Method and system for comprehensive evaluation of education based on intelligent education AI model

Abstract

The application relates to the technical field of intelligent education, and discloses a method and system for comprehensive education evaluation based on an intelligent education AI model, which comprises constructing and initializing a multi-dimensional cognitive state space model for representing the cognitive state of an evaluation object; in a dynamic evaluation cycle, the multi-dimensional cognitive state space model is updated in real time based on observation data generated by the evaluation object, and cognitive disturbance is applied according to the update result of the model; a causal path atlas is generated based on the evolution process of the multi-dimensional cognitive state space model and the applied cognitive disturbance events recorded in the dynamic evaluation cycle; diagnosis is carried out based on the causal path atlas, and a personalized intervention strategy is planned and executed. Through the construction of the multi-dimensional cognitive state space model and the causal path atlas, an adaptive closed-loop education evaluation method integrating dynamic evaluation, deep attribution diagnosis and precise personalized intervention is provided.

Description

Technical Field

[0001] This invention relates to the field of smart education technology, specifically to a method and system for comprehensive educational evaluation based on a smart education AI model. Background Technology

[0002] With the deep integration of information technology and education, using technological means to comprehensively evaluate assessment subjects has become an important direction for educational modernization. Existing educational evaluation technologies, such as online learning platforms and computerized adaptive testing systems, have improved the efficiency and breadth of evaluation to a certain extent.

[0003] However, existing educational assessment technologies still have limitations in application. Traditional assessment methods often focus on summative assessments of learning outcomes, using periodic tests or exams to obtain information about the knowledge mastery of the assessed individuals. The assessment data obtained in this way is often static and discrete, making it difficult to capture the continuous and dynamic changes in the cognitive state of the assessed individuals during the learning process. Furthermore, even if some systems can conduct online assessments, their diagnostic analysis mostly remains at the superficial level of judging whether knowledge points have been mastered. They cannot delve into the structural causes behind specific learning performance, such as whether it is due to biases in knowledge understanding, rigid problem-solving strategies, or insufficient metacognitive abilities. Therefore, they lack the ability to deeply attribute problems. This superficial diagnosis directly leads to the standardization and inefficiency of intervention measures. The intervention suggestions provided by the system are usually general resource recommendations rather than personalized and precise strategies based on the root causes. The three links of assessment, diagnosis, and intervention are isolated from each other, failing to form an adaptive closed loop that can provide real-time feedback and self-adjustment based on the intervention effect, thus limiting the overall effectiveness of educational assessment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for comprehensive educational evaluation based on a smart education AI model. This solves the problems of existing educational evaluation technologies, such as static evaluation methods, superficial diagnostic attributions, and the disconnect between evaluation, diagnosis, and intervention, which fail to form an effective adaptive closed loop.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for comprehensive educational evaluation based on a smart education AI model, comprising: S1. Construct and initialize a multidimensional cognitive state space model to characterize the cognitive state of the evaluation object.

[0006] In one specific implementation, the multidimensional cognitive state space model may include the following dimensions: Knowledge certainty dimension: used to characterize the probability distribution of the evaluation object's mastery of each knowledge point; Strategy preference dimension: used to characterize the evaluation object's tendency to choose different methods when solving problems; Metacognitive efficiency dimension: used to characterize the efficiency of the evaluation object's ability to self-monitor and regulate cognitive behavior.

[0007] S2. In the dynamic evaluation cycle, the multidimensional cognitive state space model is updated in real time based on the observation data generated by the evaluation object, and cognitive perturbation is applied according to the update result of the model.

[0008] The step of updating the multidimensional cognitive state space model in real time can employ a Bayesian filtering algorithm. This step updates the multidimensional cognitive state space model before the update. state of time As a priori state, combined with Observational data acquired at different times The posterior probability distribution is calculated, and this posterior probability distribution is used as the updated multidimensional cognitive state space model. state of time The update process can be represented by the following formula: ; in, For observation models, The priors predicted by the state transition model, The multidimensional cognitive state space model represents the state space model in The state at any given moment, Representative at The observation data acquired at any time It is a posterior probability distribution. Represents the time from initial time 1 to the current time. All observation data sequences.

[0009] The step of applying cognitive perturbation based on the updated results of the model is as follows: First, calculate the state uncertainty of the multidimensional cognitive state space model, for example, using information entropy. Quantification: ; in, For the model to be in the state space at the first The probability of a discrete state. This refers to information entropy.

[0010] Subsequently, when the state is uncertain When the preset trigger threshold is exceeded, the cognitive perturbation is applied. The step of applying the cognitive perturbation can be based on the principle of maximizing expected information gain, and can be selected from a preset set of perturbation strategies. In the process, the optimal cognitive perturbation is selected and applied. : ; in, This represents the optimal cognitive perturbation or optimal action selected after calculation. Representative set The specific cognitive perturbation scheme in it, Represents the expected value operator. Represents the current moment The state uncertainty of the multidimensional cognitive state space model Represents conditional entropy. This represents new observational data.

