc. The student quality characteristics corresponding to the comprehensive quality dimension may include the level of participation in scientific research projects, the number of competition awards, and the quality of the completed portfolio, etc. These are not listed in this embodiment. Relevant personnel can set the types of dimensions, the number of dimensions, and the types and number of student quality characteristics corresponding to each dimension.

[0021] In this embodiment, since the structured student data consists of relatively numerical, fixed fields without ambiguity, and its information is directly reflected in the numerical values ​​and fields, semantic understanding and reasoning are unnecessary. Therefore, the structured data in the student data can be directly converted into corresponding student quality characteristics based on preset mapping rules. For example, the total score in the college entrance examination in the structured student data can be directly mapped to the total score of the college entrance examination cultural courses in the academic foundation dimension. The mapping rules can be set based on experience.

[0022] In this embodiment, since unstructured text is freely expressed and has no fixed format, it cannot be directly mapped using rules. Therefore, this embodiment utilizes the attention mechanism of a multimodal large model to extract student quality features corresponding to unstructured text data in the student source data based on preset prompt words. The preset prompt words contain target prompt words used to set target dimensions and target indicators for the multimodal large model; the target indicators are the evaluation indicators corresponding to the target dimensions; the target indicators are used to construct the query vector in the attention mechanism; and the unstructured text data is used to construct the key vector and value vector in the attention mechanism.

[0023] For example, the target prompt might be: "For the 'Comprehensive Competency' dimension, please focus on the following indicators: research ability, innovative practical ability, and teamwork ability." Here, "[Comprehensive Competency]" is used to set the target dimension for the multimodal large model, while "research ability, innovative practical ability, and teamwork ability" are used to set the target indicators for the multimodal large model.

[0024] In this embodiment, after the target prompt words are input into the multimodal large model, they are first segmented and mapped to the index by the multimodal large model's word segmenter, generating a fixed-length index token sequence. Then, through the model's pre-trained word embedding layer, the index token sequence is mapped to the model's high-dimensional semantic space to obtain index word embedding vectors. Finally, mean aggregation / weighted aggregation is performed on the index word embedding vectors to generate a query vector Query with a single semantic representation. The query vector serves as the retrieval benchmark for the attention mechanism and is used to calculate the similarity with the text semantic vectors of unstructured text data in the source data, determining the degree of association between the text fragment and the evaluation index. After inputting unstructured text data into the multimodal large model, the unstructured text can first be segmented, denoised, stop word removed, and truncated (if the preprocessing process in the previous embodiment has been performed, it does not need to be performed again), generating a text sequence that conforms to the model input specifications. Then, the standardized text is input into the word segmenter to generate a token sequence. Finally, through the independent projection matrix of the model's self-attention module, the basic text representation vectors are linearly transformed to generate key vectors Key and value vectors Value. The key vector serves as a semantic matching unit, calculating similarity with the query vector Query to generate attention weights, which identify the correlation strength between each position in the text and the target indicator; the value vector serves as a feature output unit, being weighted and aggregated according to the attention weights, ultimately outputting textual semantic features related to the target indicator.

[0025] In this embodiment, in addition to the target prompts mentioned in the previous embodiment, the preset prompts also include role-setting prompts for defining the role of the multimodal large model, and format-setting prompts for defining the output format of the multimodal large model. For example, the role-setting prompt could be "You are a professional graduate admissions review expert," and the format-setting prompt could be "Required to output in structured JSON format, with output fields including: {dimensional name, evidence text fragment, confidence score (0~1 floating-point number), feature label list}." In this embodiment, the feature label list refers to the student quality features extracted by the model from the evidence text fragment. For example, the student quality features output for the comprehensive quality dimension could be quantifiable feature labels such as the level of participation in scientific research projects, the number of competition awards, the quality of the portfolio, leading national-level projects, publishing CCF-A papers, or winning international competition awards.

[0026] In this embodiment, since unstructured images are pixel matrices and have no textual structure data, they cannot be processed by rule mapping or text attention mechanisms. Therefore, in this embodiment, a multimodal large model visual encoder is used to extract the student quality features corresponding to the unstructured image data in the student source data.

[0027] As can be seen from the above, this embodiment uses preset mapping rules to transform structured data, converting standardized data such as grades, preferences, and place of origin into calculable student quality characteristics. For unstructured text data, this embodiment uses prompts containing target dimensions and indicator settings to transform evaluation indicators into query vectors with an attention mechanism. Key-value vectors are constructed from the text to achieve targeted extraction of text semantics, replacing manual reading of recommendation letters, personal statements, competition experience descriptions, etc., thus solving the problem of unstructured text not being processed automatically. This embodiment also extracts features from unstructured image data through a visual encoder, enabling automated evaluation of visual materials such as portfolios and scanned certificates. This embodiment can simultaneously process both structured and unstructured data in student data and extract relevant features, facilitating work by relevant personnel and improving evaluation efficiency.

