Education evaluation and feedback system based on artificial intelligence

Abstract

The invention, which relates to the technical field of artificial intelligence, discloses an artificial intelligence-based education evaluation and feedback system comprising a data acquisition module, a vector generation module, a prediction module, an error region positioning module and a feedback module. The system constructs a unified high-dimensional cognitive state vector by collecting answering behaviors, eye movement tracks, facial micro-expressions, voices and intonations and electroencephalogram signals of students; generating a learning evolution path map based on a dynamic Bayesian network and a causal reasoning mechanism, and predicting future learning bottleneck nodes; an error region is recognized through semantic deconstruction and graph matching, and context-associated personalized feedback content is generated in combination with a generative language model; according to the system, an evaluation feedback closed loop of cognitive state modeling, accurate identification of an erroneous region and intelligent feedback pushing is realized, and the accuracy of education evaluation and the effectiveness of intervention are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to an AI-based educational assessment and feedback system. Background Technology

[0002] Current mainstream educational assessment systems mostly assess students' learning status based on answer accuracy, behavioral trajectory analysis, or static knowledge graph matching. Although they can achieve a certain degree of automated analysis, they still have the following shortcomings: First, the assessment dimensions are singular, lacking a deep understanding of students' emotional state, cognitive load, and thinking paths; second, the feedback mechanism is fixed, often based on preset templates to output results, lacking context adaptability; third, the assessment model training relies on standard datasets, making it difficult to adapt to the dynamic differences among individual students, resulting in insufficient personalization and effectiveness of feedback.

[0003] Existing technologies have not yet solved the modeling challenges between "changes in cognitive state" and "cross-modal data fusion." In particular, when students encounter cognitive impairments or experience learning anxiety, traditional systems cannot identify their non-knowledge-based difficulties. Furthermore, existing feedback systems lack the ability to dynamically adjust content based on causal graph reasoning and generative logic, resulting in high repetition and low inspiration in the feedback content.

[0004] This invention proposes an educational assessment method that integrates micro-expression recognition, EEG data analysis, and contextual learning path modeling. For the first time, it introduces a dynamic reasoning module based on cognitive intent recognition into a teaching feedback system, and combines it with a generative language model to customize feedback schemes in real time. This allows the system to proactively predict potential errors and provide highly adaptive feedback before students are even aware of their learning difficulties. This system not only breaks through the traditional "outcome-oriented" assessment method but also establishes a cross-modal and cross-temporal cognitive regulation assessment framework, demonstrating high creativity and technological advancement. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based educational assessment and feedback system to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-based education assessment and feedback system, comprising:

[0007] The data acquisition module is used to collect multi-source heterogeneous learning behavior data of students during the learning process. The data includes answering behavior, eye movement trajectory, facial micro-expression sequence, voice tone change and low-frequency EEG signal.

[0008] The vector generation module is used to build a cross-modal embedding model, which maps heterogeneous learning behavior data to a unified high-dimensional cognitive state space and generates student cognitive state vectors.

[0009] The prediction module is used to build a student learning evolution path map based on dynamic Bayesian networks and causal reasoning mechanisms, capture the trend of cognitive state changes over time, and then predict future learning bottleneck nodes.

[0010] The misconception localization module is used to semantically deconstruct students' cognitive state vectors and link them with the course knowledge graph to generate high-confidence misconception localization results.

[0011] The feedback module is used to dynamically generate feedback content based on the misconception location results by calling a generative language model. The feedback content integrates the student's current emotional state, past misconception patterns, and learning strategy preferences to achieve context-related intelligent push.

[0012] Preferably, the vector generation module specifically includes:

[0013] A dual-channel transform attention network was used to perform temporal encoding and feature reconstruction on students’ facial micro-expression sequences and speech intonation data, respectively, to extract emotional dynamic feature vectors.

[0014] Using a graph structure attention fusion module, the emotional dynamic features are combined with low-frequency EEG signals, eye movement trajectories, and question-answering behavior data to construct a heterogeneous graph. Semantic transfer weights and temporal synchronization factors are introduced between nodes to form a cross-modal relationship graph.

[0015] Based on a self-supervised contrastive learning strategy, the cross-modal relationship graph is projected in high dimension to construct a unified cognitive state space, and student cognitive state vectors are generated through spatial alignment embedding.

