Deep learning-based multi-source practical training behavior fusion intelligent scoring method, device and equipment

Abstract

The application discloses a kind of multi-source practical training behavior fusion intelligent scoring method, device and equipment based on deep learning, it is related to deep learning technical field, including: based on practical training code interface picture, obtain code feature vector;Based on practical training course score, obtain score feature vector;Based on self-evaluation report and teacher evaluation report, obtain comment feature vector and sentiment feature vector;Based on score feature vector and code feature vector, obtain code-score fusion feature;Based on score feature vector and comment feature vector, obtain score-comment fusion feature;Based on code-score fusion feature and score-comment fusion feature, obtain multi-source practical training behavior feature;Based on no emotional practical training behavior feature and sentiment feature vector, obtain practical training comprehensive evaluation score, realizes the objectivization practical training comprehensive score based on data internal correlation driving.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a method, apparatus and device for intelligent scoring of multi-source training behavior fusion based on deep learning. Background Technology

[0002] As the vocational education practical training evaluation system develops towards process-oriented and refined directions, the types of student behavior data available in practical training scoring scenarios are increasing, including images of computer operation interfaces, historical course grades, and self-evaluation and teacher evaluation text reports. These data reflect students' comprehensive practical training abilities from different dimensions, which puts forward a technical requirement for the evaluation system to integrate multi-source heterogeneous data for comprehensive quantitative evaluation.

[0003] Current mainstream training evaluation methods include subjective human scoring, rule-based scoring, and hybrid methods. Subjective human scoring relies entirely on the teacher's personal experience, with scoring standards varying from person to person and difficult to standardize. Rule-based scoring only judges the compliance of explicit results based on fixed thresholds and preset rules, and cannot utilize auxiliary information such as historical scores, nor can it extract deep semantic features from interface images and comment texts. Hybrid methods, although combining qualitative and quantitative means, are still constrained by subjective bias and the rationality of rule setting. None of the above methods have established effective correlations between multi-source behavioral data.

[0004] Because existing methods lack the ability to learn and model the potential coupling relationships between multiple sources of data, such as interface images, historical scores, and comment texts, training scores cannot break through the limitations of single-dimensional or shallow correlations. This leads to a technical problem: how to achieve an objective comprehensive training score that does not rely on fixed rules and subjective judgments by integrating multi-source training behavior data and mining its inherent correlations.

[0005] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of this application is to provide a method, device, and equipment for intelligent scoring of multi-source training behavior fusion based on deep learning, which aims to solve the technical problem of how to achieve objective comprehensive scoring of training without relying on fixed rules and subjective judgment by fusing multi-source training behavior data and mining their inherent correlation.

[0007] To achieve the above objectives, this application proposes a deep learning-based multi-source training behavior fusion intelligent scoring method, the method comprising: Based on the images of the training code interface, code feature vectors are obtained; Based on the practical training course grades, a grade feature vector is obtained; Based on self-evaluation reports and teacher evaluation reports, we obtained comment feature vectors and affective feature vectors. Based on the performance feature vector and the code feature vector, a code-performance fusion feature is obtained; Based on the performance feature vector and the comment feature vector, a performance-comment fusion feature is obtained; Based on the code-score fusion feature and the score-comment fusion feature, the emotionless training behavior features are obtained; Based on the emotionless training behavior characteristics and the emotional feature vector, the comprehensive evaluation score of the training is obtained.

[0008] In one embodiment, obtaining the code feature vector based on the training code interface image includes: The training code interface image is standardized and cropped to obtain the target code image; The target code image is input into a preset visual model to obtain the code feature vector output by the preset visual model.

[0009] In one embodiment, obtaining the performance feature vector based on the practical training course grades includes: Normalize the practical training course grades to obtain normalized course grades; The normalized course scores are input into a two-layer nonlinear function network to obtain the score feature vector extracted by the two-layer nonlinear function network from the normalized course scores.

[0010] In one embodiment, obtaining the comment feature vector and the sentiment feature vector based on the self-evaluation report and the teacher evaluation report includes: Based on the self-evaluation report and the teacher evaluation report, the content of the evaluation report is obtained; The evaluation report content is input into a preset language model to obtain the comment feature vector output by the preset language model; The evaluation report content is input into a preset sentiment analysis tool to obtain the sentiment feature vector extracted by the preset sentiment analysis tool from the evaluation report content.

[0011] In one embodiment, obtaining the code-score fusion feature based on the performance feature vector and the code feature vector includes: Perform a linear transformation on the performance feature vector to determine the first query vector; Perform a linear transformation on the code feature vector to determine the first key vector and the first value vector; Based on the first query vector, the first key vector, and the first value vector, the first attention weight is obtained; The code feature vector and the score feature vector are weighted and summed according to the first attention weight to obtain the code-score fusion feature.

[0012] In one embodiment, obtaining the grade-comment fusion feature based on the grade feature vector and the comment feature vector includes: Perform a linear transformation on the performance feature vector to determine the second query vector; Perform a linear transformation on the comment feature vector to determine the second key vector and the second value vector; Based on the second query vector, the second key vector, and the second value vector, the second attention weight is obtained; The performance feature vector and the comment feature vector are weighted and summed according to the second attention weight to obtain the performance-comment fusion feature.

[0013] In one embodiment, obtaining emotionless training behavior features based on the code-grade fusion feature and the grade-comment fusion feature includes: Based on the code-score fusion feature and the score-comment fusion feature, the gating weight parameters are obtained; Based on the gating weight parameters, the code-score fusion features, and the score-comment fusion features, emotionless training behavior features are obtained.

