An interactive exercise automatic generation and intelligent evaluation method and system based on extended markdown

Abstract

The application discloses an interactive exercise automatic generation and intelligent evaluation method and system based on extended Markdown, and belongs to the technical field of electronic digital data processing. The method comprises the following steps: adding a special tag block on the basis of a standard Markdown grammar to define an extended Markdown format; parsing the extended Markdown document, extracting structured data, and constructing a multi-scale knowledge graph, wherein the multi-scale knowledge graph comprises a forward cognitive subgraph and an abnormal logic correlation subgraph; generating personalized exercises based on the multi-scale knowledge graph and a user's personal cognitive subgraph; parsing and rendering an interactive exercise interface according to a unified interaction protocol; obtaining and distinguishing the types of answers submitted by the user, intelligently evaluating different types of answers, obtaining a score result and evidence, and feeding back the score result and the evidence, so that the accuracy of teaching intervention, the reliability of evaluation, and the adaptability and effectiveness of the whole life cycle of the system are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to an interactive exercise automatic generation and intelligent evaluation method and system based on extended Markdown. Background Technology

[0002] With the popularization of online education and the development of artificial intelligence technology, how to achieve automated generation of exercises, intelligent assessment, and personalized learning path recommendation has become a research hotspot in the field of educational technology.

[0003] In current technologies, the level of intelligence in online education systems faces bottlenecks. First, teaching content (knowledge points, exercises, and solutions) is typically scattered across files in different formats, lacking a unified, structured data source that machines can deeply understand, resulting in a lack of high-quality input for subsequent automated processing. Second, exercise generation is often based on question bank templates or shallow semantics, making it difficult to construct cognitive models that reflect individual student weaknesses (such as specific erroneous reasoning rules), thus rendering "personalization" superficial. Third, for automatic scoring of open-ended questions, existing models are susceptible to adversarial examples, and the scoring logic may deviate from fundamental subject axioms, exhibiting poor robustness and low interpretability. Furthermore, core system components are often statically fixed, making it difficult to achieve continuous self-optimization and evolution using teaching interaction data.

[0004] Therefore, there is an urgent need for an integrated solution that can start from structured content creation and achieve in-depth cognitive diagnosis, robust intelligent evaluation, and system self-evolution. Summary of the Invention

[0005] This invention provides a method and system for automatic generation and intelligent evaluation of interactive exercises based on extended Markdown, in order to address the shortcomings of existing intelligent teaching systems in terms of personalized depth of exercise generation, robustness of evaluation, and self-evolution capabilities.

[0006] In a first aspect, embodiments of the present invention provide a method for automatic generation and intelligent evaluation of interactive exercises based on extended Markdown, including: Extended Markdown formatting is defined by adding proprietary tag blocks to the standard Markdown syntax. The extended Markdown document is parsed to extract structured data and construct a multi-scale knowledge graph, wherein the multi-scale knowledge graph includes: a positive cognitive subgraph and an abnormal logical association subgraph; Personalized exercises are generated based on the multi-scale knowledge graph and the user's personal cognitive subgraph. Based on the personalized exercises, the interactive exercise interface is parsed and rendered according to the unified interaction protocol; The system acquires and distinguishes the types of answers submitted by users, performs intelligent evaluation on different types of answers, obtains scoring results and evidence, and provides feedback on the scoring results and evidence.

[0007] Optionally, the extended Markdown document is parsed to extract structured data and construct a multi-scale knowledge graph, including: Parse the extended Markdown document to extract knowledge points, videos, exercises, and thought process templates; Construct a concept network based on the aforementioned knowledge points; Based on the exercises, videos, and thought process templates, they are associated with the knowledge points in the concept network, and a positive cognitive subgraph and an abnormal logic association subgraph are constructed. The positive cognitive subgraph and the abnormal logic association subgraph are connected by cognitive deviation weight edges.

[0008] Optionally, based on the multi-scale knowledge graph and the user's personal cognitive subgraph, personalized exercises are generated, including: Upon receiving a question request, the system invokes the user's personal cognitive subgraph, which includes the probability of knowledge mastery, a set of production rules, and metacognitive strategy labels. Based on the aforementioned personal cognitive subgraph, a cognitive model is run to generate simulation results; Based on the simulation results, the problem-generating agent combines the multi-scale knowledge graph with the course objectives to generate corresponding diagnostic and training problems.

[0009] Optionally, after generating personalized exercises based on the multi-scale knowledge graph, the method further includes: The personalized exercises are validated in multiple layers through a validation pipeline, including: knowledge accuracy validation, cognitive level matching validation, and format conformity validation.

[0010] Optionally, the system acquires and distinguishes the types of answers submitted by users, performs intelligent evaluation on different types of answers, obtains scoring results and evidence, and provides feedback on the scoring results and evidence, including: The system acquires and differentiates the types of user-submitted answers, including answers to objective questions and answers to open-ended questions. For the answers to the objective questions, they are matched with the pre-stored standard answers to determine the judgment result; For the answers to the free-response questions, the scoring assistant agent, which has been trained in adversarial robustness, automatically grades the answers according to the predefined scoring rules and outputs the scoring results and evidence. The scoring logic of the scoring assistant agent integrates a dual semantic consistency verification mechanism. The feedback generating agent combines the rating results, the evidence, and the user's thought process inferred by the metacognitive simulator to generate and output deep feedback. The deep feedback is used to explain the reasons for errors, point out thinking biases, and provide adjustment strategies.

[0011] Optionally, after acquiring and distinguishing the types of user-submitted answers, performing intelligent evaluation on different types of answers, obtaining scoring results and evidence, and providing in-depth feedback, the method further includes: Based on the scoring results and evidence, generate adversarial questions; Continuously collect interaction data from all users with various types of exercises, including the aforementioned adversarial questions; Based on the interactive data, component optimization and map update operations are performed.

[0012] Optionally, the proprietary tag blocks include: knowledge graph definition blocks, video segment blocks, exercise card blocks, free-response question scoring rule blocks, interactive instruction blocks, metadata blocks, thought process template blocks, metacognitive prompt blocks, adversarial sample tag blocks, baseline semantic tag blocks, and bias guidance tag blocks.

[0013] Secondly, embodiments of the present invention provide an interactive exercise automatic generation and intelligent evaluation system based on extended Markdown. The system is used to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown described in any embodiment of the present invention, including: Define modules to define extended Markdown formats by adding proprietary tag blocks on top of standard Markdown syntax; The parsing module is used to parse the extended Markdown document, extract structured data, and construct a multi-scale knowledge graph, wherein the multi-scale knowledge graph includes: a positive cognition subgraph and an abnormal logic association subgraph; The generation module is used to generate personalized exercises based on the multi-scale knowledge graph and the user's personal cognitive subgraph. The rendering module is used to parse and render the interactive exercise interface based on the personalized exercises and according to the unified interaction protocol. The evaluation module is used to acquire and distinguish the types of answers submitted by users, perform intelligent evaluation on different types of answers, obtain scoring results and evidence, and provide feedback on the scoring results and evidence.

[0014] Thirdly, embodiments of the present invention provide an electronic device, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown as described in any embodiment of the present invention.

[0015] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown described in any embodiment of the present invention.

[0016] The technical solution of this invention, by defining an extended Markdown format, unifies discrete teaching materials, knowledge point relationships, exercise logic, and scoring rules into a single structured document that can be deeply understood by the machine. This fundamentally solves the bottleneck problems of "separation of content and logic" and lack of high-quality input for automated processing in existing technologies, laying a solid foundation for all subsequent advanced processing. Furthermore, in the multi-scale knowledge graph constructed by parsing this document, it innovatively includes an abnormal logic association subgraph and connects it to the positive cognitive subgraph through cognitive deviation weight edges. This achieves a quantitative representation of erroneous thinking patterns and their correlation strength with correct knowledge, enabling the system to structurally "understand" the various errors students may make and their root causes.

[0017] Based on this graph and real-time updated user-specific cognitive subgraphs (including knowledge mastery probabilities, production rule sets, and metacognitive strategies), the system can deeply simulate individual students' thinking paths and weaknesses through a metacognitive simulator. This generates highly targeted diagnostic and training exercises, achieving a leap from extensive "personalization" based on group statistics to precise "personalization" based on simulation of individual thinking habits and cognitive states. In the evaluation phase, the system integrates a scoring assistant agent with a dual semantic consistency verification mechanism and a feedback generation agent. This not only provides robust and interpretable automatic grading for open-ended questions but also generates in-depth feedback pointing out specific thinking deviations based on inferred thinking paths. This effectively overcomes the problems of existing scoring models being susceptible to adversarial examples and having poor interpretability.