[0011] S3. Based on the evolution process of the multidimensional cognitive state space model recorded in the dynamic evaluation cycle and the applied cognitive perturbation events, generate a causal path map.

[0012] Specifically, this step involves constructing nodes and edges of the causal path graph based on the temporal and causal relationships between the state evolution sequence of the multidimensional cognitive state space model and the applied cognitive perturbation events. Both state mutation points and cognitive perturbation events in the state evolution sequence can constitute nodes in the graph, while the temporal associations and statistical correlations between nodes can be used to construct directed edges in the graph.

[0013] S4. Based on the causal path map, make a diagnosis, plan and execute personalized intervention strategies, and use the new observation data generated after the intervention for a new round of updates to the dynamic evaluation cycle to form a closed loop.

[0014] Specifically, the step of diagnosis based on the causal path graph involves identifying nodes with zero in-degree through graph structure analysis in the causal path graph and determining them as the root cause nodes leading to changes in cognitive state.

[0015] The steps for planning and implementing personalized intervention strategies are as follows: First, query the preset causal-intervention mapping knowledge base to map the cognitive impairment type represented by the root cause node into an intervention strategy prototype. Then, through the intervention strategy planning function, combined with the specific information of the root cause node, fill in the parameters and generate the content of the intervention strategy prototype to form a specific intervention task that includes specific resources and activities and is executable.

[0016] A second aspect of the present invention provides a system for comprehensive educational evaluation based on a smart education AI model, applicable to the method described in any of the above, comprising: The state modeling module is used to construct and initialize a multidimensional cognitive state space model that represents the cognitive state of the evaluation object.

[0017] The state update module is used to update the multidimensional cognitive state space model in real time based on the observation data generated by the evaluation object.

[0018] The perturbation generation module is used to generate and apply a cognitive perturbation when the uncertainty of the model state exceeds a preset trigger threshold after the multidimensional cognitive state space model is updated by the state update module.

[0019] The causal diagnosis module is used to generate a causal path map and identify root cause nodes based on the recorded model evolution process and cognitive perturbation events.

[0020] The intervention planning module is used to plan and generate personalized intervention strategies based on the root cause nodes for application to the evaluation object.

[0021] In this process, after the intervention planning module acts on the evaluation object, the new observation data generated is received by the state update module to form a closed loop.

[0022] This invention provides a method and system for comprehensive educational evaluation based on a smart education AI model. It has the following beneficial effects: 1. This invention realizes dynamic and proactive evaluation of cognitive state. By constructing a multidimensional cognitive state space model and using a Bayesian filtering algorithm to update the observed data in real time, this method transforms discrete evaluation interaction into continuous tracking of the cognitive state evolution process. At the same time, the introduction of a cognitive perturbation mechanism enables the system to actively detect the uncertainty region of the model, thereby achieving forward-looking and efficient evaluation, rather than passively waiting for problems to be exposed.

[0023] 2. This invention enables in-depth causal attribution diagnosis. This method is not limited to superficial judgments of evaluation performance, but rather constructs a causal path map to perform structured analysis of the evolution process and key events of cognitive state. By analyzing the map to identify root cause nodes, it can reveal the deep-seated causes leading to specific cognitive performance, providing a clear and interpretable diagnostic basis for subsequent intervention.

[0024] 3. This invention constructs an adaptive closed loop that integrates evaluation, diagnosis, and intervention. The generation of personalized intervention strategies is directly based on the diagnostic results of the root cause node, ensuring the accuracy of the intervention. Furthermore, the new observation data generated after the intervention is seamlessly fed back into the dynamic evaluation loop to drive a new round of model updates. This enables the entire system to continuously adjust and optimize itself based on the actual effects of the intervention, forming a complete closed-loop system that can truly adapt to individual changes. Attached Figure Description

[0025] Figure 1This is a flowchart illustrating the method for comprehensive educational evaluation based on a smart education AI model according to the present invention. Figure 2 This is a schematic diagram of the multidimensional cognitive state space model structure of the present invention; Figure 3 This is a schematic diagram of the Bayesian filtering update loop for the cognitive state of the present invention; Figure 4 This is a flowchart of the cognitive perturbation generation and application mechanism of the present invention; Figure 5 This is a schematic diagram of the causal path map generation process of the present invention; Figure 6 This is a schematic diagram of the causal-intervention mapping and strategy planning of the present invention; Figure 7 This is a closed-loop architecture diagram of the system for comprehensive educational evaluation based on a smart education AI model, as presented in this invention.

[0026] Among them, 10 is the state modeling module; 20 is the disturbance generation module; 30 is the state update module; 40 is the causal diagnosis module; and 50 is the intervention planning module. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see the appendix Figure 1 - Appendix Figure 6 This invention provides a method for comprehensive educational evaluation based on a smart education AI model, including: S1. Construct and initialize a multidimensional cognitive state space model of the evaluation object. This model is used to mathematically represent the cognitive state of the evaluation object from the dimensions of knowledge certainty, strategy preference and metacognitive efficiency.

