Video knowledge automatic dotting answering and mastering evaluation system based on AI large model

Abstract

The application discloses an AI large model-based video knowledge automatic dotting answering and mastering evaluation system and belongs to the technical field of AI large models, and comprises a video knowledge element extraction module, a dynamic knowledge graph construction calculation module, a learning path cognitive navigation module, an intelligent evaluation ability portrait generation module, a graph intelligent answering community distillation module and a learning achievement externalization migration guide module. According to the application, the path branch is dynamically adjusted according to the learning behavior, and the content difficulty and the individual ability are accurately matched. The navigation interface converts the abstract knowledge association into intuitive guidance through multimodal interaction such as a visual path graph, a progress bar color mark prompt and a forward-looking risk early warning, and effectively reduces the learning disorientation and cognitive overload risk. The cognitive load theory is deeply integrated, high-complexity node continuous accumulation is actively avoided in path planning, and a just-right cognitive scaffold is provided, so that the learning process is both challenging and maintains a smooth experience, and the learning autonomy is significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of AI large model technology, specifically referring to a video knowledge automatic point-and-answer and mastery assessment system based on AI large model. Background Technology

[0002] With the rapid development of online education, video has become the core carrier of knowledge transmission; however, the traditional video learning model has significant limitations: teaching videos are essentially linear and unstructured information flows, making it difficult for learners to quickly locate key knowledge points, efficiently review weak areas, or achieve deep interaction with the content; to improve learning efficiency, the industry has tried to introduce knowledge tagging technology to mark chapters or keywords on the video progress bar, but existing solutions mostly rely on manual annotation, which is costly, time-consuming, and the annotation granularity is coarse and semantically shallow, only providing title-style indexes and failing to reflect the logical connections and cognitive levels between knowledge.

[0003] However, existing video-based automatic knowledge tracking, Q&A, and mastery assessment systems still have certain shortcomings. Current technologies suffer from shallow and static knowledge extraction, making it difficult to deeply integrate the semantic relationships of multimodal information such as voice explanations, whiteboard content, and code demonstrations. Knowledge element extraction relies on manual annotation or simple keyword matching, making it susceptible to subjective bias and lacking a confidence mechanism, resulting in insufficient reliability of the initial knowledge base. Furthermore, the various stages of the system are severely fragmented; learning, practice, assessment, answering, and application lack a closed-loop linkage centered on a dynamic knowledge graph, leading to a disconnect between theoretical learning and practical application and hindering continuous data feedback and optimization. Personalized navigation capabilities are weak, path planning is rigid, and there is a lack of real-time feedback for learners. The perception and dynamic adjustment of cognitive states make it difficult to avoid cognitive overload or learning disorientation; the assessment process operates in isolation, focusing only on the correctness of answers while ignoring the deep attribution of answering behaviors, resulting in static and one-sided ability profiles that cannot effectively translate assessment results into precise intervention instructions; the Q&A system has low matching accuracy, with frequent repetitive questions, and community discussion content is scattered and disordered, failing to transform collective wisdom into structured knowledge assets through graph anchoring and distillation mechanisms; the lack of two-way linkage between courses and the community leads to the inability to accumulate and reuse high-quality Q&A, seriously restricting the sustainable enhancement of learning efficiency and knowledge transformation effects. To address this, we propose an automatic video knowledge tracking, Q&A, and mastery assessment system based on an AI big data model. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic video knowledge tracking, Q&A, and mastery assessment system based on an AI large model, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a video knowledge automatic point-and-answer and mastery assessment system based on an AI large model, including a video knowledge element extraction module, a dynamic knowledge graph construction and calculation module, a learning path cognitive navigation module, an intelligent assessment ability profile generation module, a graph intelligent Q&A community distillation module, and a learning outcome externalization and transfer guidance module; The video knowledge element extraction module extracts atomic knowledge elements with precise time points and clear types by performing frame-level and sentence-level multimodal AI analysis on the video stream. The dynamic knowledge graph construction and calculation module constructs and calculates the dynamic knowledge graph in real time based on atomic knowledge elements; The learning path cognitive navigation module dynamically generates and optimizes personalized learning paths and cognitive navigation for each student based on dynamic knowledge graphs and real-time interaction data of students. The intelligent assessment ability profile generation module generates questions nested in real-world scenarios based on the nodes and relationships of trainees in a dynamic knowledge graph, and analyzes their answering behavior to generate a refined ability profile. The graph-based intelligent Q&A community distillation module forcibly associates all student questions and community discussions with specific nodes in the dynamic knowledge graph for precise question-and-answer matching. The learning outcome externalization and transfer guidance module generates personalized micro-projects based on the learner's ability profile and dynamic knowledge graph, guides learners to complete knowledge application, and feeds back the outcome evaluation to the system.

[0006] Preferably, the video knowledge element extraction module is wirelessly connected to the dynamic knowledge graph construction and calculation module; the dynamic knowledge graph construction and calculation module is wirelessly connected to the learning path cognitive navigation module and the intelligent assessment ability profile generation module; the learning path cognitive navigation module is wirelessly connected to the intelligent assessment ability profile generation module; the intelligent assessment ability profile generation module is wirelessly connected to the graph intelligent Q&A community distillation module and the learning outcome externalization and transfer guidance module; and the graph intelligent Q&A community distillation module is wirelessly connected to the learning outcome externalization and transfer guidance module and the dynamic knowledge graph construction and calculation module.

[0007] Preferably, the video knowledge element extraction module acquires the original video, decomposes the original video stream with millisecond-level precision, performs dynamic keyframe detection on the video frame sequence, and performs in-depth analysis on effective content frames based on visual events such as sudden changes in screen content, addition of whiteboard writing, and cursor movement in the code editor, and calls multi-task visual models in parallel. The audio stream is processed by end-to-end speech recognition to generate text transcription with precise start and end timestamps, while speech activity detection is performed simultaneously to filter out silent segments; Based on the speech sentence time window, aggregate all visual analysis results within the window; Multimodal evidence is dynamically weighted using an attention mechanism: when the speech description and visual elements have a high semantic match, the semantic confidence of the segment is enhanced, generating multimodal semantic segments with a time span. Continuous video streams are segmented into semantically coherent segment units, each corresponding to a relatively complete knowledge representation. Based on semantic coherence breakpoints, long semantic segments are divided into the smallest independent semantic units. The multimodal feature vector of each semantic unit is input into a pre-trained fine-grained classifier, which outputs standardized type labels. The comprehensive confidence of knowledge elements is calculated. The implementation is as follows: , In the formula, represents the overall confidence level of knowledge elements, with a value ranging from [0, 1]. A higher value indicates a more reliable knowledge element extraction result. M represents the modality consistency score, with a value ranging from [0, 1]. P represents the probability confidence level, with a value ranging from [0, 1], and is the maximum softmax probability output by the knowledge element type classifier. C represents the contextual logical coherence score, with a value ranging from [0, 1]. Indicates the preset dynamic sharpening index; It automatically filters out units with confidence scores below a preset threshold, generates a standardized structure for each valid knowledge element, and outputs atomic knowledge elements, including the precise start and end timestamps, type tags, structured content descriptions, relationships, and confidence scores for each knowledge element.

