Recording and broadcasting class variant real-time interactive learning method and system based on AI guided question and answer

Abstract

The invention discloses a real-time interactive learning method and system for recording and broadcasting course transformation based on AI guided questions and answers, and relates to the technical field of wisdom classs.A knowledge point confusion scoring matrix M is constructed based on preset knowledge points, corresponding interactive instructions are generated for the preset knowledge points, and AI guided questions and answers are started according to the interactive instructions. Analyzing text character strings of students, updating a knowledge point confusion scoring matrix M according to an analysis result, taking the knowledge point confusion scoring matrix M as input, performing data fusion on the knowledge point confusion scoring matrix M and a JSON-format file of a preset scene template, outputting prompt word character strings of an AI role system and the JSON-format file of a finite-state machine, constructing an AI dialogue engine, and performing interactive circulation. The data tuple of the interaction round is stored as an interaction log sequence, a three-dimensional performance evaluation vector is generated, the three-dimensional performance evaluation vector is matched with a preset logic set, an action instruction is generated, and the problems of one-way infusion, knowledge disjunction and difficult application of recording and broadcasting classes are solved.

Description

Technical Field

[0001] This invention relates to the field of smart classroom technology, and in particular to a method and system for transforming recorded lessons into real-time interactive learning based on AI-guided question-and-answer. Background Technology

[0002] In the field of smart classroom online education, recorded courses serve as an important vehicle for the widespread dissemination of high-quality educational resources. However, the lack of interactivity in recorded courses has always been a core pain point that restricts learning efficiency and student participation.

[0003] Currently, Chinese invention patent application number CN113301369B discloses an interactive system and method for recording and broadcasting smart classroom videos. By constructing a three-terminal collaborative architecture involving teachers, servers, and students, basic interaction in recorded classes is achieved. Teachers simultaneously record course videos during live streaming and can review them after the live stream ends. They can also record Q&A videos addressing questions from students during the live stream, questions in the chat, or questions from the server's database of key points and difficulties. These Q&A videos are tagged with knowledge points and question summaries, and associated with corresponding question content before being uploaded to the server. When students watch the recorded videos and have questions, they can ask questions via a question module. The server matches the question text with tags and associated question content, providing the corresponding Q&A video to the student, thus enabling question answering and basic interaction in recorded classes. The aforementioned technologies fall short of meeting the deep-seated needs of smart classrooms for timely, accurate, and personalized interaction: First, the generation of Q&A videos is delayed, requiring teachers to manually record them after the live stream ends. This makes it impossible to respond immediately to real-time questions raised by students while watching the recordings. Students must wait for the server to match existing Q&A videos; if no match is found, questions cannot be resolved promptly, easily disrupting the learning pace. Second, the question matching mechanism is relatively simple, relying solely on pre-set tags and question text for matching. When student questions differ from the tags or question summaries, matching errors or omissions can occur, preventing effective Q&A videos from accurately reaching students. Third, the interaction is shallow, remaining a one-way passive interaction mode from student questioning to server-responding Q&A videos. It lacks proactive interaction between teachers and students and proactive interaction simulating real classroom scenarios, making it difficult to motivate students to think and participate actively. Fourth, the coverage of Q&A content is limited, mainly focusing on common questions and general difficulties already raised during the live stream. It does not adequately support personalized and scenario-based questions arising from differences in students' knowledge base or learning progress, failing to achieve the interactive effect of individualized instruction. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies are unable to address the core pain points of traditional recorded courses, such as one-way instruction, knowledge disconnect, and application difficulties.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for transforming pre-recorded courses into real-time interactive learning based on AI-guided question answering, comprising the following steps: Step S1: Construct a knowledge point confusion scoring matrix M based on preset knowledge points, generate corresponding interactive instructions for preset knowledge points, start AI-guided question answering according to the interactive instructions, convert the student's multimodal data into a text string, analyze the text string, and update the knowledge point confusion scoring matrix M according to the analysis results. Step S2: The knowledge point confusion scoring matrix M is taken as input and fused with the JSON format file of the preset scenario template to output the AI ​​role system prompt word string and the JSON format file of the finite state machine; Step S3: Use the AI ​​role system prompt word string as a parameter, use the JSON format file of the finite state machine as the state transition logic, build an AI dialogue engine, perform an interaction loop, and store the data tuples of the interaction rounds as an interaction log sequence. Step S4: Based on the interaction log sequence, generate a three-dimensional performance evaluation vector, match the three-dimensional performance evaluation vector with a preset logic set, and generate action instructions.

[0006] Preferably, step S1 includes the following sub-steps: Step S101: Obtain the text data of the recorded course content during playback, and deconstruct the recorded course content text data into n knowledge points based on a preset knowledge point base. Construct an n×n knowledge point confusion scoring matrix M, specifically including: Initialize an n×n zero matrix in memory as the knowledge point confusion scoring matrix M. The row index i and column index j of the knowledge point confusion scoring matrix M correspond to n knowledge points, where n is the total number of knowledge points. For each knowledge point, a corresponding interactive instruction is generated. The interactive instruction is bound to the preset knowledge point through the timestamp in the recorded video. Based on the interactive instruction, AI-guided question and answer is initiated.

[0007] Preferably, step S1 further includes the following sub-steps: Step S102: Using AI-guided question-and-answer, collect multimodal data from trainees. This multimodal data includes audio and text data. Convert the multimodal data into standardized text strings and call a preset evaluation rule function library. If the text string is determined to contain preset knowledge points... and preset knowledge points If confusion occurs, then a confusion scoring matrix will be applied to the knowledge points mentioned above. The update includes: Pre-set knowledge points and preset knowledge points Corresponding knowledge point confusion scoring matrix elements Perform atomic incrementing counting to obtain the updated knowledge point confusion scoring matrix. ; The evaluation rule function library maps preset text string patterns to correlation scores between one or more preset knowledge points; The preset text string pattern includes keyword exact matching, N-gram phrase combination matching, and cosine similarity calculation of the embedding vector of the student's text string.