[0016] Preferably, the prediction module specifically includes:

[0017] Based on the historical cognitive state vector sequence, a dynamic Bayesian network structure is constructed, where each node represents a cognitive state variable at a fixed time, and the edge weights represent the transition probabilities between states.

[0018] The knowledge point topology information and learning strategy sequence are embedded in the formed dynamic Bayesian network structure. The student learning evolution path map is generated through graph-temporal joint modeling to capture the dynamic evolution trend of their cognitive state with time and content dimensions.

[0019] Preferably, predicting future learning bottlenecks includes:

[0020] The node sequence in the student learning evolution path graph is input into a recursive graph attention network to aggregate temporal evolution features and knowledge topology to generate a set of node-level predictive representation vectors.

[0021] The bottleneck perception metric model is used to score the predicted representation vector. The bottleneck perception metric model combines the cognitive decay rate, knowledge jump rate and learning stagnation index of the node to output the confidence score of the node becoming a bottleneck in the future.

[0022] Based on the scoring results, a bottleneck risk ranking list is constructed, and high-risk learning bottleneck nodes are marked by combining the changes in the relevance of knowledge points within the prediction time window.

[0023] Preferably, the error localization module specifically includes:

[0024] A multi-layer semantic decomposition network is used to deconstruct the student cognitive state vector semantically, and the overall cognitive state vector is divided into three types of feature sub-vectors: concept cognitive factors, cognitive bias indicators, and erroneous reasoning patterns.

[0025] The sub-vectors are input into the graph attention matching module, and are aligned with the semantic embedding vectors of each knowledge node in the course knowledge graph based on similarity. A misconception risk graph is then established based on knowledge path accessibility and the propagation path of misleading factors.

[0026] Based on the misconception confidence weights of nodes in the graph and the cognitive offset of edges, multiple rounds of confidence aggregation operations are performed to filter out knowledge nodes that are highly coupled with cognitive biases, which are then output as the current misconception location results for students.

[0027] Preferably, the feedback module specifically includes:

[0028] Construct a multi-source context-aware vector, which integrates the student's current cognitive state vector, real-time emotional dynamic vector, error evolution trajectory, and personalized learning strategy embedding representation;

[0029] The error localization results and context-aware vectors are input into a generative language model based on a prompt adjustment mechanism. The target function is reconstructed through a context guidance layer to guide the generative language model to generate natural language feedback content.

[0030] The generated feedback content is scored for semantic accuracy and strategy fit using a feedback adaptability discriminator. High-adaptability feedback is retained and pushed to the user interface in a multimodal format according to students' acceptance habits.

[0031] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0032] 1. This invention constructs an AI-based educational assessment and feedback system, integrating modules such as multi-source heterogeneous data acquisition, cognitive state vector modeling, dynamic causal reasoning, and generative feedback generation to form a closed-loop architecture from "perception—assessment—prediction—feedback—optimization." The system can acquire students' behavior, emotions, and EEG data in real time, accurately identify potential pitfalls in the learning process, and combine individualized learning strategies to achieve highly adaptive feedback delivery. This significantly improves the accuracy of assessment and the intervention effect of feedback, solving problems such as assessment lag, templated feedback, and imperceptible cognitive states in traditional education systems.

[0033] 2. This invention introduces a multi-layer semantic deconstruction network, graph-structured attention fusion, a recursive graph attention prediction mechanism, and a generative language feedback model based on prompt adjustment, constructing a complete path from cognitive state vector to misconception localization and then to dynamic feedback generation. The system possesses high scalability and intelligent adaptability, not only improving the accuracy of misconception localization and the foresight of learning bottleneck prediction, but also significantly enhancing the contextual relevance and strategy fit of the feedback content. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0035] Figure 1 This is a mind map of the system modules of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1, please refer to Figure 1 As shown in this embodiment, the AI-based educational assessment and feedback system includes:

[0038] The data acquisition module is used to collect multi-source heterogeneous learning behavior data of students during the learning process. The data includes answering behavior, eye movement trajectory, facial micro-expression sequence, voice tone change and low-frequency EEG signal.

[0039] The vector generation module is used to build a cross-modal embedding model, which maps heterogeneous learning behavior data to a unified high-dimensional cognitive state space and generates student cognitive state vectors.

[0040] The prediction module is used to build a student learning evolution path map based on dynamic Bayesian networks and causal reasoning mechanisms, capture the trend of cognitive state changes over time, and then predict future learning bottleneck nodes.