[0028] In one embodiment of this application, the multimodal large model extracts the student quality features corresponding to each image in the unstructured image data of the student source data in the following manner: The image is divided into multiple image patches based on a visual encoder; local semantic feature vectors of each image patch are extracted, and the local semantic feature vectors of each image patch are aggregated to obtain a global semantic feature vector representing the overall semantics of the image; the global semantic feature vector is mapped to the text feature space based on a pre-constructed linear transformation matrix to obtain the target semantic feature vector; the linear transformation matrix is ​​used to achieve dimensional alignment between the image feature space and the text feature space; the target semantic feature vector is classified into the corresponding student quality features based on a pre-trained classification head.

[0029] In this embodiment, the image is divided into non-overlapping, ordered patch sequences according to a fixed size (e.g., 16×16). The visual encoder independently performs multi-layer self-attention operations and feedforward network transformations on each image patch, outputting a local semantic feature vector for the corresponding image patch. This vector is used to characterize the source quality information within the image patch (e.g., certificate text, award logos, portfolio creation elements, certificate issuing institution logos, etc.). Then, global average pooling or class-of-speech (CLS) token aggregation is used to weightedly fuse the local semantic feature vectors of all image patches, eliminating fragmentation bias in local features and generating a global semantic feature vector. The global semantic feature vector is a single, high-dimensional, dense vector used to carry the full visual semantics of the image.

[0030] In this embodiment, since the current global semantic feature vector is extracted from the image, it resides in the visual feature space, which differs from the text feature space of structured and unstructured text data, resulting in inconsistent spatial dimensions. Therefore, in this embodiment, a linear transformation matrix W is pre-constructed, and the global semantic feature can be represented as V, with the target semantic feature vector V' = WV. The linear transformation matrix W can be obtained through supervised training on a multimodal annotation dataset for the college admission scenario. The training optimization objective is cross-modal semantic alignment loss, ensuring that the projected visual features and text features describing the same evaluation content maintain proximity in the vector space. The loss function in this process can be the cross-entropy loss function.

[0031] In this embodiment, the pre-trained classification head can be trained on a preset training set containing multiple samples. Each sample is pre-labeled with an image semantic feature vector, an evaluation dimension, and corresponding labels for student quality features. The training objective is to minimize the regression loss / classification loss between the student quality features output by the classification head and the manually labeled standard student quality features.

[0032] As can be seen from the above, segmenting an image into image blocks and extracting local semantic features through a visual encoder, and then aggregating them into a global semantic feature vector, can eliminate the fragmentation bias of local features and fully preserve the source quality information carried by the image. Secondly, this embodiment maps the global semantic features to the text feature space through a linear transformation matrix, achieving dimensional alignment between the image and text feature spaces, and solving the problem of inconsistent computation of heterogeneous features.

[0033] In one embodiment of this application, the student quality auxiliary assessment method further includes: a process of determining the target confidence level corresponding to each student quality characteristic; the target confidence level is used to characterize the reliability of its corresponding student quality characteristic, and the target confidence level is used together with its corresponding student quality characteristic to assist in the student quality assessment of the target personnel; For each student quality feature, when that feature is extracted from unstructured text data, the process of determining the target confidence level corresponding to that feature includes: The target label description and target evidence text corresponding to the student quality feature are determined, and the semantic similarity between the label description and the target evidence text is calculated. If the semantic similarity is greater than a preset similarity threshold, the initial confidence level corresponding to the student quality feature is determined as the target confidence level of the student quality feature. If the semantic similarity is less than or equal to the preset similarity threshold, the initial confidence level corresponding to the student quality feature is multiplied by the first penalty coefficient to obtain the target confidence level of the student quality feature. The target evidence text and initial confidence level corresponding to the student quality feature are obtained in the following process: using the attention mechanism of a multimodal large model, the student quality feature corresponding to the unstructured text data in the student data is extracted based on preset prompt words.

[0034] In this embodiment, when the student quality feature is extracted from unstructured text data, referring to the aforementioned embodiment, the preset prompt words also include format setting prompt words, which limit the output fields to include: {dimension name, evidence text fragment, confidence score (0~1 floating-point number), feature label list}, where "confidence score" is the initial confidence score in this embodiment, and the target evidence text refers to the evidence text in the evidence text fragment that corresponds to the student quality feature. The target label description refers to the preset standardized expression corresponding to the student quality feature. For example, the target label description corresponding to the research ability indicator can be "possessing independent research execution ability, participating in / leading research projects and producing effective results". The essence of calculating the semantic similarity between the target evidence text and the target label description is also to calculate whether the student quality feature matches the current target label description and the degree of matching. If the semantic similarity is greater than the preset similarity threshold, the evidence text is determined to be highly matched with the indicator description, and the feature extraction is effective. In this case, the initial confidence level can be directly determined as the target confidence level. If the semantic similarity is less than or equal to the preset similarity threshold, the evidence text is determined to be insufficiently matched with the indicator description, and the feature reliability is insufficient. In this case, the product of the initial confidence level and the first penalty coefficient can be used as the target confidence level. In this embodiment, the first penalty coefficient can be 0.6, and the preset similarity threshold can be 0.75.