[0008] Preferably, step S2 includes the following sub-steps: Step S201: Query the preset character database and load the basic character setting text string; The knowledge point confusion scoring matrix is ​​scanned using a string concatenation algorithm. Find the elements of the knowledge point confusion scoring matrix. For entries exceeding a preset threshold, the knowledge point confusion scoring matrix elements are used. For entries exceeding a preset threshold, the corresponding preset text string is searched from a preset text string library; The preset text string is appended to the basic character setting text string to generate the AI ​​character system prompt string.

[0009] Preferably, step S2 further includes the following sub-steps: Step S202: Using a finite state machine algorithm, the JSON format file in the preset scenario template is parsed into a graph object G=(V,E), where V is defined as the set of state nodes and E is defined as the set of state transition rules. The set of state nodes includes a start node, a scenario node, and a termination node; The scenario node is a scenario prompt text string; The set of state transition rules includes scenario state transition rules and rollback state transition rules; The scenario state transition rule is that when the student's real-time text string is determined by the intent classifier to match the preset knowledge point of the current scenario node, the finite state machine transitions from the current scenario node to the preset next scenario node or the termination node. The fallback state transition rule is that when the student's real-time text string is determined by the intent classifier to be a non-matching preset knowledge point of the current scenario node, the finite state machine transitions from the current scenario node to the preset scenario node. Based on the elements of the knowledge point confusion scoring matrix For entries exceeding a preset threshold, a trap node is inserted into the graph object G=(V,E), and a trap state transition rule is added to the state transition set. The trap node is a correction prompt text string; The trap state transition rule is that when the student's real-time text string is determined by the intent classifier to match the knowledge point confusion scoring matrix element... When the number of entries exceeds the preset threshold, the finite state machine is forcibly transitioned from the current scenario node to the corresponding trap node; When the intent classifier determines that the student's text string matches a preset knowledge point During obfuscation, the finite state machine is forced to transition to a new trap node, and the modified graph object G=(V,E) is reserialized into a JSON format file of the finite state machine.

[0010] Preferably, step S3 includes the following sub-steps: Step S301: Use the AI ​​role system prompt word string as a preset parameter and the JSON format file of the finite state machine as the state transition logic to build an AI dialogue engine; Step S302: Perform an interactive loop based on the AI ​​dialogue engine, and record the data tuples of each interactive round. Step S303: Associate the data tuples with the session ID and the millisecond-level timestamp respectively to form an interaction log sequence.

[0011] Preferably, the interaction logic of the AI ​​dialogue engine is as follows: The system takes the student's real-time text string and current status node as input, calculates and returns a new status node through the state transition logic. The current state node includes the current start node, the current scenario node, the current trap node, and the current termination node; The prompt text string corresponding to the new status node, the dialogue history, and the prompt word string of the AI ​​role system are submitted to the AI ​​dialogue engine to generate the AI's response string; The data tuple includes the current status node, the student's real-time text string, the AI ​​role system prompt string, the new status node, and the AI's response string.

[0012] Preferably, step S4 includes the following sub-steps: Step S401, based on the interaction log sequence, generates a three-dimensional performance evaluation vector, specifically including: Based on the interaction log sequence, calculate the normalized mention frequency of the preset target knowledge point in the student's text string. The calculation expression is: ; in, To mention frequency, The time for the interactive round, For indicator functions, For the student's text string, To pre-set target knowledge points, This represents the total number of rounds in the interaction log sequence. For Boolean functions, when Mention The value is 1 if it is true, and 0 otherwise. The starting node, plot node, trap node, and termination node are taken as the set of actual state nodes; The starting node, plot node, and ending node are set as a preset optimal state node set. Based on the interaction log sequence, the accuracy rate at which the preset target knowledge point is mentioned in the preset optimal state node set is calculated, and the calculation expression is: ; in, For accuracy, This is a preset set of optimal state nodes. For logical AND, This is a preset, extremely small positive number to prevent division by zero; The path through which trainees obtain the preset target knowledge points by means of a preset set of optimal state nodes is defined as the preset optimal state path; The path through which trainees obtain the preset target knowledge points by the actual state node set is defined as the actual state path. Based on the interaction log sequence, the normalized edit distance deviation between the student's actual state path and the preset optimal state path is calculated using the following expression: ; in, To edit distance deviation, This represents the student's actual status path. To predetermine the optimal state path, Lev is the Levinstein distance algorithm, and max is the maximum value. The maximum value; The mention frequency, accuracy, and edit distance deviation are used as three-dimensional performance evaluation vectors.

[0013] Preferably, step S4 further includes the following sub-steps: Step S402 involves matching the three-dimensional performance evaluation vector with a preset logic set to generate corresponding action instructions, specifically including: If the mentioned frequency is less than the preset mentioned frequency, then a video review action command is output; If the mention frequency is greater than or equal to the preset mention frequency, and the accuracy is less than the preset accuracy, then a redo diagnostic action instruction is output. If the accuracy rate is greater than or equal to the preset accuracy rate, and the edit distance deviation is less than the preset edit distance deviation, then output a command to try a new scenario action. If the mention frequency is greater than or equal to a preset mention frequency, the accuracy is greater than or equal to a preset accuracy, and the edit distance deviation is greater than or equal to a preset edit distance deviation, then output a marker completion action command; The system will push video action instructions, diagnostic action instructions, new scenario action instructions, and completed action instructions to the learner's device.