[0041] The misconception localization module is used to semantically deconstruct students' cognitive state vectors and link them with the course knowledge graph to generate high-confidence misconception localization results.

[0042] The feedback module is used to dynamically generate feedback content based on the misconception location results by calling a generative language model. The feedback content integrates the student's current emotional state, past misconception patterns, and learning strategy preferences to achieve context-related intelligent push.

[0043] In this embodiment, the data acquisition module is used to collect multi-source heterogeneous learning behavior data in real time during students' learning activities, providing a raw data foundation for subsequent cognitive state modeling and misconception analysis. This module includes the following subsystems and implementation paths:

[0044] The question-answering behavior collection submodule uses an embedded interactive learning terminal to record students' answering time, option click paths, modification behaviors, frequency of repeatedly viewing question stems, and other operation logs in real time. The logs are stored in a timestamp format for behavior analysis and response pattern recognition.

[0045] The eye-tracking acquisition submodule deploys a miniature infrared eye-tracking device above the learning terminal to capture the student's gaze point, fixation duration, saccade path, and number of return glances. The acquisition frequency is set to 60Hz to ensure the temporal continuity of dynamically changing data and to synchronize with the screen partitioning of learning content, establishing a correspondence between gaze and knowledge points.

[0046] The facial micro-expression acquisition submodule captures sequences of students' facial expression images using a high-definition camera. Combined with a micro-expression recognition model based on a convolutional neural network, it extracts key action units (AUs) and their dynamic intensity to identify subtle emotional states such as confusion, anxiety, and understanding, providing input for subsequent emotion regulation analysis.

[0047] The speech intonation change acquisition submodule uses the terminal microphone to record students' speech data in real time when they are giving oral answers or reading tasks. It extracts tone, speech rate, pause patterns and stress distribution, and uses an acoustic emotion recognition network to label the speech emotion features, and identifies tone fluctuations and emotional tension points in the learning process.

[0048] The low-frequency EEG signal acquisition submodule uses a portable, non-invasive EEG head-mounted device (such as wearable hardware based on EEG) to collect low-frequency EEG signals such as theta waves and alpha waves while students are learning. It combines filtering and artifact removal algorithms to extract stable frequency band information and synchronizes it with event tags to assess students' attention status and cognitive load level.

[0049] All of the above sub-modules are connected to the main system's data management platform through a unified time synchronization mechanism and data interface. They are encoded in a unified format, tagged, and then sent to the feature extraction module for multimodal fusion processing to ensure the consistency of the collected data in terms of time axis and semantic level.

[0050] In a preferred embodiment of the present invention, the vector generation module is used to map multi-source heterogeneous learning behavior data to a unified high-dimensional cognitive state space and generate a student cognitive state vector (CSV). This module consists of a dual-channel transform attention network, a graph-structured attention fusion module, and a self-supervised contrastive learning mapping unit. It aims to achieve deep feature fusion and semantic alignment of emotion, behavior, and neural data across different modalities, providing a unified vector representation foundation for subsequent cognitive modeling, misconception detection, and personalized feedback.

[0051] This module first models two dynamic emotional signals that are highly correlated in the human cognitive process: facial micro-expression sequences and speech tone data.

[0052] Using continuous video frames as input, the static spatial features of the images are first extracted using ResNet-50 as the basic feature extraction network. Then, Bi-LSTM (Bidirectional Long Short-Term Memory) network is used to dynamically model the micro-expression changes in the time series. To enhance the keyframe recognition capability, a temporal channel transformation attention mechanism is introduced to automatically identify frames with drastic emotional fluctuations and increase their weights, outputting a micro-expression temporal embedding vector Vface.

[0053] The acquired speech signal is converted into a Mel spectrogram. After local speech features are extracted by the convolutional acoustic feature extraction module, the signal is input into a temporal convolutional network (TCN) combined with an attention mechanism to capture pitch fluctuation features. At the same time, auxiliary feature vectors such as basic pitch, energy, and speech rate are extracted and concatenated with the main spectrogram features through a feature fusion layer to form the speech dynamic vector Vvoice.

[0054] Finally, the feature vectors of the facial and speech channels are unified in dimension through a fully connected layer and fed into a joint embedding layer to generate the emotion dynamic feature vector Vemo.

