A method for predicting students' early performance based on sentiment large language model

Abstract

The present invention relates to the field of artificial intelligence technology, and discloses a method for predicting students' early performance based on a large emotional language model in a smart education scenario, including obtaining structured data and unstructured data related to students' academic performance and performing preprocessing; extracting structured data into structured feature vectors through a multi-layer perceptron, extracting unstructured data into emotional feature vectors through an emotional large language model, and then performing feature fusion on the structured feature vectors and the emotional feature vectors in a fusion layer, and introducing an attention mechanism to achieve feature interaction to generate fusion features that reflect students' behavior and emotions; combining multiple classifiers into a multi-classifier model through an ensemble learning method, and classifying and predicting students' academic performance based on fusion features. The present invention can accurately predict students' academic performance by fusing structured data with emotional features, helping teachers identify poorly performing students as early as possible and provide personalized learning support.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for predicting students' early performance based on a large emotional language model. Background Art

[0002] In today's information age, artificial intelligence technology, especially large language models (LLMs) technology, has made significant progress in many fields and has been widely used in natural language processing, text generation, sentiment analysis and other tasks. However, despite the excellent performance of large language models in general tasks, their application in the field of education still faces a series of challenges. Especially in the scenario of smart education, early prediction of students' academic performance is a complex and critical task, and the existing large language model technology still has many shortcomings in this application scenario.

[0003] First, in the field of education, students' academic performance is affected by many factors, including not only structured data such as academic performance, but also unstructured data such as emotional state and learning interaction. However, traditional large language models often lack in-depth understanding and accurate extraction of emotional factors when processing these multi-dimensional information. Students' emotional states during the learning process, such as anxiety, disappointment, or positive emotions, often have a significant impact on their academic performance, but existing models can usually only provide simple emotional classification when capturing these emotional features, and it is difficult to provide fine-grained emotional state analysis, which in turn affects the model's prediction accuracy of student performance.

[0004] Secondly, existing large language models mostly focus on processing text data. Although they can process structured data in certain educational scenarios (such as grades, attendance rates, etc.), they are still unable to comprehensively analyze students' learning behaviors and emotional states. For example, students' interactive texts on learning platforms, such as forum speeches and homework feedback, often contain rich emotional information and learning attitudes, which are crucial for predicting student performance. However, when integrating such unstructured data, existing models lack effective mechanisms to extract and utilize this information, resulting in unsatisfactory performance of the prediction model.

[0005] In addition, although large language models perform well in semantic understanding and generation, their personalized support and prediction capabilities in educational scenarios still need to be improved. Prediction tasks in educational scenarios require models to comprehensively consider students' structured learning data and unstructured emotional and behavioral characteristics. Existing models are usually unable to fully integrate these two types of data, resulting in low accuracy and practicality of prediction results. In actual teaching, teachers and education administrators urgently need an intelligent tool that can comprehensively analyze students' academic performance, emotional state, and learning behavior, so as to promptly identify potential problems and provide personalized intervention and support.

[0006] Based on the above background, the present invention proposes a method for predicting students' early performance based on the emotional large language model in a smart education scenario to address the deficiencies in the prior art. By introducing the emotional large language model (EmoLLM), the method can extract fine-grained emotional features from students' unstructured interactive data, and combine it with students' structured learning data to form a multi-dimensional feature fusion, thereby achieving accurate prediction of students' academic performance. In addition, the present invention uses a multi-layer perceptron (MLP) to extract features from structured data, combines emotional feature vectors, and makes predictions through a multi-classifier model to ensure the accuracy and stability of the predictions. The present invention can help teachers identify students with poor academic performance as early as possible, provide personalized learning support, and improve the teaching effect in smart education. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a method for predicting students' early performance based on a large emotional language model, aiming to improve the prediction accuracy of students' early academic performance in a smart education environment, and provide teachers with more targeted and intelligent support in a digital teaching environment.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A method for predicting students' early performance based on a sentiment large language model, comprising:

[0010] Obtain structured and unstructured data related to students’ academic performance and perform preprocessing;

[0011] The structured data is extracted through a multi-layer perceptron to obtain a structured feature vector, and the unstructured data is extracted through a large emotional language model to obtain an emotional feature vector. Then, the structured feature vector and the emotional feature vector are fused at the fusion layer, and the attention mechanism is introduced to realize the interaction of features, so as to generate fusion features that reflect students' behaviors and emotions.