[0018] Ultimately, all the above-mentioned steps do not operate in isolation. Instead, through adversarial question generation and continuous interactive data collection, a reinforcement learning closed loop is formed, driving the collaborative optimization of the metacognitive simulator, knowledge graph, question generation, and scoring model. This allows the system to continuously self-evolve its overall performance using teaching feedback data, thus systematically solving the fundamental defects of static component fixation and difficulty in iterative optimization in existing technologies. In summary, this invention constructs a complete, closed-loop, integrated solution from high-quality structured content creation, deep personalized cognitive diagnosis, robust intelligent evaluation to system self-reinforcing evolution, significantly improving the accuracy of teaching intervention, the reliability of evaluation, and the adaptability and effectiveness of the system throughout its entire lifecycle.

[0019] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 The flowchart shows an interactive exercise automatic generation and intelligent evaluation method based on extended Markdown provided in Embodiment 1 of the present invention; Figure 2 This is a framework diagram of an interactive exercise automatic generation and intelligent evaluation system based on extended Markdown, provided in Embodiment 3 of the present invention. Figure 3 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] Application Overview: In existing technologies, intelligent teaching systems face several core problems. First, content and logic are separated: teaching materials, knowledge point relationships, exercises, and scoring rules are typically stored in files or databases of different formats, lacking a unified source file that can simultaneously carry both content and deep teaching logic. This results in insufficient structuring at the starting point of the automated processing chain. Second, personalization is superficial: exercise generation is often based on group statistics or coarse student models (such as knowledge mastery), failing to construct dynamic cognitive models that incorporate students' individual thinking habits (production rules) and metacognitive characteristics. Consequently, it is difficult to generate questions that can truly diagnose specific thinking errors. Third, intelligent assessment is unreliable: especially for open-ended questions, deep learning-based scoring models are easily deceived by adversarial samples. Their scoring criteria may deviate from basic subject axioms, and the decision-making process lacks interpretability, affecting trust in teaching. Fourth, the system is static and stagnant: components such as question banks, scoring models, and student models are often static or slowly updated, failing to form a closed loop that uses teaching feedback data to drive the collaborative evolution of all components.

[0024] This invention defines an Educational Extended Markdown (EMD) format, unifying teaching content, knowledge graphs, thinking templates, and scoring rules into a single document in the form of machine-readable proprietary tag blocks, providing a structured data source for all subsequent automated processing. By parsing this document, a multi-scale dynamic knowledge graph is constructed, in which innovative "abnormal logic association subgraphs" and "cognitive deviation weighted edges" achieve structured representation and quantitative association of erroneous thinking patterns. Based on this graph and the student's individual cognitive subgraph (dynamically including knowledge probabilities, production rules, and metacognitive strategies), a metacognitive simulator is used to deeply simulate student thinking, thereby generating highly targeted diagnostic and training exercises. In the evaluation phase, a dual semantic consistency verification mechanism and a benchmark deviation penalty term are introduced to constrain the scoring assistant agent to always anchor to the fundamental axioms of the subject, ensuring the robustness and logical consistency of the evaluation. Finally, through adversarial question generation and closed-loop data collection, the metacognitive simulator, knowledge graph, exercise generation, and scoring model are continuously and collaboratively optimized, achieving overall system self-evolution. This invention ultimately achieves deep-level intelligence across the entire process, from content creation to personalized teaching, and then to system self-enhancement, significantly improving teaching accuracy, assessment reliability, and system performance throughout its lifecycle.

[0025] Example 1: Figure 1 This is a flowchart illustrating an interactive exercise automatic generation and intelligent evaluation method based on extended Markdown, provided in Embodiment 1 of the present invention. This embodiment is applicable to situations involving the automatic generation of interactive exercises and intelligent evaluation based on extended Markdown. This method can be executed by an interactive exercise automatic generation and intelligent evaluation system based on extended Markdown. Figure 1 As shown, the method includes: S110. Define extended Markdown format by adding proprietary tag blocks on top of standard Markdown syntax.

[0026] S120. Parse the extended Markdown document, extract structured data, and construct a multi-scale knowledge graph, wherein the multi-scale knowledge graph includes: a positive cognitive subgraph and an abnormal logical association subgraph.

[0027] S130. Based on the multi-scale knowledge graph and the user's personal cognitive subgraph, generate personalized exercises.

[0028] S140. Based on the personalized exercises, the interactive exercise interface is parsed and rendered according to the unified interaction protocol.

[0029] S150. Obtain and distinguish the types of answers submitted by users, perform intelligent evaluation on different types of answers, obtain scoring results and evidence, and provide feedback on the scoring results and evidence.

[0030] In this embodiment, by extending the unified Markdown content source, a multi-scale knowledge graph containing logical connections between correct and incorrect statements is constructed, realizing the generation of accurate and personalized exercises based on deep cognitive simulation; and by using a dual semantic consistency verification mechanism to ensure the robustness and interpretability of intelligent evaluation, a complete intelligent teaching closed loop is formed that can use teaching feedback data to drive the collaborative optimization of various system components and continuous self-evolution.

[0031] Example 2: The technical solution in this example is a further refinement based on the above examples.

[0032] In step S110, an extended Markdown format is defined by adding proprietary tag blocks on top of the standard Markdown syntax; In this context, standard Markdown syntax refers to the general, fundamental specifications of a lightweight markup language. It defines common text formatting rules such as # for headings, - for lists, and **text** for bolding. Its core goal is to achieve simple separation of content and style, and rapid formatting. Proprietary tag blocks refer to structured syntax elements added to standard Markdown syntax to meet the needs of specific educational fields. They are not simple tags, but blocks used to encapsulate structured content with specific educational semantics, such as defining a knowledge point, an exercise, or a video clip. Their proprietary nature lies in carrying educational logic and metadata not found in standard Markdown. Extended Markdown format refers to the final defined new file format or specification. It consists of standard Markdown syntax (as the foundation and general expression layer) and proprietary tag blocks (as the core functionality and educational semantic layer). This extended Markdown format aims to be a unified content source file that can be automatically understood and processed by subsequent system modules (such as parsing engines and knowledge graph builders).

[0033] The proprietary tagging blocks include: knowledge graph definition block, video segment block, exercise card block, free-response question scoring rules block, interactive instruction block, metadata block, thought process template block, metacognitive prompting block, adversarial sample tagging block, baseline semantic tagging block, and error guidance tagging block.

[0034] Specifically, based on standard Markdown, it fully retains all standard Markdown syntax (such as headings, lists, code blocks, links, etc.), ensuring its basic text formatting and readability. On top of the standard syntax, a series of new block syntaxes with clearly defined start and end tags are designed. These blocks typically have specific identifiers, such as three backticks followed by the block type name or other custom symbols, and can contain structured attributes (usually in YAML, JSON, or other key-value pair formats) and content. Each proprietary block is given a clear educational semantic; each new block type corresponds to a specific teaching function or data unit, enabling educational researchers to not only write text but also define teaching components, logical relationships, and assessment rules in a structured way when writing documents. The proprietary tag blocks include: Knowledge graph definition blocks: These are used to define the knowledge points and concepts of a course, and to describe the hierarchical, subordinate, and related relationships between them. They are the foundation for building the concept layer of a knowledge graph.

[0035] Video segment block: Used to associate teaching video resources. It can specify the video file identifier, the playback timestamp range, and the explanatory text or interactive instructions corresponding to the video segment.

[0036] Exercise Card Block: Used to define a complete interactive exercise. It includes the question stem, options (for multiple-choice questions), interaction type (e.g., drag, sort, connect), cognitive level (e.g., memorization, comprehension, application), and related knowledge points.

[0037] The scoring rules section for open-ended questions defines the scoring criteria, which can describe in detail the scoring points, scores, keywords, logical points, etc., providing a clear basis for subsequent automatic scoring.

[0038] Interactive instruction block: Used to define the guidance behavior of the virtual human, the logic and response rules of multimodal interactions (such as voice and gesture triggers) to enhance immersion.

[0039] Metadata block: Used to define global information at the course or chapter level, such as course objectives, difficulty level, learning objectives, author, version, etc.

[0040] The thought process template block is used to define the ideal sequence of thought steps for solving specific types of problems, as well as common erroneous thought paths among students; it provides templates for simulating and diagnosing students' problem-solving processes.

[0041] Metacognitive prompting block: This is used to define a series of question templates to guide students in self-monitoring, self-evaluation, and self-regulation during the learning process, such as "Are you sure about this conclusion?" or "Are there other methods?"

[0042] Adversarial Sample Label Block: Used to identify certain questions or options that are adversarial samples specifically designed for training system robustness, i.e. seemingly reasonable but containing logical traps or highly deceptive distractors.

[0043] Baseline semantic label block: used to mark axioms, core definitions, and unquestionable logical premises; providing a solid semantic anchor for the entire knowledge system, and preventing model baseline deviation in subsequent scoring and adversarial training.