[0029] S2 updates the multidimensional cognitive state space model in real time based on the interaction data of the evaluation object, and generates and applies a cognitive perturbation aimed at maximizing information gain when the uncertainty of the model state exceeds a preset threshold, while recording the perturbation event.

[0030] S3. After the preset evaluation session ends, a causal path map revealing changes in cognitive state is generated by combining the complete state evolution sequence of the model and the cognitive perturbation events recorded in step S2. The perturbation events are used as key anchor points for constructing causal relationships.

[0031] S4 analyzes the causal path map to identify the root cause of cognitive impairment, and plans and implements personalized intervention strategies that match the root cause to form a closed loop of evaluation and intervention.

[0032] The steps in the embodiments of the method of the present invention will be described in detail below.

[0033] The multidimensional cognitive state space model of the present invention does not perform a single-dimensional ability assessment of the evaluation object (e.g., only assessing whether the knowledge points are mastered), but rather constructs at least three mutually orthogonal and complementary cognitive dimensions to comprehensively and structurally represent the internal cognitive state of the evaluation object, thereby realizing a three-dimensional characterization of the evaluation object's cognitive ability.

[0034] Specifically, the multidimensional cognitive state space model at time points status It can be composed of composite state vectors: ; in, For knowledge determinism vectors, For policy preference vectors, For metacognitive efficiency scalar The representative of the evaluation object is in The overall cognitive state at any given moment.

[0035] By comprehensively analyzing these three dimensions, this method can distinguish different internal cognitive causes that lead to the same external manifestation (e.g., answering a question incorrectly). For example, the system can distinguish whether the error stems from a weak grasp of knowledge concepts (low knowledge certainty dimension), an inappropriate choice of problem-solving strategies (bias in strategy preference dimension), or negligence in the calculation or execution process that fails to self-correct (low metacognitive efficiency dimension). This ability to distinguish is the technical foundation for subsequent causal diagnosis and precise intervention, which is not available in existing methods that rely on single-dimensional assessment.

[0036] Furthermore, specific mathematical representations are provided for each cognitive dimension in the multidimensional cognitive state space model.

[0037] For knowledge deterministic vectors Specifically, it is implemented as a set of parameters of a Dirichlet distribution, which is required for the evaluation object. The first of the knowledge concepts For each knowledge concept, the system does not use a single probability value to describe the mastery level, but instead uses... The Dirichlet distribution of viscounts is used to characterize their respective positions. The confidence level of a preset discrete mastery level (e.g., level 1 is not mastered, level 2 is basic understanding, and level 3 is proficient application) is used to determine the knowledge determinism vector. The complete form is: ; in, For time point indexing, The total number of knowledge concepts. An index for knowledge concepts. For the first A knowledge concept at a point in time The Dirichlet distribution parameter set.

[0038] For the policy preference vector Its specific implementation is as a representation of the probability distribution. A dimensional vector. This vector describes the evaluation object's performance when facing a problem, from a range containing... The strategy preference vector represents the tendency to select a specific strategy from a set of pre-defined cognitive strategies. These pre-defined cognitive strategies may include, but are not limited to: formula application, reverse derivation, analogical reasoning, exhaustive search, or first-principles derivation. The specific form is: ; in, The total number of pre-defined cognitive strategies, Representative at The probability vector of policy selection at time step.

[0039] For metacognitive efficiency scalar Specifically, it is implemented as a scalar value that takes values ​​within a closed interval. : ; scalar value It is a quantitative assessment of the cognitive monitoring and regulation capabilities of the evaluated object. In one specific implementation method, This can be calculated using the behavioral data of the evaluation object during the interaction process, for example, It can be a function based on historical observations, whose input is the time interval between the occurrence of an erroneous behavior and the implementation of the correct corrective action. The shorter the time interval, the better the calculated result. The higher the value, in another specific implementation, It can also be the moving average success rate of the evaluated object successfully correcting itself after being prompted of an error.

[0040] In step S1, after defining the structure of the multidimensional cognitive state space model, the initial state of the model is determined. The initialization process is designed to provide a reasonable starting point for subsequent dynamic updates.

[0041] The specific implementation of this initialization process may vary depending on whether the evaluation object has available historical interaction data.

[0042] In one specific embodiment, when the system contains historical interaction data of the evaluated object, the system analyzes this data to set the initial parameters of the model, for knowledge deterministic vectors. The system statistically analyzes the evaluation subjects' performance under various knowledge concepts in historical data and maps the statistical results to the initial concentration parameters of a Dirichlet distribution. For example, the evaluation object can be in the knowledge concept. The above values, representing the number of performances at different mastery levels (e.g., completely correct, partially correct, incorrect), are combined with smoothed prior values ​​to serve as initial concentration parameters. The components of the policy preference vector. If historical data includes records of cognitive strategy usage, then the historical usage frequency of each strategy can be directly used as the initial probability distribution. For metacognitive efficiency scalar The system can calculate the average self-correction success rate or related behavioral time indicators of the evaluated objects in historical data, and use the calculation results as initial values. .