[0008] Specifically, the attention model is implemented by decomposing the video into two time-synchronized data streams: one stream is the frame-level features extracted by the visual model, including whiteboard text, code highlighting, chart content, etc., and the other stream is the text features generated by speech recognition and semantic parsing. Voice-guided visual focusing: Using voice content as a comprehension clue, the model automatically scans which areas in the current screen are highly relevant to the voice description. For example, if the voice mentions a double-checked lock, the model will focus on the highlighted code area and assign higher weights to the matching area. Visual feedback to voice verification: Based on the content of the screen, the model verifies whether the voice description is effectively supported by the screen. For example, when explaining an architecture diagram, the screen does have a corresponding diagram, which strengthens the judgment of semantic consistency. Multiple attention sub-modules run in parallel, each capturing relationships in different dimensions. For example, one head focuses on keyword matching, while another focuses on logical flow consistency, avoiding missing key information from a single perspective. Time sequence information is explicitly injected. When the attention weight is continuously concentrated in a certain period of time, the system determines it as a complete knowledge unit; if the weight decreases, the system automatically segments the boundaries of the segments.

[0009] Dynamic confidence calibration: High-match segments automatically receive high confidence, while low-match segments receive lower confidence.

[0010] Preferably, the dynamic knowledge graph construction and calculation module receives atomic knowledge elements in real time, dynamically maintains a sliding semantic window according to the time axis, and the window size is adaptively adjusted according to the course logic density. The text encoder generates semantic fingerprints. Perform incremental clustering: Calculate the cosine similarity between the new knowledge element and the fingerprints of existing nodes in the graph. The implementation is as follows: , In the formula, Represents dynamically weighted semantic similarity. Represents the semantic fingerprint vector of new knowledge elements. This represents the fingerprint vector of existing nodes in the graph. Represents the time-series decay function. This represents the time variable of the new sub-knowledge element and node. This represents the confidence collaborative enhancement function. This represents the initial confidence level between the new sub-knowledge element and the node. This represents the time-confidence weighting balance factor. for k represents the steepness coefficient. Indicates the critical time threshold; Specifically, incremental clustering processes atomic knowledge elements in the video stream in real time. Incremental clustering emphasizes single-line / streaming data processing; each new knowledge element is immediately identified as either belonging to an existing category or a newly created category, enabling dynamic updates to the knowledge graph. The system maintains an intelligent sliding semantic window along the video timeline. The window size is not fixed but adaptively adjusts based on the density of the course content. The window narrows to focus on details in theoretically dense segments, while it widens to capture more information in practical demonstration segments. The window retains only atomic knowledge elements strongly related to the currently learned segment. For each newly arrived atomic knowledge element within the window, a lightweight pre-trained text encoder is invoked to generate a high-dimensional semantic fingerprint. This fingerprint deeply encodes the core semantics, type features, and contextual clues of the knowledge element. The fingerprints of new knowledge elements are comprehensively compared with the fingerprints of all existing nodes in the graph: not only is the cosine similarity of semantic vectors calculated, but also time and confidence synergy are integrated. The judgment threshold is dynamically set according to the knowledge element type, and different similarity standards are adopted for different knowledge types. If the comprehensive similarity exceeds the threshold, it is merged into the existing node with the closest semantics; otherwise, it is judged as a new knowledge point and the node creation process is triggered. The content summary of the target node automatically integrates the key information of the new knowledge element, the time coverage range is expanded to include the timestamp of the new element, and the confidence is updated with weight.

[0011] If the similarity exceeds the type adaptive dynamic threshold, it is merged into the nearest node, and the content summary, time coverage interval and confidence are updated; otherwise, a new node is created, a unique graph ID is assigned and the attributes are initialized; for all node pairs in the window, three sources of evidence are computed in parallel, including temporal logical evidence, semantic interaction evidence and co-occurrence reinforcement evidence. The combined generated edge weights are Tz represents temporal logical evidence, S represents semantic interaction evidence, and F represents co-occurrence reinforcement evidence. This represents the Sigmoid function. for , This represents the time interval decay function. Indicates the degree of matching of logical rules. S represents a predefined logical relational schema template with semantics specific to the education domain; S is... , The weight coefficients represent the cross-attention scores in semantic interaction evidence computation. This represents the lightweight cross-attention score, with a value in the range [0, 1]. Indicates the keyword pattern matching strength; F is... , Indicates the thermal amplification factor. Indicates the student interaction popularity index; Indicates dynamic evidence weights. , Indicates the learning rate. This indicates the quality of the human feedback relationship. Indicates the quality of the relationship predicted by the model. This represents the gradient of the contribution of evidence to the prediction; after injecting new nodes and edges, lightweight calculus is triggered.

[0012] Preferably, lightweight computation performs calculations on newly injected nodes / edges and their associated local subgraphs, avoiding recalculation of the entire graph, thereby reducing computational complexity, in conjunction with the text flow; Triggering scenario: When a new node / edge is injected into the graph, only the representations of nodes and edges directly associated with the new element are updated, such as embedding vectors or weights; By combining the weighted fusion of three sources of evidence in the text to generate edge weights, the lightweight calculus can be abstracted into performing weighted fusion only on the local evidence involved in the newly added node / edge, and updating the corresponding edge weights, as follows: , In the formula, This represents the local edge weights corresponding to the newly added node / edge, and is output using lightweight computation. This represents the temporal logic evidence for the newly added part. This represents the semantic interaction evidence of the newly added part. This indicates that the co-occurrence of the newly added parts strengthens the evidence.

[0013] Preferably, the learning path cognitive navigation module integrates dynamic knowledge graph node mastery, cognitive load index, learning style profile, and current context anchor point to construct a four-dimensional real-time state vector. Centered on the student's current anchor point node, it performs a dual-threshold diffusion search to generate a lightweight quantum graph. Inject student status weights, and the node weights are: , In the formula, This represents the dynamic navigation weight of node v in the light quantum graph. This represents the original structural weights in a dynamic knowledge graph. This indicates the student's real-time knowledge mastery of node v. The function represents the knowledge of weak points; it is set to 1 when marking weak areas. Indicates node v in the light quantum graph connectivity in This indicates the preset weak reinforcement coefficient. This represents the preset structural sensitivity coefficient. express The maximum connectivity of all nodes in the algorithm.

[0014] Preferably, the learning path cognitive navigation module, based on a light quantum graph, solves for the optimal path within the light quantum graph, and is implemented as follows: , In the formula, Representing a path Overall score Indicates candidate learning paths, Represents a set of continuous edges. Representing edges in a dynamic knowledge graph Weight, L represents the preset cognitive difficulty of node v, and L represents the real-time cognitive load index of the learner. This indicates the student's real-time knowledge mastery of node v. Indicates the dynamic weighting coefficient; Progress bar enhancement layer: A semi-transparent navigation strip is overlaid above the video progress bar, and a node summary and recommendation reason are displayed when hovering over it; Side dynamic path graph: Visualized using force-guided graph The current node is highlighted with a pulse, and subsequent nodes gradually change transparency according to the recommended priority. Clicking on any node will instantly jump to the next node.