[0044] Error guidance label block: Used to identify specific types of cognitive biases or logical interferences that certain teaching content or question options are prone to cause, which helps the system to design and associate interference items in a targeted manner when generating exercises.

[0045] Once defined, the result is a complete and standardized Educational Extended Markdown (EMD) format specification. This specification includes: a technical standard document describing the language rules of EMD, serving as a guide for curriculum developers in writing teaching content; and a data exchange protocol, which is the core data interface and contract connecting human content creation and AI-automated processing. Curriculum developers can use any text editor to create .emd files containing actual course content, based on the Educational Extended Markdown format specification. These files contain all course information, from macro-level course objectives (metadata blocks) to specific knowledge points (knowledge graph definition blocks), video resources (video segment blocks), and each question and its underlying thought model, scoring criteria, and interactive logic (exercise card blocks, thought process template blocks, scoring rule blocks, etc.).

[0046] In this embodiment, by defining an extended Markdown format, the originally scattered teaching content, logical relationships, and evaluation rules are unified into a structured document that can be deeply understood by the machine. This provides a high-quality and standardized data source for subsequent knowledge graph construction, personalized exercise generation, and intelligent evaluation, fundamentally solving the bottleneck problem of the separation between teaching content and machine processing logic.

[0047] In step S120, the extended Markdown document is parsed, structured data is extracted, and a multi-scale knowledge graph is constructed, including: Parse the extended Markdown document to extract knowledge points, videos, exercises, and thought process templates; Construct a concept network based on the aforementioned knowledge points; Based on the exercises, videos, and thought process templates, they are associated with the knowledge points in the concept network, and a positive cognitive subgraph and an abnormal logic association subgraph are constructed. The positive cognitive subgraph and the abnormal logic association subgraph are connected by cognitive deviation weight edges.

[0048] Among them, knowledge points, videos, exercises, and thought process templates refer to the core teaching elements parsed from the extended Markdown document. Knowledge points can refer to core concepts, theorems, definitions, etc., in the course. Videos can refer to associated teaching video resources and their timestamp information. Exercises can refer to questions that define the question stem, options, interaction methods, correct answers, and associated knowledge points. Thought process templates define the sequence of steps for solving a certain type of problem and common incorrect reasoning paths. Conceptual networks can refer to a graph constructed based on knowledge points and their hierarchical and subordinate relationships, representing the macro-structure of the knowledge system. Positive cognitive subgraphs can refer to a graph structure representing correct knowledge logic and application, formed by associating teaching resources such as ideal problem-solving paths from exercises, videos, and thought process templates with knowledge points, under the conceptual network. Abnormal logic association subgraphs can refer to a graph structure parallel to the positive subgraph, composed of instantiations of adversarial samples, hidden fallacies, and common student error patterns, representing various erroneous thinking patterns. Cognitive deviation weighted edges refer to the edges connecting nodes in the positive cognitive subgraph and nodes in the abnormal logical association subgraph. Their weights are calculated through analysis of historical data and are used to quantify the interference intensity of a specific error pattern on a specific correct knowledge node.

[0049] Specifically, a dedicated EMD parsing engine is invoked, which incorporates complete "Extended Markdown Format" syntax rules. The engine scans the .emd file line by line, identifying all proprietary tag blocks marked with specific syntax (e.g., ::tagname::). When a start tag (e.g., ::benchmark-constraint::) is detected, the engine enters the corresponding block parsing state and continues reading subsequent content until it encounters the corresponding end tag (e.g., ::end-benchmark-constraint::). For each identified block, the engine executes specific data extraction logic based on its type, for example: From the baseline semantic constraint node label block, extract the text content of axioms or core logical premises and mark them as a structured object with a specific type (such as "axiom"). From the thought process template block, extract the step sequence of ideal problem-solving path and abnormal deduction path, with each step potentially serving as a sub-object. From the knowledge graph definition block, extract fields such as knowledge point name, ID, and parent node ID to form knowledge point objects. From the exercise card block, extract the question stem, options, correct answer, and a list of associated knowledge point IDs to form exercise objects. From the exercise card block, extract video resources and their timestamp information; from the interactive instruction block and the free-response question scoring rule block, extract interactive instructions and scoring rules.

[0050] All objects extracted from different blocks are organized into a unified, machine-readable data structure, such as a large JSON object or a structured message published to a message bus. This collection, containing various teaching elements and their relationships, is the extracted structured data, which will directly serve as input for building the knowledge graph.

[0051] The system initializes a graph database, such as Neo4j or Nebula Graph, to store nodes and edges. From the parsed structured data, it retrieves all objects representing "knowledge points." Using each knowledge point as a node, edges are created based on their hierarchical relationships (e.g., parent-child, prerequisites), forming a conceptual network representing the knowledge system architecture. Microstructures are formed by associating each exercise, video segment, and explanation (i.e., the content explained, which can be determined based on a thought process template) in the structured data with conceptual-level knowledge points. The system reads the "Associated Knowledge Point ID" attribute carried in the aforementioned teaching resource objects, creates a directed relationship edge in the graph, pointing from the microstructure resource node to the conceptual-level knowledge point node, and assigns a semantic label to this edge. For example, when writing the .emd document, the teaching and research staff have already defined the attribute for the question in the exercise card block: Associated Knowledge Point: ["Congruent Triangles", "SAS Theorem"]. When constructing the graph, the system performs the following operations: Exercise node Q001 - [Examination] - Conceptual layer node: Congruent triangles; Exercise node Q001 - [Examination] - Conceptual layer node: SAS theorem; Video node V005 - [Explanation] - Conceptual layer node: SAS theorem; Analysis node E010 - [Explanation] - Conceptual layer node: SAS theorem. At the same time, this analysis node can also be connected to exercise node Q001 through corresponding edges.

[0052] Constructing a positive cognitive subgraph: Within the microstructure layer, using the ideal problem-solving path defined in the thought process template as a blueprint, a series of nodes representing correct reasoning steps are created and connected by edges such as "derive," "next step," etc., forming one or more standard logical chains. This chain is linked to corresponding exercise and knowledge point nodes through edges. Constructing an abnormal logical association subgraph: Adversarial samples from the adversarial sample label block are instantiated as adversarial fallacy nodes as initial nodes. Hidden fallacy nodes are instantiated from the dynamic output of the adversarial generative network (GGN), a dynamically running AI component that continuously and proactively generates new erroneous logical samples (i.e., adversarial samples) designed to deceive the current scoring model during system operation. Each newly generated valid adversarial sample has its fallacy pattern extracted by the system and instantiated as a new node in the abnormal logical association subgraph—an automated, continuous, data-driven process. Repeated error patterns of students are extracted from the system interaction logs using data mining and pattern recognition algorithms and instantiated as common error nodes. For example, analysis reveals that many students mistakenly "reverse terms without changing the sign" when solving certain equations. The system abstracts this pattern and creates or reinforces a node named "Algebraic Reversal Sign Error" in the anomaly subgraph. These nodes collectively constitute the set of nodes that build the anomaly logical association subgraph.

[0053] The system analyzes which correct knowledge points / reasoning steps (located in the positive cognitive subgraph) are most likely to interfere with or confuse an abnormal node (error). For example, the erroneous node "misuse of SSA" is most likely to interfere with the correct node "application of SAS theorem".

[0054] Establishing Cognitive Deviation Weighted Edges: This involves analyzing massive amounts of historical student interaction data, performing statistical and machine learning calculations. The conditional probability or frequency of a student's answer or behavior exhibiting a specific error pattern (belonging to the abnormal logical association subgraph) when learning or applying a correct knowledge node or reasoning step (belonging to the positive cognitive subgraph) is calculated. This calculation result is quantified into weight values, creating an edge connecting the two nodes with a cognitive deviation weight attribute. A higher weight indicates a stronger and more likely association between the error and the correct knowledge. For example, a positive subgraph node: applying the SAS theorem - [cognitive deviation weight: 0.15] - an abnormal subgraph node: misusing the SSA condition; this weight is a dynamically updated key indicator. The final knowledge graph is a two-layer graph stored in a graph database, containing a "conceptual layer" and a "microstructure layer." Within the microstructure layer, there exist parallel positive cognitive subgraphs composed of "standard logical chains" and abnormal logical association subgraphs composed of "instances of various error patterns," connected by quantified "cognitive deviation weighted edges." It enables the structured representation of abnormal logical features and provides a data foundation for subsequent diagnostic exercise generation and intelligent evaluation.

[0055] In this embodiment, by parsing and expanding Markdown documents and constructing a multi-scale knowledge graph containing positive cognitive subgraphs and abnormal logical association subgraphs, a structured representation and quantitative association of knowledge points, teaching resources, and correct and incorrect thinking patterns are achieved. This provides a core data foundation for subsequent deep cognitive diagnosis, precise personalized exercise generation, and robust intelligent evaluation, thereby solving the problems of scattered teaching content and superficial machine understanding.

