Online teaching intelligent auxiliary system and method fusing course knowledge and learning condition diagnosis

Abstract

This invention discloses an intelligent online teaching assistance system and method that integrates course knowledge and learning assessment. The system includes: a course information structured representation subsystem, a student ability dynamic diagnosis subsystem, and a dual-input informed assistance generation subsystem. The course information structured representation subsystem is used to form a structured course model. The student ability dynamic diagnosis subsystem is used to dynamically, stably, and interpretably infer the student's current ability status during the semester based on the structured course model and student learning behavior data. The dual-input informed assistance generation subsystem is used by an intelligent agent based on a large language model to generate teaching assistance content using the structured course model and the student's current ability status. In this embodiment of the invention, a complete technical chain of "course information structured representation - student ability dynamic diagnosis - informed assistance content generation" is established.

Description

Technical Field

[0001] This invention relates to the field of educational assistance technology, and in particular to an intelligent online teaching assistance system and method that integrates curriculum knowledge and learning progress diagnosis. Background Technology

[0002] With the rapid development of Large Language Model (LLM) technology, generative AI, represented by GPT, Claude, DeepSeek, and Qwen, has been widely applied in higher education. Current mainstream AI teaching aids adopt a simple integration model of "knowledge base + large model": unstructured text materials such as course handouts, PPTs, and textbooks are sliced ​​and stored quantitatively, relevant segments are retrieved through cosine similarity, and answers are generated by combining the general capabilities of large language models. Intelligent agent development platforms such as Dify and Coze have further lowered the technical threshold, enabling teachers to build teaching aid intelligent agents in a low-code manner. However, the introduction of technical tools does not automatically equate to an improvement in teaching quality.

[0003] The closest existing technology to this invention is the AI ​​teaching assistance system based on retrieval enhancement generation (RAG), whose technical solution includes the following steps: (1) Knowledge base construction: unstructured text materials such as course handouts, PPTs, and textbooks are sliced ​​and vectorized, and stored in a vector database (such as Milvus and Pinecone) to construct a course knowledge base; (2) General question and answer interaction: when a student asks a question, the system performs semantic similarity retrieval in the knowledge base, recalls Top-K related text fragments, and generates an answer by combining the general capabilities of the large language model; (3) Undifferentiated assistance: this solution provides the same question and answer service to all students without distinguishing the individual cognitive state and learning progress of students.

[0004] The aforementioned existing technical solutions suffer from a profound "technical susceptibility" dilemma, which manifests in the following three interrelated defects: (1) Semantic susceptibility (lack of structured understanding): Existing technologies cannot understand the knowledge topology (such as teaching semantic relationships such as predecessors, successors, associations, and inclusions) and the logic of ability progression within the course. RAG retrieval is essentially "keyword matching + semantic similarity," without considering the teaching logic of the course design. As a result, although the AI ​​answers are semantically fluent, they may deviate from the course outline and teaching objectives, and cannot provide targeted knowledge guidance; (2) Diagnostic susceptibility (lack of dynamic cognitive diagnosis): Existing technologies do not have the ability to dynamically diagnose the level of knowledge mastery of individual students. Traditional assessment methods (such as mid-term and final exams) have serious time lags and cannot support immediate intervention in the learning process. The AI ​​system "knows nothing" about students, which makes it impossible for its auxiliary output to achieve individualized instruction; (3) Interactive susceptibility (lack of personalized interaction): Due to the lack of structured understanding of the course and dynamic perception of students' status, the AI's auxiliary output is detached from the real teaching context and cannot be upgraded from "general question and answer" to "informed assistance." Students feel that AI is an external tool that "knows a lot but doesn't understand me," rather than a collaborator that truly integrates into the teaching process, resulting in low usage and difficulty in truly improving teaching effectiveness. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides an online teaching intelligent auxiliary system and method that integrates course knowledge and learning diagnosis, and opens up a complete technical link of "structured representation of course information - dynamic diagnosis of student ability - generation of informed assistance content"; and realizes the transformation from floating to embedded, from post-evaluation to dynamic diagnosis, from undifferentiated response to informed assistance, and ensures that the diagnostic results are educationally acceptable.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide an intelligent online teaching support system that integrates course knowledge and student learning diagnosis. The system includes: a structured representation subsystem for course information, a dynamic diagnosis subsystem for student abilities, and a dual-input informed consent generation subsystem; wherein... The course information structured representation subsystem is used to transform discrete course resources into structured evidence with pedagogical semantic associations and form a structured course model. The course information structured representation subsystem includes four parts: a course knowledge graph construction module, a competency achievement system definition module, a three-level association annotation module, and a weighted cumulative calculation engine. The student ability dynamic diagnosis subsystem is used to infer the student's current ability status on a weekly basis during the semester, based on the structured course model formed by the course information structured representation subsystem and the student's learning behavior data stored on the online education platform. The student ability dynamic diagnosis subsystem includes six parts: learning behavior semantic encoding module, deep knowledge tracking model, reverse mapping module, smooth update module, coverage verification module, and weekly operation mechanism. The dual-input informed consent generation subsystem is used to obtain a structured course model from the course information structured representation subsystem and the student's current ability status from the student ability dynamic diagnosis subsystem. Based on the large language model, the intelligent agent uses the structured course model and the student's current ability status to generate corresponding teaching aids for the student. The dual-input informed consent generation subsystem includes seven parts: dual-input architecture, six-node workflow, knowledge graph-enhanced hybrid retrieval, teaching semantic processing mechanism, structured prompt word constraint, implicit scaffolding adjustment mechanism, and security filtering and auditing mechanism.

[0007] In addition, embodiments of the present invention also provide an intelligent online teaching assistance method that integrates course knowledge and learning assessment, applied to the aforementioned intelligent online teaching assistance system, the method comprising: In the course information structured representation subsystem, the discrete course resources are constructed into a course knowledge graph through the course knowledge graph construction module, the ability achievement system definition module encodes the ability dimensions defined by the course training objectives into an ability achievement system, the course knowledge graph, ability achievement system and specific teaching activities are connected through the three-level association annotation module to form a three-level association table, and the weighted cumulative calculation engine performs weighted cumulative calculation based on the three-level association table to form a structured course model; The student ability dynamic diagnosis subsystem collects students' learning behavior data on the online teaching platform, cleans it, sorts it according to timestamps to form an interaction sequence, and encodes each interaction in the interaction sequence into a structured vector containing multi-label vectors of knowledge points and answer results by combining it with the structured course model. The student ability dynamic diagnosis subsystem takes the teaching week as the granularity, inputs the structured vector into the deep knowledge tracking model, outputs the knowledge mastery probability of each knowledge point, and transforms the knowledge mastery probability into ability mastery inference through the inverse mapping mechanism. Then, through the smooth update module and the coverage verification module, it generates the student's current ability status with confidence level. The system determines whether the stated ability status is greater than or equal to a set ability status. If not, it prohibits determining the next suggested supplementary assessment activity in the next teaching week. If so, when the dual-input informed consent generation subsystem responds to student questions, it concurrently calls the course knowledge base retrieval interface and the student diagnostic data interface to inject the structured course model and ability status into the large language model prompt words after processing the teaching semantics, generating teaching aid content. Simultaneously, After re-collecting the learning behavior data generated by students after viewing the teaching aids, the data is entered into the course information structured representation subsystem and the student ability dynamic diagnosis subsystem for iterative updates of the diagnostic results.

[0008] In this embodiment of the invention, a complete technical chain is established, connecting "structured representation of course information - dynamic diagnosis of student abilities - generation of informed assistance content"; and realizing the educational acceptability of the diagnostic results from floating to embedded, from post-assessment to dynamic diagnosis, from undifferentiated response to informed assistance.