[0015] Specifically, the weighting coefficient update strategy involves collecting feedback from system users, including the difficulty level of the learning path, the mastery of knowledge points, and the cognitive load during the learning process. Based on this feedback, the rationality of the current dynamic weighting coefficient settings is analyzed. For example, if it is found that learners have excessively high cognitive load on certain paths, the weight of the cognitive load index needs to be adjusted; if the mastery of knowledge points is generally low, the weight of knowledge mastery needs to be adjusted. Based on the analysis results, the dynamic weighting coefficients are appropriately adjusted. The adjusted dynamic weighting coefficients are then applied to the system to re-evaluate the overall score of the path and the actual learning effect, verifying the effectiveness of the adjustment strategy. If the effect improves, the adjustment is retained; otherwise, the adjustment strategy is further optimized.

[0016] Preferably, the intelligent assessment ability profile generation module analyzes the student's current position, cognitive trajectory, and related topology in the dynamic knowledge graph in real time, calls the domain scenario template library, including business logic frameworks and constraints, and dynamically embeds anchor knowledge into the real workflow; it automatically injects typical challenge points through the graph problem-solution edge to generate question stems with business coherence and cognitive tension; it generates semantic-level interference options based on comparison-analogy relationships and common misconception nodes in the graph, and labels each option with the associated graph node ID and cognitive deviation type; it dynamically adjusts the scenario complexity by comprehensively considering the cognitive difficulty of the nodes, the student's current workload, and historical mastery, and accurately matches the questions with the student's abilities; and it records the entire answer chain data.

[0017] Preferably, the intelligent assessment ability profile generation module maps error patterns to a pre-built cognitive bias pattern library, associates them with misconception nodes and solution nodes in the knowledge graph, and generates an interpretable attribution report; incremental update of node mastery: refreshes the mastery level for each dynamic knowledge graph node associated with the question, implemented as follows: , In the formula, This represents the mastery level of node v after the update. This represents the level of control of node v before the update. Indicates the historical decay coefficient. Indicates the accuracy rate related to the question. This represents the efficiency adjustment coefficient. This indicates the use of a penalty coefficient; When the mastery of a node continuously falls below a preset threshold, high-risk weaknesses are automatically marked, and related reinforcement path suggestions are generated. A structured report is generated, highlighting the matching or deviation points between the answer and the nodes in the dynamic knowledge graph. Precise remedial resources are pushed, and the dynamic changes in ability are visualized. Group answer data is anonymously aggregated to identify common cognitive blind spots and feed back into the dynamic knowledge graph to update the weights of common misconception nodes and the question generation strategy library.

[0018] Preferably, the graph-based intelligent Q&A community distillation module captures the precise timestamp of the video at the moment of asking the question, the dynamic knowledge graph node ID focused on by the interface, and the semantic vector of the question text in real time. It matches the timestamp to the set of nodes with overlapping time intervals in the dynamic knowledge graph. When the highest confidence score is less than a preset threshold, it automatically associates the question with the main node of the student's current learning path and marks it as a fuzzy anchor. It generates a structured anchor record and performs a diffusion search in the dynamic knowledge graph within a preset time, centered on the main anchor node, to generate a question association subgraph containing directly related nodes, comparison nodes, and solution nodes. The domain terminology of the question and historical questions is normalized semantically matched, and the coefficients are adaptively adjusted and sorted according to the question type.

[0019] Preferably, the graph-based intelligent Q&A community distillation module pre-sets distillation trigger conditions, extracts core viewpoints through a key sentence extraction model, generates structured knowledge fragments, updates the common problem attribute library of anchor nodes, attaches distillation content and credibility tags, adds community practice relationship edges, associates them with comparison nodes, and generates derivative question nodes for frequently emerging new question patterns and links them to the main knowledge. Record students' adoption behavior of matching answers. If recommended answers are skipped frequently, the anchor path weight is dynamically reduced. If the number of questions in a single node suddenly increases and the adoption rate of answers is less than the preset threshold, a cognitive blind spot alarm is triggered and pushed to content optimization.

[0020] Preferably, the learning outcome externalization and transfer guidance module cross-analyzes the student's ability profile and dynamic knowledge graph, identifies the core knowledge clusters that need to be strengthened and are transferable, calls the domain scenario template library, injects the student's industry attributes, learning stage and interest tags, generates micro-project requirements with real business constraints, automatically generates the basic structure of the project, and embeds knowledge anchor annotations at key implementation locations. The evaluation is weighted across four dimensions: functional quality, rationality of knowledge application, innovation, and rationality of the solution process. The weights are dynamically allocated according to the project type. The evaluation results and process behavior data are written into the learner's ability profile, the node mastery and transferability labels are updated, the learning path module is replanned, anonymous project data is aggregated, high-frequency error patterns are identified, and the weights of common misconception nodes in the learning graph are strengthened. The micro-project generation strategy library is also updated.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention constructs a closed-loop intelligent learning system centered on a dynamic knowledge graph, achieving end-to-end intelligentization from in-depth analysis of video content to knowledge structuring, personalized navigation, precise ability diagnosis, collective wisdom accumulation, and practical ability transfer, significantly improving learning efficiency and knowledge conversion rate; 2. This invention captures the semantic relationships of multimodal information such as voice explanations, whiteboard content, and code demonstrations, automatically identifies knowledge types such as concept definitions, principles and mechanisms, and code examples, and filters low-quality content through a confidence mechanism to ensure the accuracy and completeness of knowledge extraction. This not only significantly reduces the cost and cycle of content processing, but also ensures the data quality of all subsequent intelligent services, making video content truly transform into computable, associative, and traceable cognitive units. At the same time, it avoids knowledge bias caused by the subjectivity of annotation and significantly improves the reliability of the system's initial knowledge base. 3. This invention intelligently predicts cognitive bottlenecks and provides micro-analysis resources in advance, dynamically adjusting path branches based on learning behavior to ensure a precise match between content difficulty and individual ability. The navigation interface uses multimodal interaction, such as visual path maps, progress bar color-coded prompts, and forward-looking risk warnings, to transform abstract knowledge into intuitive guidance, effectively reducing the risk of learning confusion and cognitive overload. It deeply integrates cognitive load theory, proactively avoiding the continuous accumulation of highly complex nodes in path planning and providing just the right cognitive scaffolding, making the learning process both challenging and smooth. It significantly enhances learning autonomy and a sense of accomplishment, helping learners build a clear cognitive map while efficiently mastering knowledge, realizing the transformation from passive reception to active construction. 4. This invention accurately identifies cognitive biases and blind spots in thinking through full-link behavior capture and deep attribution analysis, generating interpretable attribution reports and visualized ability profiles. The profiles are dynamically updated with multi-dimensional tags such as knowledge mastery, cognitive style, and strategic characteristics. The assessment results are instantly transformed into remedial resource recommendations and path optimization instructions, making each assessment an opportunity for a leap in ability. This not only enhances the educational value of the assessment but also deeply integrates the evaluation into the learning process, significantly enhancing the relevance and effectiveness of learning. 5. This invention significantly improves the efficiency and accuracy of answering questions by using multi-hop intelligent matching based on graph topology; through a dual semantic and structural verification mechanism, the system effectively intercepts duplicate questions and reduces invalid interactions; more importantly, the module transforms community discussions into structured knowledge: high-quality questions and answers are distilled and precipitated as graph node attributes or newly added relation edges, enabling seamless two-way linkage between courses and the community, allowing the learning process to naturally extend to collaborative discussions, and significantly improving problem-solving efficiency. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of the video knowledge automatic marking, answering, and mastery assessment system based on an AI large model according to the present invention; Figure 2 This invention describes the operational process of an AI-based large-scale model-based video knowledge automatic tagging, Q&A, and mastery assessment system. Figure 1 ; Figure 3 This invention describes the operational process of an AI-based large-scale model-based video knowledge automatic tagging, Q&A, and mastery assessment system. Figure 2 ; Figure 4 This invention describes the operational process of an AI-based large-scale model-based video knowledge automatic tagging, Q&A, and mastery assessment system. Figure 3 . Detailed Implementation

[0023] 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.