[0035] In one embodiment of this application, for each student quality feature, when the student quality feature is extracted based on unstructured image data, the process of determining the target confidence level corresponding to the student quality feature includes: The image quality score of the image corresponding to the student quality feature is determined. If the image quality score is greater than a preset quality score threshold, the initial confidence level corresponding to the student quality feature is determined as the target confidence level of the student quality feature. If the image quality score is less than or equal to the preset quality score threshold, the initial confidence level corresponding to the student quality feature is multiplied by a second penalty coefficient to obtain the target confidence level of the student quality feature. The initial confidence level corresponding to the student quality feature is obtained in the following process: the student quality features corresponding to unstructured image data in the student data are extracted using a multimodal large model visual encoder.

[0036] Similar to the confidence level of unstructured text data in the previous embodiments, this embodiment also requires determining the target confidence level of unstructured image data. The difference lies in the determination process of the initial confidence level. When the image feature is extracted based on unstructured image data, it can be directly determined by the process in the previous embodiments of classifying the target semantic feature vector to the corresponding student quality feature based on a pre-trained classification head. In this process, the softmax probability value output by the classification head can be directly used as the initial confidence level. Another difference lies in the adjustment process of the initial confidence level. The determination of the second penalty coefficient can be based on the image quality score of the image corresponding to the student quality feature. The BRISQUE algorithm can be used to calculate the image quality score without reference.

[0037] In this embodiment, if the image quality score is greater than a preset quality score threshold, the image is determined to meet the feature extraction requirements in terms of clarity and integrity. In this case, the initial confidence level corresponding to the student quality feature can be determined as the target confidence level for that feature. If the image quality score is less than or equal to the preset quality score threshold, the image is determined to have quality defects such as blurriness or distortion, and the features extracted by the visual encoder are prone to introducing bias or noise. Therefore, the initial confidence level can be multiplied by a second penalty coefficient to obtain a reduced target confidence level, thus reducing the weight contribution of this feature in subsequent scoring calculations. In this embodiment, the preset quality score threshold can be 40 points, and the second penalty coefficient can be 0.5.

[0038] As can be seen from the above, this embodiment constructs confidence calibration mechanisms for text and image features respectively. For text features, semantic similarity verification is used; if the evidence text matches the indicator description, the initial confidence level is retained; otherwise, it is reduced through a penalty coefficient to filter out semantically inconsistent invalid features. For image features, image quality is evaluated; if the image is satisfactory, the initial confidence level is maintained; if the image is of poor quality, the confidence level is reduced, suppressing feature bias caused by quality defects such as blurriness and distortion. The target confidence level is bound to student quality features in the evaluation, achieving adaptive calculation of high-reliability features with high weights and low-reliability features with low weights. This avoids invalid information interfering with the evaluation results, improves the reliability of automated feature verification, and enhances evaluation accuracy.

[0039] In one embodiment of this application, the target confidence level and its corresponding student quality characteristics are used to assist in the student quality assessment of the target personnel in the following manner: For each student quality feature of each dimension, the student quality feature is converted into a corresponding first score based on the target confidence level of the student quality feature and the preset scoring rules; for each dimension, the second score corresponding to each student quality feature of the dimension is determined based on the first score; the second scores corresponding to each dimension are weighted and calculated to obtain the target score; the target score is used to assist in the student quality assessment of the target personnel.

[0040] In this embodiment, the first score refers to the feature-level standardized score obtained by weighting and correcting the confidence level of the student quality characteristic target, and then converting it according to preset rules. The second score refers to the dimension-level comprehensive score obtained by aggregating and calculating the first scores corresponding to all student quality characteristics under a certain dimension, representing the overall quantitative level of student quality under the evaluation dimension. In this embodiment, the first scores of each student quality characteristic under the dimension can be weighted to obtain the second score corresponding to the dimension. The weights in the weighting calculation can be the same. The preset scoring rules can be set by the user. In this embodiment, if the student quality characteristic is extracted based on structured data, the confidence level is 1 by default. For example, if the total score of the college entrance examination is 500, the corresponding score is 70, and the confidence level is 1, then the first score is 70.