[0055] To heterogeneously fuse emotional dynamic features with other modal learning behavior data (low-frequency EEG signals, eye movement trajectories, and answering behavior) and construct a graph structure representation reflecting students' multidimensional cognitive states, the system is designed with the following module steps:

[0056] A heterogeneous graph G=(V,E) is constructed, where the node set V includes five types of data source nodes: emotion vector node Vemo, EEG signal node Veeg, eye movement behavior node Veye, question-answering trajectory node Vquiz, and auxiliary state node Vaux. Each type of node is initialized with an independent encoder for embedding representation. For example, EEG signals are composed of wavelet packet energy features + FFT frequency domain features, eye movement trajectories are composed of gaze density map convolution features, and question-answering behavior is composed of click sequence converted to RNN output features.

[0057] Semantic relation edges E are established between nodes, and two types of non-explicit weights are introduced:

[0058] Semantic transfer weight ws: used to measure the difficulty and intensity of information transfer between different modalities, estimated based on mutual information between historical data.

[0059] Timing synchronization factor wt: captures the alignment degree of timestamps of modal data, and uses DTW to calculate the alignment cost of different modalities.

[0060] The edge weights are calculated by weighting and summing the semantic transfer weights and the temporal synchronization factor.

[0061] Based on the graph structure described above, a Graph Attention Network (GAT) is used to aggregate information from heterogeneous nodes. The final representation of each node... It is calculated by averaging the weighted attention values ​​represented by its neighboring nodes. The graph attention output forms a cross-modal embedding representation Hgraph, which is the comprehensive state graph embedding of each student at the current learning stage.

[0062] To project the Hgraph onto a unified cognitive state space R and generate the final student cognitive state vector CSV, the system employs the following self-supervised contrastive learning strategy:

[0063] The state embeddings of the same student within adjacent time windows are used as positive sample pairs, while the states of different students or at different times are used as negative samples. A momentum encoder is added to enhance temporal stability.

[0064] The design employs a multilayer perceptron mapper to map graph representations (Hgraph) to a cognitive state space. After training, each graph embedding (Hgraph) is stably mapped to a unique student cognitive state vector (CSV), with its dimension d set empirically to 128 or 256 dimensions based on the training data. The CSV contains a comprehensive encoding of the student's current state, including emotion, attention, knowledge acquisition, and operational behavior.

[0065] In this invention, the prediction module is used to perform time-series modeling of students' cognitive states during the learning process, identify the evolution trend of their learning states, and predict and label potential learning bottleneck nodes based on the evolution path map.

[0066] The system first receives the student cognitive state vector sequence CSVt output from the vector generation module, where CSVt represents the student's high-dimensional cognitive state representation at time t.

[0067] To model the transition characteristics of states over time, a dynamic Bayesian network structure is constructed.

[0068] Each node Xt represents a cognitive state variable at time t, and its state is transformed into a finite set of states by CSVt through a discretization function.

[0069] The directed edge Xt-1 represents the dependency between temporally adjacent states. Its edge weight is calculated by maximum likelihood estimation, and a time window smoothing mechanism is introduced to reduce the impact of single-point abnormal fluctuations.

[0070] To enhance the model's structural adaptability, a structure learning strategy is incorporated, such as a Bayesian structure optimization algorithm based on the K2 scoring function. This allows the network topology to not only capture first-order transition probabilities but also permit the existence of some cross-temporal influence relationships. This DBN structure is used to quantify the temporal evolution of students' cognitive states, providing a time-dependent basis for subsequent graph structure modeling.

[0071] Based on the construction of a dynamic Bayesian network, the course knowledge graph and learning strategy sequence are further integrated to form a student learning evolution graph (LEG) that can represent the three-dimensional relationship of cognition-knowledge-strategy. The construction process is as follows:

[0072] In the DBN structure, each time node Xt is extended and attached to the corresponding knowledge point node Ki in the knowledge graph, forming a (Xt, Ki) tuple; by traversing the content learned by the student at time t and their cognitive state, a time-series node set NLEG is established.

[0073] The learning strategy sequence (such as repeated learning, skipping browsing, question type switching, etc.) is encoded as a strategy edge Es and embedded into the original graph to supplement the causal bridge between cognitive state transfer and knowledge point access.

[0074] The resulting LEG graph possesses a triple structural information fusion characteristic: temporal dependence, knowledge topology, and behavioral intent, providing an interpretable semantic basis for bottleneck prediction.