[0024] Example Please see Figures 1-4 As shown, the present invention provides a technical solution including a video knowledge element extraction module, a dynamic knowledge graph construction and calculation module, a learning path cognitive navigation module, an intelligent assessment ability profile generation module, a graph intelligent Q&A community distillation module, and a learning outcome externalization and transfer guidance module; The video knowledge element extraction module extracts precise and clearly categorized atomic knowledge elements by performing frame-level and sentence-level multimodal AI analysis on the video stream. The system adopts a post-review strategy, so the AI-generated atomic knowledge elements and derived content can be directly applied online without prior manual review. The manual correction process is only triggered when students report errors, which significantly reduces the cost of teacher review and ensures the efficiency of content generation. The post-review strategy shifts content quality control from pre-screening to a user-feedback-driven intelligent review mechanism. Specifically, teaching resources such as atomic knowledge elements, interactive learning mind maps, intelligent quiz questions, and Q&A content automatically generated by the system's AI model can be directly launched and used in teaching without prior manual review. The system also establishes a student feedback channel. When students mark content as incorrect, raise questions, or trigger frequent error correction actions, the system automatically marks the content as pending review and pushes it to human experts for precise correction and revision. The corrected content is immediately updated to the knowledge base and feeds back into the AI-generated model for optimization.

[0025] The dynamic knowledge graph construction and calculation module constructs and calculates the dynamic knowledge graph in real time based on atomic knowledge elements; The learning path cognitive navigation module, based on a dynamic knowledge graph and real-time interactive data of students, dynamically generates and optimizes personalized learning paths and cognitive navigation for each student; it simultaneously generates a two-layer guidance system: a general guidance interactive mind map for the entire course and a dedicated knowledge outline for each video; the guidance content is presented in an interactive mind map or outline interface, supporting node expansion and collapse, keyword search and dynamic operation, allowing students to explore the course structure independently, replacing the traditional method of teachers manually recording guidance videos.

[0026] The intelligent assessment ability profile generation module generates questions nested in real-world scenarios based on the nodes and relationships of trainees in a dynamic knowledge graph. It then analyzes the trainees' answering behavior to generate a refined ability profile. The system provides AI-powered question generation and a dual-mode assessment mechanism: it automatically generates test questions that are associated with specific time points and knowledge points, and displays them centrally on the test question tab on the right side of the video. It supports an unlock mode (the current knowledge point question must be answered correctly to continue playback) and a free mode (questions can be skipped to maintain learning flow). After answering the questions, the answer explanation is displayed immediately and is strongly associated with the video progress bar to identify the assessment segment.

[0027] The graph-based intelligent Q&A community distillation module forcibly links all student questions and community discussions to specific nodes in the dynamic knowledge graph for precise question-and-answer matching. It integrates an AI-powered intelligent Q&A assistant with the community ecosystem: the right panel of the video provides real-time answers from an AI learning assistant; the question panel dynamically refreshes historical questions associated with the current time point as the playback progresses, including question titles and details, and all questions are automatically anchored to video time points and visually marked on the progress bar; the question content is synchronized to the learning community with the complete course context, and community users can click the associated course button to jump to the precise video location with one click, achieving seamless two-way traffic redirection between the course and the community.

[0028] The learning outcome externalization and transfer guidance module generates personalized micro-projects based on the learner's ability profile and dynamic knowledge graph, guides learners to complete knowledge application, and feeds back the outcome evaluation to the system.

[0029] In this embodiment, the video knowledge element extraction module is wirelessly connected to the dynamic knowledge graph construction and calculation module; the dynamic knowledge graph construction and calculation module is wirelessly connected to the learning path cognitive navigation module and the intelligent assessment ability profile generation module; the learning path cognitive navigation module is wirelessly connected to the intelligent assessment ability profile generation module; the intelligent assessment ability profile generation module is wirelessly connected to the graph intelligent Q&A community distillation module and the learning outcome externalization and transfer guidance module; and the graph intelligent Q&A community distillation module is wirelessly connected to the learning outcome externalization and transfer guidance module and the dynamic knowledge graph construction and calculation module.

[0030] In this embodiment, the video knowledge element extraction module acquires the original video, decomposes the original video stream with millisecond-level precision, such as synchronous data streams: continuous video frame sequences, audio streams, and embedded subtitle streams, and establishes a unified timeline index. It performs dynamic keyframe detection on the video frame sequence, and based on visual events such as sudden changes in screen content, additions to whiteboard notes, and cursor activity in the code editor, it performs in-depth analysis on valid content frames and calls a multi-task visual model in parallel. For example, the OCR engine accurately recognizes text within the frame, including handwritten notes, PPT titles, and code comments, while preserving spatial location and timestamps; The scene understanding model annotates the main elements of the scene, such as architecture diagrams, debugging interfaces, and formula derivations; The visual change tracking module marks the start and end boundaries of content, such as slide transition points and highlighted areas for code modifications.

[0031] In this embodiment, the audio stream is processed by end-to-end speech recognition to generate text transcription with precise start and end timestamps, while speech activity detection is performed simultaneously to filter out silent segments. Perform fine-grained linguistic analysis on the transcribed text: Intelligent sentence boundary segmentation, combined with semantic pauses and punctuation; Dependency syntax and semantic role labeling identify subject-action-object structures, such as reflection → invocation → private constructor method; Named entity and terminology boundary recognition focuses on emerging technical concepts and method names within the course domain; Based on the speech sentence time window, aggregate all visual analysis results within the window, such as the blackboard text and corresponding charts that appear when the sentence is explained.

[0032] In this embodiment, multimodal evidence is dynamically weighted through an attention mechanism: when the voice description and visual elements are highly semantically matched, such as when the voice mentions a double-checked lock and the relevant code is highlighted in the frame, the semantic confidence of the segment is strengthened, multimodal semantic segments with time span are generated, and the continuous video stream is divided into semantically coherent segment units, with each unit corresponding to a relatively complete knowledge expression.

[0033] Based on semantic coherence breakpoints, such as pauses in speech logic and changes in screen content, long semantic segments are divided into the smallest independent semantic units. Combining visual event markers and speech stress positions, the start and end times of each unit are fine-tuned to frame-level precision. The multimodal feature vectors of each semantic unit are fused with text semantics, visual element types, and contextual logic, and input into a pre-trained fine-grained classifier to output standardized type labels: concept definition / principle mechanism / code example / comparison and analysis / application scenario / common misconceptions.