[0014] In one embodiment, obtaining the comprehensive training evaluation score based on the emotionless training behavior features and the emotion feature vector includes: The target training behavior features are obtained by concatenating the emotionless training behavior features and the emotion feature vector. The target training behavior features are input into a preset multilayer perceptron to obtain a comprehensive training evaluation score. The preset multilayer perceptron includes a first linear layer, a first nonlinear activation function layer, a second linear layer, and a second nonlinear activation function layer connected in sequence.

[0015] Furthermore, to achieve the above objectives, this application also proposes a deep learning-based multi-source training behavior fusion intelligent scoring device, which includes: The first feature extraction module is used to obtain code feature vectors based on the training code interface images; The second feature extraction module is used to obtain the performance feature vector based on the practical training course scores; The third feature extraction module is used to obtain comment feature vectors and sentiment feature vectors based on self-evaluation reports and teacher evaluation reports; The first feature fusion module is used to obtain code-score fusion features based on the score feature vector and the code feature vector; The second feature fusion module is used to obtain the grade-comment fusion feature based on the grade feature vector and the comment feature vector; The third feature fusion module is used to obtain emotionless training behavior features based on the code-score fusion features and the score-comment fusion features; The evaluation score output module is used to obtain a comprehensive evaluation score for the training based on the emotionless training behavior characteristics and the emotion feature vector.

[0016] Furthermore, to achieve the above objectives, this application also proposes a deep learning-based multi-source training behavior fusion intelligent scoring device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the deep learning-based multi-source training behavior fusion intelligent scoring method described above.

[0017] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the deep learning-based multi-source training behavior fusion intelligent scoring method described above.

[0018] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the deep learning-based multi-source training behavior fusion intelligent scoring method described above.

[0019] This application obtains code feature vectors based on training code interface images; grade feature vectors based on training course scores; comment feature vectors and sentiment feature vectors based on self-evaluation reports and teacher evaluation reports; code-grade fusion features based on the grade feature vectors and code feature vectors; grade-comment fusion features based on the grade feature vectors and comment feature vectors; emotionless training behavior features based on the code-grade fusion features and grade-comment fusion features; and a comprehensive training evaluation score based on the emotionless training behavior features and the sentiment feature vectors. By employing techniques to acquire multi-source training behavior data and extract image features, score numerical features, and comment text and sentiment features respectively, the score features can be used as guidance information to be cross-fused with code image features and comment features. The two types of fusion results are then further aggregated into a unified multi-source training behavior representation. This representation is then combined with sentiment features and input into a multilayer perceptron for end-to-end score prediction. This solves the technical problem of existing methods that lack the ability to model the potential coupling relationships between multi-source data such as interface images, historical scores, and comment texts, resulting in scores relying on fixed rules or subjective experience. This achieves objective comprehensive training score driven by the inherent correlation of data. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating an embodiment of the deep learning-based multi-source training behavior fusion intelligent scoring method of this application. Figure 2 This is a schematic diagram of the module structure of the deep learning-based multi-source training behavior fusion intelligent scoring device according to an embodiment of this application; Figure 3 This is a schematic diagram of the device structure of the hardware operating environment involved in the deep learning-based multi-source training behavior fusion intelligent scoring method in the embodiments of this application.

[0023] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0025] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0026] The main solution of this application embodiment is as follows: Based on the training code interface image, obtain the code feature vector; based on the training course score, obtain the score feature vector; based on the self-evaluation report and the teacher evaluation report, obtain the comment feature vector and the sentiment feature vector; based on the score feature vector and the code feature vector, obtain the code-score fusion feature; based on the score feature vector and the comment feature vector, obtain the score-comment fusion feature; based on the code-score fusion feature and the score-comment fusion feature, obtain the emotionless training behavior feature; based on the emotionless training behavior feature and the sentiment feature vector, obtain the comprehensive training evaluation score.

[0027] In this embodiment, for ease of description, the following description will focus on a deep learning-based multi-source training behavior fusion intelligent scoring system.

[0028] Current mainstream training evaluation methods include subjective human scoring, rule-based scoring, and hybrid methods. Subjective human scoring relies entirely on the teacher's personal experience, with varying scoring standards and difficulty in standardization. Rule-based scoring only judges the compliance of explicit results based on fixed thresholds and preset rules, failing to utilize auxiliary information such as historical scores or extract deep semantic features from interface images and comment text. Hybrid methods, while combining qualitative and quantitative approaches, are still constrained by subjective biases and the rationality of rule settings. None of these methods have established effective correlations between multi-source behavioral data.

[0029] This application provides a solution that, by acquiring multi-source training behavior data and extracting image features, score numerical features, and comment text and sentiment features respectively, enables score features to be used as guidance information for cross-fusion with code image features and comment features. The two types of fusion results are then further aggregated into a unified multi-source training behavior representation. This representation is then combined with sentiment features and input into a multilayer perceptron for end-to-end score prediction. This solves the technical problem of existing methods that lack the ability to model the potential coupling relationships between multi-source data such as interface images, historical scores, and comment texts, resulting in scores relying on fixed rules or subjective experience. This achieves objective comprehensive training score driven by the inherent correlation of data.

[0030] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a deep learning-based multi-source training behavior fusion intelligent scoring system. The following description uses a deep learning-based multi-source training behavior fusion intelligent scoring system as an example to illustrate this embodiment and the subsequent embodiments.