[0014] A pre-recorded course based on AI-guided question-and-answer is transformed into a real-time interactive learning system, including a data acquisition module, a script generation module, an interaction module, and a feedback module; The data acquisition module is used to collect students' multimodal input through AI-guided question answering, convert it into text strings, and construct a knowledge point confusion scoring matrix M. The script generation module is used to merge the knowledge point confusion scoring matrix M with the JSON format file of the preset scenario template to generate the AI ​​character system prompt word string and the JSON format file of the finite state machine; The interaction module is used to build an AI dialogue engine and perform an interaction loop, storing the data tuples of the interaction rounds as an interaction log sequence. The feedback module is used to obtain a three-dimensional performance evaluation vector based on the interaction log sequence, and generate action instructions based on a preset logic set.

[0015] The beneficial effects of this invention are as follows: By binding interactive commands to specific knowledge points in recorded courses, scenario acquisition, dynamic script generation, role-playing interaction, and feedback evaluation all revolve around the target knowledge points, preventing learning from deviating from the core and achieving precise learning with targeted anchoring of knowledge points. Relying on AI-guided question-and-answer to collect multimodal input from students, and combining it with preset knowledge points to generate dynamic scripts containing simulated role behavior profiles and multi-branch nodes, a high-fidelity practical scenario is constructed. This allows students to understand knowledge through simulated applications, overcoming the dilemma of "knowing is easy, doing is difficult." Multimodal interaction enhances the immersive learning experience, and feedback can correct students' application deviations in real time, helping them quickly master the application skills of knowledge points. The generated personalized plans combine knowledge point application with students' actual actions, forming a closed loop of learning, practice, review, and improvement. This not only improves students' internalization efficiency and application ability of recorded course knowledge but also provides technical support for the innovation of recorded course teaching models, promoting the upgrade of online education from passive reception to active practice. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a method for transforming recorded lectures into real-time interactive learning based on AI-guided question answering, as provided in one embodiment of the present invention; Figure 2 This is a basic flowchart illustrating the transformation of a pre-recorded course into a real-time interactive learning system based on AI-guided question answering, as provided in one embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1 This paper presents a method for transforming pre-recorded courses into real-time interactive learning sessions based on AI-guided question answering, including the following steps: Step S1: Construct a knowledge point confusion scoring matrix M based on preset knowledge points, generate corresponding interactive instructions for preset knowledge points, start AI-guided question and answer according to the interactive instructions, convert the student's multimodal data into a text string, analyze the text string, and update the knowledge point confusion scoring matrix M according to the analysis results.

[0019] Step S2: Take the knowledge point confusion scoring matrix M as input, and perform data fusion with the JSON format file of the preset scenario template to output the AI ​​role system prompt word string and the JSON format file of the finite state machine.

[0020] Step S3: Use the AI ​​role system prompt string as a parameter, the JSON format file of the finite state machine as the state transition logic, build the AI ​​dialogue engine, perform an interaction loop, and store the data tuples of the interaction rounds as an interaction log sequence.

[0021] Step S4: Generate a three-dimensional performance evaluation vector based on the interaction log sequence, match the three-dimensional performance evaluation vector with a preset logic set, and generate action instructions.

[0022] This invention transforms passive pre-recorded lessons into an active and personalized learning process. It establishes a complete and adaptive closed-loop learning process, from passive diagnosis to dynamically generative interaction, and then to quantitative assessment and planning. Instead of using pre-set or generic interactive scripts, it takes the individual student's blind spot knowledge point confusion scoring matrix M as input, dynamically generating AI role system prompt strings and finite state machine JSON files specific to that student. This transforms the AI ​​dialogue engine from a simple learning companion or examiner into a personalized simulator that performs situational stress tests targeting specific student weaknesses. By quantifying the student's knowledge application and strategy selection in the simulation through a three-dimensional performance evaluation vector, and generating action commands, it ensures continuous iteration and convergence of the learning process, achieving a truly personalized and measurable deep learning experience.

[0023] Step S1 includes the following sub-steps: Step S101: Obtain the text data of the recorded course content during playback, and deconstruct the recorded course content text data into n knowledge points based on a preset knowledge point base. Construct an n×n knowledge point confusion scoring matrix M, specifically including: An n×n zero matrix is ​​initialized in memory as the knowledge point confusion scoring matrix M. Row index i and column index j of the knowledge point confusion scoring matrix M correspond to n knowledge points, where n is the total number of knowledge points. Cognitive diagnosis is performed, transforming the unstructured video assets of the recorded courses into a structured knowledge graph. During runtime, AI-guided question answering, i.e., diagnostic probes, collects student responses to construct a real-time cognitive confusion matrix.

[0024] Each knowledge point generates a corresponding interactive instruction. The interactive instruction is bound to the preset knowledge point through the timestamp in the recorded video, and AI-guided question and answer is launched based on the interactive instruction.

[0025] Step S101: Before the AI-guided question-and-answer session begins, a data structure is instantiated for the student's current learning session. A zero matrix is ​​allocated and initialized in memory, where n is the total number of deconstructed knowledge points. This provides an objective benchmark and necessary data container for subsequent scoring updates. Instead of the traditional n×1 knowledge point mastery vector, an n×n confusion matrix is ​​initialized. From the outset, the confusion relationships between knowledge points are defined as quantifiable and primary diagnostic targets, rather than simply tracking isolated knowledge point mastery states. This lays the data foundation for the entire dynamic generation process. After being filled, this zero matrix transforms into a scoring matrix containing a profile of the student's individual cognitive blind spots, serving as input data for dynamic script generation. This ensures that the generation of subsequent personalized trap scripts has an objective data source.