[0034] Specifically, calculate the overall confidence level of knowledge elements. The implementation is as follows: , In the formula, represents the overall confidence level of knowledge elements, with a value ranging from [0, 1]. A higher value indicates a more reliable knowledge element extraction result. M represents the modal consistency score, with a value ranging from [0, 1], quantifying the cross-modal semantic alignment strength between speech semantics and visual content. P represents the probability confidence level, with a value ranging from [0, 1], taking the maximum softmax probability output by the knowledge element type classifier. C represents the contextual logical coherence score, with a value ranging from [0, 1]. This indicates the preset dynamic sharpening index. for , This represents the sensitivity coefficient, with a value range of [1, 3], which controls the intensity of the weakest link effect. This represents the smoothing constant.

[0035] Specifically, it automatically filters out units with confidence scores below a preset threshold, generates a standardized structure for each valid knowledge element, and outputs atomic knowledge elements, including the precise start and end timestamps, type tags, structured content descriptions, relationships, and confidence scores for each knowledge element. All generated content follows a post-review mechanism to ensure efficient deployment and continuous optimization.

[0036] In this embodiment, the dynamic knowledge graph construction and calculation module receives atomic knowledge elements in real time, dynamically maintains the sliding semantic window according to the time axis, and the window size is adaptively adjusted according to the course logic density. The text encoder generates semantic fingerprints. Perform incremental clustering: Calculate the cosine similarity between the new knowledge element and the fingerprints of existing nodes in the graph. The implementation is as follows: , In the formula, Represents dynamically weighted semantic similarity. Represents the semantic fingerprint vector of new knowledge elements. This represents the fingerprint vector of existing nodes in the graph. Represents the time-series decay function. for , This represents the time difference between the new atomic knowledge element and the midpoint of the node content. Indicates the adaptive time constant for course type. Indicates the attenuation coefficient. This represents the confidence collaborative enhancement function. for , This represents the original confidence level of the new atomic knowledge element and the node. This represents the type matching gain coefficient, with a value range of [0, 1]. Indicates a type-consistent indicator function. This represents the time-confidence weighting balance factor. for k represents the steepness coefficient. Indicates the critical time threshold, when This indicates side-mounted communication coordination, when , indicating side-time decay; The type consistency indicator function is used to determine whether the new atomic knowledge element and the node have the same type; if the types are consistent, then... If the types are inconsistent, then , for: .

[0037] Specifically, if the similarity exceeds the type adaptive dynamic threshold, it is merged into the nearest node, and the content summary, time coverage interval and confidence are updated; otherwise, a new node is created, a unique graph ID is assigned and the attributes are initialized; for all node pairs in the window, three sources of evidence are computed in parallel, including temporal logical evidence, semantic interaction evidence and co-occurrence reinforcement evidence. Temporal logic evidence Tz: Based on time intervals and the course logic rule base, such as problem description → solution interval ≤ preset time is considered a strong correlation, and output confidence of [0, 1]; Semantic interaction evidence S: The content coupling strength is calculated through a lightweight cross-attention mechanism, and the relationship type is automatically labeled by combining keyword patterns, such as difference / comparison → comparison-analogy; Co-occurrence reinforcement evidence F: Statistically calculate the co-occurrence frequency of nodes within the same video segment, and dynamically weight it by integrating the historical interaction popularity of trainees.

[0038] In this embodiment, the combined generated edge weights are: Tz represents temporal logical evidence, S represents semantic interaction evidence, and F represents co-occurrence reinforcement evidence. This represents the Sigmoid function. for , This represents the time interval decay function. Indicates the degree of matching of logical rules. S represents a predefined logical relational schema template with semantics specific to the education domain; S is... , The weight coefficients represent the cross-attention scores in semantic interaction evidence computation. This represents the lightweight cross-attention score, with a value in the range [0, 1]. Indicates the keyword pattern matching strength; F is... , Indicates the thermal amplification factor. Indicates the student interaction popularity index; Indicates dynamic evidence weights. , Indicates the learning rate. This indicates the quality of the human feedback relationship. Indicates the quality of the relationship predicted by the model. This represents the gradient of the contribution of evidence to the prediction; after injecting new nodes and edges, lightweight calculus is triggered.

[0039] In this embodiment, the learning path cognitive navigation module integrates dynamic knowledge graph node mastery, cognitive load index, learning style profile, and current context anchor point to construct a four-dimensional real-time state vector. Centered on the student's current anchor point node, it performs a dual-threshold diffusion search to generate a lightweight quantum graph. Inject student status weights, and the node weights are: , In the formula, This represents the dynamic navigation weight of node v in the light quantum graph. This represents the original structural weights in a dynamic knowledge graph. This indicates the student's real-time knowledge mastery of node v. The function represents the knowledge of weak points; it is set to 1 when marking weak areas. Indicates node v in the light quantum graph connectivity in This indicates the preset weak reinforcement coefficient. This represents the preset structural sensitivity coefficient. express The maximum connectivity of all nodes in the algorithm.

[0040] In this embodiment, the learning path cognitive navigation module, based on a light quantum graph, solves for the optimal path within the light quantum graph, and is implemented as follows: , In the formula, Representing a path The higher the overall score, the better. Indicates candidate learning paths, Represents a set of continuous edges. Representing edges in a dynamic knowledge graph Weight, L represents the preset cognitive difficulty of node v, and L represents the real-time cognitive load index of the learner. This indicates the student's real-time knowledge mastery of node v. This represents the dynamic weighting coefficient.

[0041] Specifically, the weighting coefficient update strategy involves collecting feedback from system users, including the difficulty level of the learning path, the mastery of knowledge points, and the cognitive load during the learning process. Based on this feedback, the rationality of the current dynamic weighting coefficient settings is analyzed. For example, if it is found that learners have excessively high cognitive load on certain paths, the weight of the cognitive load index needs to be adjusted; if the mastery of knowledge points is generally low, the weight of knowledge mastery needs to be adjusted. Based on the analysis results, the dynamic weighting coefficients are appropriately adjusted. The adjusted dynamic weighting coefficients are then applied to the system to re-evaluate the overall score of the path and the actual learning effect, verifying the effectiveness of the adjustment strategy. If the effect improves, the adjustment is retained; otherwise, the adjustment strategy is further optimized.

[0042] In this embodiment, the progress bar enhancement layer includes a semi-transparent navigation strip overlaid above the video progress bar, with path nodes marked by color (green = mastered, orange = to be learned, red = weak warning). Hovering over the node displays a summary and recommendation reason. Knowledge anchors include three elements: a thumbnail of the video frame, a summary title of the knowledge point, and a brief content description. The system supports both hover pop-up and fixed directory display modes. The fixed directory is presented as a clickable timeline thumbnail, allowing learners to click on any knowledge point at any time to accurately jump to the corresponding time point. The fully visualized navigation facilitates quick location and error correction.

[0043] Side dynamic path graph: Visualized using force-guided graph The current node is highlighted with a pulse, and subsequent nodes gradually change transparency according to the recommended priority. Clicking on any node will instantly jump to the next node. Based on the graph, the next high-risk node is predicted, and a light prompt pops up in the lower right corner of the video 15 seconds in advance. The system records the student's response to the navigation prompts, including the adoption rate, skip rate, and path execution completion rate. The interactive mind map supports exploration of the global course structure and focus on section outlines. Students can independently expand and collapse nodes and search for keywords to achieve a personalized learning experience.