[0031] Based on this, embodiments of this application provide a deep learning-based multi-source training behavior fusion intelligent scoring method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the deep learning-based multi-source training behavior fusion intelligent scoring method of this application.

[0032] In this embodiment, the deep learning-based multi-source training behavior fusion intelligent scoring method includes steps S10~S70: Step S10: Obtain the code feature vector based on the training code interface image; It should be noted that the training code interface image is a screenshot of the student writing code during the training process. The image records the complete interface information of the code editing area, including the code text content, code structure layout, syntax highlighting style, and code line numbers, etc. This information together constitutes an intuitive reflection of the student's training operation behavior.

[0033] Understandably, since the training code interface images contain deep behavioral characteristics such as students' coding habits, code standardization, and program implementation logic, these characteristics are difficult to quantify and extract through simple rule matching or manual review. Therefore, step S10 is performed to convert the image data into a dense vector representation, which can capture the semantic information of code behavior from a visual perspective, avoid relying on fixed rules to make shallow judgments on code content, and thus improve the accuracy of subsequent fusion scoring.

[0034] In one feasible implementation, step S10 may include: performing normalized cropping on the training code interface image to obtain a target code image; and inputting the target code image into a preset visual model to obtain the code feature vector output by the preset visual model.

[0035] It should be noted that the pre-trained visual model refers to a publicly available pre-trained visual model used to extract deep semantic features of images. This model is trained based on large-scale image-text pairing data and has the ability to map visual content into semantic vectors. For example, Contrastive Language-Image Pre-training (CLIP).

[0036] Specifically, the training code interface image is standardized and cropped to form a uniformly sized target code image. This target code image is then fed into the preset visual model as input data. The preset visual model uses its internal encoder structure to perform layer-by-layer feature abstraction on the target code image, ultimately outputting a fixed-dimensional real-valued vector. This vector is the code feature vector, used to represent the visual semantic information in the code interface. The calculation process is shown in the following formula:

[0037] Where i represents the code feature vector and CLIP represents the preset visual model.

[0038] In this embodiment, feature extraction is performed on the code interface image by using a preset visual model, which solves the problem that code images are difficult to be directly processed by the subsequent fusion module, and enables image data to participate in cross-modal feature interaction in a vectorized form.

[0039] The above are merely feasible implementations of step S10 provided in this embodiment. This embodiment does not specifically limit the specific implementation of step S10.

[0040] Step S20: Based on the practical training course grades, obtain the grade feature vector; It should be noted that the practical training course grades refer to the final exam grades of students in several courses related to the practical training content. These courses include C language course design, data structures, computer networks, algorithm design and analysis, Java language programming, database principles, web front-end basics, Java Web course design, operating systems, design patterns, software engineering, and compiler principles. These course grades can reflect the students' basic practical training abilities from the perspective of knowledge reserves.

[0041] Understandably, since the original course scores are numerical data in the form of scores, their value range and distribution vary. Directly using the original values ​​for subsequent processing will cause dimensional effects and numerical instability. Moreover, a single value is difficult to express the comprehensive correlation information between course scores. Therefore, step S20 is performed to normalize the scores and map them to a high-dimensional feature space. This can transform discrete numerical scores into continuous feature vectors, thereby avoiding the adverse effects of the scale difference of the original values ​​on model training and improving the model convergence efficiency.

[0042] In one feasible implementation, step S20 may include: normalizing the practical training course grades to obtain normalized course grades; and inputting the normalized course grades into a two-layer nonlinear function network to obtain the grade feature vector extracted by the two-layer nonlinear function network from the normalized course grades.

[0043] It should be noted that normalization refers to linearly mapping the original scores to a specified interval to eliminate numerical distribution biases caused by differences in full-mark standards and difficulty levels between different courses. A two-layer nonlinear function network refers to a computational structure consisting of two alternating linear transformation layers and nonlinear activation functions.

[0044] Specifically, the original practical training course grades are compressed into a uniform interval through linear mapping, eliminating the influence of dimensions while preserving the relative distribution of the original data. The normalized course grades are then fed into a two-layer nonlinear function network as input vectors. The first layer of the network performs a linear transformation on the input vector and introduces nonlinear mapping capabilities through an activation function. The second layer performs another linear transformation and nonlinear mapping on the output of the first layer, ultimately outputting a fixed-dimensional grade feature vector. This vector is used to represent the comprehensive ability distribution of students' course knowledge level. The calculation process is shown in the formula below:

[0045]

[0046] in, This represents the highest score value among the scores; S represents the minimum score value; S represents the original score value. This represents the normalized final exam score for the relevant course. , , and The learnable parameters are initialized using a normal distribution function; This represents a non-linear activation function.

[0047] In this embodiment, numerical scale differences are eliminated by normalization and numerical scores are mapped to high-dimensional feature representations through a two-layer nonlinear network, thus solving the problem that single numerical scores are difficult to participate in cross-modal feature fusion.

[0048] The above are merely feasible implementations of step S20 provided in this embodiment. This embodiment does not specifically limit the specific implementation of step S20.

[0049] Step S30: Based on the self-evaluation report and the teacher evaluation report, obtain the comment feature vector and the sentiment feature vector; It should be noted that the self-evaluation report is a written summary and description of the student's training process and results after the training is completed, while the teacher evaluation report is a written comment from the instructor on the student's training performance and achievements. Both are in natural language text format and contain information such as a process description of the student's training behavior, affirmation of strengths and pointing out of weaknesses.