[0041] In this embodiment, the target score is obtained by weighting the second scores of each dimension. The target score can be used to assist relevant personnel in assessing the quality of the target applicants. The weights in the weighting calculation can be the same or determined in the following way: The process involves obtaining the student origin attributes of the target personnel and determining the target student origin attribute features used to characterize the training objectives of the target personnel based on these attributes; retrieving at least one standard student origin attribute feature that matches the target student origin attribute feature from a pre-defined weight rule base; the weight rule base contains multiple standard student origin attribute features and the weight allocation ratios of multiple dimensions corresponding to each standard student origin attribute feature; and determining the weighted calculation weight of the second score for each dimension based on the weight allocation ratio corresponding to the at least one matching standard student origin attribute feature.

[0042] In this embodiment, the student source attribute can be the discipline type and the enrollment batch. The target student source attribute feature can be a triple of [discipline type, enrollment batch, training objective]. For the student source attribute (discipline type and enrollment batch) of the target personnel, there is a preset training objective based on the current enrollment scenario. For example, for the discipline type of engineering and the enrollment batch of master’s unified examination, the corresponding training objective may be scientific research application type.

[0043] In this embodiment, considering the different combinations of disciplines, enrollment batches, and training objectives, the emphasis of the corresponding evaluation dimensions should also be different. Therefore, at least one standard student attribute feature that matches the target student attribute feature can be retrieved from the preset weight rule base. For example, at least one standard student attribute feature that matches the target student attribute feature can be determined by converting the triples and the standard student attribute features in the weight rule base into vectors and calculating the cosine similarity between the vectors. In this embodiment, the standard student source attributes and the weight allocation ratios of multiple dimensions corresponding to each standard student source attribute can be, for example, as follows: Standard Triad 1: (Engineering, Master's Entrance Examination, Research and Application-oriented), with corresponding weight allocation ratios of: Academic Foundation 0.30, Student Source Structure 0.10, Application Status 0.05, Comprehensive Quality 0.35, Development Potential 0.20; Standard Triad 2: (Science, Master's Entrance Examination, Research and Innovation-oriented), with corresponding weight allocation ratios of: Academic Foundation 0.25, Student Source Structure 0.08, Application Status 0.07, Comprehensive Quality 0.30, Development Potential 0.30; Standard Triad 3: (Humanities, Master's Entrance Examination, Comprehensive Practice-oriented), with corresponding weight allocation ratios of: Academic Foundation 0.28, Student Source Structure 0.15, Application Status 0.12, Comprehensive Quality 0.25, Development Potential 0.20.

[0044] In this embodiment, if the preset weight rule library contains standard student attribute features that are completely consistent with the target student attribute features, then these standard student attribute features can be determined as the standard student attribute features that match the target student attribute features, and the weight allocation ratio corresponding to these standard student attribute features can be directly determined as the weighted calculation weight of the second score for each dimension. If the preset weight rule library does not contain standard student attribute features that are completely consistent with the target student attribute features, then the Top-K standard student attribute features can be determined as the standard student attribute features that match the target student attribute features, and the cosine similarity in the aforementioned embodiment can be used as the weighted calculation weight of the weight allocation ratio of the multiple dimensions corresponding to the Top-K standard student attribute features to obtain the weighted calculation weight of the second score for each dimension.

[0045] In one embodiment of this application, a weighted calculation is performed on the second scores corresponding to each dimension to obtain the target score, including: The second scores for each dimension are weighted to obtain an initial score. The initial score is then corrected based on a preset correction formula to obtain a target score. This correction formula includes a temperature scaling parameter and a bias term for score correction. After every N calculations of the target score for different individuals, the temperature scaling parameter and bias term are updated as follows: Obtain N historical target scores and manual review scores, and use them as a calibration set; divide the calibration set into a training set and a validation set; calibrate the temperature scaling parameters and bias terms in the correction formula based on the training set and the Prato transformation; determine the first calibration error and the second calibration error of the calibrated correction formula based on the validation set; if the second calibration error is greater than the first calibration error, maintain the temperature scaling parameters and bias terms before calibration; if the second calibration error is less than or equal to the first calibration error, update the temperature scaling parameters and bias terms based on the calibrated temperature scaling parameters and bias terms.

[0046] In this embodiment, the preset correction formula is: score_cal = Sigmoid(logit / τ + b), where score_cal represents the target score, Sigmoid is the Sigmoid activation function, logit is the initial score, τ is the temperature scaling parameter, and b is the bias term. The temperature scaling parameter is used to control the sharpening or smoothing of the score distribution in the correction formula, and the bias term is used to adjust the overall offset of the scores in the correction formula. Platt Scaling is a probabilistic calibration method based on logistic regression. The calibration set refers to the set of paired samples of historical target scores and corresponding manual review scores accumulated after N target score calculations, where N can be 50. The first calibration error and the second calibration error can be obtained by calculating the expected calibration error (ECE).