[0009] Specifically, by constructing a knowledge graph and a three-level association system, the logical structure, knowledge topology, and ability progression relationships within the course can be accurately understood. After content validity review, the S-CVI / Ave of the three-level association table reached 0.96, indicating that the mapping relationship has sufficient educational rationality; this solves the problem of "semantic levitation" in existing technologies. Through the original DKT reverse mapping mechanism, smooth update of cognitive inertia coefficient, and evidence coverage verification, it can diagnose ability status in real time using fragmented student behavioral data. The DKT model achieved an accuracy of 90.93% on the test set, demonstrating credible predictive validity in diagnosis; this solves the problem of "diagnostic levitation." Through a dual-input architecture and a multi-field structured context protocol, the large language model is driven to generate personalized content containing a complete causal chain of teaching. Students no longer face generic answers but receive precise intervention based on their own ability diagnosis results, meeting the needs of real-time dialogue teaching scenarios. This solves the "interactive levitation" problem; the introduction of the cognitive inertia coefficient makes the ability curve present a gradual evolutionary form, and the three-level confidence label of evidence coverage honestly exposes the diagnostic uncertainty, enabling teachers to make teaching decisions based on credible quantification, effectively improving teachers' trust in the AI ​​diagnostic system. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a schematic diagram of the structural composition of the online teaching intelligent auxiliary system that integrates course knowledge and learning diagnosis in an embodiment of the present invention; Figure 2 This is a partial schematic diagram of the course knowledge graph in an embodiment of the present invention; Figure 3 This is a schematic diagram of the three-level association structure in an embodiment of the present invention; Figure 4 This is a data flow diagram of the reverse mapping mechanism in this embodiment of the invention; Figure 5 This is a comparison chart of the smoothing effect of the cognitive inertia coefficient in the embodiments of the present invention; Figure 6 This is a flowchart illustrating the verification process for evidence coverage in an embodiment of the present invention; Figure 7 This is a flowchart of the six-node dual-input intelligent agent in an embodiment of the present invention; Figure 8 This is a schematic diagram of the gradual disappearance bracket adjustment in an embodiment of the present invention; Figure 9 This is a flowchart illustrating an online teaching intelligent assistance method that integrates course knowledge and learning progress diagnosis, as described in an embodiment of the present invention. Detailed Implementation

[0012] 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, and 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.

[0013] For examples, please refer to Figure 1 , Figure 1 This is a schematic diagram of the structural composition of an online teaching intelligent auxiliary system that integrates course knowledge and learning diagnosis in an embodiment of the present invention.

[0014] like Figure 1 As shown, an online teaching intelligent auxiliary system integrating course knowledge and learning diagnosis is disclosed. The system includes: a course information structured representation subsystem, a student ability dynamic diagnosis subsystem, and a dual-input informed consent generation subsystem; wherein, The course information structured representation subsystem is used to transform discrete course resources into structured evidence with pedagogical semantic associations and form a structured course model. The course information structured representation subsystem includes four parts: a course knowledge graph construction module, a competency achievement system definition module, a three-level association annotation module, and a weighted cumulative calculation engine. The student ability dynamic diagnosis subsystem is used to infer the student's current ability status on a weekly basis during the semester, based on the structured course model formed by the course information structured representation subsystem and the student's learning behavior data stored on the online education platform. The student ability dynamic diagnosis subsystem includes six parts: learning behavior semantic encoding module, deep knowledge tracking model, reverse mapping module, smooth update module, coverage verification module, and weekly operation mechanism. The dual-input informed consent generation subsystem is used to obtain a structured course model from the course information structured representation subsystem and the student's current ability status from the student ability dynamic diagnosis subsystem. Based on the large language model, the intelligent agent uses the structured course model and the student's current ability status to generate corresponding teaching aids for the student. The dual-input informed consent generation subsystem includes seven parts: dual-input architecture, six-node workflow, knowledge graph-enhanced hybrid retrieval, teaching semantic processing mechanism, structured prompt word constraint, implicit scaffolding adjustment mechanism, and security filtering and auditing mechanism.

[0015] In the specific implementation of this invention, the online teaching intelligent auxiliary system adopts a three-layer architecture, which includes, from bottom to top, an evidence layer (a course information structured representation subsystem), a diagnosis layer (a student ability dynamic diagnosis subsystem), and an intelligent agent layer (a dual-input informed consent generation subsystem); and the online teaching intelligent auxiliary system interacts with the online education platform deployed in the school.

[0016] The functional relationship of the three layers can be summarized as "evidence calibration - diagnosis positioning - agent policy making": The evidence layer (structured representation subsystem of course information) defines "what should be learned" (target space), transforming the course knowledge system, ability objectives and teaching activities into a computable structured course model; the diagnosis layer (dynamic diagnosis subsystem of student ability) answers "where have you learned now" (current state), dynamically inferring the mastery of each ability dimension based on the student's behavior sequence; the agent layer (dual-input informed assistance generation subsystem) outputs "how to help next" (assistance action), integrating course knowledge and diagnostic data to generate personalized teaching assistance content; the three layers of data flow form a closed loop: student behavior log → structured anchoring of the evidence layer → dynamic inference of the diagnosis layer → informed assistance of the agent layer → new learning behavior of students → new behavior log feedback, driving the next round of diagnosis update, enabling the system to have the adaptive evolution capability of "the more it is used, the more accurate the diagnosis, and the more personalized the assistance".

[0017] In the specific implementation of this invention, the course knowledge graph construction module adopts a hybrid construction method combining top-down and bottom-up approaches to construct the course knowledge graph. The top-down approach involves teachers pre-defining the top-level structure of knowledge modules based on the course syllabus and teaching objectives, while the bottom-up approach utilizes natural language processing technology to automatically extract fine-grained knowledge points from textbooks, lecture notes, courseware, and past exam questions. The relationships between knowledge points in the course knowledge graph are defined as four types of relationships with pedagogical semantics, including predecessor, successor, association, and inclusion relationships. The competency achievement system definition module encodes the competency dimensions defined by the course training objectives into a competency achievement system of "competency dimension—competency indicator—achievement threshold." The competency achievement system is a three-level digital structure, where each competency dimension is a comprehensive competency category defined by the course training objectives. Each competency dimension is decomposed into several observable and evaluable competency indicators, and an achievement threshold is set for each competency indicator or dimension.

[0018] Furthermore, the three-level association annotation module is used to connect the course graph constructed by the course knowledge graph construction module, the ability achievement system defined by the ability achievement system definition module, and the key components of specific teaching activities, forming a three-level association table. This three-level association table uses a single activity task as the smallest annotation unit, annotating one or more knowledge points and one or more ability points involved in each activity task, and assigning corresponding weights to each activity task and its corresponding ability point association. The weight values ​​range from [0, 1], and the sum of the weights of all ability points associated with the same activity task is normalized to 1. The weighted cumulative calculation engine is used based on the three-level association table to calculate the student's... One ability point The associated set of activity tasks is Each activity task requires specific ability points. The weight is ,and Students in the Activity tasks The performance score on This forms a structured curriculum model; then the first... The student in the first One ability point Achievement as follows: .

[0019] Specifically, the core function of the evidence layer (the course information structured representation subsystem) is to transform discrete course resources into structured evidence with pedagogical semantic associations. It includes four parts: a course knowledge graph construction module, a competency achievement system definition module, a three-level association annotation module, and a weighted cumulative calculation engine.

[0020] The course knowledge graph construction module employs a hybrid approach combining top-down and bottom-up methods. The top-down approach involves instructors pre-defining the top-level structure of knowledge modules based on the course syllabus and teaching objectives (e.g., dividing a programming course into several chapter-level knowledge modules). The bottom-up approach utilizes natural language processing (NLP) technology to automatically extract fine-grained knowledge points from textbooks, lecture notes, courseware, and past exam questions. Specifically, this includes: performing word segmentation and named entity recognition on course documents to initially extract candidate knowledge point entities; filtering high-frequency core concepts and terms through TF-IDF keyword extraction and TextRank topic term analysis; and aligning candidate knowledge points with the instructor's predefined knowledge ontology, with entities that cannot be aligned entering a manual review queue. This hybrid "NLP-assisted + manual verification" model balances extraction efficiency with knowledge accuracy.