[0056] Optionally, when a new student registers or enters the system for the first time, the system initializes a dedicated view or data area for them on the global multi-scale dynamic knowledge graph, serving as the prototype of their personal cognitive subgraph. This personal cognitive subgraph includes the knowledge mastery probability, a production rule set, and metacognitive strategy tags. The knowledge mastery probability can be initialized to a default value (e.g., 0.5, indicating uncertainty), or an initial estimate can be obtained through a short pre-diagnostic test and attached to the relevant knowledge point nodes. The production rule set is initially empty. As the student's learning behavior unfolds, the system will gradually abstract and summarize the rules belonging to that student from their interaction logs. The metacognitive strategy tags are initially set to a default value or zero, indicating that a stable strategy usage pattern has not yet been observed.

[0057] The system automatically and in real-time collects fine-grained process data throughout all interactions between students and exercises, videos, and other content. This data can include: the order of steps submitted and modification history, recording each step of the problem-solving process and its timeline; the timing and content of requests for help; the duration spent on the explanation and instructional video pages; mouse hover and drag paths (used to assist in analyzing attention and decision-making processes). Using the collected interaction data, the system continuously updates three types of markers in the individual's cognitive subgraph through the following computational model: Updating the "probability of knowledge mastery" primarily employs Bayesian knowledge tracing or similar probabilistic models. Specifically, the system treats each student's attempt to answer a question (whether it's the final answer or the correctness of intermediate steps) as an "observation." Based on the knowledge point tested by the question (associated with concept layer nodes in the knowledge graph), the model dynamically updates the Bayesian estimate of the probability of mastery for that knowledge point. For example, if a student correctly answers a question testing the Pythagorean theorem, the system increases their "probability of mastery" for that knowledge point; if they answer incorrectly, it decreases. The magnitude of the adjustment depends on model parameters such as question difficulty, guessing rate, and error rate.

[0058] The "production rule set" can be summarized and updated using sequence pattern mining and rule induction algorithms. Specifically, the system analyzes students' continuous operation sequences to find recurring "condition-action" patterns. When a pattern appears multiple times in similar situations, the system abstracts it into a "production rule," adds it to the student's personal rule set, and records its trigger frequency and success rate. For example, the system repeatedly observes that when the problem contains the geometric condition "two sides and one angle are correspondingly equal," the student always directly chooses the conclusion "triangles are congruent" (without verifying whether the angle is an included angle). The system then summarizes a potential erroneous production rule: IF condition is "two sides and one angle are correspondingly equal" THEN action is "apply the congruence theorem," and associates this rule with the student, while recording its high trigger frequency and low success rate.

[0059] Inferring and updating "metacognitive strategy labels" can be achieved through behavioral pattern recognition and calculation. Specifically, the system defines a series of mapping relationships between observable behaviors and metacognitive strategies. By analyzing interaction data, it determines whether students exhibit corresponding strategies. For example, the planning strategy is inferred by analyzing whether students frequently review the question requirements or highlight key points before starting to solve a problem. The monitoring and checking strategy is updated by identifying behaviors such as "returning to modify or rereading a step before submitting an answer" to update the label value of the "verification of conclusions" strategy. The adjustment strategy is updated by analyzing whether "asking for help" occurs after a prolonged period of hesitation or immediately, to update the tendency of their "persistence" or "resource seeking" strategies.

[0060] All updated personalized data are directly attached as attributes or weights to the corresponding nodes and edges of the global "multi-scale dynamic knowledge graph". Thus, the same global graph has countless parallel, personalized views. For example, student A's "probability of mastery" for the "SAS Theorem" node is 0.9, and its production rule includes a correct application rule; while student B's "probability of mastery" for the same node is 0.4, and it is associated with an incorrect rule of "misusing SSA".

[0061] In this embodiment, constructing a dynamic personal cognitive subgraph essentially creates a continuously learning and evolving "cognitive digital twin" for each student. By collecting behavioral data and employing BKT probability updates, sequence mining inductive rules, and pattern recognition inference strategies, the system can approximate and represent students' deep knowledge states, thinking habits, and metacognitive characteristics in real time as a computational model, laying the foundation for subsequent precise simulation and personalized intervention.

[0062] In step S130, personalized exercises are generated based on the multi-scale knowledge graph, including: Upon receiving a question request, the system invokes the user's personal cognitive subgraph, which includes the probability of knowledge mastery, a set of production rules, and metacognitive strategy labels. Based on the aforementioned personal cognitive subgraph, a cognitive model is run to generate simulation results; Based on the simulation results, the problem-generating agent combines the multi-scale knowledge graph with the course objectives to generate corresponding diagnostic and training problems.

[0063] The question-generating request refers to the instruction signal that triggers the question generation process, typically including the target user (student) identifier and the scope of the questions (such as a specific knowledge point or course chapter). The user's personal cognitive subgraph is a dynamic, weighted view derived from the global "multi-scale knowledge graph," targeting a specific user. It is a formal representation of the student's real-time cognitive state, existing as a "cognitive digital twin."

[0064] Knowledge mastery probability can refer to a quantitative estimate of the degree of mastery of each knowledge node in the "personal cognitive subgraph" (such as the probability between 0 and 1), dynamically reflecting what the user "knows" and "how well they have mastered it".

[0065] Production rule sets refer to a set of "condition-action" pairs abstracted from a user's historical behavior. They are computational models describing the user's thinking habits, representing "how the user thinks and operates in different situations," and may contain erroneous rules. Metacognitive strategy tags refer to labels that record the user's tendency and frequency of using higher-order thinking strategies such as planning, monitoring, and regulation during learning and problem-solving, reflecting "how the user manages their thinking process."

[0066] A cognitive model can refer to the computational core within a metacognitive simulator. It's a hybrid architecture combining symbolic representation (such as production rules) and neural network computation to simulate human problem-solving and metacognitive processes. Simulation results can refer to the predictive data output by the cognitive model after it runs, typically including quantitative predictions of the probability distribution of user thought paths, rule calls, obstacles, and triggering specific error nodes in the "abnormal logic association subgraph." A problem-generating agent can refer to a dedicated software module whose responsibility is to automatically construct structurally complete problem entities that meet teaching requirements based on the input objectives and constraints. A multi-scale knowledge graph can refer to a two-layer graph containing a conceptual network layer and a microstructure layer. The microstructure layer contains, in parallel, a positive cognitive subgraph (correct logic) and an abnormal logic association subgraph (error patterns), connected by cognitive deviation weighted edges. Course objectives can refer to global learning objectives defined in the "metadata block" of an extended Markdown document, or specific teaching objectives specified in a problem-generating request (such as mastering a theorem). Diagnostic exercises can refer to exercises designed to proactively expose, elicit, and precisely pinpoint specific cognitive weaknesses or erroneous rules in users. Their design deliberately targets the most likely and most disruptive erroneous logical paths users are prone to fall into. Training exercises, on the other hand, can refer to exercises designed to strengthen and consolidate correct thinking paths or beneficial metacognitive strategies, helping users internalize correct patterns and transfer them to new situations.

[0067] Specifically, when the system needs to generate exercises for a user, the metacognitive simulator queries and loads the user's personal cognitive subgraph from the graph database using the user ID. This subgraph is a weighted view of the global knowledge graph. The simulator reads all dynamic states related to the target knowledge point from this subgraph, including: knowledge mastery probability, reading the user's mastery probability value for related concept nodes, such as "SAS theorem" and "congruent triangle criteria," for example, 0.9 represents 90% mastery; production rule set, reading the "condition-action" rules formed in the user's history, such as a potentially incomplete rule, "IF the goal is to prove triangle congruence AND see two sides and one angle corresponding to each other equal THEN apply the congruence theorem" (but does not correctly distinguish between SAS and SSA); and metacognitive strategy markers, reading the frequency or tendency of the user to use specific strategies when dealing with similar problems, such as "checking condition completeness" and "backtracking verification."

[0068] The simulator acquires a potential problem framework, such as a preliminary set of known conditions for a proof problem to be generated. Using the user's individual cognitive state (i.e., probability of knowledge mastery, production rule set, and metacognitive strategy labeling) as initial parameters, it internally runs its cognitive model (neuro-symbolic architecture). The symbolic part (production rules) is responsible for the logical deduction steps, while the neural part is responsible for calculating the probability of choosing different rules at each step, allocating cognitive resources, and whether to trigger metacognitive monitoring. The simulator performs multiple simulations (similar to Monte Carlo simulations) or uses a probabilistic graphical model for computation. In each simulation, due to the probabilistic nature of rule selection, a possible thought path is generated. All simulated thought paths are statistically analyzed, and the proportion of each path that ultimately touches (i.e. uses) a specific erroneous node in the "abnormal logical association subgraph" is calculated. For example, if 150 out of 1000 simulations result in the path incorrectly applying the "SSA" node, the predicted probability of triggering that node is 15%. This yields a "probability distribution" for each anomalous node.