[0050] Understandably, since the evaluation report exists in the form of unstructured natural language text, it contains both objective descriptions of practical training behaviors and evaluative statements with subjective emotional tendencies. These two types of information contribute to the overall score of practical training in different dimensions. Therefore, by performing step S30, extracting the semantic features and emotional tendencies of the comments respectively, we can make full use of the report information from both content and attitude levels, avoiding simply viewing the text as a single semantic source and ignoring the impact of emotional information on the evaluation results.

[0051] In one feasible implementation, step S30 may include: obtaining evaluation report content based on self-evaluation reports and teacher evaluation reports; inputting the evaluation report content into a preset language model to obtain the comment feature vector output by the preset language model; and inputting the evaluation report content into a preset sentiment analysis tool to obtain the sentiment feature vector extracted by the preset sentiment analysis tool from the evaluation report content.

[0052] It should be noted that the pre-trained language model refers to a pre-trained language model that, through pre-training on a large-scale text corpus, possesses the ability to encode natural language text into semantic vectors. An example is the Bidirectional Encoder Representations from Transformers (BERT) model. The pre-trained sentiment analysis tool refers to an existing sentiment analysis tool used to identify and extract sentiment polarity information from text. An example is the multimodal sentiment analysis tool (HanLP: Han Language Processing).

[0053] Specifically, the self-evaluation report and teacher evaluation report are cleaned and merged into the evaluation report content. This content is then input into a pre-set language model for encoding. After passing through the word embedding layer and multi-layer bidirectional encoder structure within the model, a comment feature vector is obtained. This vector is used to represent the semantic content information in the report. At the same time, the evaluation report content is input into a pre-set sentiment analysis tool. This tool identifies sentiment words and sentiment modification structures in the text and outputs a sentiment feature vector. This vector is used to represent the subjective sentiment tendency in the report. The calculation process is shown below.

[0054]

[0055] Where r represents the feature vector of the comments, Bret() represents the preset language model, and Report represents the content of the evaluation report.

[0056]

[0057] Where sa represents the sentiment feature vector, HanLP() represents the preset sentiment analysis tool, and Report represents the evaluation report content.

[0058] In this embodiment, semantic features and sentiment features are extracted by using a preset language model and a preset sentiment analysis tool, respectively, which solves the problem that the content information and sentiment information in the evaluation report are processed in a singular way, resulting in the loss of the contribution of the sentiment dimension.

[0059] Step S40: Based on the performance feature vector and the code feature vector, obtain the code-performance fusion feature.

[0060] It should be noted that the code-grade fusion feature refers to the fusion expression obtained by associating and weighting code image features with grade features as guidance information. It is used to characterize the guidance relationship between course knowledge level and code behavior.

[0061] It is understandable that since students' historical course grades reflect their theoretical knowledge reserves, while their practical coding behavior reflects their ability to apply theoretical knowledge to practice, there is an inherent correlation between the two. That is, students with higher course grades tend to perform better in terms of code standardization. Therefore, by performing step S40, which uses grade features as a guide to selectively aggregate code features, prior knowledge signals can be introduced into the expression of code behavior features. This avoids treating grade information and code behavior information as independent and isolated features, thereby enhancing the information interaction between the two types of data.

[0062] In one feasible implementation, step S40 may include: performing a linear transformation on the performance feature vector to determine a first query vector; performing a linear transformation on the code feature vector to determine a first key vector and a first value vector; obtaining a first attention weight based on the first query vector, the first key vector, and the first value vector; and performing a weighted summation on the code feature vector and the performance feature vector according to the first attention weight to obtain a code-performance fusion feature.

[0063] It should be noted that the first query vector is a vector obtained by linearly transforming the performance feature vector, and is used to represent the query conditions for performance information in attention calculation. The first key vector and the first value vector are both vectors obtained by linearly transforming the code feature vector. The key vector is used to perform similarity matching with the query vector to determine the attention weight, and the value vector is used to carry the semantic information of the code to be aggregated.

[0064] Specifically, the performance feature vector is linearly mapped to obtain the query vector, and the code feature vector is linearly mapped to obtain the key vector and value vector. After similarity calculation between the query vector and the key vector, the result is normalized to obtain the first attention weight. The magnitude of this weight reflects the importance of the code feature at that position being guided by the performance feature. The value vectors at each position are weighted and aggregated using this weight to obtain the code-performance fusion feature. This application uses an attention mechanism to model the feature representation between course performance and training code features. The calculation process of the performance-guided attention mechanism is shown in the following formula.

[0065]

[0066]

[0067]

[0068]

[0069] in, , , , , , The parameters of the learnable model are initialized using a normal distribution function; s represents the performance feature vector. Indicates the size of the feature dimension; The first query vector represents the features of the relevant performance characteristics after linear transformation; i represents the code feature vector. and These represent the features of the training code interface image after linear transformation. The first key vector, It is the first value vector; This represents the relationship between the integrated training code and the relevant course grades, i.e., the code-grade fusion feature.

[0070] In this embodiment, attention-weighted aggregation of code features guided by performance features solves the problem that performance information and code behavior information cannot form an effective association when used in isolation.

[0071] The above are merely feasible implementations of step S40 provided in this embodiment. This embodiment does not specifically limit the specific implementation of step S40.

[0072] Step S50: Based on the performance feature vector and the comment feature vector, obtain the performance-comment fusion feature.