[0044] In this embodiment, the intelligent assessment capability profile generation module analyzes the trainee's current position, cognitive trajectory, and related topology in the dynamic knowledge graph in real time, calls the domain scenario template library, including business logic framework and constraints, and dynamically embeds anchor knowledge into the real workflow.

[0045] Specifically, the system automatically injects typical challenge points into the problem-solution edge of the knowledge graph, generating question stems with business coherence and cognitive tension. Based on the comparison-analogy relationships and common misconception nodes in the knowledge graph, semantic-level interference options are generated. Each option is labeled with the associated knowledge graph node ID and cognitive bias type. The complexity of the scenario is dynamically adjusted by comprehensively considering the cognitive difficulty of the nodes, the student's current workload, and historical mastery, such as adding constraints and nesting related knowledge points, to accurately match the questions with the student's abilities. The system records the entire answer chain data, including option switching heatmaps, code modification trajectories, prompt call timing, pause thinking time, and backtracking behavior. Options are matched to dynamic knowledge graph nodes, and the consistency between the selection logic chain and the knowledge graph path is analyzed. Key elements are extracted through domain-fine-tuned NLP models, and the semantic alignment and element coverage completeness with the standard answer node are calculated. The system clearly distinguishes between unlock mode and free mode: unlock mode forces students to master the current knowledge point before continuing to learn, ensuring a solid foundation; free mode respects the student's own pace, improving the smoothness and experience of learning.

[0046] In this embodiment, the intelligent assessment ability profile generation module maps error patterns to a pre-built cognitive bias pattern library, such as ignoring concurrency boundaries and confusing lifecycles, and associates them with misconception nodes and solution nodes in the knowledge graph to generate an interpretable attribution report; incremental update of node mastery: the mastery level is refreshed for each dynamic knowledge graph node associated with the question, implemented as follows: , In the formula, This represents the mastery level of node v after the update. This represents the level of control of node v before the update. Indicates the historical decay coefficient. Indicates the accuracy rate related to the question. This represents the efficiency adjustment coefficient. This indicates the use of a penalty coefficient; a multi-dimensional capability tagging system: Knowledge dimension: Label node status and weak links; Cognitive dimension: Generating dynamic labels based on behavioral clustering; Behavioral dimension: Identify policy characteristics.

[0047] Specifically, when the mastery of a node is continuously below a preset threshold or the causes of errors are highly concentrated, high-risk weaknesses are automatically marked, and suggestions for related reinforcement paths are generated. A structured report is generated, highlighting the matching or deviation points between the answer and the nodes in the dynamic knowledge graph. Precise remedial resources, such as comparative mind maps, are pushed, and the dynamic changes in ability are visualized. Group answer data is anonymously aggregated to identify common cognitive blind spots and feed back into the dynamic knowledge graph to update the weights of common misconception nodes and the question generation strategy library.

[0048] In this embodiment, the graph-based intelligent Q&A community distillation module captures the precise timestamp of the video at the moment of asking a question, the dynamic knowledge graph node ID focused on by the interface, and the semantic vector of the question text in real time. It matches the timestamp to the set of nodes with overlapping time intervals in the dynamic knowledge graph. When the highest confidence level is less than a preset threshold, it automatically associates the question with the main node of the student's current learning path and marks the fuzzy anchor, generating a structured anchor record.

[0049] Centered on the main anchor node, a diffusion search is performed within a preset time in the dynamic knowledge graph, generating a question association subgraph containing directly related nodes, comparison nodes, and solution nodes. The domain terminology of the question and historical questions is normalized to achieve semantic matching degree, and the coefficients are adaptively adjusted and sorted according to the question type. The right panel synchronously displays the AI ​​Q&A assistant's real-time dialogue window and the historical question list associated with the current time point. The question content is automatically synchronized to the learning community and carries the complete context of the course chapter video time point. If the sorted question types are greater than a preset threshold and the semantic overlap is greater than a preset threshold, the system automatically pushes historical highly matched answers and prompts that there is already an accurate answer, thus blocking the process of asking repeated questions.

[0050] In this embodiment, the graph-based intelligent Q&A community distillation module pre-sets distillation trigger conditions, such as meeting the density of expert certification and likes, or being cited ≥5 times within 72 hours. It extracts core viewpoints through a key sentence extraction model, generates structured knowledge fragments, updates the common problem attribute library of anchor nodes, attaches distillation content and credibility tags, adds community practice relationship edges, such as the applicability of solution A in scenario X, and associates them with comparison nodes. For frequently emerging new question patterns, it generates derivative question nodes and links them to the main knowledge. Seamless two-way interaction between courses and the community: Enhancements to the course platform: The question panel on the right side of the video refreshes dynamically based on the current anchor node, highlighting highly-rated community answers and expert-certified answers. The progress bar marks the hottest questions, and hovering over it displays a question summary and community popularity index.

[0051] Community-side enhancements: The question card is embedded in the course context tag, which includes the course / chapter / anchor node name; Clicking the course jump button will accurately locate the corresponding time point in the video and the corresponding graph node.

[0052] Interactive closed loop: After students supplement high-quality answers in the community, the system pushes a confirmation prompt to the course Q&A database. After being adopted, the course content is enhanced in real time. Record students' adoption behavior of matching answers. If recommended answers are skipped frequently, the anchor path weight is dynamically reduced. If the number of questions in a single node suddenly increases and the adoption rate of answers is less than the preset threshold, a cognitive blind spot alarm is triggered and pushed to content optimization.

[0053] In this embodiment, the learning outcome externalization and transfer guidance module cross-analyzes the learner's ability profile and dynamic knowledge graph, identifies the core knowledge clusters that need to be strengthened and are transferable, calls the domain scenario template library, injects the learner's industry attributes, learning stage, and interest tags, and generates micro-project requirements with real business constraints. For example, it designs a thread-safe configuration manager for high-concurrency product services, clearly marks the associated graph node ID, migration target, and deliverable form, and intelligently adjusts the constraint complexity, the number of associated knowledge points, and the density of boundary conditions based on the learner's real-time cognitive load index and historical task completion efficiency to ensure a balance between challenge and feasibility. It automatically generates the basic project structure and embeds knowledge anchor annotations at key implementation locations, such as clicking to jump to the micro-analysis of the graph node.

[0054] In this embodiment, the dynamic resource matrix aggregates three types of precise resources in real time in the sidebar. Instant Review: 30-second micro-videos or interactive mind maps linking nodes in the graph; Pitfall Avoidance Guide: Frequently Asked Errors and Correction Solutions in Community Distillation; Solution Comparison: A multi-solution decision card generated by comparing and analyzing nodes in the graph; The evaluation is weighted across four dimensions: functional quality, rationality of knowledge application, innovation, and rationality of the solution process. The weights are dynamically allocated according to the project type. The evaluation results and process behavior data are written into the learner's ability profile, the node mastery and transferability labels are updated, the learning path module is replanned, anonymous project data is aggregated, high-frequency error patterns are identified, and the weights of common misconception nodes in the learning graph are strengthened. The micro-project generation strategy library is also updated.