[0012] Through ensemble learning methods, multiple classifiers are combined into a multi-classifier model, and students' academic performance is classified and predicted based on the fusion features.

[0013] Furthermore, the structured data includes students' personal attribute data and learning behavior data.

[0014] Furthermore, the unstructured data includes interactive texts of students during the learning process.

[0015] Furthermore, the pre-processing process specifically includes:

[0016] Process missing values ​​in structured data and ensure data integrity through interpolation or mean filling; clean up data that does not conform to data logic in structured data; screen out features in structured data that are more relevant to students' academic performance than the threshold through feature importance analysis; standardize or normalize numerical features in structured data to ensure that the scales of different features are consistent;

[0017] Remove stop words, punctuation, and special characters from unstructured data.

[0018] Furthermore, extracting the structured data through a multi-layer perceptron to obtain a structured feature vector specifically includes:

[0019] The preprocessed structured data is input into the multi-layer perceptron network, the structured features are extracted layer by layer, and the structured features are nonlinearly mapped through a fully connected neural network to gradually extract deep features related to student performance; an activation function is introduced between each hidden layer to enhance the nonlinear expression ability of the multi-layer perceptron network for structured data, ensuring that the potential relationship between student attributes and academic performance can be captured in the process of multi-layer mapping; the output of the final hidden layer of the multi-layer perceptron network is used as the final structured feature to reflect the student's learning behavior profile, thereby obtaining the structured feature vector.

[0020] Furthermore, extracting the unstructured data through the emotional large language model to obtain the emotional feature vector specifically includes:

[0021] The preprocessed unstructured data is input into the emotional big language model to extract emotional features; the emotional big language model can perform emotional classification on each piece of unstructured data, generate an emotional label reflecting the emotional state of the students, and an emotional intensity score reflecting the emotional intensity of the students; the emotional features include the emotional labels and the emotional intensity scores; the emotional features of each piece of unstructured data are mapped into the emotional feature vector.

[0022] Furthermore, the structural feature vector and the emotional feature vector are fused at the fusion layer, and an attention mechanism is introduced to realize feature interaction to generate fusion features reflecting student behavior and emotion, specifically including:

[0023] The structural features and sentiment features are input into the fusion layer. The fusion layer adopts the attention mechanism to interactively fuse the structural features and sentiment features. By adaptively weighting the contribution of different features, the expression effect of the feature combination is further enhanced:

[0024] ;

[0025] in, represents the fusion layer using the attention mechanism, and is the weight parameter, represents the fusion feature corresponding to the i-th student, represents the structural features corresponding to the i-th student, Represents the sentiment feature corresponding to the i-th student.

[0026] Furthermore, the method of combining multiple classifiers into a multi-classifier model through an ensemble learning method to predict the student's academic performance specifically includes:

[0027] The fused features are input into a multi-classifier model composed of multiple classifiers. Each classifier has a different decision-making mechanism and can classify and predict students' academic performance from different angles. The classification results of multiple classifiers are summarized using the Bayesian averaging method to obtain the classification prediction results of students' academic performance corresponding to the fused features.

[0028] Furthermore, the fused features are input into multiple classifiers:

[0029] ;

[0030] represents the fusion feature corresponding to the i-th student, represents the jth classifier, It represents the academic performance classification prediction result obtained by the fusion features corresponding to the i-th student after passing through the j-th classifier;

[0031] Use Bayesian averaging to aggregate the academic performance classification predictions from multiple classifiers:

[0032]

[0033] in, is the uncertainty of the j-th classifier, is the total number of classifiers; Classify and predict the final academic performance of the i-th student.

[0034] Furthermore, the multiple classifiers include a support vector machine classifier, a K nearest neighbor classifier and a random forest classifier.