[0021] The relationships between knowledge points are defined into four types with pedagogical semantics: predecessor relationships (knowledge points that must be mastered before learning a particular knowledge point), successor relationships (inverse relationships of predecessors), association relationships (horizontal conceptual connections without strict sequential order), and inclusion relationships (hierarchical affiliation of chapters, sections, and knowledge points). Relationship annotation is completed independently by at least two course instructors and compared using consistency verification rules (e.g., strict equality of predecessor sets, Jaccard similarity of association sets not less than 0.8). Inconsistencies are re-annotated after negotiation to form refined rules until the inconsistency rate converges. The graph storage uses a graph database with an attribute graph model (e.g., Neo4j). Each knowledge point node includes attributes such as a three-level identifier, name, level, chapter, and parent node identifier, supporting path-based traversal queries (e.g., shortest prerequisite path, knowledge point reachability analysis). This provides structured course semantic constraints for the diagnostic layer and a searchable knowledge base for the agent layer. The local structure of the course knowledge graph is as follows: Figure 2 As shown.

[0022] The competency achievement system definition module encodes the competency dimensions defined by the course training objectives into a three-level digital structure of "competency dimension - competency indicator - achievement threshold". The three-level digital structure is as follows: The competency dimension is a comprehensive competency category defined by the course training objectives (for example, a programming course can be divided into five dimensions: computational thinking and abstract modeling, problem solving and programming implementation, data thinking and interdisciplinary application, logical thinking and rigorous expression, and innovative practice and self-directed learning). Each competency dimension is further decomposed into several observable and assessable competency indicators. An achievement threshold is set for each competency indicator or dimension (a percentage of 0-1, representing the accuracy or completion rate students should achieve in related activities and tasks). This is stored in a computer-readable data structure (the competency dimension table includes fields for identifier, name, description, and achievement threshold; the competency indicator table includes fields for identifier, dimension, indicator description, and measurement method). Teachers can dynamically adjust the data in a versioned manner, with adjustment records kept to ensure the traceability of achievement calculations.

[0023] The three-level association annotation module is a key component connecting the course knowledge graph, the competency achievement system, and specific teaching activities. It uses a single activity task (test, assignment, project, video, etc.) as the smallest annotation unit, annotating one or more knowledge points and one or more competency points involved in each activity task, and assigning a weight to each "activity-competency point" association. The weight ranges from 0 to 1, and the sum of the weights of all competency points associated with the same activity is normalized to 1. The weight reflects the activity's contribution to the achievement of the corresponding competency point. Annotation results are stored in an association table, whose structure includes four core fields: activity identifier, knowledge point identifier, competency point identifier, and weight. A composite primary key (activity identifier, knowledge point identifier, competency point identifier) ​​ensures the uniqueness of the many-to-many relationship, and it is linked to the activity records of the teaching platform, the knowledge point nodes of the graph database, and the competency dimension table via foreign keys. All annotation information is summarized to form a three-level association table of "knowledge point-activity task-competency point," establishing indirect associations between knowledge points and competency points through activity tasks as a bridge. This three-level association table serves three functions in the system: providing a mapping foundation for the diagnostic layer, from task prediction to ability inference; providing a structured index of course knowledge for the agent layer; and providing weighted data support for ability quantification calculation. As an optional verification method, the mapping rationality of the three-level association table can be verified through independent content review by multiple subject teachers who did not participate in its construction (e.g., using a content validity index, requiring an item-level acceptance rate of no less than 0.80), ensuring that the mapping aligns with the instructional design. Figure 1 To; a three-level association structure such as Figure 3 As shown in Table 1, the data structure of the associated table is an example of such a structure.

[0024] Table 1. Example of a three-level association table data structure for "Knowledge Point - Activity Task - Ability Point"

[0025] The weighted cumulative calculation engine calculates the number of students in the first... One ability point The associated set of activity tasks is Each activity task requires specific ability points. The weight is ,and Students in the Activity tasks The performance score on Then the first The student in the first One ability point Achievement as follows: ; The weighted cumulative calculation engine calculates the final competency achievement based on all scoreable activities at the end of the semester, which serves as a benchmark true value to verify the accuracy of the diagnosis; the weighted cumulative calculation engine also calculates partial estimates based on completed activities during the semester as a reference.

[0026] The weighted cumulative calculation engine runs at two points in time: at the end of the semester, it calculates the summative competency achievement based on all gradeable activities, serving as the baseline true value to verify the accuracy of the diagnosis; during the semester, it calculates partial estimates based on completed activities as a reference. Some estimates during the semester can deviate significantly from the true level when high-weighted activities are not completed (for example, if a competency is only assessed by the end-of-semester project, the achievement of that dimension in the first half of the semester cannot be calculated). This "cold start / data sparsity" defect is the core problem that the diagnostic layer aims to solve.

[0027] In the specific implementation of this invention, the semantic encoding module is used to extract students' learning behavior data from the database of the online teaching platform and perform data cleaning processing on the learning behavior data. The fields of the learning behavior data include student identifier, activity identifier, answer result, timestamp, and session identifier. Each interaction record is encoded into a structured vector based on a three-level association table. The data cleaning processing includes: removing abnormal records with missing key fields from the learning behavior data; retaining only the first submission for multiple submissions by the same student to the same activity task; and removing time logic anomalies from the learning behavior data. The records are as follows: student sequences with fewer than a preset number of interaction records are removed; the structured vector is defined as follows: if there are N knowledge points in the course, then task activity j is encoded as an N-dimensional binary multi-label vector, which is then concatenated with the answer result of the interaction to form an N+1-dimensional input vector; the deep knowledge tracing model is a deep knowledge tracing model based on long short-term memory network. The deep knowledge tracing model receives the student's interaction sequence, represents the student's knowledge state as a continuous latent vector and models its temporal evolution. After passing through a fully connected layer and a Sigmoid activation, it outputs the student's knowledge mastery probability for each knowledge point at the next time step.

[0028] Furthermore, the reverse mapping module is used to establish a complete mapping path from the knowledge mastery probability output by the deep knowledge tracking model to the ability inference mastery level through a three-level association table; wherein the complete mapping path includes aggregating the knowledge mastery probability into the prediction accuracy of the activity task layer, and aggregating the actual performance and predicted values ​​at the activity level into the sum evidence value of ability points through the three-level association table; wherein, the activity task In time The formula for the prediction accuracy is as follows: ; For the activity task In time The accuracy of predictions; The probability of knowledge acquisition; For the first Activity Tasks The set of knowledge points involved; define the ability points. The associated set of activity tasks is The completed subset is The incomplete subset is Then this ability point In time The overall value of evidence is: ; The smoothing update module is used to enable the capability state estimation to have cross-week time-series stability, and applies an exponential smoothing recursive update to the comprehensive evidence value as follows: ; in, For time Smoothed ability inference of mastery; The cognitive inertia coefficient represents the proportion of historical states retained. ; For time Smoothed ability inference of mastery; For ability points In time The overall value of evidence; For the first Activity tasks for the first Weights of ability points; For students in their completed first The actual performance score for the activity task is the score after normalization.

[0029] Furthermore, the coverage verification module is used to define the evidence coverage of any capability dimension corresponding to capability point k as follows: This is because, for any capability dimension that has not been covered by any activity task for more than a preset duration during the semester, and the evidence value is entirely generated by the deep knowledge tracing model. ; in, For any ability dimension, the corresponding first The key skill lies in time. Evidence coverage; For the first Activity tasks for the first Weights of ability points; For the first The set of activity tasks associated with ability points The completed subset; the weekly operation mechanism is used to run in batches on a teaching week basis, and a weekly timed task is triggered to execute the following steps in sequence: loading the full student interaction sequence up to the current week, running the deep knowledge tracking model, and updating the knowledge mastery probability of each knowledge point; calling the inverse mapping module to calculate the comprehensive evidence value of each ability point; calling the smooth update module to apply an exponential smooth recursive update to the comprehensive evidence value; calling the coverage verification module to calculate the evidence coverage and generate the current ability status of students with confidence levels, writing the current ability status of students as structured data into the memory cache, setting the validity period until the next week's update, so that the dual-input informed assistance generation subsystem can query it in real time.

[0030] Specifically, the core function of the diagnostic layer (student ability dynamic diagnostic subsystem) is to dynamically, stably, and interpretably infer the current ability status of students based on the structured curriculum model of the evidence layer and students' learning behavior data during the semester. It includes six parts: learning behavior semantic encoding module, deep knowledge tracking model, reverse mapping module, smooth update module, coverage verification module, and weekly operation mechanism.