[0043] In another specific embodiment, when the evaluation object is a new user or its historical interaction data is unavailable, the system initializes the model with no informational priors to characterize the maximum uncertainty of the initial state, for knowledge deterministic vectors. The system calculates the Dirichlet distribution concentration parameters of all knowledge concepts. All are set as constant vectors, for example (1,1,...,1), which means that at the initial moment, the system considers that the evaluation object's mastery of each knowledge concept is equally likely at any level. For the strategy preference vector... The system sets it to a uniform distribution, that is, for all All have This implies that initially, it is assumed that all cognitive strategies have the same probability of being chosen, for the metacognitive efficiency scalar. The system sets it to a preset intermediate value, for example. =0.5.

[0044] The initial state of the model after initialization is complete. It was established and serves as the starting point for subsequent dynamic update processes.

[0045] After initialization is completed in step S1, this method executes step S2 to dynamically update the multidimensional cognitive state space model. This update process is a continuous and recursive Bayesian estimation process, which aims to infer the posterior probability distribution of the cognitive state in real time based on the continuous observation data stream generated when the evaluation object interacts with the system.

[0046] This dynamic update process specifically follows the Bayesian filtering framework, at each time step... It can be broken down into two core stages: prediction and updating.

[0047] The first step is the prediction phase, the purpose of which is to build upon the prediction from the previous time step. State posterior distribution To predict the current moment State prior distribution This process is achieved through a state transition model. To achieve: ; in, and These are the cognitive state vectors for the current moment and the previous moment, respectively. Given the observation data sequence up to the previous time step, the state transition model is... It describes the natural evolution of cognitive states over time in the absence of new external interactions. for Prior state distribution at time step, for In one specific implementation, the posterior state distribution at time step can be represented as a knowledge forgetting mechanism, i.e., a knowledge deterministic vector. Dirichlet concentration parameters It will slowly decay towards an informationless prior distribution over time. In another implementation, the model may also be reflected as stability or slight drift in policy preferences or metacognitive efficiency.

[0048] Next, the update phase is executed, which utilizes the current time. Newly acquired observational data The prior distribution obtained in the prediction stage is corrected to obtain a more accurate posterior probability distribution. The calculation formula for this process is as follows: ; The core of this process is the observation model. ,for The observed data of the posterior state distribution at time t. The evaluation model refers to any quantifiable behavior generated by the object on the interactive interface, which may include, but is not limited to: the answer to a question (correct, incorrect, or partially correct), the time taken to complete a task, the number of times an auxiliary tool is clicked, or the specific operation path selected in an open-ended question. These observable behaviors were then established. With the invisible internal cognitive state The probabilistic relationship between them.

[0049] In one specific embodiment, when the observed data When evaluating an object's answer to a question, the observation model It can be defined as a logistic regression function or a more complex probabilistic graphical model, with the current cognitive state as its input. Each dimension, such as the knowledge determinism vector of knowledge points related to the question. The policy preference vector of the evaluation object in this context and its metacognitive efficiency scalar The model outputs the probability of answering correctly or incorrectly. Through this model, correct answers will positively update the relevant knowledge certainty, while errors caused by calculation mistakes will negatively update the metacognitive efficiency while updating the knowledge certainty.

[0050] In a preferred embodiment, the entire recursive Bayesian estimation process can be implemented using a particle filtering algorithm. By maintaining a set of weighted state particles, the posterior probability distribution of the state is approximated. This method can handle the nonlinear, non-Gaussian models involved in this invention, thereby more accurately tracking complex cognitive states.

[0051] In step S2, after the real-time update of the cognitive state is completed, this method further includes an active detection step, namely the generation and application of cognitive perturbations. This step is not executed after each state update, but is activated by specific triggering conditions. Its core purpose is to reduce the uncertainty of the system's cognitive state of the evaluation object in the most efficient way by actively applying optimized detection signals, thereby obtaining the most informative evidence for subsequent causal diagnosis.

[0052] First, the system performs a disturbance trigger determination. After each state update, the system calculates the posterior distribution of the current cognitive state. Overall uncertainty In one implementation, this uncertainty It can be quantified as a weighted sum of the information entropies of the components of the state vector, and the system will use this uncertainty value. With preset trigger threshold Comparison, if and only if At this time, the cognitive perturbation generation mechanism is activated. This mechanism ensures that perturbations only intervene when the system needs information most, thus avoiding unnecessary interference to the evaluation object.

[0053] Once the perturbation generation mechanism is activated, the system will perform optimal perturbation selection. The goal of this process is to select from a predefined set of perturbation strategies of various types. In this process, the perturbation strategy that maximizes the expected information gain is selected. The expected information gain is a quantification of the expected value by which the perturbation strategy can reduce state uncertainty, and its calculation formula is as follows: ; in, The selected optimal perturbation strategy. For the set of perturbation strategies Candidate perturbation strategies in the process. The uncertainty of the state before the perturbation is applied. For the evaluation object to be disturbed The response that followed For a specific observation model, In order to obtain observations Then, the posterior uncertainty of the new state obtained after Bayesian update, For all possible responses Find the expected value.