[0026] Step S1 also includes the following sub-steps: Step S102: Through AI-guided question-and-answer, multimodal data of trainees is collected. This multimodal data includes audio and text data. The multimodal data is converted into standardized text strings, and a preset evaluation rule function library is called. If the text string is determined to be a preset knowledge point... and preset knowledge points Confusion is addressed by using a knowledge point confusion scoring matrix. The update includes: Pre-set knowledge points and preset knowledge points Corresponding knowledge point confusion scoring matrix elements Perform atomic incrementing counting to obtain the updated knowledge point confusion scoring matrix. By using atomic incrementing counting operations, it is possible not only to record the frequency of obfuscation but also to quantify the weight or severity of obfuscation.

[0027] The evaluation rule function library maps preset text string patterns to association scores between one or more preset knowledge points. Through the library's hybrid analysis mechanism, it moves beyond simple right / wrong judgments, combining high-speed and high-precision rules of precise keyword matching and N-gram phrase combination with deep semantic rules of cosine similarity calculation using embedded vectors. The goal is to determine whether the learner has correctly identified the preset knowledge points. With preset knowledge points Confusing events.

[0028] The preset text string patterns include keyword exact matching, N-gram phrase combination matching, and cosine similarity calculation of the embedding vectors of the student's text string.

[0029] Step S102 involves forcibly standardizing the trainees' multimodal input data into a single text string, performing the core tasks of runtime diagnostics and data quantification, and providing a unified data format for subsequent objective analysis. The initialized zero matrix is ​​transformed into a scoring matrix filled with non-zero values. This scoring matrix is ​​no longer an empty container but a precise profile of the trainees' individual cognitive blind spots, serving as the necessary data input for dynamic script generation.

[0030] Step S1 involves a complete cognitive diagnosis and data quantification loop. This abandons the traditional model of only tracking the mastery status of n knowledge points. Instead, it initializes an n×n knowledge point confusion scoring matrix in memory at the start of the student's session. Thus, the core objective of the system design is transformed into quantifying n knowledge points. 2 The confusion relationship between pairs of knowledge points, i.e. By evaluating the rule function library, instead of simple right / wrong judgments, a hybrid analysis of keywords, N-grams, and cosine similarity of embedded vectors is used to map the student's standardized text strings to specific confusion events. Then, through atomic, incremental counting operations, the diagnostic results are quantified and populated into the knowledge point confusion scoring matrix. In this approach, the abstract cognitive state of trainees is materialized into a scoring matrix filled with objective data. This scoring matrix is ​​a precise profile of each trainee's individual cognitive blind spots and serves as the necessary data input for generating dynamic scripts.

[0031] Step S2 includes the following sub-steps: Step S201: Query the preset character database and load a basic character setting text string. A basic character setting text string is loaded as a base by querying the preset character database.

[0032] By using a string concatenation algorithm, the knowledge point confusion scoring matrix is ​​scanned. Find the elements of the knowledge point confusion scoring matrix. Entries exceeding a preset threshold will be used as elements in the knowledge point confusion scoring matrix. For entries exceeding a preset threshold, the corresponding preset text string is searched from a preset text string library. Using the knowledge point confusion scoring matrix M as diagnostic input, all confused entries exceeding the preset threshold are found by scanning the knowledge point confusion scoring matrix M. These obfuscated entries, such as pre-defined knowledge points, will be confused. and preset knowledge points Obfuscation is mapped to specific behavior instruction text in a pre-defined text string library. For example, the instruction is that when a student's discussion involves the i-th knowledge point, you must deliberately use the terminology of the j-th knowledge point in your response to test the student.

[0033] By appending a preset text string to the basic character setting text string, an AI character system prompt string is generated. This prompt text is then injected into the end of the basic character setting text string using a string concatenation algorithm, outputting a highly personalized and adversarial AI character system prompt string.

[0034] Step S201 involves dynamically constructing a behavioral blueprint for the AI ​​character, generating an objective and executable AI character system prompt string for the subsequent AI dialogue engine. Instead of loading a static or generic AI character, a diagnostic-driven text injection algorithm is executed to output a highly personalized and adversarial AI character system prompt string. This prompt string instructs the AI ​​dialogue engine, transforming the AI ​​from a neutral sparring partner into an active tester. All AI behaviors are precisely tailored to the specific cognitive blind spots exposed by the learner, ensuring the relevance and effectiveness of subsequent interactions.

[0035] Step S2 also includes the following sub-steps: Step S202: Using a finite state machine algorithm, the JSON format file in the preset scenario template is parsed into a graph object G=(V,E), where V is defined as the set of state nodes and E is defined as the set of state transition rules.

[0036] The set of state nodes includes start node, scenario node, and end node.

[0037] Context nodes are context prompt text strings used to present preset learning content to learners, ask preset questions to learners, and guide learners to perform preset operations.

[0038] The set of state transition rules includes scenario state transition rules and rollback state transition rules.

[0039] The scenario state transition rule is that when the trainee's real-time text string is determined by the intent classifier to match the preset knowledge point of the current scenario node, the finite state machine transitions from the current scenario node to the preset next scenario node or the termination node.

[0040] The fallback state transition rule is that when the student's real-time text string is determined by the intent classifier to be inconsistent with the preset knowledge point of the current scenario node, the finite state machine transitions from the current scenario node to the preset scenario node.

[0041] Based on the elements of the knowledge point confusion scoring matrix For entries exceeding the preset threshold, a trap node is inserted into the graph object G=(V,E), and a trap state transition rule is added to the state transition set.

[0042] Trap nodes are correction prompt text strings used to identify and correct students' confusion about knowledge points.

[0043] The trap state transition rule is that when the trainee's real-time text string is determined by the intent classifier to match an element of the knowledge point confusion scoring matrix... When an entry exceeds a preset threshold, the finite state machine is forcibly transitioned from the current scenario node to the corresponding trap node.