[0035] Compared with the prior art, the beneficial technical effects of the present invention are:

[0036] Deep mining of emotional features: The emotional big language model can extract emotional features from students’ unstructured data, providing an additional dimension of emotional information for the student performance prediction model.

[0037] Multi-feature fusion: The deep fusion of behavioral data and emotional data enables the model to capture the multi-dimensional characteristics of students' performance, which helps to make more accurate early performance predictions.

[0038] Intelligent feedback mechanism: The system provides feedback based on the model’s prediction results to help teachers adjust teaching strategies in a timely manner and optimize teaching results.

[0039] In summary, the present invention uses the emotional large language model to predict students' early performance in the smart education scenario, provides a more accurate and intelligent teaching support method, and significantly improves the effect and efficiency of digital teaching. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of a method in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the overall structure of the present invention for classifying and predicting students' academic performance. DETAILED DESCRIPTION

[0042] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0043] like Figure 1 As shown, a method for predicting students' early performance based on a large emotional language model in the present invention comprises the following steps:

[0044] S1, construct student data;

[0045] S2, preprocessing the constructed student data;

[0046] S3, uses the sentiment large language model to extract sentiment features;

[0047] S4, uses multi-layer perceptron to extract structured features;

[0048] S5, integrating structural features with emotional features;

[0049] S6, the fused features are predicted through a multi-classifier model.

[0050] The above steps are described in detail below.

[0051] Step S1, construct student data:

[0052] In a preferred embodiment, the present invention extracts structured data related to students' academic performance from data sources stored in CSV files. Structured data includes but is not limited to students' personal attribute data (such as age, gender, grade) and learning behavior data (such as grades in various subjects, homework submission time, cumulative learning time, etc.). These data are recorded in a standardized format to ensure the integrity and accuracy of the data, laying the foundation for subsequent data processing and model training.

[0053] In addition, the present invention further obtains unstructured data generated by students on the learning platform, including but not limited to forum speech records, Q&A interactions, and homework feedback. These unstructured data reflect the emotional state and communication interaction of students during the learning process. The emotional big language model is used to perform sentiment analysis on these unstructured data, and the emotional features are extracted as additional input information for the student academic performance prediction model.

[0054] This step provides a comprehensive data foundation for the student academic performance prediction model, covering the two dimensions of student behavior and emotion. Include students, and the structured data of each student is represented as ,in It is the number of features in the structured data, including age, gender, grade, grades in various subjects, homework completion rate, homework submission time, cumulative learning time, attendance record, etc. The above structured data features are only examples and can be increased or decreased according to needs and actual conditions.

[0055] Structured Datasets , where the structured data of the i-th student , It is structured data The mth feature in .

[0056] Unstructured dataset of students Represented as a sequence of interactive texts: , is the interactive text of the ith student on the learning platform.

[0057] Step 2: Preprocess the constructed student data:

[0058] First, the structured data is systematically cleaned, and the missing values ​​in the structured data (such as missing grades in various subjects, missing attendance records, etc.) are processed, and the integrity of the data is ensured by interpolation or mean filling. At the same time, for unreasonable data (such as negative grades or abnormally high scores, etc.), cleaning operations are taken to ensure that the data logic is reasonable and meets the specifications. During the preprocessing process, the present invention also screens out key features that are significantly related to students' academic performance (such as attendance records, homework completion rates, and grades in various subjects, etc.) through feature importance analysis to reduce redundant features and improve the training efficiency of the model. The numerical feature data is standardized (such as converted to a distribution with a mean of 0 and a standard deviation of 1) or normalized (such as scaling the data to a specific range) to ensure that the scales of different features are consistent and enhance the adaptability of the model to the data.

[0059] For interactive text in unstructured data, the present invention performs preliminary preprocessing by removing stop words, punctuation marks and special characters to reduce noise information and ensure that the emotional large language model can extract emotional feature vectors more accurately.