[0047] In this embodiment, after every N calculations of target scores for different target personnel, the update process for τ and b is triggered to avoid the degradation of calibration effect due to long-term parameter fixation. For example, the calibration set can be randomly split according to a preset ratio (e.g., 8:2), with 80% as the training set for fitting new parameters and 20% as the validation set for verifying the calibration effect. The historical target scores in the training set are used as the model input features, and the manual review scores are used as labels to construct a logistic regression model. The model parameters are fitted by maximum likelihood estimation to obtain the new temperature scaling parameter τ' and the new bias term b'.

[0048] If the second calibration error is greater than the first calibration error, it means that the calibration effect is worse than the original parameters. Keep the original τ and b unchanged and abandon this parameter update. If the second calibration error is less than or equal to the first calibration error, it means that the calibration effect is better than or equal to the original parameters. Update τ to τ' and b to b' and apply the new parameters for subsequent target score calculation.

[0049] As can be seen from the above, this embodiment combines the target confidence level with preset scoring rules to convert student quality characteristics into a first score, incorporating feature reliability into the quantitative calculation to avoid low-reliability features interfering with the score; then, it aggregates the first scores within a dimension to generate a second score, fully representing the single-dimensional student quality level; finally, based on the student attribute matching weight rule library, it dynamically determines the dimension weighting weight, adapting to different training objectives such as research application-oriented and research innovation-oriented, thus improving the accuracy of the auxiliary assessment. Furthermore, this embodiment also continuously optimizes and corrects the parameters in the formula through a dynamic calibration mechanism during the auxiliary assessment process, maintaining long-term assessment stability.

[0050] In one embodiment of this application, see Figure 2 The student quality auxiliary assessment method based on a multimodal large model can also be implemented by sequentially performing data preprocessing, multimodal feature extraction and fusion, dynamic weight scoring calibration, and finally generating an interpretable evaluation report. The core functions of each module are as follows: Module 1 defines a multi-dimensional student source evaluation index system and an evaluation benchmark. Module 1 defines five major dimension feature slots (i.e., student source quality features in the aforementioned embodiments), clarifies the target mapping space, and provides a unified evaluation dimension anchor point for subsequent data processing and feature extraction.

[0051] Module 2 performs multimodal data acquisition and preprocessing. For three types of heterogeneous source data, standardized preprocessing is performed respectively, and the output format can be directly processed by the model: structured numerical data: converted into standard numerical vectors through Z-score normalization; unstructured text: converted into token sequences after sentence segmentation, noise reduction and segmentation; image and chart data: uniformly scaled into image tensors of size 224×224.

[0052] Module 3 is the core layer for feature extraction and fusion representation of multimodal large models. It integrates the feature extraction logic of three modalities, and the key technical points are marked as "three-layer Prompt, VIT visual encoding, BRISQUE confidence, and weighted fusion". It consists of three sub-modules: LLM Text Understanding: Unstructured text is processed using a three-layer Prompt template (corresponding to the target prompt, role setting prompt, and format setting prompt in the aforementioned embodiments), and the text features and corresponding confidence scores are output in JSON format. ViT Visual Encoding: Patch encoding and global pooling are performed on the image, and visual features and confidence scores are output through a cross-modal projection layer (linear transformation matrix) W. Confidence-Weighted Fusion: Multimodal features are fused to obtain the F_fused vector through cosine similarity verification (text) and BRISQUE image quality verification, while simultaneously achieving alignment with the feature space of Module 1. In this embodiment, the F_fused vector refers to the five dimensions and the corresponding second score for each of the five dimensions.

[0053] Module four is the scoring layer for dynamic weighted scoring and online parameter calibration, consisting of two parallel scoring and calibration paths: FAISS dynamic weight retrieval: based on the "subject-batch-training objective" triplet index, it retrieves the weight configuration parameter library and obtains the dynamic weight through Top-K weighted averaging; PlattScaling online calibration: using N=50 samples as the trigger condition, it updates the temperature scaling parameter τ based on the manually reviewed sample set to achieve score distribution calibration. This module also receives calibration write-back feedback from Module five, forming a closed-loop optimization of weights and parameters.

[0054] Module 5 is the output layer for automatically generating interpretable evaluation reports. It can generate evaluation reports based on DAG inference path records, strengths and weaknesses annotations, and supplementary suggestions, containing three key types of information: DAG path records: enabling full traceability of the calculation path; strengths and weaknesses quantile annotations: automatically annotating the 80th / 20th percentile of student quality levels; and missing material suggestions: triggering supplementary prompts for features with low confidence levels, prompting the target personnel to supplement relevant material information. The inference path can be displayed as: "High score in comprehensive quality dimension, mainly based on: the supervisor in the recommendation letter clearly describes the ability to independently complete research projects, and the personal statement mentions 3 competition awards, which are highly relevant to this project."