[0073] It should be noted that the grade-comment fusion feature refers to the fusion expression obtained by using grade features as guiding information and associating and weighting the semantic features of comments, which is used to represent the correspondence between course knowledge level and report description content.

[0074] It is understandable that since students' history course grades reflect their theoretical level, while the evaluation reports of teachers and students record their specific performance in the training process, there is a potential correlation between the two. That is, students with higher grades often receive more positive descriptions in the comments. Therefore, by performing step S50, which uses the performance characteristics as a guide to selectively aggregate the comment features, quantitative performance information and qualitative comment information can be cross-modally correlated, avoiding the two types of data being simply spliced ​​together as independent features, which would lead to insufficient information utilization.

[0075] In one feasible implementation, step S50 may include: performing a linear transformation on the performance feature vector to determine a second query vector; performing a linear transformation on the comment feature vector to determine a second key vector and a second value vector; obtaining a second attention weight based on the second query vector, the second key vector, and the second value vector; and performing a weighted summation on the performance feature vector and the comment feature vector according to the second attention weight to obtain a performance-comment fusion feature.

[0076] It should be noted that the second query vector is a vector obtained by linearly transforming the performance feature vector, and is used to represent the query conditions for performance information in attention calculation. Both the second key vector and the second value vector are vectors obtained by linearly transforming the comment feature vector. The key vector is used for similarity matching with the query vector to determine attention weights, and the value vector carries the semantic information of the comments to be aggregated.

[0077] Specifically, the performance feature vector is linearly mapped to obtain the query vector, and the comment feature vector is linearly mapped to obtain the key vector and value vector. The query vector and the key vector are similar to each other and then normalized to obtain the second attention weight. The magnitude of this weight reflects the importance of the comment feature at that position being guided by the performance feature. The value vectors at each position are weighted and aggregated with this weight to obtain the performance-comment fusion feature. The calculation process is shown in the formula below.

[0078]

[0079]

[0080]

[0081]

[0082] in, , , , , , These represent the learnable model parameters, initialized using a normal distribution function. Indicates the feature dimension; s represents the performance feature vector; represents the features of the relevant performance characteristics after linear transformation, i.e., the second query vector; r represents the comment feature vector. and These represent the features of the comments after linear transformation. Second key vector, It is the second value vector; This represents the relationship between integrated comments and related course grades, i.e., grade-comment integration features.

[0083] In this embodiment, attention-weighted aggregation of performance characteristics and comment characteristics is used to guide the problem that quantitative performance information and qualitative comment information cannot be effectively correlated when used independently.

[0084] The above are merely feasible implementations of step S50 provided in this embodiment. This embodiment does not specifically limit the specific implementation of step S50.

[0085] Step S60: Based on the code-score fusion feature and the score-comment fusion feature, obtain the emotionless training behavior feature.

[0086] It should be noted that the emotionless training behavior feature refers to the comprehensive feature expression obtained by adaptively aggregating the code-grade fusion feature and the grade-comment fusion feature. This feature includes the correlation between three types of training behavior information: code behavior, course grades, and comment content.

[0087] Specifically, the code-grade fusion feature is a fusion expression of code behavior guided by grades obtained from the interaction between images and grades, while the grade-comment fusion feature is a fusion expression of comment content guided by grades obtained from the interaction between text and grades. The two types of fusion features are aggregated to form emotionless training behavior features.

[0088] It is understandable that, since the code-score fusion feature and the score-comment fusion feature express the training behavior from different data interaction perspectives, the degree of contribution of the two to the final score varies depending on the sample. Therefore, in step S60, the weight allocation of the two types of features during aggregation is dynamically adjusted through a gating mechanism. This can adaptively determine the contribution ratio of each modality fusion feature according to the sample characteristics, avoiding the loss of differential information between samples due to simple splicing with fixed weights.

[0089] In one feasible implementation, step S60 may include: obtaining gating weight parameters based on the code-grade fusion feature and the grade-comment fusion feature; and obtaining emotionless training behavior features based on the gating weight parameters, the code-grade fusion feature, and the grade-comment fusion feature.

[0090] It should be noted that the gating weight parameter is a dynamic weight coefficient obtained by transforming the code-score fusion feature and the score-comment fusion feature, and is used to control the relative proportion of the two types of features during aggregation.

[0091] Specifically, the code-grade fusion feature and the grade-comment fusion feature are concatenated and then processed by linear mapping and activation function to obtain a gating weight parameter. The value of this parameter is within a specified range and is related to the sample. The code-grade fusion feature and the grade-comment fusion feature are weighted and summed using this gating weight parameter to obtain the emotionless training behavior feature. This application adopts a dynamic gating mechanism to aggregate the feature representations learned in steps S40 and S50. The feature aggregation calculation process is shown in the following formula.

[0092]

[0093]

[0094] in, and These represent the learnable model parameters, initialized using a normal distribution function; Indicates the gating weight parameter; This represents the relationship between the integrated training code and the relevant course grades, i.e., the code-grade fusion feature; This represents the characteristics of multi-source training behavior after fusion, i.e., the characteristics of emotionless training behavior; This represents the relationship between integrated comments and related course grades, i.e., grade-comment integration features.

[0095] In this embodiment, the aggregation weights of the two types of fusion features are dynamically adjusted through a gating mechanism, which solves the problem that the differences in the contribution of different modal fusion features to the final score cannot be adaptively captured.

[0096] The above are merely feasible implementations of step S60 provided in this embodiment. This embodiment does not specifically limit the specific implementation of step S60.