[0075] The structural evolution patterns and potential bottleneck indicators between nodes are extracted from the LEG graph. A Recurrent Graph Attention Network (R-GAT) is then used for node state prediction modeling. The steps are as follows:

[0076] For each node in the graph (representing the cognitive state of a specific knowledge point at a certain moment), an initial embedding vector is constructed, which is then combined with historical states, knowledge point embeddings, and policy encoding representations.

[0077] The R-GAT model introduces a recursive structure on top of the traditional graph attention mechanism, and finally outputs the prediction vector of each node to determine its future state trend.

[0078] To identify potential future bottleneck nodes, a bottleneck awareness metric model is introduced, mapping node prediction vectors to quantifiable bottleneck risk scores. The specific steps are as follows:

[0079] Define three types of indicators:

[0080] Cognitive decay rate: This represents the rate at which a student's cognitive state regarding a knowledge point declines over time, and is calculated from the projection difference of CSVt at that node;

[0081] Knowledge jump degree: measures the topological jump distance between this node and the previous node in the knowledge graph. The larger the jump degree, the more abrupt the cognitive difficulty.

[0082] Learning stagnation index: An indicator based on repeated learning over multiple periods at a given node without significant improvement in the learning status.

[0083] The three types of indicators, after standardization, are input into the bottleneck perception measurement model, which is a machine learning model, specifically including:

[0084] Cognitive decay rate, knowledge leap rate, and learning stagnation index are converted into a comprehensive feature vector. This comprehensive feature vector is used as the input to a machine learning model. The machine learning model uses the prediction of bottleneck risk score labels for each set of comprehensive feature vectors as its prediction objective and minimizes the sum of prediction errors for all bottleneck risk score labels as its training objective. The machine learning model is trained until the sum of prediction errors converges, at which point the model training stops. The bottleneck risk score is determined based on the model output. The machine learning model is a multinomial regression model.

[0085] Sort all nodes in descending order of bottleneck risk score to form a bottleneck risk ranking list. Combined with the trend of changes in the correlation between knowledge points within the prediction time window, mark nodes with confidence scores higher than a set threshold as high-risk bottleneck nodes.

[0086] In this invention, the misconception localization module is used to structurally identify potential cognitive biases in students' current learning state. Based on deep matching between the cognitive state vector and the course knowledge graph, it outputs high-confidence misconception localization results. This module specifically consists of a multi-layer semantic decomposition network, a graph attention matching module, and a misconception risk graph construction and aggregation module. The overall implementation path is as follows:

[0087] The module receives a student cognitive state vector CSV output by the vector generation module. This vector is a high-dimensional dense representation that integrates the student's emotional state, behavioral patterns, knowledge mastery, and cognitive dynamics at the current learning stage.

[0088] To identify hidden misunderstanding structures in cognitive states, the system introduces a multi-layer semantic decomposition network. This network, through multi-head attention and multi-scale representation learning, deconstructs the original CSV vector layer by layer into three feature sub-vectors:

[0089] Conceptual cognition factor sub-vector: represents the student's level of understanding of specific concepts, including the degree of mastery of the knowledge points' structure and the accessibility of the content;

[0090] Cognitive bias index sub-vector: used to measure whether students have systematic misunderstandings, such as conceptual confusion, incorrect judgment of the scope of application of knowledge, etc.

[0091] Error reasoning pattern sub-vectors: These reflect common deviations in logical reasoning paths that students encounter during problem-solving, including reversed cause and effect, ignored conditions, or failed association transfer.

[0092] Each sub-vector is represented in a unified dimensional space, ensuring comparability in subsequent matching processes. The semantic decomposition network is self-supervised by misclassification labels from historical samples to improve decomposition quality and misjudgment tolerance.

[0093] The course knowledge graph consists of multiple knowledge point nodes and the topological relationship edges between them. Each node carries a knowledge semantic embedding vector, which is pre-trained on course content, teaching cases and student answer samples.

[0094] The three sub-vectors are used as query vectors and input into the graph attention matching module in sequence. The module performs high-dimensional semantic matching operation in the knowledge graph with each sub-vector as the center, and identifies the set of knowledge nodes that may be associated with it through attention weight allocation.

[0095] The attention matching process not only calculates the semantic similarity between vectors, but also introduces two types of structural auxiliary factors:

[0096] Knowledge path accessibility factor: Represents the density and directionality of the traversable paths of the current query subvector in the knowledge graph, used to measure the degree of association between the subvector and the knowledge structure;

[0097] Misleading factor propagation path length: This represents the shortest path length from the current query subvector that may mislead to the wrong knowledge node. The shorter the path, the higher the risk of misleading propagation.