[0055] In this embodiment, calibration write-back refers to writing the manually reviewed evaluation data output from Module 5 (automatic generation of interpretable evaluation reports) back into Module 4 (dynamic weight scoring and online parameter calibration), providing data support for iterative optimization of scoring parameters and weight configurations. For example, after Module 5 generates the evaluation report, admissions staff will manually review the evaluation results of some samples, forming paired samples of the system's target score and the manually reviewed score. At the same time, the DAG path records and quantile annotation data from Module 5 will also be included in the write-back scope. These data together constitute the current period's manually reviewed sample set. Some of the reviewed data will simultaneously optimize the weight allocation ratio corresponding to the standard student attribute characteristics, making the dynamic weights more in line with the evaluation emphasis of the actual admissions scenario.

[0056] Corresponding to the multimodal large model-based student quality auxiliary assessment method in the above embodiments, Figure 3 This is a structural block diagram of a student quality auxiliary assessment device based on a multimodal large model, provided as an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 3 The student quality auxiliary assessment device 20 based on a multimodal large model includes: a data acquisition module 21 and a feature extraction module 22.

[0057] Among them, the data acquisition module 21 is used to acquire the student source data of the target personnel; Feature extraction module 22 is used to extract features from student data using a multimodal large model in the following manner to obtain student quality features of the target personnel in multiple dimensions, with each dimension corresponding to at least one student quality feature: According to the preset mapping rules, the structured data in the student source data is converted into corresponding student source quality features; the attention mechanism of the multimodal large model is used to extract the student source quality features corresponding to the unstructured text data in the student source data based on the preset prompt words; the visual encoder of the multimodal large model is used to extract the student source quality features corresponding to the unstructured image data in the student source data. Among them, multiple dimensions of student quality characteristics are used to assist in the assessment of the student quality of target personnel; the prompt words contain target prompt words for setting target dimensions and target indicators for multimodal large models; the target indicators are the evaluation indicators corresponding to the target dimensions; the target indicators are used to construct the query vector in the attention mechanism; and unstructured text data are used to construct the key vector and value vector in the attention mechanism.

[0058] In one embodiment of this application, the feature extraction module 22 is specifically used to divide the image into multiple image blocks based on the visual encoder; extract the local semantic feature vectors of each image block, and aggregate the local semantic feature vectors of each image block to obtain a global semantic feature vector that represents the overall semantics of the image. The global semantic feature vector is mapped to the text feature space based on a pre-constructed linear transformation matrix to obtain the target semantic feature vector; the linear transformation matrix is ​​used to achieve dimensional alignment between the image feature space and the text feature space. Based on a pre-trained classification head, the target semantic feature vector is classified into the corresponding student quality features.

[0059] In one embodiment of this application, the student quality auxiliary assessment device 20 based on a multimodal large model further includes a confidence determination module for determining the target confidence level corresponding to each student quality feature. The target confidence level is used to characterize the reliability of its corresponding student quality feature, and the target confidence level is used together with its corresponding student quality feature to assist in the student quality assessment of the target personnel. For each student quality feature, when the student quality feature is extracted based on unstructured text data, the confidence determination module is specifically used to determine the target label description and target evidence text corresponding to the student quality feature, and calculate the semantic similarity between the label description and the target evidence text. If the semantic similarity is greater than the preset similarity threshold, the initial confidence level corresponding to the student quality feature is determined as the target confidence level of the student quality feature. If the semantic similarity is less than or equal to the preset similarity threshold, the initial confidence level corresponding to the student quality feature is multiplied by the first penalty coefficient to obtain the target confidence level of the student quality feature. The target evidence text and initial confidence level corresponding to the student quality feature are obtained in the following process: using the attention mechanism of a multimodal large model, the student quality feature corresponding to the unstructured text data in the student data is extracted based on preset prompt words.

[0060] In one embodiment of this application, for each student quality feature, when the student quality feature is extracted based on unstructured image data, the confidence determination module is specifically used to determine the image quality score of the image corresponding to the student quality feature. If the image quality score is greater than the preset quality score threshold, the initial confidence level corresponding to the student quality feature is determined as the target confidence level of the student quality feature. If the image quality score is less than or equal to the preset quality score threshold, the initial confidence level corresponding to the student quality feature is multiplied by the second penalty coefficient to obtain the target confidence level of the student quality feature. The initial confidence level corresponding to the student quality feature is obtained in the following process: using a multimodal large model visual encoder to extract the student quality features corresponding to unstructured image data in the student data.