[0054] set of cognitive perturbation strategies The collection includes various specific implementation methods. In one embodiment, the collection may include: Ambiguous cue perturbation: The system provides ambiguous cuees that can be interpreted semantically by multiple problem-solving strategies. The subsequent operation path of the evaluated object will directly reveal its preference for different strategies, thereby effectively reducing the strategy preference vector. Uncertainty.

[0055] Counterfactual problem perturbation: The system temporarily changes key known conditions or constraints based on the current problem and requires the evaluated object to judge the impact of the change on the conclusion. This perturbation aims to probe the evaluated object's deep understanding of the underlying physical laws or mathematical axioms of the problem, rather than rote memorization of a specific problem type, thereby effectively reducing the knowledge determinism vector. The uncertainty of intermediate and advanced mastery levels.

[0056] Resource constraint perturbation: The system temporarily disables an external auxiliary tool, such as a formula lookup tool or calculator, that the evaluation object relies on during the problem-solving process. This perturbation aims to force the evaluation object to use lower-level knowledge or alternative strategies, thereby probing its knowledge and strategy reserves and transfer capabilities.

[0057] Determining the optimal perturbation Subsequently, the system sends a command to the user terminal to apply the perturbation. At the same time, the system records the perturbation event in a structured manner. This record includes at least: the timestamp of the perturbation application, the type of perturbation used, the specific parameters of the perturbation (e.g., the modified conditions in a counterfactual problem), and the state uncertainty value that triggered the perturbation. This structured record will be used as key causal anchor data to construct a causal path graph in the subsequent S3 step.

[0058] After the preset evaluation session ends, for example, when the evaluated object completes a set of tasks or reaches the preset interaction time, this method executes step S3. The core task of this step is to generate a causal path map that reveals the internal evolutionary logic of cognitive state based on the data accumulated in the previous steps. This map aims to transform the high-dimensional, continuous state evolution data sequence generated during the evaluation process into a structured diagnostic representation that can be understood and analyzed by human experts or subsequent modules.

[0059] The technical feature of constructing this causal map is that it does not simply list state changes in chronological order, but uses the cognitive perturbation events actively applied and recorded in step S2 as key causal anchors. By analyzing the correlation between these active intervention behaviors and their subsequent state changes, this method can extract causal attributions with higher confidence from the correlations, thereby effectively revealing the underlying reasons for the learning difficulties or specific behavioral patterns of the evaluated subjects.

[0060] The input to this construction process consists of two parts of data: the first part is the complete multidimensional cognitive state space model evolution sequence of the evaluation object generated in step S2 throughout the entire session, i.e. The first part contains the posterior probability distribution information at each time point; the second part is a set of cognitive perturbation events that are also structured and recorded in step S2, and the resulting causal path graph is a directed graph: ; in: It is the set of nodes in a graph. It is the set of directed edges of a graph.

[0061] Key cognitive events (i.e., nodes) This can manifest in various forms. In one implementation, it can be a significant change in a dimension of the cognitive state of the evaluated object, such as "the certainty of mastering the knowledge concept 'vector synthesis' drops from a high-confidence level of 'proficient application' to a level of 'preliminary understanding'." In another implementation, it can also be an external event recorded by the system, such as "the system at a certain time point..." "Applying counterfactual problem perturbation" can also be a specific behavioral pattern exhibited by the evaluated object, such as "the evaluated object has used a single strategy of 'formula application' three times in a row in problems that require multi-step reasoning".

[0062] Furthermore, the specific generation algorithms for graph elements, namely nodes and edges, in the causal path graph are explained.

[0063] First, the system performs node generation, a process designed to extract discrete, meaningful key cognitive events from a continuous sequence of state evolutions and instantiate these events as nodes in a graph. Node generation includes the following two sub-steps: S31, The system traverses the state evolution sequence throughout the entire evaluation session. For each time step The system calculates the posterior state distribution at the current moment. posterior state distribution compared to the previous time step The difference between them, based on the significance of state changes to generate nodes, can be quantified in one specific implementation by calculating the Kohlbek-Leibler divergence between the two probability distributions: ; in, represent , represent When the divergence value exceeds the preset node generation threshold At that time, the system determined that at the time point A significant change in cognitive state has occurred, generating a "State Change" node. The content of this node will describe in detail which cognitive dimension (knowledge, strategy, or metacognition) has changed and what kind of change it has.

[0064] In step S32, the system reads the set of cognitive perturbation events recorded in step S2. For each perturbation event in the set, the system generates a node of type "perturbation event". The content of this node includes the type of perturbation, the time of application, and specific parameters. Through these two sub-steps, the system constructs a complete set of nodes for the graph. Nodes and edges are generated based on perturbation events and causal significance.