[0098] The two factors mentioned above participate in the adjustment process of attention weights, improving the structural rationality of the matching results and the ability to identify potential misconceptions.

[0099] After semantic matching, the system generates a misconception risk graph, where nodes are activated knowledge nodes in the graph, and edges represent the propagation paths of the aforementioned misleading factors. Each node is assigned an initial misconception confidence weight, which consists of the following three parts:

[0100] Similarity scores between cognitive bias index subvectors and node semantic embeddings;

[0101] Misleading propagation centrality of nodes in the path structure;

[0102] The frequency of this node appearing as a misconception in historical samples is weighted statistically.

[0103] Each edge is assigned a cognitive offset factor, reflecting the degree to which a student deviates from the standard cognitive path on that path. The cognitive offset is determined by the following factors: the difference between the current learning strategy and the path, the semantic span between nodes, and the abruptness of the change in cognitive state.

[0104] To ensure stable and accurate misidentification, the system performs at least two rounds of confidence aggregation calculations on the risk map. Each round of aggregation includes:

[0105] The diffusion and propagation of confidence weights between nodes;

[0106] Edge weight removal of weakly confident nodes;

[0107] Confidence enhancement for cluster center nodes.

[0108] After multiple rounds of aggregation, nodes with high confidence and activated by multiple paths will stand out significantly. The system will then select the node with the highest cognitive bias coupling as the output result for students to locate their misconceptions at the current stage.

[0109] In this invention, the feedback module is used to dynamically generate natural language feedback content with context adaptability and strategy fit based on the student's cognitive state, emotional dynamics, error evolution trajectory and personalized learning strategies, and intelligently push it in the most suitable way for the student to accept.

[0110] The student cognitive state vector CSV, output by the aforementioned vector generation module, is used as a representation of the student's overall cognitive performance at the current moment, including information such as their knowledge mastery, attention stability, and logical behavior characteristics.

[0111] By calling on the results from facial micro-expression and voice tone analysis, a real-time emotion dynamic vector is constructed. The vector content includes short-term emotion fluctuation values, the rate of change of emotion trends, and confidence values ​​of state labels such as anxiety / confusion / satisfaction.

[0112] The set of student misconception nodes for the current and recent stages is extracted from the misconception localization module. Combined with parameters such as the order of occurrence, frequency, and number of corrections, it is encoded into a misconception evolution trajectory vector, which represents the development process and ongoing risks of the current cognitive bias.

[0113] The strategies students employ in the process of learning history (such as repeated learning, skipping learning, and linking text and images) are generated into strategy vectors by the strategy encoder and embedded into a unified representation space through an individual adaptation layer.

[0114] The four types of vectors mentioned above are concatenated and then deeply fused through a nonlinear feature fusion layer (such as a residual gated fusion network) to output a multi-source context-aware vector in a unified format, abbreviated as UCA, which serves as the dynamic input basis for the feedback generation model.

[0115] Generative language models fine-tuned for educational scenarios are employed, such as autoregressive language models based on the Transformer architecture (e.g., T5 or OPT variants), which possess strong expressive power and natural language generation capabilities.

[0116] The error localization result and the UCA vector are used together as the model input. The input structure consists of two parts:

[0117] The structured prompt template section defines the tone, content scope, and style type (such as suggestive, guiding, or reflective) of the language model output.

[0118] The vector-level guidance layer uses UCA vectors as additional control signals for the hidden states in the generative language model, and adjusts the direction of attention shift during the attention calculation process through residual injection mechanism.

[0119] Traditional language models only optimize language fluency as their objective function. However, this system reconstructs the loss function by introducing three types of weighted terms: content relevance loss, cognitive fit loss, and strategy consistency loss, making the generation process more goal-oriented.

[0120] The guided model outputs structured natural language paragraphs that cover the student's main misconceptions, suggested corrective paths, and necessary learning strategy reminders. The language style is adjusted according to the context to enhance the student's understanding and acceptance.

[0121] To improve the relevance and adaptability of the final feedback, the system introduces an independent feedback adaptability discriminator, which performs the following scoring and filtering on each candidate feedback:

[0122] By using sentence vector matching and context feedback mechanisms, the generated text is evaluated to determine whether it accurately cites content related to misconceptions and whether it reasonably expresses students' emotions and cognitive biases. Semantic deviations or incomplete citations will be penalized with lower ranking.