[0055] Working principle: By breaking down the original teaching video into three data streams—video frame sequence, audio stream, and subtitle text—at millisecond level, a unified time reference is established. The system intelligently identifies visual events such as sudden changes in screen content, whiteboard updates, and code editing that trigger keyframe analysis, simultaneously performing text recognition, scene semantic annotation, and visual change tracking. The audio stream is processed by speech recognition to generate timestamped text, which is then segmented into sentences and analyzed for semantic roles. Using the speech semantic unit as a window, all visual elements within the window are integrated. A multimodal attention mechanism is used to determine the semantic consistency between speech and visuals, generating coherent semantic segments and segmenting them into the smallest knowledge units. The units are classified into knowledge types using fine-grained classification. By comprehensively considering modal consistency, model confidence, and contextual coherence to calculate reliability, low-quality content is automatically filtered out, and atomic knowledge units with accurate timing and clear types are output. All generated content adopts a post-review strategy, requiring no prior manual review before going online. Manual correction is only triggered when students report errors. The system receives atomic knowledge elements in real time and dynamically maintains the semantic analysis window based on the logical density of the course. For each new knowledge element, the system calculates its semantic similarity to existing nodes in the graph, and intelligently decides whether to merge it into an existing node or create a new node, taking into account temporal proximity and confidence co-occurrence. It also updates the content summary and time coverage simultaneously. For relationships between nodes, the system derives association evidence from three dimensions: temporal logic rules, semantic interaction strength, and student behavior co-occurrence intensity, generating edge weights with educational semantics. Injecting new nodes and edges triggers lightweight quantum graph calculus, optimizing the structure and eliminating redundancy. Student interactions such as asking and answering questions are continuously fed back to the graph, dynamically strengthening high-frequency associations and marking cognitive blind spots, allowing the knowledge network to grow and improve autonomously with the real learning process. This system transforms and forms a living knowledge hub that combines course logic with group cognitive characteristics; it deeply integrates students' real-time mastery status, cognitive load level, learning style characteristics, and current knowledge anchors to construct a dynamic cognitive state model; it performs a dual-threshold diffusion search centered on the student's current node to generate a lightweight quantum graph focusing on key needs, and dynamically weights it based on mastery gaps and the importance of node structure; the system plans personalized learning paths, balancing the logical coherence of knowledge with the rationality of cognitive load; the navigation interface overlays knowledge anchors on the video progress bar, each anchor containing a thumbnail, knowledge point title, and brief description, supporting both hover pop-up and fixed right-side directory display modes. The fixed directory is presented as a clickable timeline thumbnail, allowing students to precisely jump to the relevant sections at any time. Simultaneously, a two-tiered learning system is generated: a global interactive mind map of the entire course and a dedicated knowledge outline for each video, presented in an interactive mind map interface. This supports node expansion and collapse, keyword search, and self-exploration, replacing traditional manual recording of instructional videos by teachers, and achieving intelligent visualization and personalized navigation of the course structure. Based on the student's real-time location and learning trajectory in the dynamic knowledge graph, a domain scenario template library is invoked to seamlessly embed knowledge points into real business workflows, generating contextualized questions. The system deeply integrates problem-solving chain and common misconception nodes to design question stems and distractors. The system provides a dual-mode assessment mechanism: the unlock mode requires correct answers to the current knowledge point questions to continue learning, ensuring a solid foundation; the free mode allows skipping questions to maintain a smooth learning flow. The system features include: immediate display of answer explanations after answering questions; strong correlation between each question and video time point to identify the assessment segment; full-link capture of answering behavior; precise attribution of errors to the cognitive bias pattern library based on answer content and in-depth comparison with standard nodes in the knowledge graph; dynamic updating of knowledge point mastery based on answer quality, efficiency, and independence; construction of a multi-dimensional ability profile covering knowledge status, cognitive characteristics, and strategic tendencies; driving personalized feedback and path optimization; and anonymous aggregation of group answer data to feed back into the knowledge graph iteration; mandatory anchoring of student questions to precise nodes in the dynamic knowledge graph; triple verification by integrating video timestamps at the time of questioning, interface focus nodes, and question semantic vectors; and intelligent association to the current main learning path when confidence is insufficient.A multi-hop question subgraph is constructed around anchor nodes, and a three-dimensional sorting is performed based on content semantic matching degree, graph topological affinity, and community consensus popularity to accurately push high-quality historical answers. The right panel integrates an AI learning assistant to provide real-time Q&A, while dynamically refreshing the historical question list associated with the current time point. All questions are automatically anchored to the video time point and visually marked on the progress bar. Question content is synchronized to the learning community with complete course context. Community users can click the associated course button to jump back to the precise video position with one click, achieving seamless two-way traffic between courses and the community. High-quality community discussions are distilled and refined into structured knowledge fragments, updating graph node attributes or adding practical relationship edges, accumulating collective wisdom into system assets, and continuously monitoring adoption behavior to optimize anchoring strategies and blind spot warnings. In-depth cross-analysis of student ability profiles and dynamic knowledge graphs accurately identifies knowledge areas that need to be strengthened. The system generates personalized micro-project tasks with real business constraints by combining knowledge points and high-transfer-potential scenarios with learners' industry backgrounds and interest tags. It automatically generates project frameworks with knowledge anchor annotations, allowing learners to instantly review related knowledge point analyses by clicking on the annotations. The sidebar dynamically aggregates resources for immediate review, pitfall avoidance guides, and solution comparison cards, forming an adaptive cognitive scaffold. Upon project completion, a comprehensive evaluation is conducted across four dimensions: functional implementation, knowledge application logic, innovative thinking, and problem-solving process, generating a visual growth report and improvement suggestions. Evaluation results and process behavior data are written into the learner's competency profile in real time, triggering subsequent learning path optimization. High-performing learner achievements are anonymized and transformed into practical case nodes in the knowledge graph, enriching the application dimensions of the knowledge network. The system continuously aggregates group project data, identifies high-frequency error patterns and success strategies, and iteratively optimizes the micro-project generation logic and guidance strategy library.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0057] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A video knowledge automatic marking, answering, and mastery assessment system based on an AI large-scale model, characterized by: It includes a video knowledge element extraction module, a dynamic knowledge graph construction and calculation module, a learning path cognitive navigation module, an intelligent assessment ability profile generation module, a graph-based intelligent Q&A community distillation module, and a learning outcome externalization and transfer guidance module; The video knowledge element extraction module extracts atomic knowledge elements with precise time points and clear types by performing frame-level and sentence-level multimodal AI analysis on the video stream. The dynamic knowledge graph construction and calculation module constructs and calculates the dynamic knowledge graph in real time based on atomic knowledge elements; The learning path cognitive navigation module dynamically generates and optimizes personalized learning paths and cognitive navigation for each student based on dynamic knowledge graphs and real-time interaction data of students. The intelligent assessment ability profile generation module generates questions nested in real-world scenarios based on the nodes and relationships of trainees in a dynamic knowledge graph, and analyzes their answering behavior to generate a refined ability profile. The graph-based intelligent Q&A community distillation module forcibly associates all student questions and community discussions with specific nodes in the dynamic knowledge graph for precise question-and-answer matching. The learning outcome externalization and transfer guidance module generates personalized micro-projects based on the learner's ability profile and dynamic knowledge graph, guides learners to complete knowledge application, and feeds back the outcome evaluation to the system.