[0069] The problem-generating agent receives predictions from a metacognitive simulator. It identifies the anomalous reasoning paths with the highest cognitive deviation weights. For example, the prediction shows that the probability of triggering the "misuse of the SSA theorem" node is the highest, and the "cognitive deviation weight" edge weight between this node and the correct "application of the SAS theorem" node is also the highest. Diagnostic and training problems are generated separately.

[0070] For diagnostic problem generation: the AI ​​deliberately constructs a scenario to make the erroneous rule of "misusing SSA" seem "reasonable." For example, it generates a geometry problem where the figure is deliberately arranged such that two sides and the opposite angle of one of those sides are correspondingly equal (i.e., the SSA condition), but the two triangles are not actually congruent. The purpose of this problem is to test whether the user will fall into this pre-set, high-probability trap tailored to them.

[0071] Training exercise generation: If the goal is to consolidate the correct path, the intelligent system generates an exercise that requires clear and repeated use of the correct rule (such as SAS), and may trigger relevant "metacognitive prompts" after solving the exercise, such as "Please explain why SSA cannot be used".

[0072] In this embodiment, by deeply simulating individual students' thinking habits and weaknesses, personalization is achieved from superficial knowledge mastery to deep cognitive state. It can accurately predict and generate "diagnostic exercises" targeting specific thinking traps students are most likely to fall into, while simultaneously generating "training exercises" to reinforce correct approaches. This achieves precise teaching intervention based on cognitive models rather than question templates.

[0073] Optionally, after generating personalized exercises based on the multi-scale knowledge graph, the method further includes: The personalized exercises are validated in multiple layers through a validation pipeline, including: knowledge accuracy validation, cognitive level matching validation, and format conformity validation.

[0074] The verification pipeline can refer to an automated, pipeline-style, multi-level quality review system. It performs a series of rule and model checks on the draft "personalized exercises" output by the exercise-generating agent to ensure their quality, security, and pedagogical effectiveness. Only questions that "pass" can be officially used. Multi-level verification refers to multiple independent checking stages executed sequentially within the verification pipeline, each focusing on different quality dimensions, filtering at each level. Knowledge accuracy verification ensures that the content of the questions is absolutely correct in terms of subject matter facts and consistent with the "positive cognitive subgraph" and "baseline semantic constraint nodes" defined in the "multi-scale knowledge graph," with no factual errors. Cognitive level matching verification assesses whether the cognitive complexity of the questions matches the preset teaching objectives (such as memorization, understanding, application, and analysis), preventing questions from being too simple or beyond the curriculum. Format compliance verification checks whether the data structure, interactive instructions, and rendering tags of the questions conform to the parsing and rendering specifications of the system front-end, ensuring that the questions can be displayed and interacted with correctly. The question bank refers to the structured database used by the system to store all verified and usable exercises. Updating the knowledge graph refers to adding newly generated, qualified exercises as new "microstructure layer" nodes to the "multi-scale dynamic knowledge graph" and establishing connections with related knowledge points, thinking templates, etc., thereby enriching the system's own knowledge base.

[0075] Specifically, after the exercise generation agent produces a draft exercise, it encapsulates it into a structured data object and sends it to the verification pipeline. The verification pipeline performs multi-layered verification on the personalized exercises, including: knowledge accuracy verification, checking the consistency between the question content and the knowledge graph; cognitive level matching verification, verifying the conformity between the question difficulty and the target cognitive level; format standardization verification, checking the standardization of the question format; and storing the verified questions in the question bank and updating the knowledge graph.

[0076] The first layer: Knowledge accuracy verification. The system calls the "Knowledge Graph Query Interface" or "Rule Engine" to compare all propositions and conditions in the question with the concept layer and positive cognitive subgraph in the multi-scale knowledge graph. For example, for a geometry proof question, the system will verify whether its "conclusion to be proved" can be derived from its "known conditions" through the correct logical chain existing in the graph; for multiple-choice questions, it will check whether the error corresponding to each distractor is accurately associated with a specific node in the "abnormal logical association subgraph". Any content that contradicts authoritative knowledge in the graph or cannot be derived will result in verification failure.

[0077] The second layer: Cognitive level matching verification. The system uses a cognitive classification model (e.g., a model trained based on Bloom's Taxonomy of Educational Objectives) or rules. It analyzes the question text and the key thought processes required to solve it to determine the cognitive level it belongs to, such as memorizing facts, understanding concepts, applying rules, or analyzing principles. This determination is then compared with the "target cognitive level" specified in the question request or the "cognitive level" label in the question metadata. If there is a significant mismatch, such as a "memorization" question being generated when the target is "application," the verification fails.

[0078] The third layer: Format compliance validation. The system calls the format validator or the rendering engine's pre-parsing module. It checks the integrity of the question data structure, such as whether the question stem, options, and answer fields are missing. It verifies that the defined interaction types (drag and drop, click, input) conform to the front-end protocol and ensures the closure and validity of any extended Markdown or HTML tags. Formatting errors will prevent the question from displaying or interacting correctly on the front end, so this layer must intercept them.

[0079] If a question successfully passes all the above verification layers, the verification pipeline will tag it with "verified" and generate a unique question ID. The system will then store it in the question bank. The question bank serves as a central repository, supporting efficient retrieval by attributes such as knowledge point, difficulty, and type, for use in subsequent question creation and test paper assembly.

[0080] If any verification layer fails, the draft question will be rejected. The system will record the specific reason for the failure, such as "knowledge contradiction: conclusion C cannot be derived from conditions A and B", and may feed the error back to the "question generation agent" to adjust and optimize its subsequent generation strategy.

[0081] As a question is added to the question bank, the system triggers a knowledge graph update transaction. At the microstructure layer of the "multi-scale knowledge graph," the system creates a new "question node." Then, based on the question's attributes, it creates the following association edges: from the question node to the concept-level knowledge point node it examines, for example, -[Examined]-"Pythagorean Theorem." From the question node, it associates with the thought process template block it uses (if applicable). It associates distractors in the question with corresponding fallacies in the "abnormal logic association subgraph," for example, -[Corresponding Distractor]-"Misuse of SSA Node."

[0082] In this embodiment, through pipeline verification and knowledge graph updates, the newly generated knowledge resources of the system are preserved and become part of the global knowledge graph. When generating exercises or conducting assessments for other students in the future, this new question and its related relationships can be utilized, realizing the system's self-enrichment and continuous evolution.

[0083] In step S140, based on the personalized exercises, the interactive exercise interface is parsed and rendered according to the unified interaction protocol.

[0084] The unified interaction protocol refers to a predefined, standardized data exchange specification. It declaratively describes all elements of a question (such as the question stem, options, interaction type, layout, and associated media resources) and their behavioral logic. This protocol is typically based on JSON, XML, or a domain-specific language (DSL) and is the core contract connecting the system backend (generating questions) and the frontend (presenting the interface). The interactive question interface refers to the user-defined, directly operable interface presented to the user. It is not a static text and image page, but a complex interface dynamically composed of various interactive components (such as multiple-choice options, drag-and-drop sorting areas, and virtual experiment whiteboards), allowing users to answer questions through clicking, dragging, and input.

[0085] Specifically, the system backend, such as the question bank or exercise generation module, sends a structured data packet containing an exercise to the front-end runtime engine. This data packet follows a predefined "unified interaction protocol," which is essentially a declarative specification describing all elements of the exercise (question stem, options, interaction types, layout, associated media resources, etc.) and their behavioral logic. It can be based on JSON, XML, or a specific DSL. The core of the front-end runtime engine is a parser. It reads and parses this protocol data, identifying the component types that need to be rendered, such as "multiple-choice questions," "drag-and-drop sorting areas," and "Geometer's Sketchpad," along with their attributes. Based on the identified component types, the parser dynamically calls pre-built components from the engine's built-in interactive component library. For example, upon encountering "type": "multiple-choice," it instantiates a multiple-choice renderer and fills in the question stem and option data; upon encountering "type": "drag-and-drop," it initializes a draggable interactive area. Finally, these components are combined into a complete, interactive exercise interface presented to the user.

[0086] The unified front-end runtime engine includes a rich and extensible library of interactive components. Each interaction method (multiple-choice, drag-and-drop, sorting, matching, virtual experiment) corresponds to one or more pre-packaged UI components. These components not only handle visual presentation but also encapsulate complete interaction logic. For example, Drag-and-drop component: Manages the verification logic for drag-and-drop start, hover, and placement target areas. Virtual experiment component: Can integrate a lightweight physics engine or chemical model, responding to user operations (such as connecting circuits, mixing reagents) and providing simulation results. When parsing the "Unified Interaction Protocol," the engine only needs to select and initialize the corresponding component based on the "interactionType" field in the protocol, thereby achieving seamless support for multiple interaction methods.