[0061] In one embodiment of this application, the student quality auxiliary assessment device 20 based on a multimodal large model further includes: an auxiliary assessment module, used to convert each student quality feature of each dimension into a corresponding first score based on the target confidence level of the student quality feature and a preset scoring rule; For each dimension, the second score corresponding to that dimension is determined based on the first score corresponding to each student quality characteristic of that dimension; The target score is obtained by weighting the second scores corresponding to each dimension; the target score is used to assist in the assessment of the quality of students for the target personnel.

[0062] In one embodiment of this application, the auxiliary evaluation module is specifically used to obtain the student origin attributes of the target personnel, and determine the target student origin attribute characteristics used to characterize the training objectives of the target personnel based on the student origin attributes. Retrieve at least one standard student attribute feature that matches the target student attribute feature from the preset weight rule base; the weight rule base contains multiple standard student attribute features and the weight allocation ratio of multiple dimensions corresponding to each standard student attribute feature; The weighted calculation weight of the second score for each dimension is determined based on the weight allocation ratio corresponding to at least one matching standard student attribute feature.

[0063] In one embodiment of this application, the auxiliary evaluation module is specifically used to perform weighted calculations on the second scores of each dimension to obtain an initial score; The initial score is corrected based on a preset correction formula to obtain the target score. The correction formula contains temperature scaling parameters and bias terms used for score correction. Specifically, after every N calculations of the target scores for different individuals, the temperature scaling parameter and bias term are updated as follows: Obtain N historical target scores and manual review scores, and use them as a calibration set; divide the calibration set into a training set and a validation set; The temperature scaling parameter and bias term in the correction formula are calibrated based on the training set and the Prato transform; the first calibration error and the second calibration error of the calibrated correction formula are determined based on the validation set. If the second calibration error is greater than the first calibration error, the temperature scaling parameters and bias terms before calibration are maintained; if the second calibration error is less than or equal to the first calibration error, the temperature scaling parameters and bias terms after calibration are updated.

[0064] See Figure 4 , Figure 4 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 4 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of each module / unit in the above-described device embodiments, for example... Figure 3 The functions of the data acquisition module 21 and the feature extraction module 22 shown are illustrated.

[0065] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0066] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0067] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0068] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation method described in the auxiliary assessment method for student quality based on a multimodal large model provided in the embodiments of this application, or they can execute the implementation method of the electronic device described in the embodiments of this application, which will not be repeated here.

[0069] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0070] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., provided on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0075] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0077] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered 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 auxiliary assessment of student quality based on a multimodal large model, characterized in that, include: Obtain the student origin data of the target personnel; Using a multimodal large model, feature extraction is performed on the student source data in the following manner to obtain the student source quality features of the target personnel in multiple dimensions, with each dimension corresponding to at least one student source quality feature: The structured data in the student source data is converted into corresponding student source quality features according to the preset mapping rules; the student source quality features corresponding to the unstructured text data in the student source data are extracted based on preset prompt words using the attention mechanism of the multimodal large model; and the student source quality features corresponding to the unstructured image data in the student source data are extracted using the visual encoder of the multimodal large model. The multiple dimensions of student quality features are used to assist in assessing the student quality of target personnel; the prompt words contain target prompt words for setting target dimensions and target indicators for the multimodal large model; the target indicators are the evaluation indicators corresponding to the target dimensions; the target indicators are used to construct the query vector in the attention mechanism; and the unstructured text data are used to construct the key vector and value vector in the attention mechanism.

2. The method for auxiliary assessment of student quality based on a multimodal large model as described in claim 1, characterized in that, The multimodal large model extracts the student quality features corresponding to each image in the unstructured image data of the student data in the following manner: The image is divided into multiple image blocks based on the visual encoder; the local semantic feature vectors of each image block are extracted, and the local semantic feature vectors of each image block are aggregated to obtain a global semantic feature vector that represents the overall semantics of the image. The global semantic feature vector is mapped to the text feature space based on a pre-constructed linear transformation matrix to obtain the target semantic feature vector; the linear transformation matrix is ​​used to achieve dimensional alignment between the image feature space and the text feature space. The target semantic feature vector is classified into the corresponding student quality features based on a pre-trained classification head.

3. The method for auxiliary assessment of student quality based on a multimodal large model as described in claim 1, characterized in that, Also includes: The process of determining the target confidence level corresponding to each student quality characteristic; the target confidence level is used to characterize the reliability of its corresponding student quality characteristic, and the target confidence level, together with its corresponding student quality characteristic, is used to assist in the student quality assessment of the target personnel. For each student quality feature, when that feature is extracted from unstructured text data, the process of determining the target confidence level corresponding to that feature includes: Determine the target label description and target evidence text corresponding to the student quality characteristics, and calculate the semantic similarity between the label description and the target evidence text; If the semantic similarity is greater than the preset similarity threshold, then the initial confidence level corresponding to the student quality feature is determined as the target confidence level of the student quality feature; If the semantic similarity is less than or equal to the preset similarity threshold, the initial confidence level corresponding to the student quality feature is multiplied by the first penalty coefficient to obtain the target confidence level of the student quality feature. The target evidence text and initial confidence level corresponding to the student quality feature are obtained in the following process: using the attention mechanism of a multimodal large model, the student quality feature corresponding to the unstructured text data in the student data is extracted based on preset prompt words.