[0044] The dynamic injection mechanism for finite state machine graph objects no longer executes a fixed finite state machine algorithm, but uses the knowledge point confusion scoring matrix M as a blueprint to modify the graph object G=(V,E) in real time at the algorithm level. This modification is achieved through two key actions: first, dynamically inserting new trap nodes containing correction prompt text strings into the set of state nodes V; and second, dynamically adding trap state transition rules into the set of state transition rules E.

[0045] When the intent classifier determines that the student's text string matches a preset knowledge point During obfuscation, the finite state machine is forcibly transitioned to a new trap node, and the modified graph object G=(V,E) is reserialized into a JSON file of the finite state machine. The trap state transition rule is mandatory; that is, when the intent classifier determines that the trainee's real-time text string matches a specific preset knowledge point for obfuscation, the trap state transition rule is designed to forcibly override all scenario state transition rules and fallback state transition rules by setting the highest priority, ensuring that the finite state machine will inevitably transition to the corresponding trap node.

[0046] Step S202: A general and static preset scenario template JSON file is translated into a highly personalized and dynamic finite state machine data structure that targets the specific weaknesses of trainees. Dynamic scenario reconstruction is performed, and the finite state machine JSON format file modified by the finite state machine algorithm is output. The JSON format file is no longer a general template, but an interactive blueprint tailored for trainees and full of cognitive traps. This ensures that the execution of interactive data flow will inevitably force trainees to face and correct the specific knowledge point confusions they have exposed in specific scenarios.

[0047] Step S2 involves performing core personalized data fusion, using the knowledge point confusion scoring matrix M as key input to algorithmically modify a preset scenario template JSON file. A dual-channel dynamic injection mechanism is employed. On one hand, a string concatenation algorithm injects behavioral instructions found in a preset text string library into the AI ​​role's system prompt string, giving the AI ​​role targeted adversarial behavior; for example, the AI ​​might intentionally use confusing terminology. On the other hand, a finite state machine algorithm injects trap nodes containing corrective prompt text strings and mandatory high-priority trap state transition rules into the graph object G=(V,E) of the finite state machine JSON file. This reconstructs the interactive scenario logic itself, filling it with traps. This results in two independent but logically coordinated data files that can be executed in the next step: the AI ​​prompt string and the finite state machine JSON file. Together, these ensure that the subsequent interactive data flow execution is no longer a general and static process, but rather a customized stress test environment that forces learners to directly confront and correct their specific cognitive confusions in a concrete scenario—an environment from which they cannot avoid.

[0048] Step S3 includes the following sub-steps: Step S301: Use the AI ​​role system prompt word string as a preset parameter and the JSON format file of the finite state machine as the state transition logic to build an AI dialogue engine.

[0049] Step S301: Instantiate the AI ​​dialogue engine and state machine. Load the AI ​​role system prompt string as a fixed system prompt parameter into a large language model API client. Simultaneously, instantiate the state machine manager and load the JSON file of the finite state machine as the state transition logic of the state machine manager. This sets dual constraints, meaning instead of starting a generic AI dialogue engine, dual and immutable constraints on the AI ​​engine's behavior and logic are pre-defined at startup by loading the AI ​​role system prompt string and JSON file. The output is a configured and ready-to-execute AI dialogue engine instance, which has been fully initialized with personalized trap data.

[0050] Step S302: Perform an interactive loop based on the AI ​​dialogue engine, and record the data tuples of each interactive round.

[0051] Step S302: The AI ​​dialogue engine executes an interaction loop and records data tuples. In each interaction round, a data processing flow including state calculation and response generation is executed sequentially, and the interaction result of the round is recorded as a data tuple. The data tuple contains a set of dynamic variables for each round. Executing interaction rounds and capturing evidence is the core loop of the system, responsible for processing the complete data flow from student input to AI output. Through finite state machine-prioritized interaction logic, it is no longer a simple input / output chatbot, but a state machine-gated generative model. The architecture of calculating logic first and then generating text ensures that the dialogue will inevitably follow the designed trap path, generating data tuples in memory. These data tuples accurately record the student's decisions in each round and the AI's controlled feedback.

[0052] Step S303: Associate the data tuples with the session ID and the millisecond-level timestamp respectively to form an interaction log sequence.

[0053] Step S303 involves persisting the interaction log sequence by associating the recorded data tuples with a global session ID and a millisecond-level timestamp for that round, and appending them to persistent storage, such as a log file or database table, thus forming a complete interaction log sequence. This enriches the generated data artifacts and preserves a stateful interaction trajectory.

[0054] The interaction logic of the AI ​​dialogue engine is as follows: The system takes the student's real-time text string and the current state node as input, calculates and returns a new state node through state transition logic.

[0055] The current state nodes include the current start node, the current scenario node, the current trap node, and the current termination node.

[0056] Submit the prompt text string corresponding to the new status node, the dialogue history, and the AI ​​role system prompt word string to the AI ​​dialogue engine to generate the AI's response string.

[0057] The data tuple includes the current state node, the student's real-time text string, the AI ​​role system prompt string, the new state node, and the AI's response string.

[0058] Step S3, by performing a personalized simulation and generating an objective log, forces the combination of a finite state machine and a large language model. The finite state machine provides the rigid logic of the scenario, i.e. where the learner must go, while the AI ​​role prompts provide the adversarial behavior of the role, i.e. how the AI ​​must speak. This generates a complete sequence of interaction logs, which is no longer a simple dialogue transcription, but precise, quantifiable, and analyzable evidence of the complete decision-making process made by the learner when faced with a cognitive trap diagnosed in Step S1 and constructed in Step S2 that is tailored to the learner.

[0059] Step S4 includes the following sub-steps: Step S401: Based on the interaction log sequence, generate a three-dimensional performance evaluation vector, specifically including: Based on the interaction log sequence, calculate the normalized mention frequency of the preset target knowledge points in the learners' text strings. The calculation expression is: ; in, To mention frequency, The time for the interactive round, For indicator functions, For the student's text string, To pre-set target knowledge points, This represents the total number of rounds in the interaction log sequence. For Boolean functions, when Mention It is 1 if it is true, otherwise it is 0.