[0087] Each of the aforementioned interactive components automatically incorporates a fine-grained data acquisition listener at runtime. This listener captures all low-level events generated by the user interacting with the component. The captured raw events (such as mouse click coordinates and keyboard presses) are immediately converted into standardized log events with educational semantics. For example: Event type: Step Submission, Content: SAS Theorem Applied, Timestamp: t1. Event type: Modification History, Content: Option A Changed to Option B, Timestamp: t2. Event type: Help Request, Content: "Hint 1" Clicked, Timestamp: t3. Event type: Dwell Time, Page: Parsing Page, Duration: 120 seconds. These structured log events are sent in real-time or near real-time to the backend data processing and analysis module via a unified data bus, such as the Apache Kafka topic user-interaction-events.

[0088] The metacognitive simulator, as a continuously running background service, subscribes to and consumes user-interaction-events data streams from the front end. Internally, the simulator maintains the student's dynamic cognitive model (i.e., the state of their "personal cognitive subgraph"). Each new procedural event received is treated as an "observation" and input into a Bayesian estimation model, such as a dynamic Bayesian network or an improved knowledge-tracking model. Based on new evidence, such as the student quickly skipping key steps or repeatedly hesitating on an option, the model recalculates and updates its estimate of the student's hidden cognitive state. This includes: the most probable thought process, inferring which production rule chain the student actually used; cognitive load, estimating their current cognitive stress level based on indicators such as operating speed, error rate, and frequency of seeking help; and real-time fine-tuning of the probability of knowledge mastery.

[0089] Personalized intervention is triggered when an anomaly is detected. The metacognitive simulator has a series of preset or learned triggering rules. These rules continuously compare the real-time updated Bayesian estimation results with preset thresholds. For example, if the estimated cognitive load is greater than the threshold and there is no progress for a long time, then it is judged as a "mental deadlock".

[0090] If the production rule used in the IF prediction ∈ the error rule set THEN, it is determined as "using an error rule". If the IF detects that a student skipped all check steps THEN, it is determined as "metacognitive monitoring failure". Once a rule is triggered, the simulator immediately generates an intervention instruction. This instruction includes the intervention type (virtual human prompt / interface pop-up) and specific prompt content. The prompt content can be directly selected from the "metacognitive prompt block" associated with the exercise or dynamically generated by the simulator. This intervention instruction is pushed back to the unified front-end runtime engine in real time through a low-latency message channel, such as WebSocket. After the engine parses the instruction, if the instruction requires virtual human intervention, it activates the virtual human agent component, plays the corresponding voice and animation, and displays the dialogue content. If the instruction requires an interface prompt, it renders a prompt box at a specific location on the exercise interface, displaying a guiding question. The user's subsequent reactions to the intervention (such as viewing the prompt, answering the virtual human's questions) are collected as new "process data", thus forming a real-time teaching closed loop of "monitoring-analysis-intervention-feedback".

[0091] In this embodiment, the interface is enriched by dynamic rendering through protocol parsing, and the learning process is collected through deep data collection to provide fuel for the cognitive simulation in the backend. Ultimately, it becomes the terminal for executing personalized intervention, transforming the system's intelligent judgment into real-time teaching actions.

[0092] In step S150, the types of answers submitted by users are acquired and distinguished, and intelligent evaluation is performed on different types of answers to obtain scoring results and evidence. The scoring results and evidence are then fed back, including: The system acquires and differentiates the types of user-submitted answers, including answers to objective questions and answers to open-ended questions. For the answers to the objective questions, they are matched with the pre-stored standard answers to determine the judgment result; For the answers to the free-response questions, the scoring assistant agent, which has been trained in adversarial robustness, automatically grades the answers according to the predefined scoring rules and outputs the scoring results and evidence. The scoring logic of the scoring assistant agent integrates a dual semantic consistency verification mechanism. The feedback generating agent combines the rating results, the evidence, and the user's thought process inferred by the metacognitive simulator to generate and output deep feedback. The deep feedback is used to explain the reasons for errors, point out thinking biases, and provide adjustment strategies.

[0093] In this context, user-submitted answers refer to the content submitted by students after completing the exercises through the interactive front-end interface. Objective question answers refer to submissions for question types with clear, unique standard answers, such as multiple-choice and true / false questions. The core of its evaluation is symbol matching. Open-ended question answers refer to submissions for open-ended questions such as essays, proofs, and other essays, where students use text, formulas, etc., to explain and demonstrate their points. The core of its evaluation is semantic and logical analysis of the content. Pre-stored standard answers refer to the unique correct option or result pre-stored in the question bank for each objective question. The scoring assistant AI agent refers to a software module (AI model) specifically designed for grading open-ended questions. Its characteristic is that it undergoes adversarial robust training to enhance its ability to resist interference from confusing answers (adversarial samples). Predefined scoring rules refer to the detailed grading criteria pre-defined in the scoring rules block for open-ended questions in the extended Markdown document, including scoring points, scores, keywords, and logical points.

[0094] The dual semantic consistency verification mechanism refers to the core guarantee mechanism integrated into the scoring logic of the scoring assistant agent, including: anomaly logic matching verification, which matches the student's answer with the fallacious nodes in the anomaly logic association subgraph; if a match is found, points are deducted. Benchmark semantic backtracking verification requires that the agent's scoring reasoning must be traceable to benchmark semantic constraint nodes (such as subject axioms) to prevent "benchmark deviation" in the scoring logic. Scoring results and evidence refer to the correction conclusions output by the scoring assistant agent, including not only the score but also the evidence on which the score was derived, such as sentences in the answer that meet the scoring criteria and the specific erroneous logic triggered. The feedback generation agent can refer to a dedicated software module responsible for integrating scoring results, evidence, and the user's thought process to generate explanatory text for students. The user's thought process inferred by the metacognitive simulator refers to the sequence of production rules actually used by the user and the possible cognitive obstacles encountered when solving the problem, inferred by the metacognitive simulator based on the user's interaction data during the answering process (such as step sequence and modification traces). Deep feedback can refer to personalized guidance content generated by a feedback-generating agent. Its purpose is not only to tell you whether something is right or wrong, but also to explain the reasons for the error, point out specific thinking biases, and provide adjustment strategies for subsequent learning.

[0095] Specifically, questions are categorized into objective questions and open-ended questions based on their format. Objective questions are those with a clear, single correct answer, such as multiple-choice and true / false questions, where the core assessment is the matching of symbols in the answer. Open-ended questions require students to explain, demonstrate, or solve problems using text, formulas, or code, such as essay questions, proof questions, and compositions, where the core assessment is the semantic and logical analysis of the content.

[0096] For objective questions, students answer through interactive methods on the front-end interface (clicking options, dragging target positions). The unified front-end runtime engine captures the interaction results in real time and encapsulates them into structured data (e.g., {"question_id":"Q101", "answer": "A"}). After receiving the answer data, the system retrieves the standard answer for that question from the question bank and performs a simple string or value match. A successful match indicates a correct answer; otherwise, it indicates an incorrect answer. The judgment result (correct / incorrect), the student's answer, the time taken, and other data are written to the student's learning history database, which may trigger an update to the student's "mastery probability" of related knowledge, i.e., an update to the individual cognitive subgraph.

[0097] For open-ended questions, students complete their answers in a text input box or rich text editor and click submit. The front-end engine captures their answers (plain text or formatted text) along with information such as the question ID. The scoring assistant agent automatically grades the answers according to the scoring rules. The agent takes the question stem, scoring rules, and student answers as input, runs its internal model, and generates a preliminary score and comments. A dual semantic consistency verification mechanism is forcibly integrated into its scoring logic. This mechanism includes: anomaly logic matching verification. Simultaneously or after the agent's scoring, the system parses the student's answer text and attempts to match it with nodes and structures in the anomaly logic association subgraph. For example, if the subgraph defines a fallacy node of "inverted causality" and its common expression patterns, the system checks if such a pattern exists in the answer. If a match is successful, a deduction or alert is triggered, regardless of how good the answer appears on the surface.

[0098] Benchmark semantic backtracking verification is a constraint on the reasoning process of the scoring assistant agent. It requires that when the agent arrives at a scoring conclusion, its internal reasoning chain must consist of a traceable and logical set of steps, and these steps must ultimately be based on "benchmark semantic constraint nodes" (such as geometric axioms or physical laws) relevant to the question. The system can verify this through techniques such as attention mechanism analysis and reasoning chain extraction. If the agent's score cannot be backtracked to a benchmark node or relies on a false implicit premise, the scoring result will be considered unreliable.