[0123] The discriminator compares the suggested strategies in the generated content with the embedding vectors of students' historical learning strategies. If there is a significant mismatch in strategies (e.g., suggesting skipping exercises but students have very low acceptance), a negative label is given.

[0124] The model typically generates multiple candidate feedback texts. The system retains the first one or two with high semantic accuracy and strong strategy fit as the final feedback content and pushes them into the stage.

[0125] The system records the effectiveness of different forms of feedback (text, voice, images, etc.) received by students based on their historical usage habits, such as acceptance rate, dwell time, and click depth, and generates a feedback preference model.

[0126] The final text feedback content is converted into various forms such as voice explanation and flowchart through the speech synthesis module and the diagram generation module, and together with the original text version, they form a feedback candidate set.

[0127] The system adaptively selects appropriate feedback formats based on a feedback preference model and pushes them in real time through student learning platforms or large-screen interactive terminals. At the same time, it records students' response data to the feedback for subsequent evaluation.

[0128] Example 2: To verify the technical effectiveness of the "Artificial Intelligence-Based Educational Assessment and Feedback System" described in this invention, especially its comprehensive performance in terms of accuracy in misconception location, personalized feedback adaptation, and forward-looking prediction of bottleneck nodes, the applicant constructed an experimental platform based on a middle school mathematics learning scenario to collect data from real student samples and verify the model. The key experimental design and results are summarized below:

[0129] Eighty students were randomly selected from the first year of high school at a key high school in a certain city and divided into an experimental group (using the system of this invention) and a control group (using a conventional online homework platform), with 40 students in each group.

[0130] The experiment lasted for 14 consecutive days. Students completed daily learning tasks on mathematical knowledge points provided by the system (including videos, text reading, and practice questions), covering 38 knowledge points across two major modules: functions and geometry. The system generated daily feedback and evaluation reports.

[0131] Each student in the experimental group was equipped with a learning terminal that supports eye-tracking and EEG data acquisition. The front-facing high-definition camera and high-fidelity microphone were used to collect facial expression and voice data.

[0132] Indicator settings

[0133] For the errors identified by the system, three mathematics curriculum researchers were invited to manually verify the daily error identification results of 40 students in the experimental group. Statistics showed that 92.5% of the error identification results output by the system matched the manual assessment, a significant improvement compared to the control group platform (which only judged errors based on answer results, with an accuracy rate of approximately 74.1%).

[0134] Through behavioral log analysis and questionnaire feedback, 87.3% of the feedback from students in the experimental group was actively viewed or adopted, with audio explanations receiving the highest click-through rate. In contrast, the click-through rate for template-based feedback in the control group was only 41.8%, with most feedback being ignored. Students generally felt that "the system can point out error reasons I didn't think of" and that "the feedback is helpful."

[0135] Comparing the "high-risk bottleneck nodes within the next 3 days" predicted by the system with the actual points where students made mistakes, the accuracy rate within the prediction range was 76.4%, and the node was identified on average 2.1 days in advance, which is significantly earlier than the control group system, which could only provide feedback after the error occurred.

[0136] By comparing the test papers before and after the course, the average score of the experimental group increased from 72.1 to 86.7, an increase of 14.6 points; the average score of the control group increased by 7.4 points, an improvement rate of only about 50% of that of the experimental group. The avoidance of common mistakes in the function section was particularly effective, with the error rate decreasing by nearly 40%.

[0137] As can be seen from the above experimental data, the system of the present invention is significantly superior to the existing technology in terms of integrating multimodal data acquisition, cognitive state modeling, and error prediction, especially in the following aspects where it demonstrates significant beneficial effects:

[0138] By combining cognitive semantic deconstruction with knowledge graph matching, the accuracy of misconception localization is significantly improved; feedback based on context awareness and generative models is more personalized, has strong strategy adaptability, and is highly acceptable to students; by utilizing graph temporal reasoning mechanisms, high-risk learning nodes can be predicted in advance, allowing for early intervention and preventing knowledge gaps; overall learning efficiency and effectiveness are improved, achieving a closed loop from assessment to feedback.

[0139] In summary, this embodiment fully demonstrates the technical effects of the present invention from the perspective of empirical data, and has high practical application value and innovation.