[0060] The starting node, plot node, trap node, and ending node are treated as the set of actual state nodes.

[0061] The start node, plot node, and end node are set as the preset optimal state node set.

[0062] Based on the interaction log sequence, calculate the accuracy of the preset target knowledge points being mentioned in the preset optimal state node set. The calculation expression is: ; in, For accuracy, This is a preset set of optimal state nodes. For logical AND, This is a preset, extremely small positive number to prevent division by zero.

[0063] The path through which trainees obtain the preset target knowledge points by means of a preset set of optimal state nodes is defined as the preset optimal state path.

[0064] The path through which trainees obtain the preset target knowledge points by using the set of actual state nodes is defined as the actual state path.

[0065] Based on the interaction log sequence, the normalized edit distance deviation between the student's actual state path and the preset optimal state path is calculated. The calculation expression is as follows: ; in, To edit distance deviation, This represents the student's actual status path. To predetermine the optimal state path, Lev is the Levinstein distance algorithm, and max is the maximum value. The maximum value.

[0066] The frequency of mentions, accuracy, and edit distance deviation are used as three-dimensional performance evaluation vectors.

[0067] Step S401: The interaction log sequence is converged into a three-dimensional performance evaluation vector that can be read by the machine. The evaluation hierarchy is not a single right / wrong score, but rather a deconstruction and quantification of the learner's performance from three orthogonal dimensions. The frequency of mention quantifies the learner's knowledge awareness, the accuracy quantifies the learner's ability to apply knowledge (whether the knowledge is used at the correct node, i.e., a non-trap node), and the edit distance bias quantifies the learner's strategy efficiency (whether detours are taken, i.e., whether the learner frequently falls into trap nodes). The final result is a three-dimensional performance evaluation vector containing rich diagnostic information.

[0068] Step S4 also includes the following sub-steps: Step S402 involves matching the three-dimensional performance evaluation vector with a preset logic set to generate corresponding action instructions, specifically including: If the mention frequency is less than the preset mention frequency, a video review action command will be output.

[0069] If the frequency of mentions is greater than or equal to the preset frequency of mentions, and the accuracy is less than the preset accuracy, then a redo diagnostic action command will be output.

[0070] If the accuracy is greater than or equal to the preset accuracy, and the edit distance deviation is less than the preset edit distance deviation, then output the instruction to try a new scenario action.

[0071] If the mention frequency is greater than or equal to the preset mention frequency, the accuracy is greater than or equal to the preset accuracy, and the edit distance deviation is greater than or equal to the preset edit distance deviation, then the action instruction to mark completion is output.

[0072] The system will push video action instructions, diagnostic action instructions, new scenario action instructions, and completed action instructions to the learner's device.

[0073] Step S402 involves diagnostic decision-making and planning, translating the three-dimensional performance evaluation vector into executable action instructions. Through hierarchical diagnostic logic, it moves beyond simple parallel rule matching to a rigorous hierarchical diagnostic structure that aligns with pedagogical principles. It prioritizes assessing the most basic knowledge level, only proceeding to the application level after passing the knowledge level assessment, and finally the strategy level. This layer-by-layer diagnostic logic ensures the system always identifies the learner's most fundamental bottlenecks. Ultimately, it outputs highly targeted action instructions, guiding the learner into the next most suitable learning cycle, thus ensuring the closed-loop nature of the entire learning method.

[0074] Step S4, through quantitative evaluation and closed-loop planning, outputs highly diagnostic action instructions, closes the current diagnosis, generation and execution loop, and precisely starts the next optimal learning loop. For example, if the frequency of mention is low, the diagnosis is forced to be redone, thereby ensuring the convergence and effectiveness of the entire learning system.

[0075] Example 2, refer to Figure 2 This provides a system that transforms pre-recorded courses into real-time interactive learning systems based on AI-guided question-and-answer, including a data acquisition module, a script generation module, an interaction module, and a feedback module.

[0076] The data collection module is used to collect students' multimodal input through AI-guided question answering, convert it into text strings, and construct a knowledge point confusion scoring matrix M.

[0077] The data acquisition module performs cognitive diagnosis and data quantification, responds to interactive commands in the recorded lessons, initiates AI-guided question-and-answer sessions, and forcibly standardizes the student's multimodal input (audio or keyboard) into a single text string. It analyzes the text string using a pre-defined evaluation rule function library and atomically increments the count of identified knowledge point confusion events into a knowledge point confusion scoring matrix. The final output is the filled scoring matrix, which provides a precise profile of the student's individual cognitive blind spots and serves as data input for the script generation module.

[0078] The script generation module is used to merge the knowledge point confusion scoring matrix M with the JSON format file of the preset scenario template to generate the AI ​​character system prompt word string and the JSON format file of the finite state machine.

[0079] The script generation module performs personalized data fusion and dynamic injection. It receives the knowledge point confusion scoring matrix output by the acquisition module as diagnostic input and executes two algorithms in parallel. First, a string concatenation algorithm injects the preset behavioral instruction text mapped to high-weight confusion items in the knowledge point confusion scoring matrix into a basic character setting text, generating AI character system prompt strings. Second, a finite state machine algorithm injects trap nodes containing corrective prompt text and mandatory high-priority trap state transition rules into a graph object G=(V,E) of a preset scenario template. Finally, an executable data file is output, collectively constructing an interactive blueprint full of customized cognitive traps for the interaction module.

[0080] The interaction module is used to build an AI dialogue engine and perform interactive loops, storing the data tuples of the interactive rounds as an interaction log sequence.