[0099] During the training phase, the scoring assistant agent plays a game against the adversarial generative network. Its loss function is equal to the traditional scoring error loss plus a baseline bias penalty term. When the adversarial example induces the model to use a "skewed" logic (such as "the longer the answer, the better the answer") for scoring, the penalty term will generate a high loss as long as this logic cannot be traced back to the baseline node. This forces the model to abandon this "skewed" standard, thereby maintaining the consistency between the scoring baseline and the basic logic of the subject from the root.

[0100] In addition to providing a score, the scoring assistant agent must also output the evidence upon which its judgment is based, such as key sentences and logical chains. For example, when outputting the score, it highlights sentences in the answer that match the scoring criteria and sentences that trigger abnormal logical matches. For instance, a scoring point might be: the answer correctly mentions "conservation of energy" (sentence 2). A deduction might be: the answer contains the incorrect concept that "force is the cause of motion" (sentence 4), which falls under the "Aristotelian fallacy of motion" node. Consistency checks ensure the robustness of its internal reasoning; for example, whether the agent cites a correct principle as evidence for a higher score while ignoring statements in the answer that violate that principle. Consistency checks can be implemented using a logic rule engine or by training another consistency-based model.

[0101] After receiving the scoring results and evidence, the feedback-generating agent is activated. The metacognitive simulator infers the student's thought process based on the data from this question-and-answer process, especially which error-generating rule was used. The agent identifies specific abnormal logical paths, i.e., the corresponding fallacies. It then synthesizes this information to generate a guiding text. For example, the answer might start correctly, but in the third step, you directly applied formula A, ignoring its precondition B—a common mistake of "misusing formula A." Please check if case Z in the question satisfies precondition B. Next time, you can think like this: first list all known conditions, then check the preconditions of the formula or theorem you want to use one by one.

[0102] In this embodiment, by distinguishing between objective questions and open-ended questions, and employing an agent trained with adversarial robustness and integrated with dual semantic consistency verification for grading, the accuracy, robustness, and logical consistency of automatic scoring of open-ended answers are fundamentally improved. Simultaneously, by generating in-depth feedback based on inferred student thought processes, intelligent teaching guidance is achieved, moving beyond simple right / wrong judgments to precise explanations of the causes of errors, pointing out thought biases, and providing adjustment strategies.

[0103] Optionally, after acquiring and distinguishing the types of user-submitted answers, conducting intelligent evaluations of different types of answers, obtaining scoring results and evidence, and providing in-depth feedback, the process also includes: Based on the scoring results and evidence, generate adversarial questions; Continuously collect interaction data from all users with various types of exercises, including the aforementioned adversarial questions; Based on the interactive data, component optimization and map update operations are performed.

[0104] Among them, adversarial questions refer to challenging exercises automatically generated by the system for a specific purpose. These are mainly divided into two categories: robustness test questions, designed to test and improve the robustness of the system itself (such as scoring models and metacognitive simulators), which are usually logically complex or deceptive; and mental stress test questions, designed to target students' high-frequency or persistent mental weaknesses (corresponding to high-weight nodes in the abnormal logical association subgraph), serving as precise "cognitive bait." Interaction data refers to fine-grained, procedural behavioral data generated by users during the question-answering process, including but not limited to the order of step submissions and modification traces, the timing of requests for help, the duration of time spent on the parsing page, and mouse operation paths. This data is the fuel driving system optimization. Component optimization and graph update operations refer to specific tasks that iteratively improve the core components of the system using interaction data, mainly including: component optimization, optimizing the prediction accuracy of the metacognitive simulator, optimizing the question generation strategy of the question-generating agent, and optimizing the grading robustness of the scoring assistant agent. Graph update operations can refer to updating multi-scale knowledge graphs, such as adjusting the "estimated difficulty" of knowledge point nodes, recalculating the weights of cognitive deviation weight edges, and adding, deleting, or associating fallacies in abnormal logic association subgraphs.

[0105] Specifically, the adversarial question-generating agent is activated periodically or when triggered. Its task is to determine whether to generate "robust test questions" or "student mental stress test questions" based on the current system state.

[0106] Generate robustness test questions: Strategy: The agent analyzes historical cases of the scoring model to identify question types with low model confidence or those successfully "deceived" by adversarial examples. Simultaneously, it analyzes the prediction logs of the metacognitive simulator to identify scenarios where predictions significantly deviate from actual student behavior. Targeting these weaknesses, the agent utilizes its generative network to construct new and more extreme questions. For example, for the scoring model, it generates an extremely complex and obscure argument with a flawed core premise, or an article filled with technical jargon but lacking substance, to test whether the model can avoid being misled by form and grasp the essence. For the metacognitive simulator, it designs a question with multiple reasonable paths to solution that rely on subtle contextual understanding, to test whether the simulator can accurately predict the probability of students choosing different paths.

[0107] Generate "stress test" questions for students' thinking: Strategy: The agent scans the "abnormal logical association subgraph", especially the abnormal nodes with high "cognitive deviation weight" but lacking targeted training recently (i.e., mistakes that students are prone to make but have not been tested recently).

[0108] The agent targets high-weight anomalous nodes and constructs solutions within their topological structure. For example, if the goal is "misuse of the SSA theorem," a geometric problem is generated whose graphical conditions are extremely close to SAS, but which actually satisfies SSA and whose conclusion is invalid. This requires precise control of the problem conditions to make them the "perfect bait" for erroneous rules. The generated draft "stress test" problem must be submitted to a "dual semantic consistency verification mechanism" for review to ensure that the problem itself does not contain undefined or new factual errors. The correct solution logic of the problem can be clearly traced back to the benchmark axioms. Only problems that pass the review can be used in actual teaching to prevent the generation of invalid or harmful problems.

[0109] The system establishes an automated data pipeline that continuously collects and processes all interactive data, driving the evolution of its four main components: (1) Data collection closed loop: The system collects all users' (including users who use adversarial questions) fine-grained interaction logs, answer results, and metadata of the questions themselves into a unified data lake or stream processing platform.

[0110] (2) Optimize the metacognitive simulator by comparing the collected real student thought process data (such as step sequences and hesitation points) with the simulator's predictions at that time. By using reinforcement learning or supervised learning, adjust the parameters of the simulator's neural network or modify the trigger probability of its production rules to narrow the gap between prediction and reality, making the model closer to the real cognitive process.

[0111] (3) Update the knowledge graph, including: The difficulty and relevance are updated by dynamically adjusting the "estimated difficulty" attribute of knowledge point nodes in the knowledge graph based on a large amount of student answer data (accuracy rate, time taken). At the same time, it analyzes which knowledge points are frequently answered incorrectly or correctly by students simultaneously and updates the edge weights of the "relevance strength" between them.

[0112] The system updates the anomaly logic association subgraph based on the "stress test" questions and the answers to all questions. It recalculates the cognitive deviation weights, meaning it recalculates the weights of edges connecting the correct and incorrect subgraphs based on the latest, larger-scale data, making it more accurate in reflecting the common errors of the current student population. The system also updates the topological relationships; if certain error patterns are found to frequently occur together, new edges may be created in the anomaly subgraph to represent this association; if a historical error node has not been triggered for a long time, its weight may be reduced.

[0113] (4) Improve the problem generation and scoring model. Improve the problem generation model by using "stress test" questions and their effect data (whether they successfully induced the target error) as training samples and feeding them back to the problem generation agent so that it can learn to construct more diagnostically effective questions. Improve the scoring model by using "robust test questions" and students' answers to them as new adversarial samples and adding them to the "adversarial robust training pipeline" to conduct a new round of training for the scoring assistant agent, repair its blind spots, and improve its robustness and accuracy.

[0114] In this embodiment, a reinforcement learning closed loop is constructed to drive the system's self-evolution and continuous optimization. Targeted adversarial questions are generated using scoring evidence to test and improve the system's robustness. Furthermore, by collecting global user interaction data, core components such as the metacognitive simulator, question generation, and scoring model are dynamically optimized. Simultaneously, the weights and relationships of the knowledge graph are updated, enabling the entire teaching system to achieve iterative performance improvements based on real feedback data.

[0115] Example 3: Figure 2 This is a framework diagram of an interactive exercise automatic generation and intelligent evaluation system based on extended Markdown, provided in Embodiment 3 of the present invention. The system is used to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown described in any embodiment of the present invention. Figure 2 As shown, the system includes: Define module 210, which is used to define extended Markdown format by adding proprietary tag blocks on top of standard Markdown syntax; The parsing module 220 is used to parse the extended Markdown document, extract structured data, and construct a multi-scale knowledge graph, wherein the multi-scale knowledge graph includes: a positive cognitive subgraph and an abnormal logical association subgraph; The generation module 230 is used to generate personalized exercises based on the multi-scale knowledge graph and the user's personal cognitive subgraph. Rendering module 240 is used to parse and render the interactive exercise interface based on the personalized exercises and according to the unified interaction protocol; The evaluation module 250 is used to acquire and distinguish the types of answers submitted by users, perform intelligent evaluation on different types of answers, obtain scoring results and evidence, and provide feedback on the scoring results and evidence.