[0065] Secondly, the system constructs edges, a process designed to connect the generated nodes to represent causal relationships between events. Edge construction follows two criteria: temporal order and causal significance. For any two nodes... ,in The occurrence time was earlier than The system determines whether to establish a transaction by using the following methods. point to Directed edges: In one specific implementation, when node It is a "disturbance event" node, and the node When the next "state change" node is the one immediately following the disturbance, the system determines that the state change is directly caused by the disturbance, and therefore establishes a directed edge. The basis for this judgment is that the design goal of the perturbation is to detect and induce specific state changes. Therefore, significant state changes that are adjacent in time are attributed to the perturbation with the highest confidence.

[0066] In another, more general implementation, the system attributes the change by comparing the actual state change with the expected change at the "no intervention" baseline. Specifically, for changes occurring at a given time point... event nodes (This can be a disturbance event or a state change event), the system utilizes a state transition model. Predictions in the absence Under the influence, the state changed from The natural evolutionary distribution, when in The actual state change nodes observed at each moment When the magnitude of the change represented (e.g., KL divergence value) is significantly greater than the expected magnitude of change under the baseline evolution, the system determines the event. Regarding the event The occurrence of this has a significant causal contribution, thus establishing a directed edge. This method uses cognitive perturbation events as key causal anchors and, based on these, establishes a more credible causal chain throughout the entire history of state evolution.

[0067] After generating the causal path graph in step S3, this method executes step S4. The primary task of step S4 is to parse the causal path graph in order to automatically and objectively identify the root causes that lead to the evaluated object exhibiting cognitive difficulties or erroneous behavior in the evaluation session.

[0068] The diagnostic process for this root cause transforms the diagnosis of cognitive problems of the evaluated object into a deterministic computational process of graph theory analysis on the causal path graph. Specifically, the system locates the starting point of the causal chain by analyzing the topological structure of the graph.

[0069] In causal path map In the structure, the root cause node is explicitly defined as any node in the graph with an in-degree of zero. Such a node represents an initial event or initial state that is not caused by any other recorded event in the session, and is the source of a chain reaction of subsequent cognitive events.

[0070] The automatic identification algorithm for the root cause node includes the following sub-steps: S41, The system initializes counters used to store the in-degrees of all nodes, for the graph. Node set Each node in Their initial in-degrees are all set to 0.

[0071] S42, Systematic traversal of the graph The set of directed edges For each directed edge in the set The system will target the node The value of the in-degree counter is incremented by 1.

[0072] S43, after traversing all edges, the system traverses the node set again. The system checks each node. The final in-degree count value is obtained. All nodes with an in-degree count value of 0 are identified as root cause nodes and added to the root cause set.

[0073] Through the above algorithm steps, this method realizes the parsing of the causal path graph and can automatically output one or more root cause nodes. For example, if a node describing "low initial certainty of mastering the knowledge concept 'Newton's Second Law'" has zero in-degree in the graph and there is a directed edge pointing to a node "error when applying 'Newton's Second Law' in a multibody problem", the system will identify the former as the root cause. This diagnostic process provides clear and traceable targets for subsequent planning and precise intervention.

[0074] After identifying the root cause nodes through graph theory analysis, this method continues to perform the causal-intervention mapping and strategy planning process. The technical feature of this process is that it does not adopt a one-size-fits-all approach, but uses a pre-set knowledge base and planning function to transform the abstract cognitive diagnosis results into specific, personalized, and system-executable intervention strategies.

[0075] This process relies on a pre-built "causal-intervention mapping knowledge base," which stores a large number of mapping rules in a structured form. Each rule associates a specific type of cognitive impairment (i.e., the type of root cause node) with one or more effective intervention strategy prototypes. This knowledge base is the technical foundation for the precise intervention achieved by this invention.

[0076] In one specific implementation, the knowledge base can be a database or rule set containing multiple key-value pairs. Here, the key is a schema description of the root cause node, and the value is the corresponding intervention strategy prototype identifier. The mapping rules in this knowledge base may include, but are not limited to: If the root cause is "the confidence level of 'not mastered' or 'preliminary understanding' is too high in the knowledge certainty dimension of a specific knowledge concept", then the corresponding intervention strategy prototype is "concept relearning and tiered practice".

[0077] If the root cause is "in situations with multiple available strategies, the strategy preference vector is fixed on a single inefficient strategy", then the corresponding intervention strategy prototype is "cognitive conflict task and strategy comparison guidance".

[0078] If the root cause is "the metacognitive efficiency scalar is consistently below the preset threshold and is associated with multiple computational or executive error nodes", then the corresponding intervention strategy prototype is "self-explanation and error review training".

[0079] After obtaining the root cause node and the corresponding intervention strategy prototype, the system instantiates the prototype into an executable intervention task through an intervention strategy planning function. The role of this planning function is to fill in the parameters and generate content for the intervention strategy prototype based on the detailed information of the root cause node. This process may include the following sub-steps: S44, the system uses the content of the identified root cause node as query input and matches it in the causal-intervention mapping knowledge base to retrieve one or more matching intervention strategy prototypes.