2. The video knowledge automatic marking, answering, and mastery assessment system based on an AI large model as described in claim 1, characterized in that: The video knowledge element extraction module acquires the original video, decomposes the original video stream with millisecond-level precision, performs dynamic keyframe detection on the video frame sequence, performs in-depth analysis on the effective content frames, and calls the multi-task visual model in parallel. The audio stream is processed by end-to-end speech recognition to generate text transcription with precise start and end timestamps, while speech activity detection is performed simultaneously to filter out silent segments; Multimodal evidence is dynamically weighted through an attention mechanism: when the speech description and visual elements are highly semantically matched, the semantic confidence of the segment is enhanced to generate multimodal semantic segments with a time span; based on the semantic coherence breakpoint, the long semantic segment is divided into the smallest independent semantic unit, and the multimodal feature vector of each semantic unit is input into a pre-trained fine-grained classifier to output standardized type labels. Calculate the overall confidence level of knowledge elements ; It automatically filters out units with confidence scores below a preset threshold, generates a standardized structure for each valid knowledge element, and outputs atomic knowledge elements.

3. The video knowledge automatic point-and-answer and mastery assessment system based on AI large model as described in claim 1, characterized in that: The dynamic knowledge graph construction and calculation module receives atomic knowledge elements in real time, dynamically maintains the sliding semantic window according to the time axis, and the text encoder generates semantic fingerprints. Perform incremental clustering: Calculate the cosine similarity between the new knowledge element and the fingerprints of existing nodes in the graph. ; If the similarity exceeds the type adaptive dynamic threshold, it is merged into the nearest node, and the content summary, time coverage interval and confidence are updated; otherwise, a new node is created, a unique graph ID is assigned and the attributes are initialized. For all node pairs within the window, three sources of evidence are computed in parallel: temporal logical evidence, semantic interaction evidence, and co-occurrence reinforcement evidence. The total generated edge weight is Tz represents temporal logical evidence, S represents semantic interaction evidence, and F represents co-occurrence reinforcement evidence. This represents the Sigmoid function. The weights of the temporal logical evidence are represented. The weights representing semantic interaction evidence This indicates the weight of co-occurrence-enhanced evidence; injecting new nodes and edges triggers lightweight calculus.

4. The video knowledge automatic point-and-answer and mastery assessment system based on AI large model according to claim 1, characterized in that: The learning path cognitive navigation module integrates dynamic knowledge graph node mastery, cognitive load index, learning style profile, and current context anchor point to construct a four-dimensional real-time state vector. Centered on the student's current anchor point node, it performs a dual-threshold diffusion search to generate a lightweight quantum graph. Inject student status weights, and the node weights are: ， In the formula, This represents the dynamic navigation weight of node v in the light quantum graph. This represents the original structural weights in a dynamic knowledge graph. This indicates the student's real-time knowledge mastery of node v. The function represents the knowledge of weak points; it is set to 1 when marking weak areas. Indicates node v in the light quantum graph connectivity in This indicates the preset weak reinforcement coefficient. This represents the preset structural sensitivity coefficient. express The maximum connectivity of all nodes in the algorithm.

5. The video knowledge automatic point-and-answer and mastery assessment system based on AI large model according to claim 4, characterized in that: The learning path cognitive navigation module, based on a light quantum graph, solves for the optimal path within the light quantum graph, and is implemented as follows: ， In the formula, Representing a path Overall score Indicates candidate learning paths, Represents a set of continuous edges. Representing edges in a dynamic knowledge graph Weight, L represents the preset cognitive difficulty of node v, and L represents the real-time cognitive load index of the learner. This indicates the student's real-time knowledge mastery of node v. This represents the dynamic weighting coefficient.

6. The video knowledge automatic marking, answering, and mastery assessment system based on an AI large model according to claim 1, characterized in that: The intelligent assessment ability profile generation module analyzes the student's current position, cognitive trajectory, and related topology in the dynamic knowledge graph in real time, calls the domain scenario template library, dynamically embeds anchor knowledge into the real workflow, and dynamically adjusts the scenario complexity based on the cognitive difficulty of nodes, the student's current workload, and historical mastery, so as to accurately match the questions with the student's abilities; and records the entire answer chain data.

7. The video knowledge automatic point-and-answer and mastery assessment system based on AI large model according to claim 6, characterized in that: The intelligent assessment ability profile generation module maps error patterns to a pre-built cognitive bias pattern library, associates them with misconception nodes and solution nodes in the knowledge graph, and generates an interpretable attribution report; incremental update of node mastery: refreshes the mastery of each dynamic knowledge graph node associated with the question; When the mastery of a node continuously falls below a preset threshold, high-risk weaknesses are automatically marked, and suggestions for related reinforcement paths are generated. A structured report is generated, highlighting the matching of the answer with the dynamic knowledge graph node, and pushing precise remedial resources.

8. The video knowledge automatic marking, answering, and mastery assessment system based on an AI large model according to claim 1, characterized in that: The graph-based intelligent Q&A community distillation module captures the precise timestamp of the video at the moment of asking a question, the dynamic knowledge graph node ID focused on by the interface, and the semantic vector of the question text in real time. It matches the timestamp to the set of nodes with overlapping time intervals in the dynamic knowledge graph. When the highest confidence level is less than a preset threshold, it automatically associates it with the main node of the student's current learning path and marks it with fuzzy anchoring. Structured anchor records are generated, and a diffusion search is performed in the dynamic knowledge graph within a preset time, centered on the main anchor node. This generates a question association subgraph containing directly related nodes, comparison nodes, and solution nodes. The domain terminology of the question and historical questions is normalized to achieve semantic matching degree, and the coefficients are adaptively adjusted and sorted according to the question type.

9. The video knowledge automatic marking, answering, and mastery assessment system based on an AI large model according to claim 8, characterized in that: The graph-based intelligent Q&A community distillation module pre-sets distillation trigger conditions, updates the common question attribute library of anchor nodes, attaches distillation content and credibility tags, adds community practice relationship edges, and associates them with comparison nodes. For frequently emerging new question patterns, it generates derivative question nodes and links them to the main knowledge. It records students' adoption behavior of matching answers, and if recommended answers are frequently skipped, the anchor path weight is dynamically reduced.

10. The video knowledge automatic marking, answering, and mastery assessment system based on an AI large model according to claim 1, characterized in that: The learning outcome externalization and transfer guidance module cross-analyzes the student's ability profile and dynamic knowledge graph, identifies the core knowledge clusters that need to be strengthened and are transferable, calls the domain scenario template library, injects the student's industry attributes, learning stage and interest tags, generates micro-project requirements with real business constraints, automatically generates the basic structure of the project, and embeds knowledge anchor annotations at key implementation locations. The evaluation is weighted across four dimensions: functional quality, rationality of knowledge application, innovation, and rationality of the solution process. The weights are dynamically allocated according to the project type. The evaluation results and process behavior data are written into the learner's ability profile, the node mastery and transferability labels are updated, the learning path module is replanned, anonymous project data is aggregated, high-frequency error patterns are identified, and the weights of common misconception nodes in the learning graph are strengthened. Update the micro-project generation strategy library.

Images