[0097] Step S70: Based on the emotionless training behavior characteristics and the emotion feature vector, obtain the comprehensive training evaluation score.

[0098] It should be noted that the comprehensive evaluation score for practical training refers to the final output score used to quantify students' comprehensive practical training ability. This result is obtained by nonlinearly mapping multi-source behavioral features that incorporate emotional information using a multilayer perceptron.

[0099] Understandably, since the emotionless training behavior features only include three types of information—code behavior, course grades, and comments—while the emotional tendency contained in the evaluation report reflects the subjective attitude towards the training performance, this emotional information and the semantic information of the training content belong to different dimensions. Combining the two can more completely represent all the information in the evaluation report. Therefore, step S70 is performed, by introducing the emotional features and inputting them together into the multilayer perceptron for score prediction. This allows the final training evaluation score to simultaneously reflect both the facts of the training behavior and the subjective attitude of the evaluator, avoiding the partial loss of report information during the fusion process.

[0100] In one feasible implementation, step S70 may include: concatenating the emotionless training behavior features and the emotion feature vector to obtain target training behavior features; inputting the target training behavior features into a preset multilayer perceptron to obtain a comprehensive training evaluation score, wherein the preset multilayer perceptron includes a first linear layer, a first nonlinear activation function layer, a second linear layer, and a second nonlinear activation function layer connected in sequence.

[0101] It should be noted that the concatenation operation combines the emotionless training behavioral features with the emotional feature vectors, joining them end-to-end to create a longer vector. The pre-defined multilayer perceptron refers to a feedforward network structure consisting of two linear transformation layers and two non-linear activation function layers connected sequentially, used to map input features to the final scoring output.

[0102] Specifically, the emotionless training behavior features and the emotional feature vectors are concatenated along the feature dimension to form the target training behavior features that contain both training behavior information and emotional information, as shown in the formula below.

[0103]

[0104] in, This represents the characteristics of multi-source training behavior after fusion, i.e., the characteristics of emotionless training behavior; This represents the sentiment feature vector in the training report; This refers to the multi-source training behavior characteristics after incorporating emotional features, i.e., the target training behavior characteristics.

[0105] The training behavior characteristics of the target are sequentially transformed linearly by the first linear layer of a pre-set multilayer perceptron, a nonlinear mapping is introduced by the first nonlinear activation function layer, and a linear transformation is performed again by the second linear layer, followed by a nonlinear mapping introduced by the second nonlinear activation function layer. The result is a scalar value, which is the comprehensive training evaluation score. The calculation process for predicting the comprehensive training evaluation score is shown in the formula below:

[0106] in, , , , Represents the learnable parameters. Represents a non-linear activation function. This indicates the predicted student comprehensive evaluation score. This refers to the multi-source training behavior characteristics after incorporating emotional features, i.e., the target training behavior characteristics.

[0107] In this embodiment, by splicing emotional features and using a multilayer perceptron to complete the nonlinear rating mapping, the problem of incomplete rating information caused by the inability of a single behavioral feature to reflect the evaluation attitude is solved.

[0108] The above are merely feasible implementations of step S70 provided in this embodiment. This embodiment does not specifically limit the specific implementation of step S70.

[0109] This embodiment provides a deep learning-based intelligent scoring method for multi-source training behavior fusion. Based on the training code interface image, a code feature vector is obtained; based on the training course grades, a grade feature vector is obtained; based on self-evaluation reports and teacher evaluation reports, comment feature vectors and sentiment feature vectors are obtained; based on the grade feature vectors and the code feature vectors, a code-grade fusion feature is obtained; based on the grade feature vectors and the comment feature vectors, a grade-comment fusion feature is obtained; based on the code-grade fusion feature and the grade-comment fusion feature, an emotionless training behavior feature is obtained; based on the emotionless training behavior feature and the sentiment feature vector, a comprehensive training evaluation score is obtained. By employing techniques to acquire multi-source training behavior data and extract image features, score numerical features, and comment text and sentiment features respectively, the score features can be used as guidance information to be cross-fused with code image features and comment features. The two types of fusion results are then further aggregated into a unified multi-source training behavior representation. This representation is then combined with sentiment features and input into a multilayer perceptron for end-to-end score prediction. This solves the technical problem of existing methods that lack the ability to model the potential coupling relationships between multi-source data such as interface images, historical scores, and comment texts, resulting in scores relying on fixed rules or subjective experience. This achieves objective comprehensive training score driven by the inherent correlation of data.

[0110] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the deep learning-based multi-source training behavior fusion intelligent scoring method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0111] This application also provides a deep learning-based multi-source training behavior fusion intelligent scoring device; please refer to [reference needed]. Figure 2 The deep learning-based multi-source training behavior fusion intelligent scoring device includes: The first feature extraction module 10 is used to obtain code feature vectors based on the training code interface images; The second feature extraction module 20 is used to obtain the performance feature vector based on the practical training course scores; The third feature extraction module 30 is used to obtain comment feature vectors and sentiment feature vectors based on self-evaluation reports and teacher evaluation reports; The first feature fusion module 40 is used to obtain code-score fusion features based on the score feature vector and the code feature vector; The second feature fusion module 50 is used to obtain a grade-comment fusion feature based on the grade feature vector and the comment feature vector; The third feature fusion module 60 is used to obtain emotionless training behavior features based on the code-score fusion features and the score-comment fusion features; The evaluation score output module 70 is used to obtain a comprehensive evaluation score for the training based on the emotionless training behavior characteristics and the emotion feature vector.