[0116] Optional, parsing module 220, specifically used for: Parse the extended Markdown document to extract knowledge points, videos, exercises, and thought process templates; Construct a concept network based on the aforementioned knowledge points; Based on the exercises, videos, and thought process templates, they are associated with the knowledge points in the concept network, and a positive cognitive subgraph and an abnormal logic association subgraph are constructed. The positive cognitive subgraph and the abnormal logic association subgraph are connected by cognitive deviation weight edges.

[0117] Optionally, module 230 is generated, specifically for: Upon receiving a question request, the system invokes the user's personal cognitive subgraph, which includes the probability of knowledge mastery, a set of production rules, and metacognitive strategy labels. Based on the aforementioned personal cognitive subgraph, a cognitive model is run to generate simulation results; Based on the simulation results, the problem-generating agent combines the multi-scale knowledge graph with the course objectives to generate corresponding diagnostic and training problems.

[0118] Optionally, module 230 is also used for: The personalized exercises are validated in multiple layers through a validation pipeline, including: knowledge accuracy validation, cognitive level matching validation, and format conformity validation.

[0119] Optional, rendering module 240, specifically used for: The system acquires and differentiates the types of user-submitted answers, including answers to objective questions and answers to open-ended questions. For the answers to the objective questions, they are matched with the pre-stored standard answers to determine the judgment result; For the answers to the free-response questions, the scoring assistant agent, which has been trained in adversarial robustness, automatically grades the answers according to the predefined scoring rules and outputs the scoring results and evidence. The scoring logic of the scoring assistant agent integrates a dual semantic consistency verification mechanism. The feedback generating agent combines the rating results, the evidence, and the user's thought process inferred by the metacognitive simulator to generate and output deep feedback. The deep feedback is used to explain the reasons for errors, point out thinking biases, and provide adjustment strategies.

[0120] Optional, evaluation module 250, specifically used for: Based on the scoring results and evidence, generate adversarial questions; Continuously collect interaction data from all users with various types of exercises, including the aforementioned adversarial questions; Based on the interactive data, component optimization and map update operations are performed.

[0121] Optionally, the proprietary tag blocks include: knowledge graph definition blocks, video segment blocks, exercise card blocks, free-response question scoring rule blocks, interactive instruction blocks, metadata blocks, thought process template blocks, metacognitive prompt blocks, adversarial sample tag blocks, baseline semantic tag blocks, and bias guidance tag blocks.

[0122] The interactive exercise automatic generation and intelligent evaluation system based on extended Markdown provided in this invention can execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown provided in any embodiment of this invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0123] Example 4: Figure 3 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0124] like Figure 3 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) or a random access memory (RAM), which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0125] Processor 11 can perform various appropriate actions and processes based on computer programs stored in read-only memory (ROM) or loaded into random access memory (RAM) from storage unit 18. RAM 13 can also store various programs and data required for the operation of electronic device 10. Processor 11, ROM 12, and RAM 13 are interconnected via bus 14. Input / output (I / O) interfaces are also connected to bus 14.

[0126] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0127] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as interactive exercise automatic generation and intelligent evaluation methods based on extended Markdown.

[0128] In some embodiments, the interactive exercise automatic generation and intelligent evaluation method based on Extended Markdown can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the interactive exercise automatic generation and intelligent evaluation method based on Extended Markdown described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the interactive exercise automatic generation and intelligent evaluation method based on Extended Markdown by any other suitable means (e.g., by means of firmware).

[0129] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a memory system, at least one input device, and at least one output device, and transferring data and instructions to the memory system, the at least one input device, and the at least one output device.

[0130] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0131] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown provided by this invention. The computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, the computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0132] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD monitor)); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0133] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0134] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0135] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0136] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for automatic generation and intelligent evaluation of interactive exercises based on extended Markdown, characterized in that, include: Extended Markdown formatting is defined by adding proprietary tag blocks to the standard Markdown syntax; The extended Markdown document is parsed to extract structured data and construct a multi-scale knowledge graph, wherein the multi-scale knowledge graph includes: a positive cognition subgraph and an abnormal logic association subgraph; Personalized exercises are generated based on the multi-scale knowledge graph and the user's personal cognitive subgraph. Based on the personalized exercises, the interactive exercise interface is parsed and rendered according to the unified interaction protocol; The system acquires and distinguishes the types of answers submitted by users, performs intelligent evaluation on different types of answers, obtains scoring results and evidence, and provides feedback on the scoring results and evidence.

2. The method according to claim 1, characterized in that, The process of parsing the extended Markdown document, extracting structured data, and constructing a multi-scale knowledge graph includes: Parse the extended Markdown document to extract knowledge points, videos, exercises, and thought process templates; Construct a concept network based on the aforementioned knowledge points; Based on the exercises, videos, and thought process templates, they are associated with the knowledge points in the concept network, and a positive cognitive subgraph and an abnormal logic association subgraph are constructed. The positive cognitive subgraph and the abnormal logic association subgraph are connected by cognitive deviation weight edges.

3. The method according to claim 1, characterized in that, Based on the aforementioned multi-scale knowledge graph and the user's personal cognitive subgraph, personalized exercises are generated, including: Upon receiving a question request, the system invokes the user's personal cognitive subgraph, which includes the probability of knowledge mastery, a set of production rules, and metacognitive strategy labels. Based on the aforementioned personal cognitive subgraph, a cognitive model is run to generate simulation results; Based on the simulation results, the problem-generating agent combines the multi-scale knowledge graph with the course objectives to generate corresponding diagnostic and training problems.

4. The method according to claim 3, characterized in that, After generating personalized exercises based on the multi-scale knowledge graph, the process also includes: The personalized exercises are validated in multiple layers through a validation pipeline, including: knowledge accuracy validation, cognitive level matching validation, and format conformity validation.

5. The method according to claim 1, characterized in that, The system acquires and differentiates the types of user-submitted answers, performs intelligent evaluation on different types of answers, obtains scoring results and evidence, and provides feedback on the scoring results and evidence, including: The system acquires and differentiates the types of user-submitted answers, including answers to objective questions and answers to open-ended questions. For the answers to the objective questions, they are matched with the pre-stored standard answers to determine the judgment result; For the answers to the free-response questions, the scoring assistant agent, which has been trained in adversarial robustness, automatically grades the answers according to the predefined scoring rules and outputs the scoring results and evidence. The scoring logic of the scoring assistant agent integrates a dual semantic consistency verification mechanism. The feedback generating agent combines the rating results, the evidence, and the user's thought process inferred by the metacognitive simulator to generate and output deep feedback. The deep feedback is used to explain the reasons for errors, point out thinking biases, and provide adjustment strategies.

6. The method according to claim 1, characterized in that, After acquiring and distinguishing the types of user-submitted answers, performing intelligent evaluation on different types of answers, obtaining scoring results and evidence, and providing in-depth feedback, the process also includes: Based on the scoring results and evidence, generate adversarial questions; Continuously collect interaction data from all users with various types of exercises, including the aforementioned adversarial questions; Based on the interactive data, component optimization and map update operations are performed.

7. The method according to claim 1, characterized in that, The proprietary tag blocks include: knowledge graph definition block, video segment block, exercise card block, free-response question scoring rule block, interactive instruction block, metadata block, thought process template block, metacognitive prompt block, adversarial sample tag block, baseline semantic tag block, and bias guidance tag block.

8. An interactive exercise automatic generation and intelligent evaluation system based on extended Markdown, characterized in that, The system is used to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown as described in any one of claims 1-7, including: Define modules to define extended Markdown formats by adding proprietary tag blocks on top of standard Markdown syntax; The parsing module is used to parse the extended Markdown document, extract structured data, and construct a multi-scale knowledge graph, wherein the multi-scale knowledge graph includes: a positive cognition subgraph and an abnormal logic association subgraph; The generation module is used to generate personalized exercises based on the multi-scale knowledge graph and the user's personal cognitive subgraph. The rendering module is used to parse and render the interactive exercise interface based on the personalized exercises and according to the unified interaction protocol. The evaluation module is used to acquire and distinguish the types of answers submitted by users, perform intelligent evaluation on different types of answers, obtain scoring results and evidence, and provide feedback on the scoring results and evidence.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the interactive exercise automatic generation and intelligent evaluation method based on extended Markdown as described in any one of claims 1-7.

Images