[0031] The learning behavior semantic encoding module extracts student learning behavior data (fields include student identifier, activity identifier, answer result, timestamp, session identifier, etc.) from the online teaching platform database, and performs data cleaning: removing abnormal records with missing key fields; retaining only the first submission for multiple submissions by the same student to the same activity; removing records with abnormal time logic; excluding student sequences with too few interaction records (e.g., less than 5); subsequently, based on a three-level association table, each interaction record is encoded into a structured vector: assuming the course has N knowledge points, activity j is encoded as an N-dimensional binary multi-label vector (the knowledge points involved in the activity correspond to the bits). (Set to 1 otherwise to 0), and then concatenate it with the response result of that interaction (binary values ​​for correct / incorrect, or normalized completion degree) to form an N+1 dimensional input vector; each student's interaction record is sorted by timestamp to form a variable-length sequence, and the sequence length is standardized by padding or truncation (e.g., 20), and segmented by teaching week. The interaction record of each week serves as the basic data unit for diagnosis that week; the reason for using weekly granularity is that too high a diagnosis frequency (e.g., daily) will lead to excessive prediction fluctuations, while too low a frequency (e.g., monthly) will not meet the needs of timely intervention. Weekly granularity achieves a balance between timeliness and stability, and matches the teaching rhythm of university courses on a weekly basis. This encoding method allows the evolution of the hidden state of the deep model to be traced back to the process of mastering specific knowledge concepts, rather than meaningless numerical signals.

[0032] The Deep Knowledge Tracking (DKT) model uses a Long Short-Term Memory (LSTM)-based deep knowledge tracking model as its sequence modeling kernel. The DKT model receives the student's interaction sequence, represents the student's knowledge state as a continuous hidden vector, models its temporal evolution, and outputs a probability vector m(i,t) of the student's mastery of each knowledge point at the next time step after a fully connected layer and sigmoid activation. A specific implementation parameter is as follows: a two-layer LSTM structure with a hidden layer dimension of 64, an input dimension of N+1, and an output dimension of N. The LSTM is followed by a Dropout layer (dropout rate of 0.1), the loss function is cross-entropy, the optimizer is Adam, and an early stopping strategy is employed. The model training adopts a strategy of "pre-training with historical data + fine-tuning for the new semester": before the start of the semester, the model is pre-trained and the parameters are saved using historical interaction data accumulated in the same course in previous semesters (such as the last three years) on the platform; after the start of the new semester, the model is quickly adapted to the initial data of the new semester with a low learning rate, and then the parameters are kept fixed during the semester, and only forward inference is performed to ensure the temporal stability and attribution clarity of the diagnostic results; the dataset is divided into training set, validation set and test set (such as 80% / 10% / 10%) according to the student dimension, and all sequences of each student are completely included in the same subset to prevent the leakage of the same student's answer pattern across sets.

[0033] The reverse mapping module addresses the issue where the standard deep knowledge tracing model outputs the mastery probability of each knowledge point, remaining at the task-level prediction level and creating a semantic gap with the ability goals focused on in teaching practice. The reverse mapping module establishes a complete mapping path from the mastery probability of knowledge points to the inferred mastery level through a three-level association table, including two steps: The first step is to aggregate knowledge point-level data to the activity-level data: This involves aggregating the probability of knowledge point mastery into the prediction accuracy at the activity level. This is based on linear aggregation of knowledge coverage for activity tasks. In time The prediction accuracy is: ; For the activity task In time The accuracy of predictions; The probability of knowledge acquisition; For the first Activity Tasks The set of knowledge points involved.

[0034] The second step of the reverse mapping is to aggregate the actual performance and predicted values ​​at the activity level into a comprehensive evidence value for the capability point through a three-level association table; let the capability point be... The associated set of activity tasks is The completed subset is The incomplete subset is Then this ability point In time The overall value of evidence is: ; The system uses students' actual normalized scores for completed activities and the prediction accuracy obtained in the first step for incomplete activities. The technical advantage of this hybrid evidence mechanism is that, in the early part of the semester when many activities are incomplete, sequential prediction fills the evidence gaps, allowing ability inference to avoid passively waiting for all activities to finish, thus overcoming the time lag inherent in traditional ability assessment. As the semester progresses and activity coverage becomes more complete, actual scores gradually dominate, and predicted items naturally dilute, requiring no additional weighting adjustments. Therefore, the level of ability inference mastery becomes the core variable output from the diagnostic layer to the agent layer, realizing the educational semantic transformation from "predicting whether the next question can be answered" to "inferring the current level of mastery in a certain ability dimension." The complete data flow of the reverse mapping mechanism is as follows: Figure 4 As shown.

[0035] The smoothing update module is used to make the capability state estimation have cross-week time-series stability. It applies an exponential smoothing recursive update to the comprehensive evidence value as follows: ; in, For time Smoothed ability inference of mastery level; The cognitive inertia coefficient represents the proportion of historical states retained. ; For time Smoothed ability inference of mastery level; For ability points In time The overall value of evidence.

[0036] This recursive formula is equivalent to applying a first-order low-pass filter to the evidence sequence, with its cutoff frequency controlled by β: it significantly attenuates high-frequency noise with periods shorter than one week (single-task errors, temporary guesses, and prediction variance inflation during the cold start phase), while preserving the true capability trend on a semester scale; without this mechanism ( =0) will trigger three risks, such as: the amplification of prediction noise at the beginning of the semester, resulting in non-monotonic jumps in the ability profile of adjacent weeks; a single assessment error causing a cliff drop by lowering the ability estimate with full weight; and unpredictable jumps in the diagnostic results damaging teachers' trust in the system.

[0037] The cognitive inertia coefficient was determined using a two-stage procedure: "offline calibration during the calibration semester - solidification during the formal semester." For example, during the calibration semester, which is independent of the formal intervention, a grid search was performed within the candidate interval (e.g., [0.3, 0.9], step size 0.05) to determine the optimal value, with the goal of maximizing the correlation coefficient between the inferred mastery level after smoothing in the early part of the semester and the summative ability achievement level (benchmark true value) at the end of the semester. Simultaneously record RMSE and MAE; if the correlation difference among multiple candidate values ​​is less than a preset threshold (e.g., 0.01), select the value closest to the center of the interval; after calibration... In the formal semester, this parameter is fixed as a hard-coded parameter and is not adjusted with week or individual differences to avoid introducing time-varying confounding variables; the cross-week smoothing effect of having or not having a cognitive inertia coefficient is compared to... Figure 5 As shown.

[0038] The coverage verification module addresses the risk of "false precision" arising from certain capability dimensions not being covered by any activities over a long period of the semester, where their evidence values ​​consist almost entirely of predictive items. This module defines the evidence coverage of capability point k as follows: ; in, For any ability dimension, the corresponding first The key skill lies in time. Evidence coverage; For the first Activity tasks for the first Weights of ability points; For the first The set of activity tasks associated with ability points For the completed subset; This refers to the weighted contribution percentage of completed activities to this capability dimension, ranging from 0 to 1. The system sets a minimum coverage threshold (e.g., 0.3) and establishes a three-level confidence rating system, such as: when 0≤λ<0.3, it is low confidence ("limited inference", this dimension value is a mixed estimate of prediction and historical state, for reference only, marked in gray at the front end); when 0.3≤λ<0.7, it is medium confidence ("preliminary diagnosis", there is some experimental evidence, it is recommended to combine classroom observation for comprehensive judgment, marked in yellow at the front end); when 0.7≤λ≤1, it is high confidence ("sufficient evidence", experimental data dominates, can be directly used for teaching decisions, marked in green at the front end).

[0039] This mechanism shifts the focus of data sparsity issues from the model's internal processes to honest identification at the system boundary. For example, low-coverage dimensions do not disappear or become falsely precise; instead, they explicitly indicate insufficient evidence, prompting teachers to fill the evidence gaps through supplementary assessment activities, forming a closed loop of "diagnosis-feedback-supplementation." Simultaneously, the system records the complete computational chain from knowledge point prediction to ability inference, allowing teachers to trace the source activities and weights of each ability value, achieving an interpretable transformation of black-box predictions. The coverage verification process and its linkage with the agent's response are as follows: Figure 6 As shown in Table 2, the linkage rules between the three confidence levels and each link in the system are as follows.