[0112] The deep learning-based multi-source training behavior fusion intelligent scoring device provided in this application employs the deep learning-based multi-source training behavior fusion intelligent scoring method described in the above embodiments. It addresses the technical problem of how to achieve objective comprehensive training scoring without relying on fixed rules and subjective judgment by fusing multi-source training behavior data and mining their inherent correlations. Compared with existing technologies, the beneficial effects of the deep learning-based multi-source training behavior fusion intelligent scoring device provided in this application are the same as those of the deep learning-based multi-source training behavior fusion intelligent scoring method provided in the above embodiments. Furthermore, other technical features in the deep learning-based multi-source training behavior fusion intelligent scoring device are the same as those disclosed in the methods of the above embodiments, and will not be elaborated upon here.

[0113] The first feature extraction module 10 is further used to perform standardized cropping on the training code interface image to obtain a target code image; and to input the target code image into a preset visual model to obtain the code feature vector output by the preset visual model.

[0114] The second feature extraction module 20 is also used to normalize the practical training course scores to obtain normalized course scores; and input the normalized course scores into a two-layer nonlinear function network to obtain the score feature vector extracted by the two-layer nonlinear function network from the normalized course scores.

[0115] The third feature extraction module 30 is further configured to obtain evaluation report content based on the self-evaluation report and the teacher evaluation report; input the evaluation report content into a preset language model to obtain the comment feature vector output by the preset language model; and input the evaluation report content into a preset sentiment analysis tool to obtain the sentiment feature vector extracted by the preset sentiment analysis tool from the evaluation report content.

[0116] The first feature fusion module 40 is further configured to perform a linear transformation on the score feature vector to determine a first query vector; perform a linear transformation on the code feature vector to determine a first key vector and a first value vector; obtain a first attention weight based on the first query vector, the first key vector, and the first value vector; and perform a weighted summation on the code feature vector and the score feature vector according to the first attention weight to obtain a code-score fusion feature.

[0117] The second feature fusion module 50 is further configured to perform a linear transformation on the grade feature vector to determine a second query vector; perform a linear transformation on the comment feature vector to determine a second key vector and a second value vector; obtain a second attention weight based on the second query vector, the second key vector, and the second value vector; and perform a weighted summation on the grade feature vector and the comment feature vector according to the second attention weight to obtain a grade-comment fusion feature.

[0118] The third feature fusion module 60 is further configured to obtain gating weight parameters based on the code-score fusion features and the score-comment fusion features; and to obtain emotionless training behavior features based on the gating weight parameters, the code-score fusion features, and the score-comment fusion features.

[0119] The evaluation score output module 70 is also used to concatenate the emotionless training behavior features and the emotion feature vector to obtain the target training behavior features; input the target training behavior features into a preset multilayer perceptron to obtain a comprehensive training evaluation score, wherein the preset multilayer perceptron includes a first linear layer, a first nonlinear activation function layer, a second linear layer, and a second nonlinear activation function layer connected in sequence.

[0120] This application provides a deep learning-based multi-source training behavior fusion intelligent scoring device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the deep learning-based multi-source training behavior fusion intelligent scoring method in Embodiment 1 above.

[0121] The following is for reference. Figure 3 This document illustrates a structural diagram of a deep learning-based multi-source training behavior fusion intelligent scoring device suitable for implementing embodiments of this application. The deep learning-based multi-source training behavior fusion intelligent scoring device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), tablets, PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 3 The deep learning-based multi-source training behavior fusion intelligent scoring device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0122] like Figure 3 As shown, the deep learning-based multi-source training behavior fusion intelligent scoring device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the deep learning-based multi-source training behavior fusion intelligent scoring device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the deep learning-based multi-source training behavior fusion intelligent scoring device to wirelessly or wiredly communicate with other devices to exchange data. Although the figure shows a deep learning-based multi-source training behavior fusion intelligent scoring device with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0123] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0124] The deep learning-based multi-source training behavior fusion intelligent scoring device provided in this application, employing the deep learning-based multi-source training behavior fusion intelligent scoring method described in the above embodiments, solves the technical problem of how to achieve objective comprehensive training scoring without relying on fixed rules and subjective judgment by fusing multi-source training behavior data and mining their inherent correlations. Compared with the prior art, the beneficial effects of the deep learning-based multi-source training behavior fusion intelligent scoring device provided in this application are the same as those of the deep learning-based multi-source training behavior fusion intelligent scoring method provided in the above embodiments, and other technical features in this deep learning-based multi-source training behavior fusion intelligent scoring device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0125] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0126] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0127] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the deep learning-based multi-source training behavior fusion intelligent scoring method in the above embodiments.

[0128] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0129] The aforementioned computer-readable storage medium may be included in a deep learning-based multi-source training behavior fusion intelligent scoring device; or it may exist independently and not be assembled into a deep learning-based multi-source training behavior fusion intelligent scoring device.

[0130] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a deep learning-based multi-source training behavior fusion intelligent scoring device, the deep learning-based multi-source training behavior fusion intelligent scoring device: obtains a code feature vector based on the training code interface image; obtains a performance feature vector based on the training course score; obtains a comment feature vector and an emotion feature vector based on the self-evaluation report and the teacher evaluation report; obtains a code-performance fusion feature based on the performance feature vector and the code feature vector; obtains a performance-comment fusion feature based on the performance feature vector and the comment feature vector; obtains an emotionless training behavior feature based on the code-performance fusion feature and the performance-comment fusion feature; and obtains a comprehensive training evaluation score based on the emotionless training behavior feature and the emotion feature vector.