[0081] The interaction module executes a double-constraint interaction loop. First, it instantiates the AI ​​dialogue engine and state machine manager, loading the AI ​​character system prompts and finite state machine JSON file output by the script generation module as rigid constraints. Then, it executes the core finite state machine-priority interaction logic. In each round, the student's input string is first calculated by the state machine manager based on the finite state machine JSON file to generate a new state. This new state is then used to constrain the AI ​​dialogue engine, generating the AI's response string. Finally, it outputs an analyzable sequence of interaction logs, which serves as input to the feedback module.

[0082] The feedback module is used to obtain a three-dimensional performance evaluation vector based on the interaction log sequence and generate action instructions based on a preset logic set.

[0083] The feedback module performs quantitative evaluation and closed-loop planning. It receives the interaction log sequence output from the interaction module and executes a three-dimensional performance evaluation algorithm. This algorithm converges the learner's mention frequency in the logs, application accuracy (whether the strategy was applied at non-trap nodes), and strategy efficiency (the edit distance deviation between the path and the optimal path) into a three-dimensional performance evaluation vector. The hierarchical rule engine then applies the vector to preset thresholds layer by layer, from the knowledge layer to the application layer and then to the strategy layer, performing attribution judgments. Finally, it outputs a single, clear, and executable action command, closing the current diagnosis, generation, and execution loop and precisely initiating the next optimal learning loop.

[0084] By binding interactive commands to specific knowledge points in recorded courses, scenario acquisition, dynamic script generation, role-playing interaction, and feedback assessment all revolve around the target knowledge points, preventing learning from deviating from the core and achieving precise learning with targeted anchoring of knowledge points. Relying on AI-guided question-and-answer to collect multimodal input from students, combined with preset knowledge points, a dynamic script containing simulated role behavior profiles and multi-branch nodes is generated, constructing a high-fidelity practical scenario. This allows students to understand knowledge through simulated application, overcoming the dilemma of knowing is easy but doing is difficult. Multimodal interaction enhances the immersive learning experience, and feedback can correct students' application deviations in real time, helping them quickly master the application skills of knowledge points. The generated personalized plans combine knowledge point application with students' actual actions, forming a closed loop of learning, practice, review, and improvement. This not only improves students' internalization efficiency and application ability of recorded course knowledge but also provides technical support for the innovation of recorded course teaching models, promoting the upgrade of online education from passive reception to active practice.

[0085] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-guided question-and-answer-based method for transforming a recorded course into real-time interactive learning, characterized in that, Comprise the following steps: Step S1, based on the preset knowledge point construction knowledge point confusion score matrix M, for the preset knowledge point generates corresponding interactive instruction, according to interactive instruction start AI guide type question and answer, the multimodal data of student is converted into text string, text string is analyzed, and the knowledge point confusion score matrix M is updated according to the analysis result; Step S2, the knowledge point confusion score matrix M is regarded as input, and the JSON format file of the preset scene template is data fusion, and the AI character system prompt word string and the JSON format file of finite state machine are output; Step S3, the AI character system prompt word string is regarded as parameter, the JSON format file of finite state machine is regarded as state transition logic, constructs AI dialogue engine, and carries out interactive cycle, and the data tuple of interactive round is stored as interactive log sequence; Step S4, according to the interactive log sequence, generate three-dimensional performance evaluation vector, the three-dimensional performance evaluation vector is matched with the preset logic set, and action instruction is generated.

2. The AI-guided Q&A-based recorded course turning into real-time interactive learning method according to claim 1, characterized in that, The step S1 includes the following sub-steps: In step S101, the recorded-broadcast course content text data during playing is acquired, and the recorded-broadcast course content text data is deconstructed into n knowledge points based on a preset knowledge point library , an n*n knowledge point confusion score matrix M is constructed, specifically including: Initialize n×n zero matrix in memory as knowledge point confusion score matrix M, the row index i and column index j of the knowledge point confusion score matrix M correspond to n knowledge points, wherein n is the total number of knowledge points; For each knowledge point, generate corresponding interactive instruction, the interactive instruction is bound with the preset knowledge point through the timestamp in the recording and broadcasting video, and the AI guide type question is started based on the interactive instruction.

3. The AI-guided Q&A based recorded class turning into real-time interactive learning method according to claim 2, characterized in that, The step S1 further includes the following sub-steps: Step S102, through AI guided question and answer, multi-modal data of the student is collected, the multi-modal data includes audio data and text data, the multi-modal data is converted into a standardized text string, a preset evaluation rule function library is called, if the text string is determined to confuse the preset knowledge point and the preset knowledge point , the knowledge point confusion score matrix is updated, specifically including: Pre-set knowledge points and preset knowledge points Corresponding knowledge point confusion scoring matrix elements Perform atomic incrementing counting to obtain the updated knowledge point confusion scoring matrix. ; The evaluation rule function library maps the preset text string mode to the association score between one or more preset knowledge points; The preset text string mode includes keyword exact match, N-gram phrase combination matching and cosine similarity calculation of the embedding vector of the text string of the student.

4. The AI-guided Q&A based recorded course turning into real-time interactive learning method according to claim 3, characterized in that, The step S2 includes the following sub-steps: Step S201, query the preset role database, load the basic role setting text string; The knowledge point confusion score matrix is scanned by a string concatenation algorithm to find entries of the knowledge point confusion score matrix greater than a preset threshold, and corresponding preset text strings are found from a preset text string library for entries of the knowledge point confusion score matrix greater than the preset threshold. The preset text string is spliced to the basic role setting text string, and the AI character system prompt word string is generated.