[0080] S45. For each retrieved intervention strategy prototype, the intervention strategy planning function is activated. This function generates a specific intervention task based on the specific information contained in the root cause node (e.g., which knowledge concept or which erroneous strategy is involved) and the current overall cognitive state of the evaluated subject. For example, if the matched prototype is "concept relearning and tiered practice" and the root cause involves the knowledge concept of "vector composition", the planning function will extract basic explanation videos about "vector composition" from the teaching resource library and select three practice questions of increasing difficulty, all labeled "vector composition", from the question database to form a specific intervention task package. Similarly, if the matched prototype is "cognitive conflict task", the planning function will generate or select a specially designed problem situation that will inevitably cause the inefficient strategies commonly used by the evaluated subject to fail, as the core of the intervention task.

[0081] Through this process, the system automatically plans from abstract diagnosis to specific intervention programs, ensuring the relevance and effectiveness of subsequent interventions.

[0082] In step S4, once a specific, personalized intervention strategy is planned and generated, this method immediately implements the intervention and re-evaluates the intervention effect, thus forming a complete adaptive learning loop.

[0083] The specific implementation of this process includes the following sub-steps: S46, the system pushes the instantiated specific intervention task generated in the previous process to the user terminal device used by the evaluation object through the network interface, and presents it on the interactive interface of the device. During the entire process of the evaluation object performing the intervention task, the system of the present invention continuously captures all the interactive data generated. The interactive data may include, but is not limited to: the evaluation object's answer to the exercise questions in the intervention task, the answering time, the viewing time of the teaching video, the record of pause or replay behavior, or the sequence of operation steps taken in the open intervention task.

[0084] S47 After the intervention task is completed or the preset checkpoint is reached, the interaction data captured in sub-step S46 is recaptured by the system and formatted into a new observation data sequence. This new observation data sequence is then used as new input and returned to step S2 in the method flow of this invention, namely the real-time update process of cognitive state. The multidimensional cognitive state space model will use this batch of new observation data to conduct a new round of evaluation of the cognitive state of the evaluation object after receiving intervention, thereby updating its posterior probability distribution.

[0085] Through the above steps, this method achieves a complete dynamic closed loop of "evaluation-diagnosis-intervention-reassessment". The system can quantitatively evaluate the actual effect of the intervention (for example, by comparing the changes in the knowledge certainty distribution of the target knowledge points before and after the intervention), and based on the evaluation results, decide whether to continue consolidation and enter a new learning stage, or to plan a new round of intervention for new or residual cognitive impairments. This adaptive and continuously optimized closed-loop system is the core technology that ensures that this invention can achieve efficient personalized education.

[0086] The system for comprehensive educational evaluation based on a smart education AI model described below can be referred to in conjunction with the method for comprehensive educational evaluation based on a smart education AI model described above.

[0087] See attached document Figure 7 A system for comprehensive educational evaluation based on a smart education AI model. The system can be deployed on a server and interact with user terminals via a network. The system includes: a state modeling module 10, a disturbance generation module 20, a state update module 30, a causal diagnosis module 40, and an intervention planning module 50.

[0088] The state modeling module 10 is used to construct and maintain a multidimensional cognitive state space model for the evaluation object. This multidimensional cognitive state space model is a mathematical representation of the internal cognitive state of the evaluation object.

[0089] At the start of an evaluation session, the state modeling module 10 initializes the multidimensional cognitive state space model based on the historical data or preset priors of the evaluation object, and sets the initial state. Output to status update module 30.

[0090] The state update module 30 is used to perform real-time inference and update of the multidimensional cognitive state space model based on the continuous interactive data stream received from the user terminal. After receiving new observation data, the state update module 30 uses a Bayesian filtering framework to update the state. The posterior probability distribution is updated, and the updated state uncertainty value is passed to the perturbation generation module 20.

[0091] The disturbance generation module 20 is used to generate and apply a cognitive disturbance to the user terminal when the state uncertainty value output by the state update module 30 exceeds a preset threshold. The disturbance generation module 20 selects the optimal disturbance from a predefined set of disturbance strategies according to the principle of maximizing expected information gain, and sends the disturbance instruction to the user terminal.

[0092] The causal diagnosis module 40 is used to receive a complete state evolution sequence from the state update module 30 after an evaluation session ends. The causal diagnosis module 40 constructs a causal path map that reveals changes in cognitive state by tracing back the state evolution history and the perturbation events applied by the perturbation generation module 20, and outputs the causal path map to the intervention planning module 50.

[0093] The intervention planning module 50 is used to automatically plan personalized intervention strategies based on the causal path map output by the causal diagnosis module 40. The intervention planning module 50 matches the optimal intervention strategy for the root node of the fundamental cognitive impairment identified in the causal path map in the preset causal-intervention mapping knowledge base, generates an intervention task instruction, and sends the instruction to the user terminal for execution.