[0060] Step 3: Use the big emotional language model to extract emotional features:

[0061] The present invention inputs the preprocessed unstructured data into the emotional large language model EmoLLM to extract emotional features. The emotional large language model can classify the emotions of each interactive text, generate positive, negative or neutral emotional labels, and use it as one of the important input features of the performance prediction model to reflect the emotional state of the students. In addition, the emotional large language model can also generate emotional intensity scores to quantify the specific manifestations of emotions, such as the intensity of emotions such as anxiety, anger, and disappointment. These emotional intensity scores serve as inputs of multidimensional emotional features, providing a fine-grained emotional representation for the performance prediction model, helping the performance prediction model to fully capture the dynamics of students' emotional changes. Furthermore, the emotional large language model maps the emotional state of each interactive text into a high-dimensional emotional feature vector, which contains rich emotional information and represents a complex emotional relationship structure. The embedded vector will be used as an input feature to provide additional emotional dimension support for the performance prediction model.

[0062] Emotional features include emotional labels (positive, negative, neutral, etc.) and their corresponding emotion intensity scores, thereby obtaining a more comprehensive picture of students’ emotional states.

[0063] Step 4: Use a multi-layer perceptron to extract structured features:

[0064] The present invention inputs the structured data features into the multi-layer perceptron (MLP) network to extract key features layer by layer. The multi-layer perceptron performs nonlinear mapping on the structured data through a fully connected neural network, and gradually extracts deep features related to the student's performance. The activation function (such as ReLU) is introduced between each hidden layer to enhance the nonlinear expression ability of the multi-layer perceptron for structured data, ensuring that the performance prediction model can capture the potential relationship between students' personal attributes and academic performance during the multi-layer mapping process. The output of the multi-layer perceptron is a set of vectorized structured features, which reflect the students' learning behavior profile and lay the foundation for the subsequent feature fusion process.

[0065] Standardized structured data The process of inputting to the multilayer perceptron for nonlinear mapping is as follows:

[0066] ;

[0067] ;

[0068] in, represents the output of the neural network layer l for the structured data of the ith student, that is, the activation value of the neural network layer l, represents the weight matrix from the lth layer to the l+1th layer of the neural network, represents the bias term from the lth layer to the l+1th layer of the neural network, Represents the total number of layers in the neural network.

[0069] The final hidden layer output is used as a structured feature vector :

[0070] ;

[0071] in, Represents the structured features corresponding to the i-th student.

[0072] Step 5: Fusion of structural features and sentiment features:

[0073] In this step, the present invention inputs the structural features and the sentiment features into the fusion layer. The fusion layer can use the attention mechanism to interactively fuse the two types of features. By adaptively weighting the contributions of different features, the attention mechanism further enhances the expression effect of the feature combination.

[0074] ;

[0075] in, is the sentiment feature corresponding to the i-th student, and is the relevant weight parameter, Represents the fusion features corresponding to the i-th student.

[0076] Finally, the fusion feature matrix is ​​obtained: . Indicates the dimension of fused features.

[0077] The fusion layer further combines the structural features and the emotional features to generate the final fusion features. This fusion feature not only contains the structural attribute information of the students, but also integrates the expression of the emotional dimension, so that the performance prediction model can comprehensively analyze the multi-dimensional feature information of the students, thereby improving the accuracy of the prediction.

[0078] Step 6: Use the fusion features to predict through a multi-classifier model:

[0079] In a preferred embodiment, the present invention proposes a multi-classifier model consisting of a support vector machine (SVM) classifier, a K-nearest neighbor (KNN) classifier and a random forest (RF) classifier, namely a performance prediction model. Each classifier has a unique decision-making mechanism and can perform classification predictions of students' academic performance from different angles. The support vector machine classifier, the K-nearest neighbor classifier and the random forest classifier are combined by the Bagging method. This method adopts a voting method to summarize the prediction results of each classifier based on the independent prediction of each classifier to enhance the robustness and accuracy of the performance prediction model. Finally, the performance prediction model outputs the classification prediction results of the students' academic performance. The results can be used to provide personalized learning support for students, help teachers identify poorly performing students as early as possible, and then take targeted intervention measures to improve students' academic performance.