[0131] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0132] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0133] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0134] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned deep learning-based multi-source training behavior fusion intelligent scoring method. This method solves the technical problem of how to achieve objective comprehensive training scoring without relying on fixed rules and subjective judgment by fusing multi-source training behavior data and mining their inherent correlations. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the deep learning-based multi-source training behavior fusion intelligent scoring method provided in the above embodiments, and will not be elaborated upon here.

[0135] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the deep learning-based multi-source training behavior fusion intelligent scoring method described above.

[0136] The computer program product provided in this application can solve the technical problem of how to achieve objective comprehensive training scoring without relying on fixed rules and subjective judgment by fusing multi-source training behavior data and mining their inherent correlations. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the deep learning-based multi-source training behavior fusion intelligent scoring method provided in the above embodiments, and will not be repeated here.

[0137] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A deep learning-based intelligent scoring method for multi-source training behavior fusion, characterized in that, The method includes: Based on the images of the training code interface, code feature vectors are obtained; Based on the practical training course grades, a grade feature vector is obtained; Based on self-evaluation reports and teacher evaluation reports, we obtained comment feature vectors and affective feature vectors. Based on the performance feature vector and the code feature vector, a code-performance fusion feature is obtained; Based on the performance feature vector and the comment feature vector, a performance-comment fusion feature is obtained; Based on the code-score fusion feature and the score-comment fusion feature, the emotionless training behavior features are obtained; Based on the emotionless training behavior characteristics and the emotional feature vector, the comprehensive evaluation score of the training is obtained.

2. The method as described in claim 1, characterized in that, The code feature vector obtained based on the training code interface image includes: The training code interface image is standardized and cropped to obtain the target code image; The target code image is input into a preset visual model to obtain the code feature vector output by the preset visual model.

3. The method as described in claim 1, characterized in that, The performance feature vector obtained based on practical training course grades includes: Normalize the practical training course grades to obtain normalized course grades; The normalized course scores are input into a two-layer nonlinear function network to obtain the score feature vector extracted by the two-layer nonlinear function network from the normalized course scores.

4. The method as described in claim 1, characterized in that, The evaluation feature vector and sentiment feature vector, obtained based on self-evaluation reports and teacher evaluation reports, include: Based on the self-evaluation report and the teacher evaluation report, the content of the evaluation report is obtained; The evaluation report content is input into a preset language model to obtain the comment feature vector output by the preset language model; The evaluation report content is input into a preset sentiment analysis tool to obtain the sentiment feature vector extracted by the preset sentiment analysis tool from the evaluation report content.

5. The method as described in claim 1, characterized in that, The process of obtaining code-performance fusion features based on the performance feature vector and the code feature vector includes: Perform a linear transformation on the performance feature vector to determine the first query vector; Perform a linear transformation on the code feature vector to determine the first key vector and the first value vector; Based on the first query vector, the first key vector, and the first value vector, the first attention weight is obtained; The code feature vector and the score feature vector are weighted and summed according to the first attention weight to obtain the code-score fusion feature.

6. The method as described in claim 1, characterized in that, The process of obtaining grade-comment fusion features based on the grade feature vector and the comment feature vector includes: Perform a linear transformation on the performance feature vector to determine the second query vector; Perform a linear transformation on the comment feature vector to determine the second key vector and the second value vector; Based on the second query vector, the second key vector, and the second value vector, the second attention weight is obtained; The performance feature vector and the comment feature vector are weighted and summed according to the second attention weight to obtain the performance-comment fusion feature.

7. The method as described in claim 1, characterized in that, The emotionless training behavior characteristics obtained based on the code-grade fusion features and the grade-comment fusion features include: Based on the code-score fusion feature and the score-comment fusion feature, the gating weight parameters are obtained; Based on the gating weight parameters, the code-score fusion features, and the score-comment fusion features, emotionless training behavior features are obtained.

8. The method as described in claim 1, characterized in that, The comprehensive evaluation score for the training is obtained based on the emotionless training behavior characteristics and the emotion feature vector, including: The target training behavior features are obtained by concatenating the emotionless training behavior features and the emotion feature vector. The target training behavior features are input into a preset multilayer perceptron to obtain a comprehensive training evaluation score. The preset multilayer perceptron includes a first linear layer, a first nonlinear activation function layer, a second linear layer, and a second nonlinear activation function layer connected in sequence.

9. A deep learning-based intelligent scoring device for multi-source training behavior fusion, characterized in that, The device includes: The first feature extraction module is used to obtain code feature vectors based on the training code interface images; The second feature extraction module is used to obtain the performance feature vector based on the practical training course scores; The third feature extraction module is used to obtain comment feature vectors and sentiment feature vectors based on self-evaluation reports and teacher evaluation reports; The first feature fusion module is used to obtain code-score fusion features based on the score feature vector and the code feature vector; The second feature fusion module is used to obtain the grade-comment fusion feature based on the grade feature vector and the comment feature vector; The third feature fusion module is used to obtain emotionless training behavior features based on the code-score fusion features and the score-comment fusion features; The evaluation score output module is used to obtain a comprehensive evaluation score for the training based on the emotionless training behavior characteristics and the emotion feature vector.

10. A deep learning-based multi-source training behavior fusion intelligent scoring device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the deep learning-based multi-source training behavior fusion intelligent scoring method as described in any one of claims 1 to 8.

Images