[0094] In the overall operation of the system, the state modeling module 10 provides an initial framework for the cognitive state, the state update module 30 and the disturbance generation module 20 constitute a dynamic evaluation loop of "detection-update", the causal diagnosis module 40 performs in-depth attribution of this loop process, and the intervention planning module 50 outputs precise intervention in a closed loop based on the attribution results, thereby forming a complete and adaptive comprehensive evaluation and intervention process.

[0095] The system in this embodiment can be used to execute the above method embodiments, and its principle and technical effect are similar, so they will not be described again here.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for comprehensive educational evaluation based on a smart education AI model, characterized in that, include: S1. Construct and initialize a multidimensional cognitive state space model to characterize the cognitive state of the evaluation object; S2. In the dynamic evaluation cycle, the multidimensional cognitive state space model is updated in real time based on the observation data generated by the evaluation object, and cognitive perturbation is applied according to the update result of the model. S3. Based on the evolution process of the multidimensional cognitive state space model recorded in the dynamic evaluation cycle and the applied cognitive perturbation events, generate a causal path map; S4. Based on the causal path map, make a diagnosis, plan and execute personalized intervention strategies, and use the new observation data generated after the intervention for a new round of updates to the dynamic evaluation cycle to form a closed loop.

2. The method for comprehensive educational evaluation based on a smart education AI model according to claim 1, characterized in that, The multidimensional cognitive state space model includes at least the following three dimensions: A knowledge certainty dimension used to characterize the degree of mastery of each knowledge point by the evaluation subject; The strategy preference dimension is used to characterize the tendency of the evaluated object to choose methods when solving problems; Metacognitive efficiency dimension is used to characterize the self-monitoring and regulation capabilities of the evaluated object.

3. The method for comprehensive educational evaluation based on a smart education AI model according to claim 1, characterized in that, The step of updating the multidimensional cognitive state space model in real time in step S2 is as follows: A Bayesian filtering algorithm is used to take the multidimensional cognitive state space model before the update as the prior state, and calculate the posterior probability distribution by combining it with the observation data at the current moment. The posterior probability distribution is then used as the updated multidimensional cognitive state space model.

4. The method for comprehensive educational evaluation based on a smart education AI model according to claim 1, characterized in that, The step of applying cognitive perturbation based on the updated model in step S2 is as follows: Calculate the state uncertainty of the multidimensional cognitive state space model; When the state uncertainty exceeds a preset trigger threshold, the cognitive perturbation is applied.

5. The method for comprehensive educational evaluation based on a smart education AI model according to claim 4, characterized in that, The specific steps for applying the cognitive perturbation in step S2 are as follows: Based on the principle of maximizing expected information gain, the optimal cognitive perturbation is selected and applied from a pre-set set of perturbation strategies.

6. The method for comprehensive educational evaluation based on a smart education AI model according to claim 1, characterized in that, The specific steps for generating the causal path graph in step S3 are as follows: Based on the temporal and causal relationships between the state evolution sequence of the multidimensional cognitive state space model and the applied cognitive perturbation events, the nodes and edges of the causal path graph are constructed.

7. The method for comprehensive educational evaluation based on a smart education AI model according to claim 1, characterized in that, The specific steps of diagnosis based on the causal path map in step S4 are as follows: In the causal path graph, nodes with an in-degree of zero are identified and determined as the root cause nodes leading to changes in cognitive state.

8. The method for comprehensive educational evaluation based on a smart education AI model according to claim 7, characterized in that, The specific steps for planning and implementing personalized intervention strategies in step S4 are as follows: The root cause node is mapped to an intervention strategy prototype by querying a pre-defined causal-intervention mapping knowledge base.

9. The method for comprehensive educational evaluation based on a smart education AI model according to claim 8, characterized in that, The step of planning and implementing personalized intervention strategies in step S4 is further specified as follows: By using the intervention strategy planning function and combining the specific information of the root cause node, the intervention strategy prototype is filled with parameters and content is generated to form an executable specific intervention task.

10. A system for comprehensive educational evaluation based on a smart education AI model, characterized in that: The method for comprehensive educational evaluation based on a smart education AI model, as described in any one of claims 1-9, includes: The state modeling module is used to construct and initialize a multidimensional cognitive state space model for representing the cognitive state of the evaluation object. The state update module is used to update the multidimensional cognitive state space model in real time based on the observation data generated by the evaluation object. The perturbation generation module is used to generate and apply a cognitive perturbation when the uncertainty of the model state exceeds a preset trigger threshold after the multidimensional cognitive state space model is updated. The causal diagnosis module is used to generate a causal path map and identify root cause nodes based on the recorded model evolution process and cognitive perturbation events. The intervention planning module is used to plan and generate personalized intervention strategies based on the root cause nodes for application to the evaluation object; In this process, after the intervention planning module acts on the evaluation object, the new observation data generated is received by the state update module to form a closed loop.

Images