[0140] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. An AI-based educational assessment and feedback system, characterized by: include: The data acquisition module is used to collect multi-source heterogeneous learning behavior data of students during the learning process. The data includes answering behavior, eye movement trajectory, facial micro-expression sequence, voice tone change and low-frequency EEG signal. The vector generation module is used to build a cross-modal embedding model, which maps heterogeneous learning behavior data to a unified high-dimensional cognitive state space and generates student cognitive state vectors. The prediction module is used to build a student learning evolution path map based on dynamic Bayesian networks and causal reasoning mechanisms, capture the trend of cognitive state changes over time, and then predict future learning bottleneck nodes. The misconception localization module is used to semantically deconstruct students' cognitive state vectors and link them with the course knowledge graph to generate high-confidence misconception localization results. The feedback module is used to dynamically generate feedback content based on the misconception location results by calling a generative language model. The feedback content integrates the student's current emotional state, past misconception patterns, and learning strategy preferences to achieve context-related intelligent push.

2. The AI-based educational assessment and feedback system according to claim 1, characterized in that: The vector generation module specifically includes: A dual-channel transform attention network was used to perform temporal encoding and feature reconstruction on students’ facial micro-expression sequences and speech intonation data, respectively, to extract emotional dynamic feature vectors. Using a graph structure attention fusion module, the emotional dynamic features are combined with low-frequency EEG signals, eye movement trajectories, and question-answering behavior data to construct a heterogeneous graph. Semantic transfer weights and temporal synchronization factors are introduced between nodes to form a cross-modal relationship graph. Based on a self-supervised contrastive learning strategy, the cross-modal relationship graph is projected in high dimension to construct a unified cognitive state space, and student cognitive state vectors are generated through spatial alignment embedding.

3. The AI-based educational assessment and feedback system according to claim 1, characterized in that: The prediction module specifically includes: Based on the historical cognitive state vector sequence, a dynamic Bayesian network structure is constructed, where each node represents a cognitive state variable at a fixed time, and the edge weights represent the transition probabilities between states. The knowledge point topology information and learning strategy sequence are embedded in the formed dynamic Bayesian network structure. The student learning evolution path map is generated through graph-temporal joint modeling to capture the dynamic evolution trend of their cognitive state with time and content dimensions.

4. The AI-based educational assessment and feedback system according to claim 3, characterized in that: Predicting future learning bottlenecks, specifically including: The node sequence in the student learning evolution path graph is input into a recursive graph attention network to aggregate temporal evolution features and knowledge topology to generate a set of node-level predictive representation vectors. The bottleneck perception metric model is used to score the predicted representation vector. The bottleneck perception metric model combines the cognitive decay rate, knowledge jump rate and learning stagnation index of the node to output the confidence score of the node becoming a bottleneck in the future. Based on the scoring results, a bottleneck risk ranking list is constructed, and high-risk learning bottleneck nodes are marked by combining the changes in the relevance of knowledge points within the prediction time window.

5. The AI-based educational assessment and feedback system according to claim 1, characterized in that: The misconception localization module specifically includes: A multi-layer semantic decomposition network is used to deconstruct the student cognitive state vector semantically, and the overall cognitive state vector is divided into three types of feature sub-vectors: concept cognitive factors, cognitive bias indicators, and erroneous reasoning patterns. The sub-vectors are input into the graph attention matching module and aligned with the semantic embedding vectors of each knowledge node in the course knowledge graph based on similarity. A misconception risk graph is then established based on knowledge path accessibility and the propagation path of misleading factors. Based on the misconception confidence weights of nodes in the graph and the cognitive offset of edges, multiple rounds of confidence aggregation operations are performed to filter out knowledge nodes that are highly coupled with cognitive biases, which are then output as the current misconception location results for students.

6. The AI-based educational assessment and feedback system according to claim 1, characterized in that: The feedback module specifically includes: Construct a multi-source context-aware vector, which integrates the student's current cognitive state vector, real-time emotional dynamic vector, error evolution trajectory, and personalized learning strategy embedding representation; The error localization results and context-aware vectors are input into a generative language model based on a prompt adjustment mechanism. The target function is reconstructed through a context guidance layer to guide the generative language model to generate natural language feedback content. The generated feedback content is scored for semantic accuracy and strategy fit using a feedback adaptability discriminator. High-adaptability feedback is retained and pushed to the user interface in a multimodal format according to students' acceptance habits.

Images