[0080] Input the fusion features into a multi-classifier model consisting of multiple classifiers (SVM, KNN, RF):

[0081] ;

[0082] represents the jth classifier, represents the academic performance classification prediction result obtained by the fusion feature corresponding to the i-th student after passing through the j-th classifier; in the preferred embodiment, there are three classifiers, .

[0083] Use Bayesian averaging to aggregate classification results:

[0084]

[0085] in, To represent the uncertainty of the j-th classifier, Represents the final classification prediction result of the academic performance of the i-th student.

[0086] The collection of N students' academic performance classification prediction results .

[0087] In a preferred embodiment, the present invention can be used to predict whether the i-th student can pass the final examination. If you cannot pass .

[0088] The system regularly updates the knowledge base to ensure the accuracy and timeliness of the data used. As the emotional language model's ability to process student interaction data gradually increases, the system can dynamically adjust the prediction results of student performance and help teachers keep abreast of students' learning dynamics through personalized feedback.

[0089] Bagging is the abbreviation of Bootstrap Aggregating, which is a powerful ensemble learning method that combines multiple basic classifiers to improve prediction accuracy and robustness. The multi-classifier model proposed in this embodiment consists of three basic classifiers: support vector machine, K nearest neighbor and random forest. The ensemble learning method enables the present invention to take advantage of the diversity of these classifiers and use their unique decision-making capabilities to perform more comprehensive classification. The present invention will briefly summarize the basic classifiers used in the ensemble below.

[0090] Support Vector Machine (SVM) is a powerful and widely used classifier suitable for both binary and multi-class classification tasks. It seeks to find an optimal hyperplane that maximizes the separation of different classes in the feature space, making it very effective in capturing complex decision boundaries.

[0091] K-nearest neighbors (KNN) is a non-parametric, instance-based classifier. It classifies instances by considering the class labels of the k nearest neighbors in the feature space. K-nearest neighbors is intuitive, easy to implement, and well suited for multi-class classification tasks.

[0092] Random forest is an ensemble learning method based on decision trees. It builds multiple decision trees during training and combines their predictions by voting to arrive at the final classification. Random forest is very powerful and can handle large datasets and effectively handle noisy data.

[0093] Bayesian averaging is an ensemble learning technique used to combine the predictions of multiple models in a principled and probabilistic manner. Unlike traditional averaging methods, Bayesian averaging takes into account the predictions of the models as well as their uncertainty. It assigns weights to the predictions of each model based on its performance and reliability, resulting in a more robust and accurate overall prediction. The overall prediction is obtained by calculating the weighted average of the individual model predictions. The Bayesian averaging method not only improves the overall accuracy of the ensemble, but also provides a measure of uncertainty for the final prediction. This uncertainty estimate is very valuable in situations where prediction reliability is critical.

[0094] The present invention utilizes the emotion large language model EmoLLM to extract unstructured emotion features, combines structured personal attribute data and learning behavior data, and makes accurate predictions of students' early performance through multi-layer analysis of feature fusion and multi-classifier models.

[0095] The present invention provides an effective student academic performance prediction solution with high accuracy and multi-dimensional feature support, which provides an important basis for personalized support and academic performance improvement of students in smart education scenarios. It has the following advantages:

[0096] 1) Deeper emotional understanding: The emotional large language model can capture complex emotional information, generate high-dimensional emotional feature vectors, and provide richer emotional representation than simple emotional labels and intensities.

[0097] 2) Simplify the modeling process: The sentiment features extracted by the large sentiment language model can be directly used as model input without the need for additional time series modeling, thus simplifying the model structure.

[0098] 3) Efficient feature fusion: By fusing structured data with sentiment features, the model can simultaneously capture students’ learning behaviors and sentiment changes, thereby more accurately predicting whether they will drop a course or pass an exam.

[0099] This sentiment analysis method based on the large emotional language model can provide stronger sentiment understanding capabilities for the model predicting student learning behavior, improving the accuracy and practicality of the model.