[0040] Table 2. Correlation between Diagnostic Confidence Level and System Response

[0041] The weekly operation mechanism means that the student ability dynamic diagnosis subsystem will run in batches on a teaching week basis. A weekly scheduled task (such as early Sunday morning) will be triggered and execute the following steps in sequence: loading the full student interaction sequence up to the current week; running the knowledge tracking model to infer and update the mastery probability of each knowledge point; calling the inverse mapping module to calculate the comprehensive evidence value of each ability point; performing cognitive inertia smoothing update; calculating evidence coverage and generating confidence level.

[0042] The diagnostic results (including the inferred mastery level, confidence level, trend direction, and list of weak knowledge points for each ability point) are written as structured data to an in-memory cache (such as Redis), with an expiration date set until the next week's update, for the intelligent agent layer (student ability dynamic diagnostic subsystem) to query in real time with millisecond latency. The cold start handling strategy at the beginning of the semester includes: using historical data from previous semesters to pre-train the model, so that it has initial predictive ability from the start of the semester; for students without interaction records, the historical average achievement of students in each ability dimension in previous semesters of the course is used as the initial ability status (the default value of 0.5 is used and clearly marked when there is no historical data); the coverage of diagnostic results is naturally low in the first few weeks of the semester, and the system forces the display of the "inference limited" label; the diagnostic layer (student ability dynamic diagnostic subsystem) is deployed as an independent microservice, providing standardized query interfaces (latest diagnostic interface, weak point Top-N interface, trend sequence interface), with a response time target of less than 100 milliseconds for P95 latency; even if the diagnostic service fails, the platform's basic teaching functions will not be affected.

[0043] In the specific implementation of this invention, the dual-input architecture involves the dual-input informed consent generation subsystem simultaneously accessing two types of information sources. One type is a course knowledge base, which comprises three types of knowledge constructed based on the course information structured representation subsystem. These three types of knowledge are stored in a vector database format: a knowledge point knowledge base, an activity task knowledge base, and a competency system knowledge base. The other type is student diagnostic data, which consists of the student competency dynamic diagnosis subsystem outputting the inferred mastery level of each competency point, a list of weak knowledge points, and competency development trends. The six-node workflow is the arrangement of six sequentially executed nodes for each response of the dual-input informed consent generation subsystem. Node one receives a question request from the front end of the online teaching platform and parses the student identifier and question meaning. The diagram shows: Node 2 is used to call the interface of the student ability dynamic diagnosis subsystem, input the student identifier, and obtain the structured data of the corresponding student's ability status; Node 3 is used to encode the question text through an embedded model, perform an approximate nearest neighbor search in the course knowledge vector library, and recall relevant knowledge fragments with preset rankings; Node 4 is used to render the role setting, search results, formatted diagnostic data, dialogue history, and user questions into complete prompt words through a template engine; Node 5 is used to send the assembled prompt words to the locally deployed large language model for reasoning and generating an answer; Node 6 is used to perform content filtering on the generated answer, add hyperlinks to the knowledge points referenced in the answer pointing to the corresponding nodes in the course knowledge graph, and generate learning behavior data based on the question-answer pair along with the current diagnostic data snapshot, and write it into the online teaching platform.

[0044] Furthermore, the knowledge graph-enhanced hybrid retrieval is used to introduce knowledge graph enhancement into the retrieval process. In the offline indexing stage, the course handouts are divided into paragraphs, the core knowledge point entities involved in each paragraph are extracted, and the graph database is queried to obtain the neighboring entities and corresponding relationship paths of the core knowledge point entities. The original paragraphs and the extended entity context are jointly encoded into vector index items. In the online retrieval stage, named entity recognition is first performed on the user's question to extract knowledge point entities. Then, the graph database is queried to obtain the neighboring entities and relationship paths to form an extended query set. The original question and the extended query set are combined for vector retrieval. The retrieval results are merged and the results with a preset ranking are input into the large language model. The teaching semantic processing mechanism is used to process the output teaching appropriateness through three mechanisms before injecting prompt words. The three mechanisms include an achievement rate conversion mechanism, an inference credibility hierarchical response mechanism, and a weakness attribution to causal chain integrity mechanism.

[0045] Furthermore, the structured prompt word constraint is used to transform teaching constraints into executable generation rules when the prompt words are solidified into six structured paragraphs. The generation rules include role definition paragraphs, ability boundary paragraphs, achievement level labeling paragraphs, causal chain integrity paragraphs, and safety constraint paragraphs. The fading scaffolding adjustment mechanism is used to dynamically drive the intervention intensity of the dual-input informed aid generation subsystem based on the diagnostic data of the student ability dynamic diagnosis subsystem. When the student's inferred mastery of a certain ability dimension is lower than the first preset threshold, the dual-input informed aid generation subsystem provides detailed explanations, prior concept reviews, and step-by-step examples. When the student's mastery exceeds the second preset threshold, the intervention intensity is automatically reduced, and the system transitions from "explainer" to "answerer," only intervening when the student actively asks questions. The safety filtering and auditing mechanism is used to perform two-layer safety filtering on the output of the dual-input informed aid generation subsystem. The first layer of safety filtering is keyword filtering, and the second layer of safety filtering is semantic classification filtering.

[0046] Specifically, the core function of the intelligent agent layer (dual-input informed assistance generation subsystem) is to enable intelligent agents based on large language models to "know both the course and the students" when generating teaching assistance content. It includes seven parts: dual-input architecture, six-node workflow, knowledge graph-enhanced hybrid retrieval, teaching semantic processing mechanism, structured prompt word constraint, implicit scaffolding adjustment and security audit mechanism.

[0047] The dual-input architecture allows the dual-input informed consent generation subsystem to simultaneously access two types of information sources. The first is a course knowledge base, which is not simply a collection of unstructured document slices, but rather a three-tiered knowledge base built on an evidence-layer structured course model: a knowledge point knowledge base (storing knowledge point entities and their predecessors, successors, associations, and inclusion relationships), an activity task knowledge base (storing the associations and weights between activities and knowledge points and ability points), and an ability system knowledge base (storing ability dimension definitions, indicators, and achievement standards), all stored in the form of a vector database. The second is student diagnostic data, namely the inferred mastery level of each ability point, a list of weak knowledge points, and ability development trends output by the diagnostic layer, which is obtained in real time through a standardized data interface.

[0048] The integration of the two types of information follows three principles: student-question-driven (conceptual questions are primarily driven by knowledge base retrieval results, while "how should I learn?" questions are primarily driven by diagnostic data); knowledge-first (factual answers prioritize knowledge base content, and diagnostic data must not cover or negate course content); diagnostic enhancement (personalized suggestions are added on top of knowledge-based answers); and the use of diagnostic data follows the principle of "quoting rather than restating": the agent naturally integrates diagnostic information into the answer (e.g., "Based on your recent learning, you need to strengthen your practice of recursion-related programming problems"), rather than mechanically reading out data values.

[0049] The six-node workflow is a dual-input informed consent generation subsystem where each response is orchestrated into six sequentially executed nodes. Node 1 (the start node) receives the question request from the platform frontend, parses the student identifier and question intent, and verifies the student's identity via the platform's OAuth authentication interface to ensure that dialogue permissions match course access permissions. Node 2 (the diagnostic data acquisition node) calls the diagnostic layer interface, inputs the student identifier, and retrieves the student's latest structured ability status data (inferred mastery levels for each ability point, a list of weak knowledge points sorted in ascending order of mastery level, and trend direction). Because the diagnostic results are cached in an in-memory database, this node's call latency is in the millisecond range. Node 3 (the knowledge base retrieval node)... The workflow consists of six nodes: Node 1 (the first node), Node 2 (the second node), Node 3 (the third node), and Node 4 (the fourth node). Node 1 encodes the question text using an embedding model (e.g., BGE-M3, vector dimension 1024), performs near nearest neighbor retrieval (e.g., based on FAISS index) in the course knowledge vector base, and retrieves Top-K (e.g., Top-5) related knowledge fragments. Node 2 (the fifth node) renders the system role settings, retrieval results, formatted diagnostic data, dialogue history, and user questions into complete prompts using a template engine (e.g., Jinja2). This node is the core hub for dual-input fusion. Node 3 (the fifth node) sends the assembled prompts to a locally deployed large language model for inference generation (with parameters such as temperature 0.7 and maximum output length 2048 tokens to balance generation consistency and richness). Node 4 (the sixth node) performs content filtering on the generated answer, adds hyperlinks to the corresponding nodes in the course knowledge graph for the knowledge points referenced in the answer, and writes the question-answer pair along with a snapshot of the diagnostic data at that time back to the platform's dialogue log for teaching review and subsequent analysis. The six-node workflow is structured as follows: Figure 7 As shown in Table 3, the inputs, outputs, and core functions of each node are defined.