4. The method for auxiliary assessment of student quality based on a multimodal large model as described in claim 3, characterized in that, For each student quality feature, when that feature is extracted from unstructured image data, the process of determining the target confidence level corresponding to that feature includes: Determine the image quality score of the image corresponding to the student quality characteristic; If the image quality score is greater than the preset quality score threshold, then the initial confidence level corresponding to the student quality feature is determined as the target confidence level of the student quality feature. If the image quality score is less than or equal to the preset quality score threshold, the initial confidence level corresponding to the student quality feature is multiplied by the second penalty coefficient to obtain the target confidence level of the student quality feature. The initial confidence level corresponding to the student quality feature is obtained in the following process: using the visual encoder of the multimodal large model to extract the student quality feature corresponding to the unstructured image data in the student data.

5. The method for auxiliary assessment of student quality based on a multimodal large model as described in claim 3 or 4, characterized in that, The target confidence level and its corresponding student quality characteristics are used to assist in the student quality assessment of the target personnel in the following ways: For each student quality feature in each dimension, the student quality feature is converted into a corresponding first score based on the target confidence level of that student quality feature and the preset scoring rules. For each dimension, the second score corresponding to that dimension is determined based on the first score corresponding to each student quality characteristic of that dimension; The target score is obtained by weighting the second scores corresponding to each dimension; the target score is used to assist in the assessment of the quality of students for the target personnel.

6. The method for auxiliary assessment of student quality based on a multimodal large model as described in claim 5, characterized in that, The process of determining the weights for the second score corresponding to each dimension includes: Obtain the student origin attributes of the target personnel, and determine the target student origin attribute characteristics used to characterize the training objectives of the target personnel based on the student origin attributes; Retrieve at least one standard student attribute feature that matches the target student attribute feature from a preset weight rule base; the weight rule base contains multiple standard student attribute features and the weight allocation ratio of the multiple dimensions corresponding to each standard student attribute feature; The weighted calculation weight of the second score for each dimension is determined based on the weight allocation ratio corresponding to at least one matching standard student attribute feature.

7. The method for auxiliary assessment of student quality based on a multimodal large model as described in claim 5, characterized in that, The weighted calculation of the second scores corresponding to each dimension to obtain the target score includes: The initial score is obtained by weighting the second scores of each dimension. The initial score is corrected based on a preset correction formula to obtain the target score. The correction formula contains a temperature scaling parameter and a bias term for score correction. Specifically, after every N calculations of the target scores for different individuals, the temperature scaling parameter and the bias term are updated in the following manner: Obtain the historical target scores and manual review scores from the N times, and use them as a calibration set; divide the calibration set into a training set and a validation set; The temperature scaling parameter and the bias term in the correction formula are calibrated based on the training set and the Prato transform; the first calibration error and the second calibration error of the calibrated correction formula are determined based on the validation set. If the second calibration error is greater than the first calibration error, the temperature scaling parameters and the bias term are maintained before calibration; if the second calibration error is less than or equal to the first calibration error, the temperature scaling parameters and the bias term are updated based on the calibrated temperature scaling parameters and the bias term.

8. A student quality auxiliary assessment device based on a multimodal large model, characterized in that, include: The data acquisition module is used to acquire the student origin data of the target personnel; The feature extraction module is used to extract features from the student source data using a multimodal large model in the following manner, to obtain the student source quality features of the target personnel in multiple dimensions, with each dimension corresponding to at least one student source quality feature: The structured data in the student source data is converted into corresponding student source quality features according to the preset mapping rules; the student source quality features corresponding to the unstructured text data in the student source data are extracted based on preset prompt words using the attention mechanism of the multimodal large model; and the student source quality features corresponding to the unstructured image data in the student source data are extracted using the visual encoder of the multimodal large model. The multiple dimensions of student quality features are used to assist in assessing the student quality of target personnel; the prompt words contain target prompt words for setting target dimensions and target indicators for the multimodal large model; the target indicators are the evaluation indicators corresponding to the target dimensions; the target indicators are used to construct the query vector in the attention mechanism; and the unstructured text data are used to construct the key vector and value vector in the attention mechanism.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 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 steps of the method as described in any one of claims 1 to 7.

Images