[0100] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0101] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for predicting students' early performance based on a large emotional language model, characterized in that: include: Obtain structured and unstructured data related to students’ academic performance and pre-process them; structured data includes students’ personal attribute data and learning behavior data, and unstructured data includes students’ interactive texts during the learning process; The structured data is extracted through a multi-layer perceptron to obtain a structured feature vector, and the unstructured data is extracted through a large emotional language model to obtain an emotional feature vector. Then, the structured feature vector and the emotional feature vector are fused at the fusion layer, and the attention mechanism is introduced to realize the interaction of features, so as to generate fusion features that reflect students' behaviors and emotions. Through the ensemble learning method, multiple classifiers are combined into a multi-classifier model, and the students' academic performance is classified and predicted based on the fusion features: the fusion features are input into the multi-classifier model composed of multiple classifiers. Each classifier has a different decision-making mechanism and can classify and predict the students' academic performance from different angles: ; represents the fusion feature corresponding to the i-th student, represents the jth classifier, It represents the academic performance classification prediction result obtained by the fusion features corresponding to the i-th student after passing through the j-th classifier; The classification results of multiple classifiers are summarized using the Bayesian average method to obtain the classification prediction results of the students' academic performance corresponding to the fusion features: ; in, is the uncertainty of the j-th classifier, is the total number of classifiers; Classify and predict the final academic performance of the i-th student.

2. The method for predicting students' early performance based on the large emotional language model according to claim 1 is characterized in that: The pre-processing process specifically includes: Process missing values ​​in structured data and ensure data integrity through interpolation or mean filling; clean up data that does not conform to data logic in structured data; screen out features in structured data that are more relevant to students' academic performance than the threshold through feature importance analysis; standardize or normalize numerical features in structured data to ensure that the scales of different features are consistent; Remove stop words, punctuation, and special characters from unstructured data.

3. The method for predicting students' early performance based on the large emotional language model according to claim 1 is characterized in that: The step of extracting the structured data through a multi-layer perceptron to obtain a structured feature vector specifically includes: The preprocessed structured data is input into the multi-layer perceptron network, the structured features are extracted layer by layer, and the structured features are nonlinearly mapped through a fully connected neural network to gradually extract deep features related to student performance; an activation function is introduced between each hidden layer to enhance the nonlinear expression ability of the multi-layer perceptron network for structured data, ensuring that the potential relationship between student attributes and academic performance can be captured in the process of multi-layer mapping; the output of the final hidden layer of the multi-layer perceptron network is used as the final structured feature to reflect the student's learning behavior profile, thereby obtaining the structured feature vector.

4. The method for predicting students' early performance based on the large emotional language model according to claim 1 is characterized in that: The step of extracting the unstructured data through the emotional language model to obtain the emotional feature vector specifically includes: The preprocessed unstructured data is input into the emotional big language model to extract emotional features; the emotional big language model can perform emotional classification on each piece of unstructured data, generate an emotional label reflecting the emotional state of the students, and an emotional intensity score reflecting the emotional intensity of the students; the emotional features include the emotional labels and the emotional intensity scores; the emotional features of each piece of unstructured data are mapped into the emotional feature vector.

5. The method for predicting students' early performance based on the large emotional language model according to claim 1 is characterized in that: The structured feature vector and the emotional feature vector are fused at the fusion layer, and the attention mechanism is introduced to realize the interaction of features to generate fusion features that reflect the student's behavior and emotions, specifically including: The structural features and sentiment features are input into the fusion layer. The fusion layer adopts the attention mechanism to interactively fuse the structural features and sentiment features. By adaptively weighting the contribution of different features, the expression effect of the feature combination is further enhanced: ； in, represents the fusion layer using the attention mechanism, and is the weight parameter, represents the fusion feature corresponding to the i-th student, represents the structural features corresponding to the i-th student, Represents the sentiment feature corresponding to the i-th student.

6. The method for predicting students' early performance based on the large emotional language model according to claim 1 is characterized in that: The multiple classifiers include a support vector machine classifier, a K nearest neighbor classifier, and a random forest classifier.

Images