[0050] Table 3. Definition of Six-Node Workflow for Intelligent Agents

[0051] To overcome the omissions of semantically related but superficially lexically different content by pure vector similarity retrieval (e.g., students describing "recursion" as "calling themselves"), knowledge graph enhancement is introduced into the retrieval enhancement generation process: In the offline indexing stage, course handouts are segmented into paragraphs, and the core knowledge point entities involved in each paragraph are extracted. The graph database is queried to obtain the neighboring entities (predecessors, successors, and related knowledge points) and relational paths of these entities. The original paragraphs and the extended entity context are jointly encoded into vector index items. In the online retrieval stage, named entity recognition is first performed on the user's question to extract knowledge point entities. Then, the graph database is queried to obtain neighboring entities and relational paths to form an extended query set. Finally, the original question and the extended entity set are combined for vector retrieval, merged, sorted, and the Top-K results are input into the large language model. This hybrid strategy of "entity expansion + vector retrieval" significantly improves retrieval recall and the grasp of the deep logic of course knowledge.

[0052] The instructional semantic processing mechanism involves not simply concatenating the two input messages, but processing them through three mechanisms before the input prompts are given, to ensure the appropriateness of the output for instructional purposes: Mechanism 1: Achievement Rate Conversion Mechanism. To avoid exposing students to absolute inference values ​​without a reference frame (such as "Your logical thinking score is 0.42"), the system converts the smoothed inference mastery level into an achievement rate relative to the class average. ; ; in, For the total number of students in the class, the achievement rate is mapped to four levels of semantic labels: reaching or exceeding the class average (providing consolidation suggestions or removing scaffolding), approaching the class average (suggesting maintaining the current pace), below the class average (providing targeted attribution and practice suggestions), and significantly below the class average (providing a systematic remedial path). This mapping table is encoded as the agent's decision rules to ensure that feedback uses relative semantic labels rather than raw scores. The mapping relationship between achievement rate intervals and semantic labels is shown in Table 4.

[0053] Table 4 Achievement Rate - Semantic Tag Mapping Table

[0054] Mechanism Two is a graded response mechanism based on inference confidence. The depth of diagnosis, the granularity of suggestions, and the coverage of evidence in the agent's response are linked: At high confidence (λ≥0.7), a complete attribution analysis can be performed, tracing back to specific knowledge modules; at medium confidence (0.3≤λ<0.7), only a brief analysis is performed, and continued observation is recommended; at low confidence (λ<0.3), the inference limitation status must be clearly stated, and definitive suggestions or specific practice instructions are prohibited. For students whose learning profiles have not yet been established (with very little effective interaction data), the agent does not output any diagnostic suggestions or use negative attribution statements. Instead, it states the status as "Learning profile is being built" and provides clear first-step action guidance.

[0055] Mechanism 3, the weak attribution and causal chain integrity mechanism, requires that each learning suggestion from the agent must contain the four elements of a complete causal chain: ability signal → knowledge attribution → specific practice → expected effect. Specifically, when a low achievement rate is identified in a certain ability dimension, the system traces the knowledge module with the highest weight in that dimension and that has been published along the three-level association table, and generates a suggestion that includes attribution basis, practice content and effect prediction, avoiding fragmented statements such as "Logical thinking score 37 points, it is recommended to write an if statement".

[0056] The structured prompts are fixed into six structured paragraphs, transforming teaching constraints into executable generation rules: Role Definition Paragraph (clearly defining the agent as a teaching aid AI, with tasks limited to answering knowledge questions, providing ability status feedback, and offering personalized learning suggestions); Ability Boundary Paragraph (clearly defining what the agent knows—course syllabus, textbooks, courseware, and exercise solutions; what it doesn't know—extracurricular knowledge, student personal information, and grade assessments; guiding students to consult teachers when outside the scope); Achievement Rate Expression Paragraph (absolutely prohibiting the exposure of raw ability values; achievement rates and semantic labels must be used); Confidence Labeling Paragraph (each ability dimension diagnosis must be accompanied by a confidence label; definitive suggestions are prohibited when confidence is low); Complete Causal Chain Paragraph (suggestions must include four elements: ability signal, knowledge attribution, specific practice, and expected effect); Safety Constraint Paragraph (prohibiting the agent from completing assignments or exams for students, prohibiting the evaluation of individual student abilities, prohibiting the fabrication of diagnostic data, and providing encouragement when students exhibit negative emotions).

[0057] The gradual scaffolding adjustment mechanism dynamically drives the intervention intensity of the agent based on diagnostic data: when a student's inferred mastery of a certain ability dimension falls below a preset threshold, the agent provides detailed explanations, a review of prior concepts, and step-by-step examples; when mastery exceeds the preset threshold (e.g., 0.70), the intervention intensity is automatically reduced, shifting from "explainer" to "answerer," intervening only when the student actively asks questions, thus gradually removing the scaffolding and preventing students from developing cognitive dependence. Supporting strategies include: differentiated explanation strategies (using a concise style and skipping basic introductions for dimensions with high mastery; increasing the depth of explanation for dimensions with low mastery); and follow-up questioning strategies (when a student's question is ambiguous, first checking the mastery of relevant prior knowledge points in the diagnostic data, then asking targeted follow-up questions to precisely pinpoint the difficulties). The adjustment logic of the gradual scaffolding is as follows: Figure 8 As shown.

[0058] The system employs a security filtering and auditing mechanism, requiring the intelligent agent's output to undergo a two-layer security filtering process: the first layer is keyword filtering, which detects academic misconduct-inducing content (such as statements like "I'll give you the answer directly" or "I'll help you finish writing the code") based on a sensitive word database in the education field; the second layer is semantic classification filtering, which identifies the category of the answer intent (normal teaching assistance / potential academic misconduct / beyond the scope of the course / harmful content) based on a lightweight classification model (such as a BERT-like model), triggering interception or manual review for abnormal categories; all dialogue records are stored in a structured form in the school's database, with fields including student identifier, conversation identifier, question, answer, diagnostic data snapshot during the dialogue, timestamp, and model version, supporting retrieval and export by course, student, and time period, and the logs are anonymized and archived after one semester; the system also has operation monitoring (average response latency, request volume, error rate, number of filter triggers, satisfaction score), automatically alarming and switching to a backup model when the error rate or latency exceeds the threshold.

[0059] The three layers are loosely coupled and integrated through standardized interfaces, with inter-layer interfaces and a closed-loop operation mechanism. The evidence layer provides: an activity-knowledge point mapping interface (input activity identifier, return a list of related knowledge points), a capability-activity mapping interface (input capability point identifier, return a list of associated activities and their weights), and a capability threshold query interface. The diagnosis layer provides: a latest diagnosis interface (returns the latest inferred mastery level, confidence level, and trend direction for each capability point), a weakness interface (returns the N capability points and knowledge points with the lowest mastery levels), and a trend interface (returns the state sequence of a specified capability point over the past few weeks). The returned data includes timestamps to allow the caller to determine data freshness. The agent layer provides a dialogue service interface (input student identifier, course identifier, question text, and dialogue history, return answer text, answer type, and recommended resource list). The table supports both long-connection real-time dialogue and short-connection single question-and-answer interfaces, as well as dialogue history interfaces. The online teaching platform provides behavior log writing interfaces (student identifier, activity identifier, answer result, timestamp, and session identifier) ​​to the evidence layer. The interfaces adopt RESTful style, JSON data format, API key authentication, and tiered rate limiting. The resource items in the recommended resource list include resource type, resource identifier, and name. The front end directly links to the corresponding content on the platform via the resource identifier, ensuring the operability of learning suggestions. The three-layer data closed-loop path is "student behavior log → evidence layer structured anchoring → diagnostic layer dynamic inference → agent layer informed assistance → new student behavior → log feedback". Upgrades to the internal implementation of any layer (such as knowledge tracking model replacement or knowledge base storage scheme change) do not affect other layers; only the interface contract needs to remain unchanged.