5. The AI-guided Q&A based recorded class turning into real-time interactive learning method according to claim 4, characterized in that, The step S2 further includes the following sub-steps: Step S202, the JSON format file in the preset scene template is parsed into graph object G=(V,E) by finite state machine algorithm, define V as state node set, define E as state transition rule set; The state node set includes start node, scene node and end node; The scene node is a scene prompt text string; The state transition rule set includes scene state transition rule and backtracking state transition rule; The scene state transition rule is that when the real-time text string of the student is determined by the intent classifier as matching the preset knowledge point of the current scene node, the finite state machine is transferred from the current scene node to the preset next scene node or end node; The backtracking state transition rule is that when the real-time text string of the student is determined by the intent classifier as not matching the preset knowledge point of the current scene node, the finite state machine is transferred from the current scene node to the preset scene node. According to the knowledge point confusion score matrix element If the entry is greater than the preset threshold value, a trap node is inserted in the graph object G=(V, E), and a trap state transition rule is added to the state transition set. The trap node is a correction prompt text string; The trap state transition rule is that when the real-time text string of the student is determined by the intent classifier to match the knowledge point confusion score matrix element When the entry is greater than a preset threshold, the finite state machine is forcibly transferred from the current scenario node to the corresponding trap node. When the intent classifier judges the text string of the trainee as matching the preset knowledge point When the confusion occurs, the finite state machine is forced to transfer to a new trap node, and the modified graph object G=(V,E) is re-serialized into a JSON format file of the finite state machine.

6. The AI-guided Q&A based recorded class turning into real-time interactive learning method according to claim 5, characterized in that, The step S3 comprises the following sub-steps: Step S301, taking the AI role system prompt word string as a preset parameter, taking a JSON format file of a finite state machine as a state transition logic, and constructing an AI dialogue engine; Step S302, performing an interactive cycle according to the AI dialogue engine, and recording a data tuple of each interactive round in each interactive round; Step S303, associating the data tuple with a session ID and a millisecond-level timestamp respectively to form an interactive log sequence.

7. The AI-guided Q&A based recorded class turning into real-time interactive learning method according to claim 6, characterized in that, The interactive logic of the AI dialogue engine is: Taking a real-time text string of a student and a current state node as inputs, calculating and returning a new state node through the state transition logic; The current state node comprises a current start node, a current scenario node, a current trap node and a current end node; Submitting a prompt text string corresponding to the new state node, a dialogue history and the AI role system prompt word string to the AI dialogue engine to generate an AI response string; The data tuple comprises the current state node, the real-time text string of the student, the AI role system prompt word string, the new state node and the AI response string.

8. The AI-guided Q&A based recorded class turning into real-time interactive learning method of claim 7, wherein, The step S4 comprises the following sub-steps: Step S401, generating a three-dimensional performance evaluation vector according to the interactive log sequence, specifically comprising: According to the interactive log sequence, calculating a normalized mention frequency of a preset target knowledge point in the text string of the student, and the calculation expression is: ； in, To mention frequency, The time for the interactive round, For indicator functions, For the student's text string, To pre-set target knowledge points, This represents the total number of rounds in the interaction log sequence. For Boolean functions, when Mention The value is 1 if it is true, and 0 otherwise. Taking the start node, the scenario node, the trap node and the end node as an actual state node set; Taking the start node, the scenario node and the end node as a preset optimal state node set; According to the interactive log sequence, calculating an accuracy rate of the preset target knowledge point being mentioned in the preset optimal state node set, and the calculation expression is: ； wherein, is the accuracy rate, is a preset optimal state node set, is a logical and, is a preset minimum normal number for preventing division by zero; Defining a path of the student obtaining the preset target knowledge point through the preset optimal state node set as a preset optimal state path; Defining a path of the student obtaining the preset target knowledge point through the actual state node set as an actual state path; According to the interactive log sequence, calculating a normalized edit distance deviation between the actual state path of the student and the preset optimal state path, and the calculation expression is: ； wherein, is the edit distance deviation, is the actual state path of the student, is the preset optimal state path, Lev is the Levenshtein distance algorithm, max is the maximum value of . Taking the mention frequency, the accuracy rate and the edit distance deviation as a three-dimensional performance evaluation vector.

9. The AI-guided Q&A based recorded class turning into real-time interactive learning method according to claim 8, characterized in that, The step S4 further comprises the following sub-steps: Step S402, matching the three-dimensional performance evaluation vector with a preset logic set to generate a corresponding action instruction, specifically comprising: If the mention frequency is less than a preset mention frequency, output a review video action instruction; If the mention frequency is greater than or equal to the preset mention frequency, and the accuracy rate is less than a preset accuracy rate, output a redo diagnosis action instruction; If the accuracy rate is greater than or equal to the preset accuracy rate, and the edit distance deviation is less than a preset edit distance deviation, output a new scenario attempt action instruction; If the mentioned frequency is greater than or equal to a preset mentioned frequency, the accuracy is greater than or equal to a preset accuracy, and the edit distance deviation is greater than or equal to a preset edit distance deviation, an action instruction of marking completion is output. The review video action instruction, the redo diagnosis action instruction, the try new scene action instruction and the marking completion action instruction are pushed to the student end.

10. An AI-guided question and answer-based recorded course into real-time interactive learning system, applied to an AI-guided question and answer-based recorded course into real-time interactive learning method according to any one of claims 1-9, characterized in that, The system comprises a collection module, a script generation module, an interaction module and a feedback module; The collection module is configured to collect multi-modal input of the student through AI guided question and answer, convert the multi-modal input into a text string, and construct a knowledge point confusion score matrix M; The script generation module is configured to fuse the knowledge point confusion score matrix M and a JSON format file of a preset scene template, generate an AI role system prompt word string and a JSON format file of a finite state machine; The interaction module is configured to construct an AI dialogue engine and perform an interaction cycle, and store data tuples of an interaction round as an interaction log sequence; The feedback module is configured to obtain a three-dimensional performance evaluation vector according to the interaction log sequence, and generate an action instruction according to a preset logic set.

Images