[0060] In this embodiment of the invention, a complete technical chain is established, connecting "structured representation of course information - dynamic diagnosis of student abilities - generation of informed assistance content"; achieving a shift from detached to embedded, from post-assessment to dynamic diagnosis, from undifferentiated to informed assistance, and ensuring that the diagnostic results are educationally acceptable.

[0061] By constructing a knowledge graph and a three-level association system, we can accurately understand the internal logical structure, knowledge topology, and ability progression relationships of the course. After content validity review, the S-CVI / Ave of the three-level association table reached 0.96, indicating that the mapping relationship has sufficient educational rationality. This solves the problem of "semantic levitation" in existing technologies.

[0062] Through the original DKT inverse mapping mechanism, smooth update of cognitive inertia coefficient, and evidence coverage verification, it is possible to diagnose students' ability status in real time using fragmented behavioral data; the DKT model achieves an accuracy of 90.93% on the test set, and the diagnosis has reliable predictive validity; this solves the problem of "diagnostic susceptibility".

[0063] By employing a dual-input architecture and a multi-field structured context protocol, a large language model is driven to generate personalized content containing a complete causal chain of instruction. Students no longer face generic answers but receive precise interventions based on their own ability diagnostic results, meeting the needs of real-time dialogue teaching scenarios. This solves the problem of "interaction floating".

[0064] The introduction of the cognitive inertia coefficient enables the ability curve to exhibit a gradual evolutionary form, and the three-level confidence label of evidence coverage honestly exposes diagnostic uncertainty, enabling teachers to make teaching decisions based on credible quantifiable evidence, effectively improving teachers' trust in the AI ​​diagnostic system.

[0065] For examples, please refer to Figure 9 , Figure 9 This is a flowchart illustrating the online teaching intelligent assistance method that integrates structured representation of course knowledge with dynamic diagnosis of learning progress in an embodiment of the present invention.

[0066] like Figure 9 As shown, an intelligent online teaching assistance method integrating course knowledge and learning diagnosis is applied to any of the above-mentioned intelligent online teaching assistance systems. The method includes: S901: In the course information structured representation subsystem, the discrete course resources are constructed into a course knowledge graph through the course knowledge graph construction module, the ability achievement system definition module encodes the ability dimensions defined by the course training objectives into an ability achievement system, the course knowledge graph, ability achievement system and specific teaching activities are connected through the three-level association annotation module to form a three-level association table, and the weighted cumulative calculation engine performs weighted cumulative calculation based on the three-level association table to form a structured course model; S902: The student ability dynamic diagnosis subsystem collects students' learning behavior data on the online teaching platform, cleans it, sorts it according to timestamps to form an interaction sequence, and encodes each interaction in the interaction sequence into a structured vector containing multi-label vectors of knowledge points and answer results by combining it with the structured course model. S903: The student ability dynamic diagnosis subsystem takes the teaching week as the granularity, inputs the structured vector into the deep knowledge tracking model, outputs the knowledge mastery probability of each knowledge point, and transforms the knowledge mastery probability into ability mastery inference through the inverse mapping mechanism. Then, through the smooth update module and the coverage verification module, it generates the current ability status of students with confidence level. S904: Determine whether the stated ability status is greater than or equal to the set ability status. If not, prohibit determining the next suggested supplementary assessment activity in the next teaching week. If yes, when the dual-input informed consent generation subsystem responds to student questions, it concurrently calls the course knowledge base retrieval interface and the student diagnostic data interface to inject the structured course model and ability status into the large language model prompt words after processing with teaching semantics, generating teaching aid content. Simultaneously... S905: After re-collecting the learning behavior data generated by students after viewing the teaching aids, the data is entered into the course information structured representation subsystem and the student ability dynamic diagnosis subsystem for iterative updating of the diagnostic results.

[0067] For implementation examples of the method section, please refer to the system section, which will not be repeated here.

[0068] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0069] Furthermore, the above provides a detailed description of an online teaching intelligent auxiliary system and method integrating course knowledge and learning diagnosis provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An online teaching intelligent auxiliary system that integrates course knowledge and learning assessment, characterized in that, The system includes: a structured representation subsystem for course information, a dynamic diagnosis subsystem for student abilities, and a dual-input informed consent generation subsystem; wherein... The course information structured representation subsystem is used to transform discrete course resources into structured evidence with pedagogical semantic associations and form a structured course model. The course information structured representation subsystem includes four parts: a course knowledge graph construction module, a competency achievement system definition module, a three-level association annotation module, and a weighted cumulative calculation engine. The student ability dynamic diagnosis subsystem is used to infer the student's current ability status on a weekly basis during the semester, based on the structured course model formed by the course information structured representation subsystem and the student's learning behavior data stored on the online education platform. The student ability dynamic diagnosis subsystem includes six parts: learning behavior semantic encoding module, deep knowledge tracking model, reverse mapping module, smooth update module, coverage verification module, and weekly operation mechanism. The dual-input informed consent generation subsystem is used to obtain a structured course model from the course information structured representation subsystem and the student's current ability status from the student ability dynamic diagnosis subsystem. Based on the large language model, the intelligent agent uses the structured course model and the student's current ability status to generate corresponding teaching aids for the student. The dual-input informed consent generation subsystem includes seven parts: dual-input architecture, six-node workflow, knowledge graph-enhanced hybrid retrieval, teaching semantic processing mechanism, structured prompt word constraint, implicit scaffolding adjustment mechanism, and security filtering and auditing mechanism.

2. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The course knowledge graph construction module employs a hybrid construction method combining top-down and bottom-up approaches. The top-down approach involves teachers pre-defining the top-level structure of knowledge modules based on the course syllabus and teaching objectives. The bottom-up approach utilizes natural language processing technology to automatically extract fine-grained knowledge points from textbooks, lecture notes, courseware, and past exam questions. The relationships between knowledge points in the course knowledge graph are defined as four types of relationships with pedagogical semantics, including predecessor, successor, association, and inclusion relationships. The capability achievement system definition module is a capability achievement system that encodes the capability dimensions defined by the course training objectives into "capability dimension - capability indicator - achievement threshold". The capability achievement system is a three-level digital structure, wherein the capability dimension is a comprehensive capability category defined by the course training objectives. Each capability dimension is decomposed into several observable and evaluable capability indicators, and an achievement threshold is set for each capability indicator or dimension.

3. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The three-level association annotation module is used to connect the course graph constructed by the course knowledge graph construction module, the ability achievement system defined by the ability achievement system definition module, and the key components of specific teaching activities, and form a three-level association table. The three-level association table takes a single activity task as the smallest annotation unit, annotates one or more knowledge points involved and one or more ability points served by each activity task, and assigns a corresponding weight to each activity task and its corresponding ability point association, wherein the weight value ranges from [0, 1], and the sum of the weights of all ability points associated with the same activity task is normalized to 1. The weighted cumulative calculation engine is used to calculate the student's... based on the three-level association table. One ability point The associated set of activity tasks is Each activity task requires specific ability points. The weight is ,and Students in Activity tasks The performance score on This forms a structured curriculum model; then the first... The student in the first One ability point Achievement as follows: 。 4. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The semantic encoding module is used to extract students' learning behavior data from the database of the online teaching platform and perform data cleaning processing on the learning behavior data. The fields of the learning behavior data include student identifier, activity identifier, answer result, timestamp and session identifier. Each interaction record is encoded into a structured vector based on a three-level association table; The data cleaning process includes: removing abnormal records with missing key fields from the learning behavior data; retaining only the first submission for multiple submissions by the same student to the same activity task; removing records with abnormal time logic from the learning behavior data; removing student sequences with fewer than a preset number of interaction records; the structured vector: assuming the course has N knowledge points, the task activity j is encoded as an N-dimensional binary multi-label vector, which is then concatenated with the answer result of the interaction to form an N+1-dimensional input vector; The deep knowledge tracing model is a deep knowledge tracing model based on long short-term memory network. The deep knowledge tracing model receives the student's interaction sequence, represents the student's knowledge state as a continuous latent vector and models its temporal evolution. After passing through a fully connected layer and sigmoid activation, it outputs the student's knowledge mastery probability for each knowledge point in the next time step.

5. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The reverse mapping module is used to establish a complete mapping path from the knowledge mastery probability output by the deep knowledge tracking model to the ability inference mastery level through a three-level association table; The complete mapping path includes aggregating the knowledge mastery probability into the prediction accuracy of the activity task layer, and aggregating the actual performance and prediction values ​​at the activity level into the sum evidence value of the ability points through a three-level association table. Among them, the activity tasks In time The formula for the prediction accuracy is as follows: ； For the activity task In time The accuracy of predictions; The probability of knowledge acquisition; For the first Activity Tasks The set of knowledge points involved; Set capability points The associated set of activity tasks is The completed subset is The incomplete subset is Then this ability point In time The overall value of evidence is: ； The smoothing update module is used to enable the capability state estimation to have cross-week time-series stability, and applies an exponential smoothing recursive update to the comprehensive evidence value as follows: ； in, For time Smoothed ability inference of mastery; The cognitive inertia coefficient represents the proportion of historical states retained. ; For time Smoothed ability inference of mastery level; For ability points In time The overall value of evidence; For the first Activity tasks for the first Weights of ability points; For students in their completed first The actual performance score on the activity task.

6. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The coverage verification module is used to define the evidence coverage of ability point k corresponding to any ability dimension as follows: when any ability dimension has not been covered by any activity task for more than a preset period of the semester, so that the evidence value is completely generated by the deep knowledge tracing model. ； in, For any ability dimension, the corresponding first The key skill lies in time. The coverage of evidence; For the first Activity tasks for the first Weights of ability points; For the first The set of activity tasks associated with ability points For the completed subset; The weekly operation mechanism is used to run in batches on a teaching week basis, with a weekly scheduled task triggering the following steps: loading the full student interaction sequence up to the current week; running the deep knowledge tracking model to update and obtain the knowledge mastery probability of each knowledge point; calling the reverse mapping module to calculate the comprehensive evidence value of each ability point; calling the smooth update module to apply exponential smooth recursive update to the comprehensive evidence value; calling the coverage verification module to calculate the evidence coverage and generate the student's current ability status with confidence level; writing the student's current ability status as structured data into the memory cache and setting the validity period until the next week's update for real-time querying by the dual-input informed consent generation subsystem.

7. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The dual-input architecture allows the dual-input informed assistance generation subsystem to simultaneously access two types of heterogeneous information sources. One type is a course knowledge base, which consists of three types of knowledge constructed based on the course information structured representation subsystem. These three types of knowledge are stored in a vector database format and are a knowledge point knowledge base, an activity task knowledge base, and a competency system knowledge base. The other type is student diagnostic data, which consists of the student competency dynamic diagnosis subsystem outputting the inferred mastery level of each competency point, a list of weak knowledge points, and competency development trends. The six-node workflow is that the dual-input informed assistance generation subsystem is arranged into six sequentially executed nodes for each response; Node 1 receives question requests from the front end of the online teaching platform and parses the student identifier and question intent. Node 2 calls the interface of the student ability dynamic diagnosis subsystem, inputs the student identifier, and obtains the structured data of the corresponding student's ability status. Node 3 encodes the question text using an embedding model, performs an approximate nearest neighbor search in the course knowledge vector library, and recalls relevant knowledge fragments with preset rankings. Node 4 renders the role settings, search results, formatted diagnostic data, dialogue history, and user questions into complete prompt words using a template engine. Node 5 sends the assembled prompt words to a locally deployed large language model for reasoning and generating an answer. Node 6 performs content filtering on the generated answer, adds hyperlinks to the knowledge points referenced in the answer pointing to the corresponding nodes in the course knowledge graph, and generates learning behavior data based on the question-answer pair along with the current diagnostic data snapshot, and writes it into the online teaching platform.

8. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The knowledge graph-enhanced hybrid retrieval is used to introduce knowledge graph enhancement into the retrieval process of Zeng Qian; During the offline indexing phase, the course handouts are divided into paragraphs, the core knowledge point entities involved in each paragraph are extracted, the graph database is queried to obtain the neighboring entities of the core knowledge point entities and the corresponding relationship paths, and the original paragraphs and the extended entity context are jointly encoded into vector index items. In the online retrieval phase, named entity recognition is first performed on the user's question to extract knowledge point entities. Then, the graph database is queried to obtain neighbor entities and relationship paths to form an extended query set. The original question and the extended query set are then combined for vector retrieval. The retrieval results are merged and the results with the preset ranking are input into the large language model. The instructional semantic processing mechanism is used to process the input prompts through three mechanisms to ensure the appropriateness of the output instruction; the three mechanisms include an achievement rate conversion mechanism, an inference credibility grading response mechanism, and a weakness attribution to causal chain integrity mechanism.

9. The online teaching intelligent auxiliary system according to claim 1, characterized in that, The structured cue word constraints are used to transform inter-teaching constraints into executable generation rules when the cue words are fixed into six structured paragraphs. The generation rules include role definition paragraphs, ability boundary paragraphs, achievement labeling paragraphs, causal chain integrity paragraphs, and safety constraint paragraphs. The gradual scaffolding adjustment mechanism is used in the dual-input informed consent generation subsystem where the intervention intensity is dynamically driven by the diagnostic data of the student ability dynamic diagnosis subsystem. When the student's inferred mastery of a certain ability dimension is lower than the first preset threshold, the dual-input informed consent generation subsystem provides detailed explanations, prior concept reviews, and step-by-step examples. When the student's mastery exceeds the second preset threshold, the intervention intensity is automatically reduced, and the system transitions from "explainer" to "answerer," only intervening when the student actively asks questions. The security filtering and auditing mechanism is used to perform two-layer security filtering on the output of the dual-input informed assistance generation subsystem. The first layer of security filtering is keyword filtering, and the second layer of security filtering is semantic classification filtering.

10. An intelligent online teaching assistance method that integrates course knowledge and learning assessment, characterized in that, The method, applied to the online teaching intelligent auxiliary system according to any one of claims 1-9, comprises: In the course information structured representation subsystem, the discrete course resources are constructed into a course knowledge graph through the course knowledge graph construction module, the ability achievement system definition module encodes the ability dimensions defined by the course training objectives into an ability achievement system, the course knowledge graph, ability achievement system and specific teaching activities are connected through the three-level association annotation module to form a three-level association table, and the weighted cumulative calculation engine performs weighted cumulative calculation based on the three-level association table to form a structured course model; The student ability dynamic diagnosis subsystem collects students' learning behavior data on the online teaching platform, cleans it, sorts it according to timestamps to form an interaction sequence, and encodes each interaction in the interaction sequence into a structured vector containing multi-label vectors of knowledge points and answer results by combining it with the structured course model. The student ability dynamic diagnosis subsystem takes the teaching week as the granularity, inputs the structured vector into the deep knowledge tracking model, outputs the knowledge mastery probability of each knowledge point, and transforms the knowledge mastery probability into ability mastery inference through the inverse mapping mechanism. Then, through the smooth update module and the coverage verification module, it generates the student's current ability status with confidence level. The system determines whether the stated ability status is greater than or equal to a set ability status. If not, it prohibits determining the next suggested supplementary assessment activity in the next teaching week. If so, when the dual-input informed consent generation subsystem responds to student questions, it concurrently calls the course knowledge base retrieval interface and the student diagnostic data interface to inject the structured course model and ability status into the large language model prompt words after processing the teaching semantics, generating teaching aid content. Simultaneously, After re-collecting the learning behavior data generated by students after viewing the teaching aids, the data is entered into the course information structured representation subsystem and the student ability dynamic diagnosis subsystem for iterative updates of the diagnostic results.

Images